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 CPU load, unless cfs_rq belongs to a non-root task_group */
2761 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2762 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2764 return decayed || removed;
2767 /* Update task and its cfs_rq load average */
2768 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2770 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2771 u64 now = cfs_rq_clock_task(cfs_rq);
2772 int cpu = cpu_of(rq_of(cfs_rq));
2775 * Track task load average for carrying it to new CPU after migrated, and
2776 * track group sched_entity load average for task_h_load calc in migration
2778 __update_load_avg(now, cpu, &se->avg,
2779 se->on_rq * scale_load_down(se->load.weight),
2780 cfs_rq->curr == se, NULL);
2782 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2783 update_tg_load_avg(cfs_rq, 0);
2785 if (entity_is_task(se))
2786 trace_sched_load_avg_task(task_of(se), &se->avg);
2789 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2791 if (!sched_feat(ATTACH_AGE_LOAD))
2795 * If we got migrated (either between CPUs or between cgroups) we'll
2796 * have aged the average right before clearing @last_update_time.
2798 if (se->avg.last_update_time) {
2799 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2800 &se->avg, 0, 0, NULL);
2803 * XXX: we could have just aged the entire load away if we've been
2804 * absent from the fair class for too long.
2809 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2810 cfs_rq->avg.load_avg += se->avg.load_avg;
2811 cfs_rq->avg.load_sum += se->avg.load_sum;
2812 cfs_rq->avg.util_avg += se->avg.util_avg;
2813 cfs_rq->avg.util_sum += se->avg.util_sum;
2816 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2818 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2819 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2820 cfs_rq->curr == se, NULL);
2822 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2823 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2824 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2825 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2828 /* Add the load generated by se into cfs_rq's load average */
2830 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2832 struct sched_avg *sa = &se->avg;
2833 u64 now = cfs_rq_clock_task(cfs_rq);
2834 int migrated, decayed;
2836 migrated = !sa->last_update_time;
2838 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2839 se->on_rq * scale_load_down(se->load.weight),
2840 cfs_rq->curr == se, NULL);
2843 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2845 cfs_rq->runnable_load_avg += sa->load_avg;
2846 cfs_rq->runnable_load_sum += sa->load_sum;
2849 attach_entity_load_avg(cfs_rq, se);
2851 if (decayed || migrated)
2852 update_tg_load_avg(cfs_rq, 0);
2855 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2857 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2859 update_load_avg(se, 1);
2861 cfs_rq->runnable_load_avg =
2862 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2863 cfs_rq->runnable_load_sum =
2864 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2867 #ifndef CONFIG_64BIT
2868 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2870 u64 last_update_time_copy;
2871 u64 last_update_time;
2874 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2876 last_update_time = cfs_rq->avg.last_update_time;
2877 } while (last_update_time != last_update_time_copy);
2879 return last_update_time;
2882 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2884 return cfs_rq->avg.last_update_time;
2889 * Task first catches up with cfs_rq, and then subtract
2890 * itself from the cfs_rq (task must be off the queue now).
2892 void remove_entity_load_avg(struct sched_entity *se)
2894 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2895 u64 last_update_time;
2898 * Newly created task or never used group entity should not be removed
2899 * from its (source) cfs_rq
2901 if (se->avg.last_update_time == 0)
2904 last_update_time = cfs_rq_last_update_time(cfs_rq);
2906 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2907 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2908 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2912 * Update the rq's load with the elapsed running time before entering
2913 * idle. if the last scheduled task is not a CFS task, idle_enter will
2914 * be the only way to update the runnable statistic.
2916 void idle_enter_fair(struct rq *this_rq)
2921 * Update the rq's load with the elapsed idle time before a task is
2922 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2923 * be the only way to update the runnable statistic.
2925 void idle_exit_fair(struct rq *this_rq)
2929 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2931 return cfs_rq->runnable_load_avg;
2934 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2936 return cfs_rq->avg.load_avg;
2939 static int idle_balance(struct rq *this_rq);
2941 #else /* CONFIG_SMP */
2943 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2945 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2947 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2948 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2951 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2953 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2955 static inline int idle_balance(struct rq *rq)
2960 #endif /* CONFIG_SMP */
2962 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2964 #ifdef CONFIG_SCHEDSTATS
2965 struct task_struct *tsk = NULL;
2967 if (entity_is_task(se))
2970 if (se->statistics.sleep_start) {
2971 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2976 if (unlikely(delta > se->statistics.sleep_max))
2977 se->statistics.sleep_max = delta;
2979 se->statistics.sleep_start = 0;
2980 se->statistics.sum_sleep_runtime += delta;
2983 account_scheduler_latency(tsk, delta >> 10, 1);
2984 trace_sched_stat_sleep(tsk, delta);
2987 if (se->statistics.block_start) {
2988 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2993 if (unlikely(delta > se->statistics.block_max))
2994 se->statistics.block_max = delta;
2996 se->statistics.block_start = 0;
2997 se->statistics.sum_sleep_runtime += delta;
3000 if (tsk->in_iowait) {
3001 se->statistics.iowait_sum += delta;
3002 se->statistics.iowait_count++;
3003 trace_sched_stat_iowait(tsk, delta);
3006 trace_sched_stat_blocked(tsk, delta);
3007 trace_sched_blocked_reason(tsk);
3010 * Blocking time is in units of nanosecs, so shift by
3011 * 20 to get a milliseconds-range estimation of the
3012 * amount of time that the task spent sleeping:
3014 if (unlikely(prof_on == SLEEP_PROFILING)) {
3015 profile_hits(SLEEP_PROFILING,
3016 (void *)get_wchan(tsk),
3019 account_scheduler_latency(tsk, delta >> 10, 0);
3025 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3027 #ifdef CONFIG_SCHED_DEBUG
3028 s64 d = se->vruntime - cfs_rq->min_vruntime;
3033 if (d > 3*sysctl_sched_latency)
3034 schedstat_inc(cfs_rq, nr_spread_over);
3039 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3041 u64 vruntime = cfs_rq->min_vruntime;
3044 * The 'current' period is already promised to the current tasks,
3045 * however the extra weight of the new task will slow them down a
3046 * little, place the new task so that it fits in the slot that
3047 * stays open at the end.
3049 if (initial && sched_feat(START_DEBIT))
3050 vruntime += sched_vslice(cfs_rq, se);
3052 /* sleeps up to a single latency don't count. */
3054 unsigned long thresh = sysctl_sched_latency;
3057 * Halve their sleep time's effect, to allow
3058 * for a gentler effect of sleepers:
3060 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3066 /* ensure we never gain time by being placed backwards. */
3067 se->vruntime = max_vruntime(se->vruntime, vruntime);
3070 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3073 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3076 * Update the normalized vruntime before updating min_vruntime
3077 * through calling update_curr().
3079 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3080 se->vruntime += cfs_rq->min_vruntime;
3083 * Update run-time statistics of the 'current'.
3085 update_curr(cfs_rq);
3086 enqueue_entity_load_avg(cfs_rq, se);
3087 account_entity_enqueue(cfs_rq, se);
3088 update_cfs_shares(cfs_rq);
3090 if (flags & ENQUEUE_WAKEUP) {
3091 place_entity(cfs_rq, se, 0);
3092 enqueue_sleeper(cfs_rq, se);
3095 update_stats_enqueue(cfs_rq, se);
3096 check_spread(cfs_rq, se);
3097 if (se != cfs_rq->curr)
3098 __enqueue_entity(cfs_rq, se);
3101 if (cfs_rq->nr_running == 1) {
3102 list_add_leaf_cfs_rq(cfs_rq);
3103 check_enqueue_throttle(cfs_rq);
3107 static void __clear_buddies_last(struct sched_entity *se)
3109 for_each_sched_entity(se) {
3110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3111 if (cfs_rq->last != se)
3114 cfs_rq->last = NULL;
3118 static void __clear_buddies_next(struct sched_entity *se)
3120 for_each_sched_entity(se) {
3121 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3122 if (cfs_rq->next != se)
3125 cfs_rq->next = NULL;
3129 static void __clear_buddies_skip(struct sched_entity *se)
3131 for_each_sched_entity(se) {
3132 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3133 if (cfs_rq->skip != se)
3136 cfs_rq->skip = NULL;
3140 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3142 if (cfs_rq->last == se)
3143 __clear_buddies_last(se);
3145 if (cfs_rq->next == se)
3146 __clear_buddies_next(se);
3148 if (cfs_rq->skip == se)
3149 __clear_buddies_skip(se);
3152 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3155 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3158 * Update run-time statistics of the 'current'.
3160 update_curr(cfs_rq);
3161 dequeue_entity_load_avg(cfs_rq, se);
3163 update_stats_dequeue(cfs_rq, se);
3164 if (flags & DEQUEUE_SLEEP) {
3165 #ifdef CONFIG_SCHEDSTATS
3166 if (entity_is_task(se)) {
3167 struct task_struct *tsk = task_of(se);
3169 if (tsk->state & TASK_INTERRUPTIBLE)
3170 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3171 if (tsk->state & TASK_UNINTERRUPTIBLE)
3172 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3177 clear_buddies(cfs_rq, se);
3179 if (se != cfs_rq->curr)
3180 __dequeue_entity(cfs_rq, se);
3182 account_entity_dequeue(cfs_rq, se);
3185 * Normalize the entity after updating the min_vruntime because the
3186 * update can refer to the ->curr item and we need to reflect this
3187 * movement in our normalized position.
3189 if (!(flags & DEQUEUE_SLEEP))
3190 se->vruntime -= cfs_rq->min_vruntime;
3192 /* return excess runtime on last dequeue */
3193 return_cfs_rq_runtime(cfs_rq);
3195 update_min_vruntime(cfs_rq);
3196 update_cfs_shares(cfs_rq);
3200 * Preempt the current task with a newly woken task if needed:
3203 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3205 unsigned long ideal_runtime, delta_exec;
3206 struct sched_entity *se;
3209 ideal_runtime = sched_slice(cfs_rq, curr);
3210 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3211 if (delta_exec > ideal_runtime) {
3212 resched_curr(rq_of(cfs_rq));
3214 * The current task ran long enough, ensure it doesn't get
3215 * re-elected due to buddy favours.
3217 clear_buddies(cfs_rq, curr);
3222 * Ensure that a task that missed wakeup preemption by a
3223 * narrow margin doesn't have to wait for a full slice.
3224 * This also mitigates buddy induced latencies under load.
3226 if (delta_exec < sysctl_sched_min_granularity)
3229 se = __pick_first_entity(cfs_rq);
3230 delta = curr->vruntime - se->vruntime;
3235 if (delta > ideal_runtime)
3236 resched_curr(rq_of(cfs_rq));
3240 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3242 /* 'current' is not kept within the tree. */
3245 * Any task has to be enqueued before it get to execute on
3246 * a CPU. So account for the time it spent waiting on the
3249 update_stats_wait_end(cfs_rq, se);
3250 __dequeue_entity(cfs_rq, se);
3251 update_load_avg(se, 1);
3254 update_stats_curr_start(cfs_rq, se);
3256 #ifdef CONFIG_SCHEDSTATS
3258 * Track our maximum slice length, if the CPU's load is at
3259 * least twice that of our own weight (i.e. dont track it
3260 * when there are only lesser-weight tasks around):
3262 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3263 se->statistics.slice_max = max(se->statistics.slice_max,
3264 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3267 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3271 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3274 * Pick the next process, keeping these things in mind, in this order:
3275 * 1) keep things fair between processes/task groups
3276 * 2) pick the "next" process, since someone really wants that to run
3277 * 3) pick the "last" process, for cache locality
3278 * 4) do not run the "skip" process, if something else is available
3280 static struct sched_entity *
3281 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3283 struct sched_entity *left = __pick_first_entity(cfs_rq);
3284 struct sched_entity *se;
3287 * If curr is set we have to see if its left of the leftmost entity
3288 * still in the tree, provided there was anything in the tree at all.
3290 if (!left || (curr && entity_before(curr, left)))
3293 se = left; /* ideally we run the leftmost entity */
3296 * Avoid running the skip buddy, if running something else can
3297 * be done without getting too unfair.
3299 if (cfs_rq->skip == se) {
3300 struct sched_entity *second;
3303 second = __pick_first_entity(cfs_rq);
3305 second = __pick_next_entity(se);
3306 if (!second || (curr && entity_before(curr, second)))
3310 if (second && wakeup_preempt_entity(second, left) < 1)
3315 * Prefer last buddy, try to return the CPU to a preempted task.
3317 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3321 * Someone really wants this to run. If it's not unfair, run it.
3323 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3326 clear_buddies(cfs_rq, se);
3331 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3333 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3336 * If still on the runqueue then deactivate_task()
3337 * was not called and update_curr() has to be done:
3340 update_curr(cfs_rq);
3342 /* throttle cfs_rqs exceeding runtime */
3343 check_cfs_rq_runtime(cfs_rq);
3345 check_spread(cfs_rq, prev);
3347 update_stats_wait_start(cfs_rq, prev);
3348 /* Put 'current' back into the tree. */
3349 __enqueue_entity(cfs_rq, prev);
3350 /* in !on_rq case, update occurred at dequeue */
3351 update_load_avg(prev, 0);
3353 cfs_rq->curr = NULL;
3357 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3360 * Update run-time statistics of the 'current'.
3362 update_curr(cfs_rq);
3365 * Ensure that runnable average is periodically updated.
3367 update_load_avg(curr, 1);
3368 update_cfs_shares(cfs_rq);
3370 #ifdef CONFIG_SCHED_HRTICK
3372 * queued ticks are scheduled to match the slice, so don't bother
3373 * validating it and just reschedule.
3376 resched_curr(rq_of(cfs_rq));
3380 * don't let the period tick interfere with the hrtick preemption
3382 if (!sched_feat(DOUBLE_TICK) &&
3383 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3387 if (cfs_rq->nr_running > 1)
3388 check_preempt_tick(cfs_rq, curr);
3392 /**************************************************
3393 * CFS bandwidth control machinery
3396 #ifdef CONFIG_CFS_BANDWIDTH
3398 #ifdef HAVE_JUMP_LABEL
3399 static struct static_key __cfs_bandwidth_used;
3401 static inline bool cfs_bandwidth_used(void)
3403 return static_key_false(&__cfs_bandwidth_used);
3406 void cfs_bandwidth_usage_inc(void)
3408 static_key_slow_inc(&__cfs_bandwidth_used);
3411 void cfs_bandwidth_usage_dec(void)
3413 static_key_slow_dec(&__cfs_bandwidth_used);
3415 #else /* HAVE_JUMP_LABEL */
3416 static bool cfs_bandwidth_used(void)
3421 void cfs_bandwidth_usage_inc(void) {}
3422 void cfs_bandwidth_usage_dec(void) {}
3423 #endif /* HAVE_JUMP_LABEL */
3426 * default period for cfs group bandwidth.
3427 * default: 0.1s, units: nanoseconds
3429 static inline u64 default_cfs_period(void)
3431 return 100000000ULL;
3434 static inline u64 sched_cfs_bandwidth_slice(void)
3436 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3440 * Replenish runtime according to assigned quota and update expiration time.
3441 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3442 * additional synchronization around rq->lock.
3444 * requires cfs_b->lock
3446 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3450 if (cfs_b->quota == RUNTIME_INF)
3453 now = sched_clock_cpu(smp_processor_id());
3454 cfs_b->runtime = cfs_b->quota;
3455 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3458 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3460 return &tg->cfs_bandwidth;
3463 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3464 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3466 if (unlikely(cfs_rq->throttle_count))
3467 return cfs_rq->throttled_clock_task;
3469 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3472 /* returns 0 on failure to allocate runtime */
3473 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3475 struct task_group *tg = cfs_rq->tg;
3476 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3477 u64 amount = 0, min_amount, expires;
3479 /* note: this is a positive sum as runtime_remaining <= 0 */
3480 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3482 raw_spin_lock(&cfs_b->lock);
3483 if (cfs_b->quota == RUNTIME_INF)
3484 amount = min_amount;
3486 start_cfs_bandwidth(cfs_b);
3488 if (cfs_b->runtime > 0) {
3489 amount = min(cfs_b->runtime, min_amount);
3490 cfs_b->runtime -= amount;
3494 expires = cfs_b->runtime_expires;
3495 raw_spin_unlock(&cfs_b->lock);
3497 cfs_rq->runtime_remaining += amount;
3499 * we may have advanced our local expiration to account for allowed
3500 * spread between our sched_clock and the one on which runtime was
3503 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3504 cfs_rq->runtime_expires = expires;
3506 return cfs_rq->runtime_remaining > 0;
3510 * Note: This depends on the synchronization provided by sched_clock and the
3511 * fact that rq->clock snapshots this value.
3513 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3515 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3517 /* if the deadline is ahead of our clock, nothing to do */
3518 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3521 if (cfs_rq->runtime_remaining < 0)
3525 * If the local deadline has passed we have to consider the
3526 * possibility that our sched_clock is 'fast' and the global deadline
3527 * has not truly expired.
3529 * Fortunately we can check determine whether this the case by checking
3530 * whether the global deadline has advanced. It is valid to compare
3531 * cfs_b->runtime_expires without any locks since we only care about
3532 * exact equality, so a partial write will still work.
3535 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3536 /* extend local deadline, drift is bounded above by 2 ticks */
3537 cfs_rq->runtime_expires += TICK_NSEC;
3539 /* global deadline is ahead, expiration has passed */
3540 cfs_rq->runtime_remaining = 0;
3544 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3546 /* dock delta_exec before expiring quota (as it could span periods) */
3547 cfs_rq->runtime_remaining -= delta_exec;
3548 expire_cfs_rq_runtime(cfs_rq);
3550 if (likely(cfs_rq->runtime_remaining > 0))
3554 * if we're unable to extend our runtime we resched so that the active
3555 * hierarchy can be throttled
3557 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3558 resched_curr(rq_of(cfs_rq));
3561 static __always_inline
3562 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3564 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3567 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3570 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3572 return cfs_bandwidth_used() && cfs_rq->throttled;
3575 /* check whether cfs_rq, or any parent, is throttled */
3576 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3578 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3582 * Ensure that neither of the group entities corresponding to src_cpu or
3583 * dest_cpu are members of a throttled hierarchy when performing group
3584 * load-balance operations.
3586 static inline int throttled_lb_pair(struct task_group *tg,
3587 int src_cpu, int dest_cpu)
3589 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3591 src_cfs_rq = tg->cfs_rq[src_cpu];
3592 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3594 return throttled_hierarchy(src_cfs_rq) ||
3595 throttled_hierarchy(dest_cfs_rq);
3598 /* updated child weight may affect parent so we have to do this bottom up */
3599 static int tg_unthrottle_up(struct task_group *tg, void *data)
3601 struct rq *rq = data;
3602 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3604 cfs_rq->throttle_count--;
3606 if (!cfs_rq->throttle_count) {
3607 /* adjust cfs_rq_clock_task() */
3608 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3609 cfs_rq->throttled_clock_task;
3616 static int tg_throttle_down(struct task_group *tg, void *data)
3618 struct rq *rq = data;
3619 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3621 /* group is entering throttled state, stop time */
3622 if (!cfs_rq->throttle_count)
3623 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3624 cfs_rq->throttle_count++;
3629 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3631 struct rq *rq = rq_of(cfs_rq);
3632 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3633 struct sched_entity *se;
3634 long task_delta, dequeue = 1;
3637 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3639 /* freeze hierarchy runnable averages while throttled */
3641 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3644 task_delta = cfs_rq->h_nr_running;
3645 for_each_sched_entity(se) {
3646 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3647 /* throttled entity or throttle-on-deactivate */
3652 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3653 qcfs_rq->h_nr_running -= task_delta;
3655 if (qcfs_rq->load.weight)
3660 sub_nr_running(rq, task_delta);
3662 cfs_rq->throttled = 1;
3663 cfs_rq->throttled_clock = rq_clock(rq);
3664 raw_spin_lock(&cfs_b->lock);
3665 empty = list_empty(&cfs_b->throttled_cfs_rq);
3668 * Add to the _head_ of the list, so that an already-started
3669 * distribute_cfs_runtime will not see us
3671 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3674 * If we're the first throttled task, make sure the bandwidth
3678 start_cfs_bandwidth(cfs_b);
3680 raw_spin_unlock(&cfs_b->lock);
3683 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3685 struct rq *rq = rq_of(cfs_rq);
3686 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3687 struct sched_entity *se;
3691 se = cfs_rq->tg->se[cpu_of(rq)];
3693 cfs_rq->throttled = 0;
3695 update_rq_clock(rq);
3697 raw_spin_lock(&cfs_b->lock);
3698 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3699 list_del_rcu(&cfs_rq->throttled_list);
3700 raw_spin_unlock(&cfs_b->lock);
3702 /* update hierarchical throttle state */
3703 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3705 if (!cfs_rq->load.weight)
3708 task_delta = cfs_rq->h_nr_running;
3709 for_each_sched_entity(se) {
3713 cfs_rq = cfs_rq_of(se);
3715 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3716 cfs_rq->h_nr_running += task_delta;
3718 if (cfs_rq_throttled(cfs_rq))
3723 add_nr_running(rq, task_delta);
3725 /* determine whether we need to wake up potentially idle cpu */
3726 if (rq->curr == rq->idle && rq->cfs.nr_running)
3730 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3731 u64 remaining, u64 expires)
3733 struct cfs_rq *cfs_rq;
3735 u64 starting_runtime = remaining;
3738 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3740 struct rq *rq = rq_of(cfs_rq);
3742 raw_spin_lock(&rq->lock);
3743 if (!cfs_rq_throttled(cfs_rq))
3746 runtime = -cfs_rq->runtime_remaining + 1;
3747 if (runtime > remaining)
3748 runtime = remaining;
3749 remaining -= runtime;
3751 cfs_rq->runtime_remaining += runtime;
3752 cfs_rq->runtime_expires = expires;
3754 /* we check whether we're throttled above */
3755 if (cfs_rq->runtime_remaining > 0)
3756 unthrottle_cfs_rq(cfs_rq);
3759 raw_spin_unlock(&rq->lock);
3766 return starting_runtime - remaining;
3770 * Responsible for refilling a task_group's bandwidth and unthrottling its
3771 * cfs_rqs as appropriate. If there has been no activity within the last
3772 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3773 * used to track this state.
3775 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3777 u64 runtime, runtime_expires;
3780 /* no need to continue the timer with no bandwidth constraint */
3781 if (cfs_b->quota == RUNTIME_INF)
3782 goto out_deactivate;
3784 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3785 cfs_b->nr_periods += overrun;
3788 * idle depends on !throttled (for the case of a large deficit), and if
3789 * we're going inactive then everything else can be deferred
3791 if (cfs_b->idle && !throttled)
3792 goto out_deactivate;
3794 __refill_cfs_bandwidth_runtime(cfs_b);
3797 /* mark as potentially idle for the upcoming period */
3802 /* account preceding periods in which throttling occurred */
3803 cfs_b->nr_throttled += overrun;
3805 runtime_expires = cfs_b->runtime_expires;
3808 * This check is repeated as we are holding onto the new bandwidth while
3809 * we unthrottle. This can potentially race with an unthrottled group
3810 * trying to acquire new bandwidth from the global pool. This can result
3811 * in us over-using our runtime if it is all used during this loop, but
3812 * only by limited amounts in that extreme case.
3814 while (throttled && cfs_b->runtime > 0) {
3815 runtime = cfs_b->runtime;
3816 raw_spin_unlock(&cfs_b->lock);
3817 /* we can't nest cfs_b->lock while distributing bandwidth */
3818 runtime = distribute_cfs_runtime(cfs_b, runtime,
3820 raw_spin_lock(&cfs_b->lock);
3822 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3824 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3828 * While we are ensured activity in the period following an
3829 * unthrottle, this also covers the case in which the new bandwidth is
3830 * insufficient to cover the existing bandwidth deficit. (Forcing the
3831 * timer to remain active while there are any throttled entities.)
3841 /* a cfs_rq won't donate quota below this amount */
3842 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3843 /* minimum remaining period time to redistribute slack quota */
3844 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3845 /* how long we wait to gather additional slack before distributing */
3846 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3849 * Are we near the end of the current quota period?
3851 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3852 * hrtimer base being cleared by hrtimer_start. In the case of
3853 * migrate_hrtimers, base is never cleared, so we are fine.
3855 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3857 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3860 /* if the call-back is running a quota refresh is already occurring */
3861 if (hrtimer_callback_running(refresh_timer))
3864 /* is a quota refresh about to occur? */
3865 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3866 if (remaining < min_expire)
3872 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3874 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3876 /* if there's a quota refresh soon don't bother with slack */
3877 if (runtime_refresh_within(cfs_b, min_left))
3880 hrtimer_start(&cfs_b->slack_timer,
3881 ns_to_ktime(cfs_bandwidth_slack_period),
3885 /* we know any runtime found here is valid as update_curr() precedes return */
3886 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3888 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3889 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3891 if (slack_runtime <= 0)
3894 raw_spin_lock(&cfs_b->lock);
3895 if (cfs_b->quota != RUNTIME_INF &&
3896 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3897 cfs_b->runtime += slack_runtime;
3899 /* we are under rq->lock, defer unthrottling using a timer */
3900 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3901 !list_empty(&cfs_b->throttled_cfs_rq))
3902 start_cfs_slack_bandwidth(cfs_b);
3904 raw_spin_unlock(&cfs_b->lock);
3906 /* even if it's not valid for return we don't want to try again */
3907 cfs_rq->runtime_remaining -= slack_runtime;
3910 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3912 if (!cfs_bandwidth_used())
3915 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3918 __return_cfs_rq_runtime(cfs_rq);
3922 * This is done with a timer (instead of inline with bandwidth return) since
3923 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3925 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3927 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3930 /* confirm we're still not at a refresh boundary */
3931 raw_spin_lock(&cfs_b->lock);
3932 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3933 raw_spin_unlock(&cfs_b->lock);
3937 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3938 runtime = cfs_b->runtime;
3940 expires = cfs_b->runtime_expires;
3941 raw_spin_unlock(&cfs_b->lock);
3946 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3948 raw_spin_lock(&cfs_b->lock);
3949 if (expires == cfs_b->runtime_expires)
3950 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3951 raw_spin_unlock(&cfs_b->lock);
3955 * When a group wakes up we want to make sure that its quota is not already
3956 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3957 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3959 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3961 if (!cfs_bandwidth_used())
3964 /* an active group must be handled by the update_curr()->put() path */
3965 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3968 /* ensure the group is not already throttled */
3969 if (cfs_rq_throttled(cfs_rq))
3972 /* update runtime allocation */
3973 account_cfs_rq_runtime(cfs_rq, 0);
3974 if (cfs_rq->runtime_remaining <= 0)
3975 throttle_cfs_rq(cfs_rq);
3978 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3979 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3981 if (!cfs_bandwidth_used())
3984 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3988 * it's possible for a throttled entity to be forced into a running
3989 * state (e.g. set_curr_task), in this case we're finished.
3991 if (cfs_rq_throttled(cfs_rq))
3994 throttle_cfs_rq(cfs_rq);
3998 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4000 struct cfs_bandwidth *cfs_b =
4001 container_of(timer, struct cfs_bandwidth, slack_timer);
4003 do_sched_cfs_slack_timer(cfs_b);
4005 return HRTIMER_NORESTART;
4008 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4010 struct cfs_bandwidth *cfs_b =
4011 container_of(timer, struct cfs_bandwidth, period_timer);
4015 raw_spin_lock(&cfs_b->lock);
4017 overrun = hrtimer_forward_now(timer, cfs_b->period);
4021 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4024 cfs_b->period_active = 0;
4025 raw_spin_unlock(&cfs_b->lock);
4027 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4030 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4032 raw_spin_lock_init(&cfs_b->lock);
4034 cfs_b->quota = RUNTIME_INF;
4035 cfs_b->period = ns_to_ktime(default_cfs_period());
4037 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4038 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4039 cfs_b->period_timer.function = sched_cfs_period_timer;
4040 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4041 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4044 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4046 cfs_rq->runtime_enabled = 0;
4047 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4050 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4052 lockdep_assert_held(&cfs_b->lock);
4054 if (!cfs_b->period_active) {
4055 cfs_b->period_active = 1;
4056 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4057 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4061 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4063 /* init_cfs_bandwidth() was not called */
4064 if (!cfs_b->throttled_cfs_rq.next)
4067 hrtimer_cancel(&cfs_b->period_timer);
4068 hrtimer_cancel(&cfs_b->slack_timer);
4071 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4073 struct cfs_rq *cfs_rq;
4075 for_each_leaf_cfs_rq(rq, cfs_rq) {
4076 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4078 raw_spin_lock(&cfs_b->lock);
4079 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4080 raw_spin_unlock(&cfs_b->lock);
4084 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4086 struct cfs_rq *cfs_rq;
4088 for_each_leaf_cfs_rq(rq, cfs_rq) {
4089 if (!cfs_rq->runtime_enabled)
4093 * clock_task is not advancing so we just need to make sure
4094 * there's some valid quota amount
4096 cfs_rq->runtime_remaining = 1;
4098 * Offline rq is schedulable till cpu is completely disabled
4099 * in take_cpu_down(), so we prevent new cfs throttling here.
4101 cfs_rq->runtime_enabled = 0;
4103 if (cfs_rq_throttled(cfs_rq))
4104 unthrottle_cfs_rq(cfs_rq);
4108 #else /* CONFIG_CFS_BANDWIDTH */
4109 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4111 return rq_clock_task(rq_of(cfs_rq));
4114 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4115 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4116 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4117 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4119 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4124 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4129 static inline int throttled_lb_pair(struct task_group *tg,
4130 int src_cpu, int dest_cpu)
4135 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4137 #ifdef CONFIG_FAIR_GROUP_SCHED
4138 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4141 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4145 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4146 static inline void update_runtime_enabled(struct rq *rq) {}
4147 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4149 #endif /* CONFIG_CFS_BANDWIDTH */
4151 /**************************************************
4152 * CFS operations on tasks:
4155 #ifdef CONFIG_SCHED_HRTICK
4156 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4158 struct sched_entity *se = &p->se;
4159 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4161 WARN_ON(task_rq(p) != rq);
4163 if (cfs_rq->nr_running > 1) {
4164 u64 slice = sched_slice(cfs_rq, se);
4165 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4166 s64 delta = slice - ran;
4173 hrtick_start(rq, delta);
4178 * called from enqueue/dequeue and updates the hrtick when the
4179 * current task is from our class and nr_running is low enough
4182 static void hrtick_update(struct rq *rq)
4184 struct task_struct *curr = rq->curr;
4186 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4189 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4190 hrtick_start_fair(rq, curr);
4192 #else /* !CONFIG_SCHED_HRTICK */
4194 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4198 static inline void hrtick_update(struct rq *rq)
4204 static bool cpu_overutilized(int cpu);
4205 static inline unsigned long boosted_cpu_util(int cpu);
4207 #define boosted_cpu_util(cpu) cpu_util(cpu)
4211 static void update_capacity_of(int cpu)
4213 unsigned long req_cap;
4218 /* Convert scale-invariant capacity to cpu. */
4219 req_cap = boosted_cpu_util(cpu);
4220 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4221 set_cfs_cpu_capacity(cpu, true, req_cap);
4226 * The enqueue_task method is called before nr_running is
4227 * increased. Here we update the fair scheduling stats and
4228 * then put the task into the rbtree:
4231 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4233 struct cfs_rq *cfs_rq;
4234 struct sched_entity *se = &p->se;
4236 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4237 int task_wakeup = flags & ENQUEUE_WAKEUP;
4240 for_each_sched_entity(se) {
4243 cfs_rq = cfs_rq_of(se);
4244 enqueue_entity(cfs_rq, se, flags);
4247 * end evaluation on encountering a throttled cfs_rq
4249 * note: in the case of encountering a throttled cfs_rq we will
4250 * post the final h_nr_running increment below.
4252 if (cfs_rq_throttled(cfs_rq))
4254 cfs_rq->h_nr_running++;
4255 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4257 flags = ENQUEUE_WAKEUP;
4260 for_each_sched_entity(se) {
4261 cfs_rq = cfs_rq_of(se);
4262 cfs_rq->h_nr_running++;
4263 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4265 if (cfs_rq_throttled(cfs_rq))
4268 update_load_avg(se, 1);
4269 update_cfs_shares(cfs_rq);
4273 add_nr_running(rq, 1);
4278 * Update SchedTune accounting.
4280 * We do it before updating the CPU capacity to ensure the
4281 * boost value of the current task is accounted for in the
4282 * selection of the OPP.
4284 * We do it also in the case where we enqueue a throttled task;
4285 * we could argue that a throttled task should not boost a CPU,
4287 * a) properly implementing CPU boosting considering throttled
4288 * tasks will increase a lot the complexity of the solution
4289 * b) it's not easy to quantify the benefits introduced by
4290 * such a more complex solution.
4291 * Thus, for the time being we go for the simple solution and boost
4292 * also for throttled RQs.
4294 schedtune_enqueue_task(p, cpu_of(rq));
4297 walt_inc_cumulative_runnable_avg(rq, p);
4298 if (!task_new && !rq->rd->overutilized &&
4299 cpu_overutilized(rq->cpu)) {
4300 rq->rd->overutilized = true;
4301 trace_sched_overutilized(true);
4305 * We want to potentially trigger a freq switch
4306 * request only for tasks that are waking up; this is
4307 * because we get here also during load balancing, but
4308 * in these cases it seems wise to trigger as single
4309 * request after load balancing is done.
4311 if (task_new || task_wakeup)
4312 update_capacity_of(cpu_of(rq));
4315 #endif /* CONFIG_SMP */
4319 static void set_next_buddy(struct sched_entity *se);
4322 * The dequeue_task method is called before nr_running is
4323 * decreased. We remove the task from the rbtree and
4324 * update the fair scheduling stats:
4326 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4328 struct cfs_rq *cfs_rq;
4329 struct sched_entity *se = &p->se;
4330 int task_sleep = flags & DEQUEUE_SLEEP;
4332 for_each_sched_entity(se) {
4333 cfs_rq = cfs_rq_of(se);
4334 dequeue_entity(cfs_rq, se, flags);
4337 * end evaluation on encountering a throttled cfs_rq
4339 * note: in the case of encountering a throttled cfs_rq we will
4340 * post the final h_nr_running decrement below.
4342 if (cfs_rq_throttled(cfs_rq))
4344 cfs_rq->h_nr_running--;
4345 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4347 /* Don't dequeue parent if it has other entities besides us */
4348 if (cfs_rq->load.weight) {
4350 * Bias pick_next to pick a task from this cfs_rq, as
4351 * p is sleeping when it is within its sched_slice.
4353 if (task_sleep && parent_entity(se))
4354 set_next_buddy(parent_entity(se));
4356 /* avoid re-evaluating load for this entity */
4357 se = parent_entity(se);
4360 flags |= DEQUEUE_SLEEP;
4363 for_each_sched_entity(se) {
4364 cfs_rq = cfs_rq_of(se);
4365 cfs_rq->h_nr_running--;
4366 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4368 if (cfs_rq_throttled(cfs_rq))
4371 update_load_avg(se, 1);
4372 update_cfs_shares(cfs_rq);
4376 sub_nr_running(rq, 1);
4381 * Update SchedTune accounting
4383 * We do it before updating the CPU capacity to ensure the
4384 * boost value of the current task is accounted for in the
4385 * selection of the OPP.
4387 schedtune_dequeue_task(p, cpu_of(rq));
4390 walt_dec_cumulative_runnable_avg(rq, p);
4393 * We want to potentially trigger a freq switch
4394 * request only for tasks that are going to sleep;
4395 * this is because we get here also during load
4396 * balancing, but in these cases it seems wise to
4397 * trigger as single request after load balancing is
4401 if (rq->cfs.nr_running)
4402 update_capacity_of(cpu_of(rq));
4403 else if (sched_freq())
4404 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4408 #endif /* CONFIG_SMP */
4416 * per rq 'load' arrray crap; XXX kill this.
4420 * The exact cpuload at various idx values, calculated at every tick would be
4421 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4423 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4424 * on nth tick when cpu may be busy, then we have:
4425 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4426 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4428 * decay_load_missed() below does efficient calculation of
4429 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4430 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4432 * The calculation is approximated on a 128 point scale.
4433 * degrade_zero_ticks is the number of ticks after which load at any
4434 * particular idx is approximated to be zero.
4435 * degrade_factor is a precomputed table, a row for each load idx.
4436 * Each column corresponds to degradation factor for a power of two ticks,
4437 * based on 128 point scale.
4439 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4440 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4442 * With this power of 2 load factors, we can degrade the load n times
4443 * by looking at 1 bits in n and doing as many mult/shift instead of
4444 * n mult/shifts needed by the exact degradation.
4446 #define DEGRADE_SHIFT 7
4447 static const unsigned char
4448 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4449 static const unsigned char
4450 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4451 {0, 0, 0, 0, 0, 0, 0, 0},
4452 {64, 32, 8, 0, 0, 0, 0, 0},
4453 {96, 72, 40, 12, 1, 0, 0},
4454 {112, 98, 75, 43, 15, 1, 0},
4455 {120, 112, 98, 76, 45, 16, 2} };
4458 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4459 * would be when CPU is idle and so we just decay the old load without
4460 * adding any new load.
4462 static unsigned long
4463 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4467 if (!missed_updates)
4470 if (missed_updates >= degrade_zero_ticks[idx])
4474 return load >> missed_updates;
4476 while (missed_updates) {
4477 if (missed_updates % 2)
4478 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4480 missed_updates >>= 1;
4487 * Update rq->cpu_load[] statistics. This function is usually called every
4488 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4489 * every tick. We fix it up based on jiffies.
4491 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4492 unsigned long pending_updates)
4496 this_rq->nr_load_updates++;
4498 /* Update our load: */
4499 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4500 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4501 unsigned long old_load, new_load;
4503 /* scale is effectively 1 << i now, and >> i divides by scale */
4505 old_load = this_rq->cpu_load[i];
4506 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4507 new_load = this_load;
4509 * Round up the averaging division if load is increasing. This
4510 * prevents us from getting stuck on 9 if the load is 10, for
4513 if (new_load > old_load)
4514 new_load += scale - 1;
4516 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4519 sched_avg_update(this_rq);
4522 /* Used instead of source_load when we know the type == 0 */
4523 static unsigned long weighted_cpuload(const int cpu)
4525 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4528 #ifdef CONFIG_NO_HZ_COMMON
4530 * There is no sane way to deal with nohz on smp when using jiffies because the
4531 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4532 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4534 * Therefore we cannot use the delta approach from the regular tick since that
4535 * would seriously skew the load calculation. However we'll make do for those
4536 * updates happening while idle (nohz_idle_balance) or coming out of idle
4537 * (tick_nohz_idle_exit).
4539 * This means we might still be one tick off for nohz periods.
4543 * Called from nohz_idle_balance() to update the load ratings before doing the
4546 static void update_idle_cpu_load(struct rq *this_rq)
4548 unsigned long curr_jiffies = READ_ONCE(jiffies);
4549 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4550 unsigned long pending_updates;
4553 * bail if there's load or we're actually up-to-date.
4555 if (load || curr_jiffies == this_rq->last_load_update_tick)
4558 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4559 this_rq->last_load_update_tick = curr_jiffies;
4561 __update_cpu_load(this_rq, load, pending_updates);
4565 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4567 void update_cpu_load_nohz(void)
4569 struct rq *this_rq = this_rq();
4570 unsigned long curr_jiffies = READ_ONCE(jiffies);
4571 unsigned long pending_updates;
4573 if (curr_jiffies == this_rq->last_load_update_tick)
4576 raw_spin_lock(&this_rq->lock);
4577 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4578 if (pending_updates) {
4579 this_rq->last_load_update_tick = curr_jiffies;
4581 * We were idle, this means load 0, the current load might be
4582 * !0 due to remote wakeups and the sort.
4584 __update_cpu_load(this_rq, 0, pending_updates);
4586 raw_spin_unlock(&this_rq->lock);
4588 #endif /* CONFIG_NO_HZ */
4591 * Called from scheduler_tick()
4593 void update_cpu_load_active(struct rq *this_rq)
4595 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4597 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4599 this_rq->last_load_update_tick = jiffies;
4600 __update_cpu_load(this_rq, load, 1);
4604 * Return a low guess at the load of a migration-source cpu weighted
4605 * according to the scheduling class and "nice" value.
4607 * We want to under-estimate the load of migration sources, to
4608 * balance conservatively.
4610 static unsigned long source_load(int cpu, int type)
4612 struct rq *rq = cpu_rq(cpu);
4613 unsigned long total = weighted_cpuload(cpu);
4615 if (type == 0 || !sched_feat(LB_BIAS))
4618 return min(rq->cpu_load[type-1], total);
4622 * Return a high guess at the load of a migration-target cpu weighted
4623 * according to the scheduling class and "nice" value.
4625 static unsigned long target_load(int cpu, int type)
4627 struct rq *rq = cpu_rq(cpu);
4628 unsigned long total = weighted_cpuload(cpu);
4630 if (type == 0 || !sched_feat(LB_BIAS))
4633 return max(rq->cpu_load[type-1], total);
4637 static unsigned long cpu_avg_load_per_task(int cpu)
4639 struct rq *rq = cpu_rq(cpu);
4640 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4641 unsigned long load_avg = weighted_cpuload(cpu);
4644 return load_avg / nr_running;
4649 static void record_wakee(struct task_struct *p)
4652 * Rough decay (wiping) for cost saving, don't worry
4653 * about the boundary, really active task won't care
4656 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4657 current->wakee_flips >>= 1;
4658 current->wakee_flip_decay_ts = jiffies;
4661 if (current->last_wakee != p) {
4662 current->last_wakee = p;
4663 current->wakee_flips++;
4667 static void task_waking_fair(struct task_struct *p)
4669 struct sched_entity *se = &p->se;
4670 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4673 #ifndef CONFIG_64BIT
4674 u64 min_vruntime_copy;
4677 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4679 min_vruntime = cfs_rq->min_vruntime;
4680 } while (min_vruntime != min_vruntime_copy);
4682 min_vruntime = cfs_rq->min_vruntime;
4685 se->vruntime -= min_vruntime;
4689 #ifdef CONFIG_FAIR_GROUP_SCHED
4691 * effective_load() calculates the load change as seen from the root_task_group
4693 * Adding load to a group doesn't make a group heavier, but can cause movement
4694 * of group shares between cpus. Assuming the shares were perfectly aligned one
4695 * can calculate the shift in shares.
4697 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4698 * on this @cpu and results in a total addition (subtraction) of @wg to the
4699 * total group weight.
4701 * Given a runqueue weight distribution (rw_i) we can compute a shares
4702 * distribution (s_i) using:
4704 * s_i = rw_i / \Sum rw_j (1)
4706 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4707 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4708 * shares distribution (s_i):
4710 * rw_i = { 2, 4, 1, 0 }
4711 * s_i = { 2/7, 4/7, 1/7, 0 }
4713 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4714 * task used to run on and the CPU the waker is running on), we need to
4715 * compute the effect of waking a task on either CPU and, in case of a sync
4716 * wakeup, compute the effect of the current task going to sleep.
4718 * So for a change of @wl to the local @cpu with an overall group weight change
4719 * of @wl we can compute the new shares distribution (s'_i) using:
4721 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4723 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4724 * differences in waking a task to CPU 0. The additional task changes the
4725 * weight and shares distributions like:
4727 * rw'_i = { 3, 4, 1, 0 }
4728 * s'_i = { 3/8, 4/8, 1/8, 0 }
4730 * We can then compute the difference in effective weight by using:
4732 * dw_i = S * (s'_i - s_i) (3)
4734 * Where 'S' is the group weight as seen by its parent.
4736 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4737 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4738 * 4/7) times the weight of the group.
4740 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4742 struct sched_entity *se = tg->se[cpu];
4744 if (!tg->parent) /* the trivial, non-cgroup case */
4747 for_each_sched_entity(se) {
4748 struct cfs_rq *cfs_rq = se->my_q;
4749 long W, w = cfs_rq_load_avg(cfs_rq);
4754 * W = @wg + \Sum rw_j
4756 W = wg + atomic_long_read(&tg->load_avg);
4758 /* Ensure \Sum rw_j >= rw_i */
4759 W -= cfs_rq->tg_load_avg_contrib;
4768 * wl = S * s'_i; see (2)
4771 wl = (w * (long)tg->shares) / W;
4776 * Per the above, wl is the new se->load.weight value; since
4777 * those are clipped to [MIN_SHARES, ...) do so now. See
4778 * calc_cfs_shares().
4780 if (wl < MIN_SHARES)
4784 * wl = dw_i = S * (s'_i - s_i); see (3)
4786 wl -= se->avg.load_avg;
4789 * Recursively apply this logic to all parent groups to compute
4790 * the final effective load change on the root group. Since
4791 * only the @tg group gets extra weight, all parent groups can
4792 * only redistribute existing shares. @wl is the shift in shares
4793 * resulting from this level per the above.
4802 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4810 * Returns the current capacity of cpu after applying both
4811 * cpu and freq scaling.
4813 unsigned long capacity_curr_of(int cpu)
4815 return cpu_rq(cpu)->cpu_capacity_orig *
4816 arch_scale_freq_capacity(NULL, cpu)
4817 >> SCHED_CAPACITY_SHIFT;
4820 static inline bool energy_aware(void)
4822 return sched_feat(ENERGY_AWARE);
4826 struct sched_group *sg_top;
4827 struct sched_group *sg_cap;
4834 struct task_struct *task;
4849 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4850 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4851 * energy calculations. Using the scale-invariant util returned by
4852 * cpu_util() and approximating scale-invariant util by:
4854 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4856 * the normalized util can be found using the specific capacity.
4858 * capacity = capacity_orig * curr_freq/max_freq
4860 * norm_util = running_time/time ~ util/capacity
4862 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4864 int util = __cpu_util(cpu, delta);
4866 if (util >= capacity)
4867 return SCHED_CAPACITY_SCALE;
4869 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4872 static int calc_util_delta(struct energy_env *eenv, int cpu)
4874 if (cpu == eenv->src_cpu)
4875 return -eenv->util_delta;
4876 if (cpu == eenv->dst_cpu)
4877 return eenv->util_delta;
4882 unsigned long group_max_util(struct energy_env *eenv)
4885 unsigned long max_util = 0;
4887 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4888 delta = calc_util_delta(eenv, i);
4889 max_util = max(max_util, __cpu_util(i, delta));
4896 * group_norm_util() returns the approximated group util relative to it's
4897 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4898 * energy calculations. Since task executions may or may not overlap in time in
4899 * the group the true normalized util is between max(cpu_norm_util(i)) and
4900 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4901 * latter is used as the estimate as it leads to a more pessimistic energy
4902 * estimate (more busy).
4905 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4908 unsigned long util_sum = 0;
4909 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4911 for_each_cpu(i, sched_group_cpus(sg)) {
4912 delta = calc_util_delta(eenv, i);
4913 util_sum += __cpu_norm_util(i, capacity, delta);
4916 if (util_sum > SCHED_CAPACITY_SCALE)
4917 return SCHED_CAPACITY_SCALE;
4921 static int find_new_capacity(struct energy_env *eenv,
4922 const struct sched_group_energy * const sge)
4925 unsigned long util = group_max_util(eenv);
4927 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4928 if (sge->cap_states[idx].cap >= util)
4932 eenv->cap_idx = idx;
4937 static int group_idle_state(struct sched_group *sg)
4939 int i, state = INT_MAX;
4941 /* Find the shallowest idle state in the sched group. */
4942 for_each_cpu(i, sched_group_cpus(sg))
4943 state = min(state, idle_get_state_idx(cpu_rq(i)));
4945 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4952 * sched_group_energy(): Computes the absolute energy consumption of cpus
4953 * belonging to the sched_group including shared resources shared only by
4954 * members of the group. Iterates over all cpus in the hierarchy below the
4955 * sched_group starting from the bottom working it's way up before going to
4956 * the next cpu until all cpus are covered at all levels. The current
4957 * implementation is likely to gather the same util statistics multiple times.
4958 * This can probably be done in a faster but more complex way.
4959 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4961 static int sched_group_energy(struct energy_env *eenv)
4963 struct sched_domain *sd;
4964 int cpu, total_energy = 0;
4965 struct cpumask visit_cpus;
4966 struct sched_group *sg;
4968 WARN_ON(!eenv->sg_top->sge);
4970 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4972 while (!cpumask_empty(&visit_cpus)) {
4973 struct sched_group *sg_shared_cap = NULL;
4975 cpu = cpumask_first(&visit_cpus);
4978 * Is the group utilization affected by cpus outside this
4981 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4985 * We most probably raced with hotplug; returning a
4986 * wrong energy estimation is better than entering an
4992 sg_shared_cap = sd->parent->groups;
4994 for_each_domain(cpu, sd) {
4997 /* Has this sched_domain already been visited? */
4998 if (sd->child && group_first_cpu(sg) != cpu)
5002 unsigned long group_util;
5003 int sg_busy_energy, sg_idle_energy;
5004 int cap_idx, idle_idx;
5006 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5007 eenv->sg_cap = sg_shared_cap;
5011 cap_idx = find_new_capacity(eenv, sg->sge);
5013 if (sg->group_weight == 1) {
5014 /* Remove capacity of src CPU (before task move) */
5015 if (eenv->util_delta == 0 &&
5016 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5017 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5018 eenv->cap.delta -= eenv->cap.before;
5020 /* Add capacity of dst CPU (after task move) */
5021 if (eenv->util_delta != 0 &&
5022 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5023 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5024 eenv->cap.delta += eenv->cap.after;
5028 idle_idx = group_idle_state(sg);
5029 group_util = group_norm_util(eenv, sg);
5030 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5031 >> SCHED_CAPACITY_SHIFT;
5032 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5033 * sg->sge->idle_states[idle_idx].power)
5034 >> SCHED_CAPACITY_SHIFT;
5036 total_energy += sg_busy_energy + sg_idle_energy;
5039 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5041 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5044 } while (sg = sg->next, sg != sd->groups);
5047 cpumask_clear_cpu(cpu, &visit_cpus);
5051 eenv->energy = total_energy;
5055 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5057 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5061 * energy_diff(): Estimate the energy impact of changing the utilization
5062 * distribution. eenv specifies the change: utilisation amount, source, and
5063 * destination cpu. Source or destination cpu may be -1 in which case the
5064 * utilization is removed from or added to the system (e.g. task wake-up). If
5065 * both are specified, the utilization is migrated.
5067 static inline int __energy_diff(struct energy_env *eenv)
5069 struct sched_domain *sd;
5070 struct sched_group *sg;
5071 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5073 struct energy_env eenv_before = {
5075 .src_cpu = eenv->src_cpu,
5076 .dst_cpu = eenv->dst_cpu,
5077 .nrg = { 0, 0, 0, 0},
5081 if (eenv->src_cpu == eenv->dst_cpu)
5084 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5085 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5088 return 0; /* Error */
5093 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5094 eenv_before.sg_top = eenv->sg_top = sg;
5096 if (sched_group_energy(&eenv_before))
5097 return 0; /* Invalid result abort */
5098 energy_before += eenv_before.energy;
5100 /* Keep track of SRC cpu (before) capacity */
5101 eenv->cap.before = eenv_before.cap.before;
5102 eenv->cap.delta = eenv_before.cap.delta;
5104 if (sched_group_energy(eenv))
5105 return 0; /* Invalid result abort */
5106 energy_after += eenv->energy;
5108 } while (sg = sg->next, sg != sd->groups);
5110 eenv->nrg.before = energy_before;
5111 eenv->nrg.after = energy_after;
5112 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5115 trace_sched_energy_diff(eenv->task,
5116 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5117 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5118 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5119 eenv->nrg.delta, eenv->payoff);
5121 return eenv->nrg.diff;
5124 #ifdef CONFIG_SCHED_TUNE
5126 struct target_nrg schedtune_target_nrg;
5129 * System energy normalization
5130 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5131 * corresponding to the specified energy variation.
5134 normalize_energy(int energy_diff)
5137 #ifdef CONFIG_SCHED_DEBUG
5140 /* Check for boundaries */
5141 max_delta = schedtune_target_nrg.max_power;
5142 max_delta -= schedtune_target_nrg.min_power;
5143 WARN_ON(abs(energy_diff) >= max_delta);
5146 /* Do scaling using positive numbers to increase the range */
5147 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5149 /* Scale by energy magnitude */
5150 normalized_nrg <<= SCHED_LOAD_SHIFT;
5152 /* Normalize on max energy for target platform */
5153 normalized_nrg = reciprocal_divide(
5154 normalized_nrg, schedtune_target_nrg.rdiv);
5156 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5160 energy_diff(struct energy_env *eenv)
5162 int boost = schedtune_task_boost(eenv->task);
5165 /* Conpute "absolute" energy diff */
5166 __energy_diff(eenv);
5168 /* Return energy diff when boost margin is 0 */
5170 return eenv->nrg.diff;
5172 /* Compute normalized energy diff */
5173 nrg_delta = normalize_energy(eenv->nrg.diff);
5174 eenv->nrg.delta = nrg_delta;
5176 eenv->payoff = schedtune_accept_deltas(
5182 * When SchedTune is enabled, the energy_diff() function will return
5183 * the computed energy payoff value. Since the energy_diff() return
5184 * value is expected to be negative by its callers, this evaluation
5185 * function return a negative value each time the evaluation return a
5186 * positive payoff, which is the condition for the acceptance of
5187 * a scheduling decision
5189 return -eenv->payoff;
5191 #else /* CONFIG_SCHED_TUNE */
5192 #define energy_diff(eenv) __energy_diff(eenv)
5196 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5197 * A waker of many should wake a different task than the one last awakened
5198 * at a frequency roughly N times higher than one of its wakees. In order
5199 * to determine whether we should let the load spread vs consolodating to
5200 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5201 * partner, and a factor of lls_size higher frequency in the other. With
5202 * both conditions met, we can be relatively sure that the relationship is
5203 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5204 * being client/server, worker/dispatcher, interrupt source or whatever is
5205 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5207 static int wake_wide(struct task_struct *p)
5209 unsigned int master = current->wakee_flips;
5210 unsigned int slave = p->wakee_flips;
5211 int factor = this_cpu_read(sd_llc_size);
5214 swap(master, slave);
5215 if (slave < factor || master < slave * factor)
5220 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5222 s64 this_load, load;
5223 s64 this_eff_load, prev_eff_load;
5224 int idx, this_cpu, prev_cpu;
5225 struct task_group *tg;
5226 unsigned long weight;
5230 this_cpu = smp_processor_id();
5231 prev_cpu = task_cpu(p);
5232 load = source_load(prev_cpu, idx);
5233 this_load = target_load(this_cpu, idx);
5236 * If sync wakeup then subtract the (maximum possible)
5237 * effect of the currently running task from the load
5238 * of the current CPU:
5241 tg = task_group(current);
5242 weight = current->se.avg.load_avg;
5244 this_load += effective_load(tg, this_cpu, -weight, -weight);
5245 load += effective_load(tg, prev_cpu, 0, -weight);
5249 weight = p->se.avg.load_avg;
5252 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5253 * due to the sync cause above having dropped this_load to 0, we'll
5254 * always have an imbalance, but there's really nothing you can do
5255 * about that, so that's good too.
5257 * Otherwise check if either cpus are near enough in load to allow this
5258 * task to be woken on this_cpu.
5260 this_eff_load = 100;
5261 this_eff_load *= capacity_of(prev_cpu);
5263 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5264 prev_eff_load *= capacity_of(this_cpu);
5266 if (this_load > 0) {
5267 this_eff_load *= this_load +
5268 effective_load(tg, this_cpu, weight, weight);
5270 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5273 balanced = this_eff_load <= prev_eff_load;
5275 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5280 schedstat_inc(sd, ttwu_move_affine);
5281 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5286 static inline unsigned long task_util(struct task_struct *p)
5288 #ifdef CONFIG_SCHED_WALT
5289 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5290 unsigned long demand = p->ravg.demand;
5291 return (demand << 10) / walt_ravg_window;
5294 return p->se.avg.util_avg;
5297 unsigned int capacity_margin = 1280; /* ~20% margin */
5299 static inline unsigned long boosted_task_util(struct task_struct *task);
5301 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5303 unsigned long capacity = capacity_of(cpu);
5305 util += boosted_task_util(p);
5307 return (capacity * 1024) > (util * capacity_margin);
5310 static inline bool task_fits_max(struct task_struct *p, int cpu)
5312 unsigned long capacity = capacity_of(cpu);
5313 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5315 if (capacity == max_capacity)
5318 if (capacity * capacity_margin > max_capacity * 1024)
5321 return __task_fits(p, cpu, 0);
5324 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5326 return __task_fits(p, cpu, cpu_util(cpu));
5329 static bool cpu_overutilized(int cpu)
5331 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5334 #ifdef CONFIG_SCHED_TUNE
5337 schedtune_margin(unsigned long signal, long boost)
5339 long long margin = 0;
5342 * Signal proportional compensation (SPC)
5344 * The Boost (B) value is used to compute a Margin (M) which is
5345 * proportional to the complement of the original Signal (S):
5346 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5347 * M = B * S, if B is negative
5348 * The obtained M could be used by the caller to "boost" S.
5351 margin = SCHED_LOAD_SCALE - signal;
5354 margin = -signal * boost;
5356 * Fast integer division by constant:
5357 * Constant : (C) = 100
5358 * Precision : 0.1% (P) = 0.1
5359 * Reference : C * 100 / P (R) = 100000
5362 * Shift bits : ceil(log(R,2)) (S) = 17
5363 * Mult const : round(2^S/C) (M) = 1311
5376 schedtune_cpu_margin(unsigned long util, int cpu)
5378 int boost = schedtune_cpu_boost(cpu);
5383 return schedtune_margin(util, boost);
5387 schedtune_task_margin(struct task_struct *task)
5389 int boost = schedtune_task_boost(task);
5396 util = task_util(task);
5397 margin = schedtune_margin(util, boost);
5402 #else /* CONFIG_SCHED_TUNE */
5405 schedtune_cpu_margin(unsigned long util, int cpu)
5411 schedtune_task_margin(struct task_struct *task)
5416 #endif /* CONFIG_SCHED_TUNE */
5418 static inline unsigned long
5419 boosted_cpu_util(int cpu)
5421 unsigned long util = cpu_util(cpu);
5422 long margin = schedtune_cpu_margin(util, cpu);
5424 trace_sched_boost_cpu(cpu, util, margin);
5426 return util + margin;
5429 static inline unsigned long
5430 boosted_task_util(struct task_struct *task)
5432 unsigned long util = task_util(task);
5433 long margin = schedtune_task_margin(task);
5435 trace_sched_boost_task(task, util, margin);
5437 return util + margin;
5441 * find_idlest_group finds and returns the least busy CPU group within the
5444 static struct sched_group *
5445 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5446 int this_cpu, int sd_flag)
5448 struct sched_group *idlest = NULL, *group = sd->groups;
5449 struct sched_group *fit_group = NULL, *spare_group = NULL;
5450 unsigned long min_load = ULONG_MAX, this_load = 0;
5451 unsigned long fit_capacity = ULONG_MAX;
5452 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5453 int load_idx = sd->forkexec_idx;
5454 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5456 if (sd_flag & SD_BALANCE_WAKE)
5457 load_idx = sd->wake_idx;
5460 unsigned long load, avg_load, spare_capacity;
5464 /* Skip over this group if it has no CPUs allowed */
5465 if (!cpumask_intersects(sched_group_cpus(group),
5466 tsk_cpus_allowed(p)))
5469 local_group = cpumask_test_cpu(this_cpu,
5470 sched_group_cpus(group));
5472 /* Tally up the load of all CPUs in the group */
5475 for_each_cpu(i, sched_group_cpus(group)) {
5476 /* Bias balancing toward cpus of our domain */
5478 load = source_load(i, load_idx);
5480 load = target_load(i, load_idx);
5485 * Look for most energy-efficient group that can fit
5486 * that can fit the task.
5488 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5489 fit_capacity = capacity_of(i);
5494 * Look for group which has most spare capacity on a
5497 spare_capacity = capacity_of(i) - cpu_util(i);
5498 if (spare_capacity > max_spare_capacity) {
5499 max_spare_capacity = spare_capacity;
5500 spare_group = group;
5504 /* Adjust by relative CPU capacity of the group */
5505 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5508 this_load = avg_load;
5509 } else if (avg_load < min_load) {
5510 min_load = avg_load;
5513 } while (group = group->next, group != sd->groups);
5521 if (!idlest || 100*this_load < imbalance*min_load)
5527 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5530 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5532 unsigned long load, min_load = ULONG_MAX;
5533 unsigned int min_exit_latency = UINT_MAX;
5534 u64 latest_idle_timestamp = 0;
5535 int least_loaded_cpu = this_cpu;
5536 int shallowest_idle_cpu = -1;
5539 /* Traverse only the allowed CPUs */
5540 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5541 if (task_fits_spare(p, i)) {
5542 struct rq *rq = cpu_rq(i);
5543 struct cpuidle_state *idle = idle_get_state(rq);
5544 if (idle && idle->exit_latency < min_exit_latency) {
5546 * We give priority to a CPU whose idle state
5547 * has the smallest exit latency irrespective
5548 * of any idle timestamp.
5550 min_exit_latency = idle->exit_latency;
5551 latest_idle_timestamp = rq->idle_stamp;
5552 shallowest_idle_cpu = i;
5553 } else if (idle_cpu(i) &&
5554 (!idle || idle->exit_latency == min_exit_latency) &&
5555 rq->idle_stamp > latest_idle_timestamp) {
5557 * If equal or no active idle state, then
5558 * the most recently idled CPU might have
5561 latest_idle_timestamp = rq->idle_stamp;
5562 shallowest_idle_cpu = i;
5563 } else if (shallowest_idle_cpu == -1) {
5565 * If we haven't found an idle CPU yet
5566 * pick a non-idle one that can fit the task as
5569 shallowest_idle_cpu = i;
5571 } else if (shallowest_idle_cpu == -1) {
5572 load = weighted_cpuload(i);
5573 if (load < min_load || (load == min_load && i == this_cpu)) {
5575 least_loaded_cpu = i;
5580 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5584 * Try and locate an idle CPU in the sched_domain.
5586 static int select_idle_sibling(struct task_struct *p, int target)
5588 struct sched_domain *sd;
5589 struct sched_group *sg;
5590 int i = task_cpu(p);
5592 int best_idle_cstate = -1;
5593 int best_idle_capacity = INT_MAX;
5595 if (!sysctl_sched_cstate_aware) {
5596 if (idle_cpu(target))
5600 * If the prevous cpu is cache affine and idle, don't be stupid.
5602 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5607 * Otherwise, iterate the domains and find an elegible idle cpu.
5609 sd = rcu_dereference(per_cpu(sd_llc, target));
5610 for_each_lower_domain(sd) {
5613 if (!cpumask_intersects(sched_group_cpus(sg),
5614 tsk_cpus_allowed(p)))
5617 if (sysctl_sched_cstate_aware) {
5618 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5619 struct rq *rq = cpu_rq(i);
5620 int idle_idx = idle_get_state_idx(rq);
5621 unsigned long new_usage = boosted_task_util(p);
5622 unsigned long capacity_orig = capacity_orig_of(i);
5623 if (new_usage > capacity_orig || !idle_cpu(i))
5626 if (i == target && new_usage <= capacity_curr_of(target))
5629 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5631 best_idle_cstate = idle_idx;
5632 best_idle_capacity = capacity_orig;
5636 for_each_cpu(i, sched_group_cpus(sg)) {
5637 if (i == target || !idle_cpu(i))
5641 target = cpumask_first_and(sched_group_cpus(sg),
5642 tsk_cpus_allowed(p));
5647 } while (sg != sd->groups);
5656 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5659 int target_cpu = -1;
5660 int target_util = 0;
5661 int backup_capacity = 0;
5662 int best_idle_cpu = -1;
5663 int best_idle_cstate = INT_MAX;
5664 int backup_cpu = -1;
5665 unsigned long task_util_boosted, new_util;
5667 task_util_boosted = boosted_task_util(p);
5668 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5674 * Iterate from higher cpus for boosted tasks.
5676 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5678 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5682 * p's blocked utilization is still accounted for on prev_cpu
5683 * so prev_cpu will receive a negative bias due to the double
5684 * accounting. However, the blocked utilization may be zero.
5686 new_util = cpu_util(i) + task_util_boosted;
5689 * Ensure minimum capacity to grant the required boost.
5690 * The target CPU can be already at a capacity level higher
5691 * than the one required to boost the task.
5693 if (new_util > capacity_orig_of(i))
5696 #ifdef CONFIG_SCHED_WALT
5697 if (walt_cpu_high_irqload(i))
5701 * Unconditionally favoring tasks that prefer idle cpus to
5704 if (idle_cpu(i) && prefer_idle) {
5705 if (best_idle_cpu < 0)
5710 cur_capacity = capacity_curr_of(i);
5712 idle_idx = idle_get_state_idx(rq);
5714 if (new_util < cur_capacity) {
5715 if (cpu_rq(i)->nr_running) {
5717 /* Find a target cpu with highest
5720 if (target_util == 0 ||
5721 target_util < new_util) {
5723 target_util = new_util;
5726 /* Find a target cpu with lowest
5729 if (target_util == 0 ||
5730 target_util > new_util) {
5732 target_util = new_util;
5735 } else if (!prefer_idle) {
5736 if (best_idle_cpu < 0 ||
5737 (sysctl_sched_cstate_aware &&
5738 best_idle_cstate > idle_idx)) {
5739 best_idle_cstate = idle_idx;
5743 } else if (backup_capacity == 0 ||
5744 backup_capacity > cur_capacity) {
5745 // Find a backup cpu with least capacity.
5746 backup_capacity = cur_capacity;
5751 if (prefer_idle && best_idle_cpu >= 0)
5752 target_cpu = best_idle_cpu;
5753 else if (target_cpu < 0)
5754 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5759 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5761 struct sched_domain *sd;
5762 struct sched_group *sg, *sg_target;
5763 int target_max_cap = INT_MAX;
5764 int target_cpu = task_cpu(p);
5765 unsigned long task_util_boosted, new_util;
5768 if (sysctl_sched_sync_hint_enable && sync) {
5769 int cpu = smp_processor_id();
5770 cpumask_t search_cpus;
5771 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5772 if (cpumask_test_cpu(cpu, &search_cpus))
5776 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5784 if (sysctl_sched_is_big_little) {
5787 * Find group with sufficient capacity. We only get here if no cpu is
5788 * overutilized. We may end up overutilizing a cpu by adding the task,
5789 * but that should not be any worse than select_idle_sibling().
5790 * load_balance() should sort it out later as we get above the tipping
5794 /* Assuming all cpus are the same in group */
5795 int max_cap_cpu = group_first_cpu(sg);
5798 * Assume smaller max capacity means more energy-efficient.
5799 * Ideally we should query the energy model for the right
5800 * answer but it easily ends up in an exhaustive search.
5802 if (capacity_of(max_cap_cpu) < target_max_cap &&
5803 task_fits_max(p, max_cap_cpu)) {
5805 target_max_cap = capacity_of(max_cap_cpu);
5807 } while (sg = sg->next, sg != sd->groups);
5809 task_util_boosted = boosted_task_util(p);
5810 /* Find cpu with sufficient capacity */
5811 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5813 * p's blocked utilization is still accounted for on prev_cpu
5814 * so prev_cpu will receive a negative bias due to the double
5815 * accounting. However, the blocked utilization may be zero.
5817 new_util = cpu_util(i) + task_util_boosted;
5820 * Ensure minimum capacity to grant the required boost.
5821 * The target CPU can be already at a capacity level higher
5822 * than the one required to boost the task.
5824 if (new_util > capacity_orig_of(i))
5827 if (new_util < capacity_curr_of(i)) {
5829 if (cpu_rq(i)->nr_running)
5833 /* cpu has capacity at higher OPP, keep it as fallback */
5834 if (target_cpu == task_cpu(p))
5839 * Find a cpu with sufficient capacity
5841 #ifdef CONFIG_CGROUP_SCHEDTUNE
5842 bool boosted = schedtune_task_boost(p) > 0;
5843 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5846 bool prefer_idle = 0;
5848 int tmp_target = find_best_target(p, boosted, prefer_idle);
5849 if (tmp_target >= 0) {
5850 target_cpu = tmp_target;
5851 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5856 if (target_cpu != task_cpu(p)) {
5857 struct energy_env eenv = {
5858 .util_delta = task_util(p),
5859 .src_cpu = task_cpu(p),
5860 .dst_cpu = target_cpu,
5864 /* Not enough spare capacity on previous cpu */
5865 if (cpu_overutilized(task_cpu(p)))
5868 if (energy_diff(&eenv) >= 0)
5876 * select_task_rq_fair: Select target runqueue for the waking task in domains
5877 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5878 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5880 * Balances load by selecting the idlest cpu in the idlest group, or under
5881 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5883 * Returns the target cpu number.
5885 * preempt must be disabled.
5888 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5890 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5891 int cpu = smp_processor_id();
5892 int new_cpu = prev_cpu;
5893 int want_affine = 0;
5894 int sync = wake_flags & WF_SYNC;
5896 if (sd_flag & SD_BALANCE_WAKE)
5897 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5898 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5902 for_each_domain(cpu, tmp) {
5903 if (!(tmp->flags & SD_LOAD_BALANCE))
5907 * If both cpu and prev_cpu are part of this domain,
5908 * cpu is a valid SD_WAKE_AFFINE target.
5910 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5911 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5916 if (tmp->flags & sd_flag)
5918 else if (!want_affine)
5923 sd = NULL; /* Prefer wake_affine over balance flags */
5924 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5929 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5930 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5931 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5932 new_cpu = select_idle_sibling(p, new_cpu);
5935 struct sched_group *group;
5938 if (!(sd->flags & sd_flag)) {
5943 group = find_idlest_group(sd, p, cpu, sd_flag);
5949 new_cpu = find_idlest_cpu(group, p, cpu);
5950 if (new_cpu == -1 || new_cpu == cpu) {
5951 /* Now try balancing at a lower domain level of cpu */
5956 /* Now try balancing at a lower domain level of new_cpu */
5958 weight = sd->span_weight;
5960 for_each_domain(cpu, tmp) {
5961 if (weight <= tmp->span_weight)
5963 if (tmp->flags & sd_flag)
5966 /* while loop will break here if sd == NULL */
5974 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5975 * cfs_rq_of(p) references at time of call are still valid and identify the
5976 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5977 * other assumptions, including the state of rq->lock, should be made.
5979 static void migrate_task_rq_fair(struct task_struct *p)
5982 * We are supposed to update the task to "current" time, then its up to date
5983 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5984 * what current time is, so simply throw away the out-of-date time. This
5985 * will result in the wakee task is less decayed, but giving the wakee more
5986 * load sounds not bad.
5988 remove_entity_load_avg(&p->se);
5990 /* Tell new CPU we are migrated */
5991 p->se.avg.last_update_time = 0;
5993 /* We have migrated, no longer consider this task hot */
5994 p->se.exec_start = 0;
5997 static void task_dead_fair(struct task_struct *p)
5999 remove_entity_load_avg(&p->se);
6002 #define task_fits_max(p, cpu) true
6003 #endif /* CONFIG_SMP */
6005 static unsigned long
6006 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6008 unsigned long gran = sysctl_sched_wakeup_granularity;
6011 * Since its curr running now, convert the gran from real-time
6012 * to virtual-time in his units.
6014 * By using 'se' instead of 'curr' we penalize light tasks, so
6015 * they get preempted easier. That is, if 'se' < 'curr' then
6016 * the resulting gran will be larger, therefore penalizing the
6017 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6018 * be smaller, again penalizing the lighter task.
6020 * This is especially important for buddies when the leftmost
6021 * task is higher priority than the buddy.
6023 return calc_delta_fair(gran, se);
6027 * Should 'se' preempt 'curr'.
6041 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6043 s64 gran, vdiff = curr->vruntime - se->vruntime;
6048 gran = wakeup_gran(curr, se);
6055 static void set_last_buddy(struct sched_entity *se)
6057 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6060 for_each_sched_entity(se)
6061 cfs_rq_of(se)->last = se;
6064 static void set_next_buddy(struct sched_entity *se)
6066 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6069 for_each_sched_entity(se)
6070 cfs_rq_of(se)->next = se;
6073 static void set_skip_buddy(struct sched_entity *se)
6075 for_each_sched_entity(se)
6076 cfs_rq_of(se)->skip = se;
6080 * Preempt the current task with a newly woken task if needed:
6082 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6084 struct task_struct *curr = rq->curr;
6085 struct sched_entity *se = &curr->se, *pse = &p->se;
6086 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6087 int scale = cfs_rq->nr_running >= sched_nr_latency;
6088 int next_buddy_marked = 0;
6090 if (unlikely(se == pse))
6094 * This is possible from callers such as attach_tasks(), in which we
6095 * unconditionally check_prempt_curr() after an enqueue (which may have
6096 * lead to a throttle). This both saves work and prevents false
6097 * next-buddy nomination below.
6099 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6102 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6103 set_next_buddy(pse);
6104 next_buddy_marked = 1;
6108 * We can come here with TIF_NEED_RESCHED already set from new task
6111 * Note: this also catches the edge-case of curr being in a throttled
6112 * group (e.g. via set_curr_task), since update_curr() (in the
6113 * enqueue of curr) will have resulted in resched being set. This
6114 * prevents us from potentially nominating it as a false LAST_BUDDY
6117 if (test_tsk_need_resched(curr))
6120 /* Idle tasks are by definition preempted by non-idle tasks. */
6121 if (unlikely(curr->policy == SCHED_IDLE) &&
6122 likely(p->policy != SCHED_IDLE))
6126 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6127 * is driven by the tick):
6129 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6132 find_matching_se(&se, &pse);
6133 update_curr(cfs_rq_of(se));
6135 if (wakeup_preempt_entity(se, pse) == 1) {
6137 * Bias pick_next to pick the sched entity that is
6138 * triggering this preemption.
6140 if (!next_buddy_marked)
6141 set_next_buddy(pse);
6150 * Only set the backward buddy when the current task is still
6151 * on the rq. This can happen when a wakeup gets interleaved
6152 * with schedule on the ->pre_schedule() or idle_balance()
6153 * point, either of which can * drop the rq lock.
6155 * Also, during early boot the idle thread is in the fair class,
6156 * for obvious reasons its a bad idea to schedule back to it.
6158 if (unlikely(!se->on_rq || curr == rq->idle))
6161 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6165 static struct task_struct *
6166 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6168 struct cfs_rq *cfs_rq = &rq->cfs;
6169 struct sched_entity *se;
6170 struct task_struct *p;
6174 #ifdef CONFIG_FAIR_GROUP_SCHED
6175 if (!cfs_rq->nr_running)
6178 if (prev->sched_class != &fair_sched_class)
6182 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6183 * likely that a next task is from the same cgroup as the current.
6185 * Therefore attempt to avoid putting and setting the entire cgroup
6186 * hierarchy, only change the part that actually changes.
6190 struct sched_entity *curr = cfs_rq->curr;
6193 * Since we got here without doing put_prev_entity() we also
6194 * have to consider cfs_rq->curr. If it is still a runnable
6195 * entity, update_curr() will update its vruntime, otherwise
6196 * forget we've ever seen it.
6200 update_curr(cfs_rq);
6205 * This call to check_cfs_rq_runtime() will do the
6206 * throttle and dequeue its entity in the parent(s).
6207 * Therefore the 'simple' nr_running test will indeed
6210 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6214 se = pick_next_entity(cfs_rq, curr);
6215 cfs_rq = group_cfs_rq(se);
6221 * Since we haven't yet done put_prev_entity and if the selected task
6222 * is a different task than we started out with, try and touch the
6223 * least amount of cfs_rqs.
6226 struct sched_entity *pse = &prev->se;
6228 while (!(cfs_rq = is_same_group(se, pse))) {
6229 int se_depth = se->depth;
6230 int pse_depth = pse->depth;
6232 if (se_depth <= pse_depth) {
6233 put_prev_entity(cfs_rq_of(pse), pse);
6234 pse = parent_entity(pse);
6236 if (se_depth >= pse_depth) {
6237 set_next_entity(cfs_rq_of(se), se);
6238 se = parent_entity(se);
6242 put_prev_entity(cfs_rq, pse);
6243 set_next_entity(cfs_rq, se);
6246 if (hrtick_enabled(rq))
6247 hrtick_start_fair(rq, p);
6249 rq->misfit_task = !task_fits_max(p, rq->cpu);
6256 if (!cfs_rq->nr_running)
6259 put_prev_task(rq, prev);
6262 se = pick_next_entity(cfs_rq, NULL);
6263 set_next_entity(cfs_rq, se);
6264 cfs_rq = group_cfs_rq(se);
6269 if (hrtick_enabled(rq))
6270 hrtick_start_fair(rq, p);
6272 rq->misfit_task = !task_fits_max(p, rq->cpu);
6277 rq->misfit_task = 0;
6279 * This is OK, because current is on_cpu, which avoids it being picked
6280 * for load-balance and preemption/IRQs are still disabled avoiding
6281 * further scheduler activity on it and we're being very careful to
6282 * re-start the picking loop.
6284 lockdep_unpin_lock(&rq->lock);
6285 new_tasks = idle_balance(rq);
6286 lockdep_pin_lock(&rq->lock);
6288 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6289 * possible for any higher priority task to appear. In that case we
6290 * must re-start the pick_next_entity() loop.
6302 * Account for a descheduled task:
6304 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6306 struct sched_entity *se = &prev->se;
6307 struct cfs_rq *cfs_rq;
6309 for_each_sched_entity(se) {
6310 cfs_rq = cfs_rq_of(se);
6311 put_prev_entity(cfs_rq, se);
6316 * sched_yield() is very simple
6318 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6320 static void yield_task_fair(struct rq *rq)
6322 struct task_struct *curr = rq->curr;
6323 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6324 struct sched_entity *se = &curr->se;
6327 * Are we the only task in the tree?
6329 if (unlikely(rq->nr_running == 1))
6332 clear_buddies(cfs_rq, se);
6334 if (curr->policy != SCHED_BATCH) {
6335 update_rq_clock(rq);
6337 * Update run-time statistics of the 'current'.
6339 update_curr(cfs_rq);
6341 * Tell update_rq_clock() that we've just updated,
6342 * so we don't do microscopic update in schedule()
6343 * and double the fastpath cost.
6345 rq_clock_skip_update(rq, true);
6351 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6353 struct sched_entity *se = &p->se;
6355 /* throttled hierarchies are not runnable */
6356 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6359 /* Tell the scheduler that we'd really like pse to run next. */
6362 yield_task_fair(rq);
6368 /**************************************************
6369 * Fair scheduling class load-balancing methods.
6373 * The purpose of load-balancing is to achieve the same basic fairness the
6374 * per-cpu scheduler provides, namely provide a proportional amount of compute
6375 * time to each task. This is expressed in the following equation:
6377 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6379 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6380 * W_i,0 is defined as:
6382 * W_i,0 = \Sum_j w_i,j (2)
6384 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6385 * is derived from the nice value as per prio_to_weight[].
6387 * The weight average is an exponential decay average of the instantaneous
6390 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6392 * C_i is the compute capacity of cpu i, typically it is the
6393 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6394 * can also include other factors [XXX].
6396 * To achieve this balance we define a measure of imbalance which follows
6397 * directly from (1):
6399 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6401 * We them move tasks around to minimize the imbalance. In the continuous
6402 * function space it is obvious this converges, in the discrete case we get
6403 * a few fun cases generally called infeasible weight scenarios.
6406 * - infeasible weights;
6407 * - local vs global optima in the discrete case. ]
6412 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6413 * for all i,j solution, we create a tree of cpus that follows the hardware
6414 * topology where each level pairs two lower groups (or better). This results
6415 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6416 * tree to only the first of the previous level and we decrease the frequency
6417 * of load-balance at each level inv. proportional to the number of cpus in
6423 * \Sum { --- * --- * 2^i } = O(n) (5)
6425 * `- size of each group
6426 * | | `- number of cpus doing load-balance
6428 * `- sum over all levels
6430 * Coupled with a limit on how many tasks we can migrate every balance pass,
6431 * this makes (5) the runtime complexity of the balancer.
6433 * An important property here is that each CPU is still (indirectly) connected
6434 * to every other cpu in at most O(log n) steps:
6436 * The adjacency matrix of the resulting graph is given by:
6439 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6442 * And you'll find that:
6444 * A^(log_2 n)_i,j != 0 for all i,j (7)
6446 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6447 * The task movement gives a factor of O(m), giving a convergence complexity
6450 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6455 * In order to avoid CPUs going idle while there's still work to do, new idle
6456 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6457 * tree itself instead of relying on other CPUs to bring it work.
6459 * This adds some complexity to both (5) and (8) but it reduces the total idle
6467 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6470 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6475 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6477 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6479 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6482 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6483 * rewrite all of this once again.]
6486 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6488 enum fbq_type { regular, remote, all };
6497 #define LBF_ALL_PINNED 0x01
6498 #define LBF_NEED_BREAK 0x02
6499 #define LBF_DST_PINNED 0x04
6500 #define LBF_SOME_PINNED 0x08
6503 struct sched_domain *sd;
6511 struct cpumask *dst_grpmask;
6513 enum cpu_idle_type idle;
6515 unsigned int src_grp_nr_running;
6516 /* The set of CPUs under consideration for load-balancing */
6517 struct cpumask *cpus;
6522 unsigned int loop_break;
6523 unsigned int loop_max;
6525 enum fbq_type fbq_type;
6526 enum group_type busiest_group_type;
6527 struct list_head tasks;
6531 * Is this task likely cache-hot:
6533 static int task_hot(struct task_struct *p, struct lb_env *env)
6537 lockdep_assert_held(&env->src_rq->lock);
6539 if (p->sched_class != &fair_sched_class)
6542 if (unlikely(p->policy == SCHED_IDLE))
6546 * Buddy candidates are cache hot:
6548 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6549 (&p->se == cfs_rq_of(&p->se)->next ||
6550 &p->se == cfs_rq_of(&p->se)->last))
6553 if (sysctl_sched_migration_cost == -1)
6555 if (sysctl_sched_migration_cost == 0)
6558 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6560 return delta < (s64)sysctl_sched_migration_cost;
6563 #ifdef CONFIG_NUMA_BALANCING
6565 * Returns 1, if task migration degrades locality
6566 * Returns 0, if task migration improves locality i.e migration preferred.
6567 * Returns -1, if task migration is not affected by locality.
6569 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6571 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6572 unsigned long src_faults, dst_faults;
6573 int src_nid, dst_nid;
6575 if (!static_branch_likely(&sched_numa_balancing))
6578 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6581 src_nid = cpu_to_node(env->src_cpu);
6582 dst_nid = cpu_to_node(env->dst_cpu);
6584 if (src_nid == dst_nid)
6587 /* Migrating away from the preferred node is always bad. */
6588 if (src_nid == p->numa_preferred_nid) {
6589 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6595 /* Encourage migration to the preferred node. */
6596 if (dst_nid == p->numa_preferred_nid)
6600 src_faults = group_faults(p, src_nid);
6601 dst_faults = group_faults(p, dst_nid);
6603 src_faults = task_faults(p, src_nid);
6604 dst_faults = task_faults(p, dst_nid);
6607 return dst_faults < src_faults;
6611 static inline int migrate_degrades_locality(struct task_struct *p,
6619 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6622 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6626 lockdep_assert_held(&env->src_rq->lock);
6629 * We do not migrate tasks that are:
6630 * 1) throttled_lb_pair, or
6631 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6632 * 3) running (obviously), or
6633 * 4) are cache-hot on their current CPU.
6635 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6638 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6641 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6643 env->flags |= LBF_SOME_PINNED;
6646 * Remember if this task can be migrated to any other cpu in
6647 * our sched_group. We may want to revisit it if we couldn't
6648 * meet load balance goals by pulling other tasks on src_cpu.
6650 * Also avoid computing new_dst_cpu if we have already computed
6651 * one in current iteration.
6653 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6656 /* Prevent to re-select dst_cpu via env's cpus */
6657 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6658 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6659 env->flags |= LBF_DST_PINNED;
6660 env->new_dst_cpu = cpu;
6668 /* Record that we found atleast one task that could run on dst_cpu */
6669 env->flags &= ~LBF_ALL_PINNED;
6671 if (task_running(env->src_rq, p)) {
6672 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6677 * Aggressive migration if:
6678 * 1) destination numa is preferred
6679 * 2) task is cache cold, or
6680 * 3) too many balance attempts have failed.
6682 tsk_cache_hot = migrate_degrades_locality(p, env);
6683 if (tsk_cache_hot == -1)
6684 tsk_cache_hot = task_hot(p, env);
6686 if (tsk_cache_hot <= 0 ||
6687 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6688 if (tsk_cache_hot == 1) {
6689 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6690 schedstat_inc(p, se.statistics.nr_forced_migrations);
6695 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6700 * detach_task() -- detach the task for the migration specified in env
6702 static void detach_task(struct task_struct *p, struct lb_env *env)
6704 lockdep_assert_held(&env->src_rq->lock);
6706 deactivate_task(env->src_rq, p, 0);
6707 p->on_rq = TASK_ON_RQ_MIGRATING;
6708 double_lock_balance(env->src_rq, env->dst_rq);
6709 set_task_cpu(p, env->dst_cpu);
6710 double_unlock_balance(env->src_rq, env->dst_rq);
6714 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6715 * part of active balancing operations within "domain".
6717 * Returns a task if successful and NULL otherwise.
6719 static struct task_struct *detach_one_task(struct lb_env *env)
6721 struct task_struct *p, *n;
6723 lockdep_assert_held(&env->src_rq->lock);
6725 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6726 if (!can_migrate_task(p, env))
6729 detach_task(p, env);
6732 * Right now, this is only the second place where
6733 * lb_gained[env->idle] is updated (other is detach_tasks)
6734 * so we can safely collect stats here rather than
6735 * inside detach_tasks().
6737 schedstat_inc(env->sd, lb_gained[env->idle]);
6743 static const unsigned int sched_nr_migrate_break = 32;
6746 * detach_tasks() -- tries to detach up to imbalance weighted load from
6747 * busiest_rq, as part of a balancing operation within domain "sd".
6749 * Returns number of detached tasks if successful and 0 otherwise.
6751 static int detach_tasks(struct lb_env *env)
6753 struct list_head *tasks = &env->src_rq->cfs_tasks;
6754 struct task_struct *p;
6758 lockdep_assert_held(&env->src_rq->lock);
6760 if (env->imbalance <= 0)
6763 while (!list_empty(tasks)) {
6765 * We don't want to steal all, otherwise we may be treated likewise,
6766 * which could at worst lead to a livelock crash.
6768 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6771 p = list_first_entry(tasks, struct task_struct, se.group_node);
6774 /* We've more or less seen every task there is, call it quits */
6775 if (env->loop > env->loop_max)
6778 /* take a breather every nr_migrate tasks */
6779 if (env->loop > env->loop_break) {
6780 env->loop_break += sched_nr_migrate_break;
6781 env->flags |= LBF_NEED_BREAK;
6785 if (!can_migrate_task(p, env))
6788 load = task_h_load(p);
6790 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6793 if ((load / 2) > env->imbalance)
6796 detach_task(p, env);
6797 list_add(&p->se.group_node, &env->tasks);
6800 env->imbalance -= load;
6802 #ifdef CONFIG_PREEMPT
6804 * NEWIDLE balancing is a source of latency, so preemptible
6805 * kernels will stop after the first task is detached to minimize
6806 * the critical section.
6808 if (env->idle == CPU_NEWLY_IDLE)
6813 * We only want to steal up to the prescribed amount of
6816 if (env->imbalance <= 0)
6821 list_move_tail(&p->se.group_node, tasks);
6825 * Right now, this is one of only two places we collect this stat
6826 * so we can safely collect detach_one_task() stats here rather
6827 * than inside detach_one_task().
6829 schedstat_add(env->sd, lb_gained[env->idle], detached);
6835 * attach_task() -- attach the task detached by detach_task() to its new rq.
6837 static void attach_task(struct rq *rq, struct task_struct *p)
6839 lockdep_assert_held(&rq->lock);
6841 BUG_ON(task_rq(p) != rq);
6842 p->on_rq = TASK_ON_RQ_QUEUED;
6843 activate_task(rq, p, 0);
6844 check_preempt_curr(rq, p, 0);
6848 * attach_one_task() -- attaches the task returned from detach_one_task() to
6851 static void attach_one_task(struct rq *rq, struct task_struct *p)
6853 raw_spin_lock(&rq->lock);
6856 * We want to potentially raise target_cpu's OPP.
6858 update_capacity_of(cpu_of(rq));
6859 raw_spin_unlock(&rq->lock);
6863 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6866 static void attach_tasks(struct lb_env *env)
6868 struct list_head *tasks = &env->tasks;
6869 struct task_struct *p;
6871 raw_spin_lock(&env->dst_rq->lock);
6873 while (!list_empty(tasks)) {
6874 p = list_first_entry(tasks, struct task_struct, se.group_node);
6875 list_del_init(&p->se.group_node);
6877 attach_task(env->dst_rq, p);
6881 * We want to potentially raise env.dst_cpu's OPP.
6883 update_capacity_of(env->dst_cpu);
6885 raw_spin_unlock(&env->dst_rq->lock);
6888 #ifdef CONFIG_FAIR_GROUP_SCHED
6889 static void update_blocked_averages(int cpu)
6891 struct rq *rq = cpu_rq(cpu);
6892 struct cfs_rq *cfs_rq;
6893 unsigned long flags;
6895 raw_spin_lock_irqsave(&rq->lock, flags);
6896 update_rq_clock(rq);
6899 * Iterates the task_group tree in a bottom up fashion, see
6900 * list_add_leaf_cfs_rq() for details.
6902 for_each_leaf_cfs_rq(rq, cfs_rq) {
6903 /* throttled entities do not contribute to load */
6904 if (throttled_hierarchy(cfs_rq))
6907 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6908 update_tg_load_avg(cfs_rq, 0);
6910 raw_spin_unlock_irqrestore(&rq->lock, flags);
6914 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6915 * This needs to be done in a top-down fashion because the load of a child
6916 * group is a fraction of its parents load.
6918 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6920 struct rq *rq = rq_of(cfs_rq);
6921 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6922 unsigned long now = jiffies;
6925 if (cfs_rq->last_h_load_update == now)
6928 cfs_rq->h_load_next = NULL;
6929 for_each_sched_entity(se) {
6930 cfs_rq = cfs_rq_of(se);
6931 cfs_rq->h_load_next = se;
6932 if (cfs_rq->last_h_load_update == now)
6937 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6938 cfs_rq->last_h_load_update = now;
6941 while ((se = cfs_rq->h_load_next) != NULL) {
6942 load = cfs_rq->h_load;
6943 load = div64_ul(load * se->avg.load_avg,
6944 cfs_rq_load_avg(cfs_rq) + 1);
6945 cfs_rq = group_cfs_rq(se);
6946 cfs_rq->h_load = load;
6947 cfs_rq->last_h_load_update = now;
6951 static unsigned long task_h_load(struct task_struct *p)
6953 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6955 update_cfs_rq_h_load(cfs_rq);
6956 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6957 cfs_rq_load_avg(cfs_rq) + 1);
6960 static inline void update_blocked_averages(int cpu)
6962 struct rq *rq = cpu_rq(cpu);
6963 struct cfs_rq *cfs_rq = &rq->cfs;
6964 unsigned long flags;
6966 raw_spin_lock_irqsave(&rq->lock, flags);
6967 update_rq_clock(rq);
6968 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6969 raw_spin_unlock_irqrestore(&rq->lock, flags);
6972 static unsigned long task_h_load(struct task_struct *p)
6974 return p->se.avg.load_avg;
6978 /********** Helpers for find_busiest_group ************************/
6981 * sg_lb_stats - stats of a sched_group required for load_balancing
6983 struct sg_lb_stats {
6984 unsigned long avg_load; /*Avg load across the CPUs of the group */
6985 unsigned long group_load; /* Total load over the CPUs of the group */
6986 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6987 unsigned long load_per_task;
6988 unsigned long group_capacity;
6989 unsigned long group_util; /* Total utilization of the group */
6990 unsigned int sum_nr_running; /* Nr tasks running in the group */
6991 unsigned int idle_cpus;
6992 unsigned int group_weight;
6993 enum group_type group_type;
6994 int group_no_capacity;
6995 int group_misfit_task; /* A cpu has a task too big for its capacity */
6996 #ifdef CONFIG_NUMA_BALANCING
6997 unsigned int nr_numa_running;
6998 unsigned int nr_preferred_running;
7003 * sd_lb_stats - Structure to store the statistics of a sched_domain
7004 * during load balancing.
7006 struct sd_lb_stats {
7007 struct sched_group *busiest; /* Busiest group in this sd */
7008 struct sched_group *local; /* Local group in this sd */
7009 unsigned long total_load; /* Total load of all groups in sd */
7010 unsigned long total_capacity; /* Total capacity of all groups in sd */
7011 unsigned long avg_load; /* Average load across all groups in sd */
7013 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7014 struct sg_lb_stats local_stat; /* Statistics of the local group */
7017 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7020 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7021 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7022 * We must however clear busiest_stat::avg_load because
7023 * update_sd_pick_busiest() reads this before assignment.
7025 *sds = (struct sd_lb_stats){
7029 .total_capacity = 0UL,
7032 .sum_nr_running = 0,
7033 .group_type = group_other,
7039 * get_sd_load_idx - Obtain the load index for a given sched domain.
7040 * @sd: The sched_domain whose load_idx is to be obtained.
7041 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7043 * Return: The load index.
7045 static inline int get_sd_load_idx(struct sched_domain *sd,
7046 enum cpu_idle_type idle)
7052 load_idx = sd->busy_idx;
7055 case CPU_NEWLY_IDLE:
7056 load_idx = sd->newidle_idx;
7059 load_idx = sd->idle_idx;
7066 static unsigned long scale_rt_capacity(int cpu)
7068 struct rq *rq = cpu_rq(cpu);
7069 u64 total, used, age_stamp, avg;
7073 * Since we're reading these variables without serialization make sure
7074 * we read them once before doing sanity checks on them.
7076 age_stamp = READ_ONCE(rq->age_stamp);
7077 avg = READ_ONCE(rq->rt_avg);
7078 delta = __rq_clock_broken(rq) - age_stamp;
7080 if (unlikely(delta < 0))
7083 total = sched_avg_period() + delta;
7085 used = div_u64(avg, total);
7088 * deadline bandwidth is defined at system level so we must
7089 * weight this bandwidth with the max capacity of the system.
7090 * As a reminder, avg_bw is 20bits width and
7091 * scale_cpu_capacity is 10 bits width
7093 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7095 if (likely(used < SCHED_CAPACITY_SCALE))
7096 return SCHED_CAPACITY_SCALE - used;
7101 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7103 raw_spin_lock_init(&mcc->lock);
7108 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7110 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7111 struct sched_group *sdg = sd->groups;
7112 struct max_cpu_capacity *mcc;
7113 unsigned long max_capacity;
7115 unsigned long flags;
7117 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7119 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7121 raw_spin_lock_irqsave(&mcc->lock, flags);
7122 max_capacity = mcc->val;
7123 max_cap_cpu = mcc->cpu;
7125 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7126 (max_capacity < capacity)) {
7127 mcc->val = capacity;
7129 #ifdef CONFIG_SCHED_DEBUG
7130 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7131 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7136 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7138 skip_unlock: __attribute__ ((unused));
7139 capacity *= scale_rt_capacity(cpu);
7140 capacity >>= SCHED_CAPACITY_SHIFT;
7145 cpu_rq(cpu)->cpu_capacity = capacity;
7146 sdg->sgc->capacity = capacity;
7147 sdg->sgc->max_capacity = capacity;
7150 void update_group_capacity(struct sched_domain *sd, int cpu)
7152 struct sched_domain *child = sd->child;
7153 struct sched_group *group, *sdg = sd->groups;
7154 unsigned long capacity, max_capacity;
7155 unsigned long interval;
7157 interval = msecs_to_jiffies(sd->balance_interval);
7158 interval = clamp(interval, 1UL, max_load_balance_interval);
7159 sdg->sgc->next_update = jiffies + interval;
7162 update_cpu_capacity(sd, cpu);
7169 if (child->flags & SD_OVERLAP) {
7171 * SD_OVERLAP domains cannot assume that child groups
7172 * span the current group.
7175 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7176 struct sched_group_capacity *sgc;
7177 struct rq *rq = cpu_rq(cpu);
7180 * build_sched_domains() -> init_sched_groups_capacity()
7181 * gets here before we've attached the domains to the
7184 * Use capacity_of(), which is set irrespective of domains
7185 * in update_cpu_capacity().
7187 * This avoids capacity from being 0 and
7188 * causing divide-by-zero issues on boot.
7190 if (unlikely(!rq->sd)) {
7191 capacity += capacity_of(cpu);
7193 sgc = rq->sd->groups->sgc;
7194 capacity += sgc->capacity;
7197 max_capacity = max(capacity, max_capacity);
7201 * !SD_OVERLAP domains can assume that child groups
7202 * span the current group.
7205 group = child->groups;
7207 struct sched_group_capacity *sgc = group->sgc;
7209 capacity += sgc->capacity;
7210 max_capacity = max(sgc->max_capacity, max_capacity);
7211 group = group->next;
7212 } while (group != child->groups);
7215 sdg->sgc->capacity = capacity;
7216 sdg->sgc->max_capacity = max_capacity;
7220 * Check whether the capacity of the rq has been noticeably reduced by side
7221 * activity. The imbalance_pct is used for the threshold.
7222 * Return true is the capacity is reduced
7225 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7227 return ((rq->cpu_capacity * sd->imbalance_pct) <
7228 (rq->cpu_capacity_orig * 100));
7232 * Group imbalance indicates (and tries to solve) the problem where balancing
7233 * groups is inadequate due to tsk_cpus_allowed() constraints.
7235 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7236 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7239 * { 0 1 2 3 } { 4 5 6 7 }
7242 * If we were to balance group-wise we'd place two tasks in the first group and
7243 * two tasks in the second group. Clearly this is undesired as it will overload
7244 * cpu 3 and leave one of the cpus in the second group unused.
7246 * The current solution to this issue is detecting the skew in the first group
7247 * by noticing the lower domain failed to reach balance and had difficulty
7248 * moving tasks due to affinity constraints.
7250 * When this is so detected; this group becomes a candidate for busiest; see
7251 * update_sd_pick_busiest(). And calculate_imbalance() and
7252 * find_busiest_group() avoid some of the usual balance conditions to allow it
7253 * to create an effective group imbalance.
7255 * This is a somewhat tricky proposition since the next run might not find the
7256 * group imbalance and decide the groups need to be balanced again. A most
7257 * subtle and fragile situation.
7260 static inline int sg_imbalanced(struct sched_group *group)
7262 return group->sgc->imbalance;
7266 * group_has_capacity returns true if the group has spare capacity that could
7267 * be used by some tasks.
7268 * We consider that a group has spare capacity if the * number of task is
7269 * smaller than the number of CPUs or if the utilization is lower than the
7270 * available capacity for CFS tasks.
7271 * For the latter, we use a threshold to stabilize the state, to take into
7272 * account the variance of the tasks' load and to return true if the available
7273 * capacity in meaningful for the load balancer.
7274 * As an example, an available capacity of 1% can appear but it doesn't make
7275 * any benefit for the load balance.
7278 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7280 if (sgs->sum_nr_running < sgs->group_weight)
7283 if ((sgs->group_capacity * 100) >
7284 (sgs->group_util * env->sd->imbalance_pct))
7291 * group_is_overloaded returns true if the group has more tasks than it can
7293 * group_is_overloaded is not equals to !group_has_capacity because a group
7294 * with the exact right number of tasks, has no more spare capacity but is not
7295 * overloaded so both group_has_capacity and group_is_overloaded return
7299 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7301 if (sgs->sum_nr_running <= sgs->group_weight)
7304 if ((sgs->group_capacity * 100) <
7305 (sgs->group_util * env->sd->imbalance_pct))
7313 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7314 * per-cpu capacity than sched_group ref.
7317 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7319 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7320 ref->sgc->max_capacity;
7324 group_type group_classify(struct sched_group *group,
7325 struct sg_lb_stats *sgs)
7327 if (sgs->group_no_capacity)
7328 return group_overloaded;
7330 if (sg_imbalanced(group))
7331 return group_imbalanced;
7333 if (sgs->group_misfit_task)
7334 return group_misfit_task;
7340 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7341 * @env: The load balancing environment.
7342 * @group: sched_group whose statistics are to be updated.
7343 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7344 * @local_group: Does group contain this_cpu.
7345 * @sgs: variable to hold the statistics for this group.
7346 * @overload: Indicate more than one runnable task for any CPU.
7347 * @overutilized: Indicate overutilization for any CPU.
7349 static inline void update_sg_lb_stats(struct lb_env *env,
7350 struct sched_group *group, int load_idx,
7351 int local_group, struct sg_lb_stats *sgs,
7352 bool *overload, bool *overutilized)
7357 memset(sgs, 0, sizeof(*sgs));
7359 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7360 struct rq *rq = cpu_rq(i);
7362 /* Bias balancing toward cpus of our domain */
7364 load = target_load(i, load_idx);
7366 load = source_load(i, load_idx);
7368 sgs->group_load += load;
7369 sgs->group_util += cpu_util(i);
7370 sgs->sum_nr_running += rq->cfs.h_nr_running;
7372 nr_running = rq->nr_running;
7376 #ifdef CONFIG_NUMA_BALANCING
7377 sgs->nr_numa_running += rq->nr_numa_running;
7378 sgs->nr_preferred_running += rq->nr_preferred_running;
7380 sgs->sum_weighted_load += weighted_cpuload(i);
7382 * No need to call idle_cpu() if nr_running is not 0
7384 if (!nr_running && idle_cpu(i))
7387 if (cpu_overutilized(i)) {
7388 *overutilized = true;
7389 if (!sgs->group_misfit_task && rq->misfit_task)
7390 sgs->group_misfit_task = capacity_of(i);
7394 /* Adjust by relative CPU capacity of the group */
7395 sgs->group_capacity = group->sgc->capacity;
7396 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7398 if (sgs->sum_nr_running)
7399 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7401 sgs->group_weight = group->group_weight;
7403 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7404 sgs->group_type = group_classify(group, sgs);
7408 * update_sd_pick_busiest - return 1 on busiest group
7409 * @env: The load balancing environment.
7410 * @sds: sched_domain statistics
7411 * @sg: sched_group candidate to be checked for being the busiest
7412 * @sgs: sched_group statistics
7414 * Determine if @sg is a busier group than the previously selected
7417 * Return: %true if @sg is a busier group than the previously selected
7418 * busiest group. %false otherwise.
7420 static bool update_sd_pick_busiest(struct lb_env *env,
7421 struct sd_lb_stats *sds,
7422 struct sched_group *sg,
7423 struct sg_lb_stats *sgs)
7425 struct sg_lb_stats *busiest = &sds->busiest_stat;
7427 if (sgs->group_type > busiest->group_type)
7430 if (sgs->group_type < busiest->group_type)
7434 * Candidate sg doesn't face any serious load-balance problems
7435 * so don't pick it if the local sg is already filled up.
7437 if (sgs->group_type == group_other &&
7438 !group_has_capacity(env, &sds->local_stat))
7441 if (sgs->avg_load <= busiest->avg_load)
7445 * Candiate sg has no more than one task per cpu and has higher
7446 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7448 if (sgs->sum_nr_running <= sgs->group_weight &&
7449 group_smaller_cpu_capacity(sds->local, sg))
7452 /* This is the busiest node in its class. */
7453 if (!(env->sd->flags & SD_ASYM_PACKING))
7457 * ASYM_PACKING needs to move all the work to the lowest
7458 * numbered CPUs in the group, therefore mark all groups
7459 * higher than ourself as busy.
7461 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7465 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7472 #ifdef CONFIG_NUMA_BALANCING
7473 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7475 if (sgs->sum_nr_running > sgs->nr_numa_running)
7477 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7482 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7484 if (rq->nr_running > rq->nr_numa_running)
7486 if (rq->nr_running > rq->nr_preferred_running)
7491 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7496 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7500 #endif /* CONFIG_NUMA_BALANCING */
7503 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7504 * @env: The load balancing environment.
7505 * @sds: variable to hold the statistics for this sched_domain.
7507 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7509 struct sched_domain *child = env->sd->child;
7510 struct sched_group *sg = env->sd->groups;
7511 struct sg_lb_stats tmp_sgs;
7512 int load_idx, prefer_sibling = 0;
7513 bool overload = false, overutilized = false;
7515 if (child && child->flags & SD_PREFER_SIBLING)
7518 load_idx = get_sd_load_idx(env->sd, env->idle);
7521 struct sg_lb_stats *sgs = &tmp_sgs;
7524 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7527 sgs = &sds->local_stat;
7529 if (env->idle != CPU_NEWLY_IDLE ||
7530 time_after_eq(jiffies, sg->sgc->next_update))
7531 update_group_capacity(env->sd, env->dst_cpu);
7534 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7535 &overload, &overutilized);
7541 * In case the child domain prefers tasks go to siblings
7542 * first, lower the sg capacity so that we'll try
7543 * and move all the excess tasks away. We lower the capacity
7544 * of a group only if the local group has the capacity to fit
7545 * these excess tasks. The extra check prevents the case where
7546 * you always pull from the heaviest group when it is already
7547 * under-utilized (possible with a large weight task outweighs
7548 * the tasks on the system).
7550 if (prefer_sibling && sds->local &&
7551 group_has_capacity(env, &sds->local_stat) &&
7552 (sgs->sum_nr_running > 1)) {
7553 sgs->group_no_capacity = 1;
7554 sgs->group_type = group_classify(sg, sgs);
7558 * Ignore task groups with misfit tasks if local group has no
7559 * capacity or if per-cpu capacity isn't higher.
7561 if (sgs->group_type == group_misfit_task &&
7562 (!group_has_capacity(env, &sds->local_stat) ||
7563 !group_smaller_cpu_capacity(sg, sds->local)))
7564 sgs->group_type = group_other;
7566 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7568 sds->busiest_stat = *sgs;
7572 /* Now, start updating sd_lb_stats */
7573 sds->total_load += sgs->group_load;
7574 sds->total_capacity += sgs->group_capacity;
7577 } while (sg != env->sd->groups);
7579 if (env->sd->flags & SD_NUMA)
7580 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7582 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7584 if (!env->sd->parent) {
7585 /* update overload indicator if we are at root domain */
7586 if (env->dst_rq->rd->overload != overload)
7587 env->dst_rq->rd->overload = overload;
7589 /* Update over-utilization (tipping point, U >= 0) indicator */
7590 if (env->dst_rq->rd->overutilized != overutilized) {
7591 env->dst_rq->rd->overutilized = overutilized;
7592 trace_sched_overutilized(overutilized);
7595 if (!env->dst_rq->rd->overutilized && overutilized) {
7596 env->dst_rq->rd->overutilized = true;
7597 trace_sched_overutilized(true);
7604 * check_asym_packing - Check to see if the group is packed into the
7607 * This is primarily intended to used at the sibling level. Some
7608 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7609 * case of POWER7, it can move to lower SMT modes only when higher
7610 * threads are idle. When in lower SMT modes, the threads will
7611 * perform better since they share less core resources. Hence when we
7612 * have idle threads, we want them to be the higher ones.
7614 * This packing function is run on idle threads. It checks to see if
7615 * the busiest CPU in this domain (core in the P7 case) has a higher
7616 * CPU number than the packing function is being run on. Here we are
7617 * assuming lower CPU number will be equivalent to lower a SMT thread
7620 * Return: 1 when packing is required and a task should be moved to
7621 * this CPU. The amount of the imbalance is returned in *imbalance.
7623 * @env: The load balancing environment.
7624 * @sds: Statistics of the sched_domain which is to be packed
7626 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7630 if (!(env->sd->flags & SD_ASYM_PACKING))
7636 busiest_cpu = group_first_cpu(sds->busiest);
7637 if (env->dst_cpu > busiest_cpu)
7640 env->imbalance = DIV_ROUND_CLOSEST(
7641 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7642 SCHED_CAPACITY_SCALE);
7648 * fix_small_imbalance - Calculate the minor imbalance that exists
7649 * amongst the groups of a sched_domain, during
7651 * @env: The load balancing environment.
7652 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7655 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7657 unsigned long tmp, capa_now = 0, capa_move = 0;
7658 unsigned int imbn = 2;
7659 unsigned long scaled_busy_load_per_task;
7660 struct sg_lb_stats *local, *busiest;
7662 local = &sds->local_stat;
7663 busiest = &sds->busiest_stat;
7665 if (!local->sum_nr_running)
7666 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7667 else if (busiest->load_per_task > local->load_per_task)
7670 scaled_busy_load_per_task =
7671 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7672 busiest->group_capacity;
7674 if (busiest->avg_load + scaled_busy_load_per_task >=
7675 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7676 env->imbalance = busiest->load_per_task;
7681 * OK, we don't have enough imbalance to justify moving tasks,
7682 * however we may be able to increase total CPU capacity used by
7686 capa_now += busiest->group_capacity *
7687 min(busiest->load_per_task, busiest->avg_load);
7688 capa_now += local->group_capacity *
7689 min(local->load_per_task, local->avg_load);
7690 capa_now /= SCHED_CAPACITY_SCALE;
7692 /* Amount of load we'd subtract */
7693 if (busiest->avg_load > scaled_busy_load_per_task) {
7694 capa_move += busiest->group_capacity *
7695 min(busiest->load_per_task,
7696 busiest->avg_load - scaled_busy_load_per_task);
7699 /* Amount of load we'd add */
7700 if (busiest->avg_load * busiest->group_capacity <
7701 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7702 tmp = (busiest->avg_load * busiest->group_capacity) /
7703 local->group_capacity;
7705 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7706 local->group_capacity;
7708 capa_move += local->group_capacity *
7709 min(local->load_per_task, local->avg_load + tmp);
7710 capa_move /= SCHED_CAPACITY_SCALE;
7712 /* Move if we gain throughput */
7713 if (capa_move > capa_now)
7714 env->imbalance = busiest->load_per_task;
7718 * calculate_imbalance - Calculate the amount of imbalance present within the
7719 * groups of a given sched_domain during load balance.
7720 * @env: load balance environment
7721 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7723 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7725 unsigned long max_pull, load_above_capacity = ~0UL;
7726 struct sg_lb_stats *local, *busiest;
7728 local = &sds->local_stat;
7729 busiest = &sds->busiest_stat;
7731 if (busiest->group_type == group_imbalanced) {
7733 * In the group_imb case we cannot rely on group-wide averages
7734 * to ensure cpu-load equilibrium, look at wider averages. XXX
7736 busiest->load_per_task =
7737 min(busiest->load_per_task, sds->avg_load);
7741 * In the presence of smp nice balancing, certain scenarios can have
7742 * max load less than avg load(as we skip the groups at or below
7743 * its cpu_capacity, while calculating max_load..)
7745 if (busiest->avg_load <= sds->avg_load ||
7746 local->avg_load >= sds->avg_load) {
7747 /* Misfitting tasks should be migrated in any case */
7748 if (busiest->group_type == group_misfit_task) {
7749 env->imbalance = busiest->group_misfit_task;
7754 * Busiest group is overloaded, local is not, use the spare
7755 * cycles to maximize throughput
7757 if (busiest->group_type == group_overloaded &&
7758 local->group_type <= group_misfit_task) {
7759 env->imbalance = busiest->load_per_task;
7764 return fix_small_imbalance(env, sds);
7768 * If there aren't any idle cpus, avoid creating some.
7770 if (busiest->group_type == group_overloaded &&
7771 local->group_type == group_overloaded) {
7772 load_above_capacity = busiest->sum_nr_running *
7774 if (load_above_capacity > busiest->group_capacity)
7775 load_above_capacity -= busiest->group_capacity;
7777 load_above_capacity = ~0UL;
7781 * We're trying to get all the cpus to the average_load, so we don't
7782 * want to push ourselves above the average load, nor do we wish to
7783 * reduce the max loaded cpu below the average load. At the same time,
7784 * we also don't want to reduce the group load below the group capacity
7785 * (so that we can implement power-savings policies etc). Thus we look
7786 * for the minimum possible imbalance.
7788 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7790 /* How much load to actually move to equalise the imbalance */
7791 env->imbalance = min(
7792 max_pull * busiest->group_capacity,
7793 (sds->avg_load - local->avg_load) * local->group_capacity
7794 ) / SCHED_CAPACITY_SCALE;
7796 /* Boost imbalance to allow misfit task to be balanced. */
7797 if (busiest->group_type == group_misfit_task)
7798 env->imbalance = max_t(long, env->imbalance,
7799 busiest->group_misfit_task);
7802 * if *imbalance is less than the average load per runnable task
7803 * there is no guarantee that any tasks will be moved so we'll have
7804 * a think about bumping its value to force at least one task to be
7807 if (env->imbalance < busiest->load_per_task)
7808 return fix_small_imbalance(env, sds);
7811 /******* find_busiest_group() helpers end here *********************/
7814 * find_busiest_group - Returns the busiest group within the sched_domain
7815 * if there is an imbalance. If there isn't an imbalance, and
7816 * the user has opted for power-savings, it returns a group whose
7817 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7818 * such a group exists.
7820 * Also calculates the amount of weighted load which should be moved
7821 * to restore balance.
7823 * @env: The load balancing environment.
7825 * Return: - The busiest group if imbalance exists.
7826 * - If no imbalance and user has opted for power-savings balance,
7827 * return the least loaded group whose CPUs can be
7828 * put to idle by rebalancing its tasks onto our group.
7830 static struct sched_group *find_busiest_group(struct lb_env *env)
7832 struct sg_lb_stats *local, *busiest;
7833 struct sd_lb_stats sds;
7835 init_sd_lb_stats(&sds);
7838 * Compute the various statistics relavent for load balancing at
7841 update_sd_lb_stats(env, &sds);
7843 if (energy_aware() && !env->dst_rq->rd->overutilized)
7846 local = &sds.local_stat;
7847 busiest = &sds.busiest_stat;
7849 /* ASYM feature bypasses nice load balance check */
7850 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7851 check_asym_packing(env, &sds))
7854 /* There is no busy sibling group to pull tasks from */
7855 if (!sds.busiest || busiest->sum_nr_running == 0)
7858 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7859 / sds.total_capacity;
7862 * If the busiest group is imbalanced the below checks don't
7863 * work because they assume all things are equal, which typically
7864 * isn't true due to cpus_allowed constraints and the like.
7866 if (busiest->group_type == group_imbalanced)
7869 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7870 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7871 busiest->group_no_capacity)
7874 /* Misfitting tasks should be dealt with regardless of the avg load */
7875 if (busiest->group_type == group_misfit_task) {
7880 * If the local group is busier than the selected busiest group
7881 * don't try and pull any tasks.
7883 if (local->avg_load >= busiest->avg_load)
7887 * Don't pull any tasks if this group is already above the domain
7890 if (local->avg_load >= sds.avg_load)
7893 if (env->idle == CPU_IDLE) {
7895 * This cpu is idle. If the busiest group is not overloaded
7896 * and there is no imbalance between this and busiest group
7897 * wrt idle cpus, it is balanced. The imbalance becomes
7898 * significant if the diff is greater than 1 otherwise we
7899 * might end up to just move the imbalance on another group
7901 if ((busiest->group_type != group_overloaded) &&
7902 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7903 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7907 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7908 * imbalance_pct to be conservative.
7910 if (100 * busiest->avg_load <=
7911 env->sd->imbalance_pct * local->avg_load)
7916 env->busiest_group_type = busiest->group_type;
7917 /* Looks like there is an imbalance. Compute it */
7918 calculate_imbalance(env, &sds);
7927 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7929 static struct rq *find_busiest_queue(struct lb_env *env,
7930 struct sched_group *group)
7932 struct rq *busiest = NULL, *rq;
7933 unsigned long busiest_load = 0, busiest_capacity = 1;
7936 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7937 unsigned long capacity, wl;
7941 rt = fbq_classify_rq(rq);
7944 * We classify groups/runqueues into three groups:
7945 * - regular: there are !numa tasks
7946 * - remote: there are numa tasks that run on the 'wrong' node
7947 * - all: there is no distinction
7949 * In order to avoid migrating ideally placed numa tasks,
7950 * ignore those when there's better options.
7952 * If we ignore the actual busiest queue to migrate another
7953 * task, the next balance pass can still reduce the busiest
7954 * queue by moving tasks around inside the node.
7956 * If we cannot move enough load due to this classification
7957 * the next pass will adjust the group classification and
7958 * allow migration of more tasks.
7960 * Both cases only affect the total convergence complexity.
7962 if (rt > env->fbq_type)
7965 capacity = capacity_of(i);
7967 wl = weighted_cpuload(i);
7970 * When comparing with imbalance, use weighted_cpuload()
7971 * which is not scaled with the cpu capacity.
7974 if (rq->nr_running == 1 && wl > env->imbalance &&
7975 !check_cpu_capacity(rq, env->sd) &&
7976 env->busiest_group_type != group_misfit_task)
7980 * For the load comparisons with the other cpu's, consider
7981 * the weighted_cpuload() scaled with the cpu capacity, so
7982 * that the load can be moved away from the cpu that is
7983 * potentially running at a lower capacity.
7985 * Thus we're looking for max(wl_i / capacity_i), crosswise
7986 * multiplication to rid ourselves of the division works out
7987 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7988 * our previous maximum.
7990 if (wl * busiest_capacity > busiest_load * capacity) {
7992 busiest_capacity = capacity;
8001 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8002 * so long as it is large enough.
8004 #define MAX_PINNED_INTERVAL 512
8006 /* Working cpumask for load_balance and load_balance_newidle. */
8007 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8009 static int need_active_balance(struct lb_env *env)
8011 struct sched_domain *sd = env->sd;
8013 if (env->idle == CPU_NEWLY_IDLE) {
8016 * ASYM_PACKING needs to force migrate tasks from busy but
8017 * higher numbered CPUs in order to pack all tasks in the
8018 * lowest numbered CPUs.
8020 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8025 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8026 * It's worth migrating the task if the src_cpu's capacity is reduced
8027 * because of other sched_class or IRQs if more capacity stays
8028 * available on dst_cpu.
8030 if ((env->idle != CPU_NOT_IDLE) &&
8031 (env->src_rq->cfs.h_nr_running == 1)) {
8032 if ((check_cpu_capacity(env->src_rq, sd)) &&
8033 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8037 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8038 env->src_rq->cfs.h_nr_running == 1 &&
8039 cpu_overutilized(env->src_cpu) &&
8040 !cpu_overutilized(env->dst_cpu)) {
8044 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8047 static int active_load_balance_cpu_stop(void *data);
8049 static int should_we_balance(struct lb_env *env)
8051 struct sched_group *sg = env->sd->groups;
8052 struct cpumask *sg_cpus, *sg_mask;
8053 int cpu, balance_cpu = -1;
8056 * In the newly idle case, we will allow all the cpu's
8057 * to do the newly idle load balance.
8059 if (env->idle == CPU_NEWLY_IDLE)
8062 sg_cpus = sched_group_cpus(sg);
8063 sg_mask = sched_group_mask(sg);
8064 /* Try to find first idle cpu */
8065 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8066 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8073 if (balance_cpu == -1)
8074 balance_cpu = group_balance_cpu(sg);
8077 * First idle cpu or the first cpu(busiest) in this sched group
8078 * is eligible for doing load balancing at this and above domains.
8080 return balance_cpu == env->dst_cpu;
8084 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8085 * tasks if there is an imbalance.
8087 static int load_balance(int this_cpu, struct rq *this_rq,
8088 struct sched_domain *sd, enum cpu_idle_type idle,
8089 int *continue_balancing)
8091 int ld_moved, cur_ld_moved, active_balance = 0;
8092 struct sched_domain *sd_parent = sd->parent;
8093 struct sched_group *group;
8095 unsigned long flags;
8096 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8098 struct lb_env env = {
8100 .dst_cpu = this_cpu,
8102 .dst_grpmask = sched_group_cpus(sd->groups),
8104 .loop_break = sched_nr_migrate_break,
8107 .tasks = LIST_HEAD_INIT(env.tasks),
8111 * For NEWLY_IDLE load_balancing, we don't need to consider
8112 * other cpus in our group
8114 if (idle == CPU_NEWLY_IDLE)
8115 env.dst_grpmask = NULL;
8117 cpumask_copy(cpus, cpu_active_mask);
8119 schedstat_inc(sd, lb_count[idle]);
8122 if (!should_we_balance(&env)) {
8123 *continue_balancing = 0;
8127 group = find_busiest_group(&env);
8129 schedstat_inc(sd, lb_nobusyg[idle]);
8133 busiest = find_busiest_queue(&env, group);
8135 schedstat_inc(sd, lb_nobusyq[idle]);
8139 BUG_ON(busiest == env.dst_rq);
8141 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8143 env.src_cpu = busiest->cpu;
8144 env.src_rq = busiest;
8147 if (busiest->nr_running > 1) {
8149 * Attempt to move tasks. If find_busiest_group has found
8150 * an imbalance but busiest->nr_running <= 1, the group is
8151 * still unbalanced. ld_moved simply stays zero, so it is
8152 * correctly treated as an imbalance.
8154 env.flags |= LBF_ALL_PINNED;
8155 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8158 raw_spin_lock_irqsave(&busiest->lock, flags);
8161 * cur_ld_moved - load moved in current iteration
8162 * ld_moved - cumulative load moved across iterations
8164 cur_ld_moved = detach_tasks(&env);
8166 * We want to potentially lower env.src_cpu's OPP.
8169 update_capacity_of(env.src_cpu);
8172 * We've detached some tasks from busiest_rq. Every
8173 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8174 * unlock busiest->lock, and we are able to be sure
8175 * that nobody can manipulate the tasks in parallel.
8176 * See task_rq_lock() family for the details.
8179 raw_spin_unlock(&busiest->lock);
8183 ld_moved += cur_ld_moved;
8186 local_irq_restore(flags);
8188 if (env.flags & LBF_NEED_BREAK) {
8189 env.flags &= ~LBF_NEED_BREAK;
8194 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8195 * us and move them to an alternate dst_cpu in our sched_group
8196 * where they can run. The upper limit on how many times we
8197 * iterate on same src_cpu is dependent on number of cpus in our
8200 * This changes load balance semantics a bit on who can move
8201 * load to a given_cpu. In addition to the given_cpu itself
8202 * (or a ilb_cpu acting on its behalf where given_cpu is
8203 * nohz-idle), we now have balance_cpu in a position to move
8204 * load to given_cpu. In rare situations, this may cause
8205 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8206 * _independently_ and at _same_ time to move some load to
8207 * given_cpu) causing exceess load to be moved to given_cpu.
8208 * This however should not happen so much in practice and
8209 * moreover subsequent load balance cycles should correct the
8210 * excess load moved.
8212 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8214 /* Prevent to re-select dst_cpu via env's cpus */
8215 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8217 env.dst_rq = cpu_rq(env.new_dst_cpu);
8218 env.dst_cpu = env.new_dst_cpu;
8219 env.flags &= ~LBF_DST_PINNED;
8221 env.loop_break = sched_nr_migrate_break;
8224 * Go back to "more_balance" rather than "redo" since we
8225 * need to continue with same src_cpu.
8231 * We failed to reach balance because of affinity.
8234 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8236 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8237 *group_imbalance = 1;
8240 /* All tasks on this runqueue were pinned by CPU affinity */
8241 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8242 cpumask_clear_cpu(cpu_of(busiest), cpus);
8243 if (!cpumask_empty(cpus)) {
8245 env.loop_break = sched_nr_migrate_break;
8248 goto out_all_pinned;
8253 schedstat_inc(sd, lb_failed[idle]);
8255 * Increment the failure counter only on periodic balance.
8256 * We do not want newidle balance, which can be very
8257 * frequent, pollute the failure counter causing
8258 * excessive cache_hot migrations and active balances.
8260 if (idle != CPU_NEWLY_IDLE)
8261 if (env.src_grp_nr_running > 1)
8262 sd->nr_balance_failed++;
8264 if (need_active_balance(&env)) {
8265 raw_spin_lock_irqsave(&busiest->lock, flags);
8267 /* don't kick the active_load_balance_cpu_stop,
8268 * if the curr task on busiest cpu can't be
8271 if (!cpumask_test_cpu(this_cpu,
8272 tsk_cpus_allowed(busiest->curr))) {
8273 raw_spin_unlock_irqrestore(&busiest->lock,
8275 env.flags |= LBF_ALL_PINNED;
8276 goto out_one_pinned;
8280 * ->active_balance synchronizes accesses to
8281 * ->active_balance_work. Once set, it's cleared
8282 * only after active load balance is finished.
8284 if (!busiest->active_balance) {
8285 busiest->active_balance = 1;
8286 busiest->push_cpu = this_cpu;
8289 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8291 if (active_balance) {
8292 stop_one_cpu_nowait(cpu_of(busiest),
8293 active_load_balance_cpu_stop, busiest,
8294 &busiest->active_balance_work);
8298 * We've kicked active balancing, reset the failure
8301 sd->nr_balance_failed = sd->cache_nice_tries+1;
8304 sd->nr_balance_failed = 0;
8306 if (likely(!active_balance)) {
8307 /* We were unbalanced, so reset the balancing interval */
8308 sd->balance_interval = sd->min_interval;
8311 * If we've begun active balancing, start to back off. This
8312 * case may not be covered by the all_pinned logic if there
8313 * is only 1 task on the busy runqueue (because we don't call
8316 if (sd->balance_interval < sd->max_interval)
8317 sd->balance_interval *= 2;
8324 * We reach balance although we may have faced some affinity
8325 * constraints. Clear the imbalance flag if it was set.
8328 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8330 if (*group_imbalance)
8331 *group_imbalance = 0;
8336 * We reach balance because all tasks are pinned at this level so
8337 * we can't migrate them. Let the imbalance flag set so parent level
8338 * can try to migrate them.
8340 schedstat_inc(sd, lb_balanced[idle]);
8342 sd->nr_balance_failed = 0;
8345 /* tune up the balancing interval */
8346 if (((env.flags & LBF_ALL_PINNED) &&
8347 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8348 (sd->balance_interval < sd->max_interval))
8349 sd->balance_interval *= 2;
8356 static inline unsigned long
8357 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8359 unsigned long interval = sd->balance_interval;
8362 interval *= sd->busy_factor;
8364 /* scale ms to jiffies */
8365 interval = msecs_to_jiffies(interval);
8366 interval = clamp(interval, 1UL, max_load_balance_interval);
8372 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8374 unsigned long interval, next;
8376 interval = get_sd_balance_interval(sd, cpu_busy);
8377 next = sd->last_balance + interval;
8379 if (time_after(*next_balance, next))
8380 *next_balance = next;
8384 * idle_balance is called by schedule() if this_cpu is about to become
8385 * idle. Attempts to pull tasks from other CPUs.
8387 static int idle_balance(struct rq *this_rq)
8389 unsigned long next_balance = jiffies + HZ;
8390 int this_cpu = this_rq->cpu;
8391 struct sched_domain *sd;
8392 int pulled_task = 0;
8394 long removed_util=0;
8396 idle_enter_fair(this_rq);
8399 * We must set idle_stamp _before_ calling idle_balance(), such that we
8400 * measure the duration of idle_balance() as idle time.
8402 this_rq->idle_stamp = rq_clock(this_rq);
8404 if (!energy_aware() &&
8405 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8406 !this_rq->rd->overload)) {
8408 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8410 update_next_balance(sd, 0, &next_balance);
8416 raw_spin_unlock(&this_rq->lock);
8419 * If removed_util_avg is !0 we most probably migrated some task away
8420 * from this_cpu. In this case we might be willing to trigger an OPP
8421 * update, but we want to do so if we don't find anybody else to pull
8422 * here (we will trigger an OPP update with the pulled task's enqueue
8425 * Record removed_util before calling update_blocked_averages, and use
8426 * it below (before returning) to see if an OPP update is required.
8428 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8429 update_blocked_averages(this_cpu);
8431 for_each_domain(this_cpu, sd) {
8432 int continue_balancing = 1;
8433 u64 t0, domain_cost;
8435 if (!(sd->flags & SD_LOAD_BALANCE))
8438 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8439 update_next_balance(sd, 0, &next_balance);
8443 if (sd->flags & SD_BALANCE_NEWIDLE) {
8444 t0 = sched_clock_cpu(this_cpu);
8446 pulled_task = load_balance(this_cpu, this_rq,
8448 &continue_balancing);
8450 domain_cost = sched_clock_cpu(this_cpu) - t0;
8451 if (domain_cost > sd->max_newidle_lb_cost)
8452 sd->max_newidle_lb_cost = domain_cost;
8454 curr_cost += domain_cost;
8457 update_next_balance(sd, 0, &next_balance);
8460 * Stop searching for tasks to pull if there are
8461 * now runnable tasks on this rq.
8463 if (pulled_task || this_rq->nr_running > 0)
8468 raw_spin_lock(&this_rq->lock);
8470 if (curr_cost > this_rq->max_idle_balance_cost)
8471 this_rq->max_idle_balance_cost = curr_cost;
8474 * While browsing the domains, we released the rq lock, a task could
8475 * have been enqueued in the meantime. Since we're not going idle,
8476 * pretend we pulled a task.
8478 if (this_rq->cfs.h_nr_running && !pulled_task)
8482 /* Move the next balance forward */
8483 if (time_after(this_rq->next_balance, next_balance))
8484 this_rq->next_balance = next_balance;
8486 /* Is there a task of a high priority class? */
8487 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8491 idle_exit_fair(this_rq);
8492 this_rq->idle_stamp = 0;
8493 } else if (removed_util) {
8495 * No task pulled and someone has been migrated away.
8496 * Good case to trigger an OPP update.
8498 update_capacity_of(this_cpu);
8505 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8506 * running tasks off the busiest CPU onto idle CPUs. It requires at
8507 * least 1 task to be running on each physical CPU where possible, and
8508 * avoids physical / logical imbalances.
8510 static int active_load_balance_cpu_stop(void *data)
8512 struct rq *busiest_rq = data;
8513 int busiest_cpu = cpu_of(busiest_rq);
8514 int target_cpu = busiest_rq->push_cpu;
8515 struct rq *target_rq = cpu_rq(target_cpu);
8516 struct sched_domain *sd;
8517 struct task_struct *p = NULL;
8519 raw_spin_lock_irq(&busiest_rq->lock);
8521 /* make sure the requested cpu hasn't gone down in the meantime */
8522 if (unlikely(busiest_cpu != smp_processor_id() ||
8523 !busiest_rq->active_balance))
8526 /* Is there any task to move? */
8527 if (busiest_rq->nr_running <= 1)
8531 * This condition is "impossible", if it occurs
8532 * we need to fix it. Originally reported by
8533 * Bjorn Helgaas on a 128-cpu setup.
8535 BUG_ON(busiest_rq == target_rq);
8537 /* Search for an sd spanning us and the target CPU. */
8539 for_each_domain(target_cpu, sd) {
8540 if ((sd->flags & SD_LOAD_BALANCE) &&
8541 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8546 struct lb_env env = {
8548 .dst_cpu = target_cpu,
8549 .dst_rq = target_rq,
8550 .src_cpu = busiest_rq->cpu,
8551 .src_rq = busiest_rq,
8555 schedstat_inc(sd, alb_count);
8557 p = detach_one_task(&env);
8559 schedstat_inc(sd, alb_pushed);
8561 * We want to potentially lower env.src_cpu's OPP.
8563 update_capacity_of(env.src_cpu);
8566 schedstat_inc(sd, alb_failed);
8570 busiest_rq->active_balance = 0;
8571 raw_spin_unlock(&busiest_rq->lock);
8574 attach_one_task(target_rq, p);
8581 static inline int on_null_domain(struct rq *rq)
8583 return unlikely(!rcu_dereference_sched(rq->sd));
8586 #ifdef CONFIG_NO_HZ_COMMON
8588 * idle load balancing details
8589 * - When one of the busy CPUs notice that there may be an idle rebalancing
8590 * needed, they will kick the idle load balancer, which then does idle
8591 * load balancing for all the idle CPUs.
8594 cpumask_var_t idle_cpus_mask;
8596 unsigned long next_balance; /* in jiffy units */
8597 } nohz ____cacheline_aligned;
8599 static inline int find_new_ilb(void)
8601 int ilb = cpumask_first(nohz.idle_cpus_mask);
8603 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8610 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8611 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8612 * CPU (if there is one).
8614 static void nohz_balancer_kick(void)
8618 nohz.next_balance++;
8620 ilb_cpu = find_new_ilb();
8622 if (ilb_cpu >= nr_cpu_ids)
8625 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8628 * Use smp_send_reschedule() instead of resched_cpu().
8629 * This way we generate a sched IPI on the target cpu which
8630 * is idle. And the softirq performing nohz idle load balance
8631 * will be run before returning from the IPI.
8633 smp_send_reschedule(ilb_cpu);
8637 static inline void nohz_balance_exit_idle(int cpu)
8639 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8641 * Completely isolated CPUs don't ever set, so we must test.
8643 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8644 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8645 atomic_dec(&nohz.nr_cpus);
8647 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8651 static inline void set_cpu_sd_state_busy(void)
8653 struct sched_domain *sd;
8654 int cpu = smp_processor_id();
8657 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8659 if (!sd || !sd->nohz_idle)
8663 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8668 void set_cpu_sd_state_idle(void)
8670 struct sched_domain *sd;
8671 int cpu = smp_processor_id();
8674 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8676 if (!sd || sd->nohz_idle)
8680 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8686 * This routine will record that the cpu is going idle with tick stopped.
8687 * This info will be used in performing idle load balancing in the future.
8689 void nohz_balance_enter_idle(int cpu)
8692 * If this cpu is going down, then nothing needs to be done.
8694 if (!cpu_active(cpu))
8697 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8701 * If we're a completely isolated CPU, we don't play.
8703 if (on_null_domain(cpu_rq(cpu)))
8706 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8707 atomic_inc(&nohz.nr_cpus);
8708 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8711 static int sched_ilb_notifier(struct notifier_block *nfb,
8712 unsigned long action, void *hcpu)
8714 switch (action & ~CPU_TASKS_FROZEN) {
8716 nohz_balance_exit_idle(smp_processor_id());
8724 static DEFINE_SPINLOCK(balancing);
8727 * Scale the max load_balance interval with the number of CPUs in the system.
8728 * This trades load-balance latency on larger machines for less cross talk.
8730 void update_max_interval(void)
8732 max_load_balance_interval = HZ*num_online_cpus()/10;
8736 * It checks each scheduling domain to see if it is due to be balanced,
8737 * and initiates a balancing operation if so.
8739 * Balancing parameters are set up in init_sched_domains.
8741 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8743 int continue_balancing = 1;
8745 unsigned long interval;
8746 struct sched_domain *sd;
8747 /* Earliest time when we have to do rebalance again */
8748 unsigned long next_balance = jiffies + 60*HZ;
8749 int update_next_balance = 0;
8750 int need_serialize, need_decay = 0;
8753 update_blocked_averages(cpu);
8756 for_each_domain(cpu, sd) {
8758 * Decay the newidle max times here because this is a regular
8759 * visit to all the domains. Decay ~1% per second.
8761 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8762 sd->max_newidle_lb_cost =
8763 (sd->max_newidle_lb_cost * 253) / 256;
8764 sd->next_decay_max_lb_cost = jiffies + HZ;
8767 max_cost += sd->max_newidle_lb_cost;
8769 if (!(sd->flags & SD_LOAD_BALANCE))
8773 * Stop the load balance at this level. There is another
8774 * CPU in our sched group which is doing load balancing more
8777 if (!continue_balancing) {
8783 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8785 need_serialize = sd->flags & SD_SERIALIZE;
8786 if (need_serialize) {
8787 if (!spin_trylock(&balancing))
8791 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8792 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8794 * The LBF_DST_PINNED logic could have changed
8795 * env->dst_cpu, so we can't know our idle
8796 * state even if we migrated tasks. Update it.
8798 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8800 sd->last_balance = jiffies;
8801 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8804 spin_unlock(&balancing);
8806 if (time_after(next_balance, sd->last_balance + interval)) {
8807 next_balance = sd->last_balance + interval;
8808 update_next_balance = 1;
8813 * Ensure the rq-wide value also decays but keep it at a
8814 * reasonable floor to avoid funnies with rq->avg_idle.
8816 rq->max_idle_balance_cost =
8817 max((u64)sysctl_sched_migration_cost, max_cost);
8822 * next_balance will be updated only when there is a need.
8823 * When the cpu is attached to null domain for ex, it will not be
8826 if (likely(update_next_balance)) {
8827 rq->next_balance = next_balance;
8829 #ifdef CONFIG_NO_HZ_COMMON
8831 * If this CPU has been elected to perform the nohz idle
8832 * balance. Other idle CPUs have already rebalanced with
8833 * nohz_idle_balance() and nohz.next_balance has been
8834 * updated accordingly. This CPU is now running the idle load
8835 * balance for itself and we need to update the
8836 * nohz.next_balance accordingly.
8838 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8839 nohz.next_balance = rq->next_balance;
8844 #ifdef CONFIG_NO_HZ_COMMON
8846 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8847 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8849 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8851 int this_cpu = this_rq->cpu;
8854 /* Earliest time when we have to do rebalance again */
8855 unsigned long next_balance = jiffies + 60*HZ;
8856 int update_next_balance = 0;
8858 if (idle != CPU_IDLE ||
8859 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8862 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8863 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8867 * If this cpu gets work to do, stop the load balancing
8868 * work being done for other cpus. Next load
8869 * balancing owner will pick it up.
8874 rq = cpu_rq(balance_cpu);
8877 * If time for next balance is due,
8880 if (time_after_eq(jiffies, rq->next_balance)) {
8881 raw_spin_lock_irq(&rq->lock);
8882 update_rq_clock(rq);
8883 update_idle_cpu_load(rq);
8884 raw_spin_unlock_irq(&rq->lock);
8885 rebalance_domains(rq, CPU_IDLE);
8888 if (time_after(next_balance, rq->next_balance)) {
8889 next_balance = rq->next_balance;
8890 update_next_balance = 1;
8895 * next_balance will be updated only when there is a need.
8896 * When the CPU is attached to null domain for ex, it will not be
8899 if (likely(update_next_balance))
8900 nohz.next_balance = next_balance;
8902 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8906 * Current heuristic for kicking the idle load balancer in the presence
8907 * of an idle cpu in the system.
8908 * - This rq has more than one task.
8909 * - This rq has at least one CFS task and the capacity of the CPU is
8910 * significantly reduced because of RT tasks or IRQs.
8911 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8912 * multiple busy cpu.
8913 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8914 * domain span are idle.
8916 static inline bool nohz_kick_needed(struct rq *rq)
8918 unsigned long now = jiffies;
8919 struct sched_domain *sd;
8920 struct sched_group_capacity *sgc;
8921 int nr_busy, cpu = rq->cpu;
8924 if (unlikely(rq->idle_balance))
8928 * We may be recently in ticked or tickless idle mode. At the first
8929 * busy tick after returning from idle, we will update the busy stats.
8931 set_cpu_sd_state_busy();
8932 nohz_balance_exit_idle(cpu);
8935 * None are in tickless mode and hence no need for NOHZ idle load
8938 if (likely(!atomic_read(&nohz.nr_cpus)))
8941 if (time_before(now, nohz.next_balance))
8944 if (rq->nr_running >= 2 &&
8945 (!energy_aware() || cpu_overutilized(cpu)))
8949 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8950 if (sd && !energy_aware()) {
8951 sgc = sd->groups->sgc;
8952 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8961 sd = rcu_dereference(rq->sd);
8963 if ((rq->cfs.h_nr_running >= 1) &&
8964 check_cpu_capacity(rq, sd)) {
8970 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8971 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8972 sched_domain_span(sd)) < cpu)) {
8982 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8986 * run_rebalance_domains is triggered when needed from the scheduler tick.
8987 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8989 static void run_rebalance_domains(struct softirq_action *h)
8991 struct rq *this_rq = this_rq();
8992 enum cpu_idle_type idle = this_rq->idle_balance ?
8993 CPU_IDLE : CPU_NOT_IDLE;
8996 * If this cpu has a pending nohz_balance_kick, then do the
8997 * balancing on behalf of the other idle cpus whose ticks are
8998 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8999 * give the idle cpus a chance to load balance. Else we may
9000 * load balance only within the local sched_domain hierarchy
9001 * and abort nohz_idle_balance altogether if we pull some load.
9003 nohz_idle_balance(this_rq, idle);
9004 rebalance_domains(this_rq, idle);
9008 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9010 void trigger_load_balance(struct rq *rq)
9012 /* Don't need to rebalance while attached to NULL domain */
9013 if (unlikely(on_null_domain(rq)))
9016 if (time_after_eq(jiffies, rq->next_balance))
9017 raise_softirq(SCHED_SOFTIRQ);
9018 #ifdef CONFIG_NO_HZ_COMMON
9019 if (nohz_kick_needed(rq))
9020 nohz_balancer_kick();
9024 static void rq_online_fair(struct rq *rq)
9028 update_runtime_enabled(rq);
9031 static void rq_offline_fair(struct rq *rq)
9035 /* Ensure any throttled groups are reachable by pick_next_task */
9036 unthrottle_offline_cfs_rqs(rq);
9039 #endif /* CONFIG_SMP */
9042 * scheduler tick hitting a task of our scheduling class:
9044 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9046 struct cfs_rq *cfs_rq;
9047 struct sched_entity *se = &curr->se;
9049 for_each_sched_entity(se) {
9050 cfs_rq = cfs_rq_of(se);
9051 entity_tick(cfs_rq, se, queued);
9054 if (static_branch_unlikely(&sched_numa_balancing))
9055 task_tick_numa(rq, curr);
9058 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9059 rq->rd->overutilized = true;
9060 trace_sched_overutilized(true);
9063 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9069 * called on fork with the child task as argument from the parent's context
9070 * - child not yet on the tasklist
9071 * - preemption disabled
9073 static void task_fork_fair(struct task_struct *p)
9075 struct cfs_rq *cfs_rq;
9076 struct sched_entity *se = &p->se, *curr;
9077 int this_cpu = smp_processor_id();
9078 struct rq *rq = this_rq();
9079 unsigned long flags;
9081 raw_spin_lock_irqsave(&rq->lock, flags);
9083 update_rq_clock(rq);
9085 cfs_rq = task_cfs_rq(current);
9086 curr = cfs_rq->curr;
9089 * Not only the cpu but also the task_group of the parent might have
9090 * been changed after parent->se.parent,cfs_rq were copied to
9091 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9092 * of child point to valid ones.
9095 __set_task_cpu(p, this_cpu);
9098 update_curr(cfs_rq);
9101 se->vruntime = curr->vruntime;
9102 place_entity(cfs_rq, se, 1);
9104 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9106 * Upon rescheduling, sched_class::put_prev_task() will place
9107 * 'current' within the tree based on its new key value.
9109 swap(curr->vruntime, se->vruntime);
9113 se->vruntime -= cfs_rq->min_vruntime;
9115 raw_spin_unlock_irqrestore(&rq->lock, flags);
9119 * Priority of the task has changed. Check to see if we preempt
9123 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9125 if (!task_on_rq_queued(p))
9129 * Reschedule if we are currently running on this runqueue and
9130 * our priority decreased, or if we are not currently running on
9131 * this runqueue and our priority is higher than the current's
9133 if (rq->curr == p) {
9134 if (p->prio > oldprio)
9137 check_preempt_curr(rq, p, 0);
9140 static inline bool vruntime_normalized(struct task_struct *p)
9142 struct sched_entity *se = &p->se;
9145 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9146 * the dequeue_entity(.flags=0) will already have normalized the
9153 * When !on_rq, vruntime of the task has usually NOT been normalized.
9154 * But there are some cases where it has already been normalized:
9156 * - A forked child which is waiting for being woken up by
9157 * wake_up_new_task().
9158 * - A task which has been woken up by try_to_wake_up() and
9159 * waiting for actually being woken up by sched_ttwu_pending().
9161 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9167 static void detach_task_cfs_rq(struct task_struct *p)
9169 struct sched_entity *se = &p->se;
9170 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9172 if (!vruntime_normalized(p)) {
9174 * Fix up our vruntime so that the current sleep doesn't
9175 * cause 'unlimited' sleep bonus.
9177 place_entity(cfs_rq, se, 0);
9178 se->vruntime -= cfs_rq->min_vruntime;
9181 /* Catch up with the cfs_rq and remove our load when we leave */
9182 detach_entity_load_avg(cfs_rq, se);
9185 static void attach_task_cfs_rq(struct task_struct *p)
9187 struct sched_entity *se = &p->se;
9188 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9192 * Since the real-depth could have been changed (only FAIR
9193 * class maintain depth value), reset depth properly.
9195 se->depth = se->parent ? se->parent->depth + 1 : 0;
9198 /* Synchronize task with its cfs_rq */
9199 attach_entity_load_avg(cfs_rq, se);
9201 if (!vruntime_normalized(p))
9202 se->vruntime += cfs_rq->min_vruntime;
9205 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9207 detach_task_cfs_rq(p);
9210 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9212 attach_task_cfs_rq(p);
9214 if (task_on_rq_queued(p)) {
9216 * We were most likely switched from sched_rt, so
9217 * kick off the schedule if running, otherwise just see
9218 * if we can still preempt the current task.
9223 check_preempt_curr(rq, p, 0);
9227 /* Account for a task changing its policy or group.
9229 * This routine is mostly called to set cfs_rq->curr field when a task
9230 * migrates between groups/classes.
9232 static void set_curr_task_fair(struct rq *rq)
9234 struct sched_entity *se = &rq->curr->se;
9236 for_each_sched_entity(se) {
9237 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9239 set_next_entity(cfs_rq, se);
9240 /* ensure bandwidth has been allocated on our new cfs_rq */
9241 account_cfs_rq_runtime(cfs_rq, 0);
9245 void init_cfs_rq(struct cfs_rq *cfs_rq)
9247 cfs_rq->tasks_timeline = RB_ROOT;
9248 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9249 #ifndef CONFIG_64BIT
9250 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9253 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9254 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9258 #ifdef CONFIG_FAIR_GROUP_SCHED
9259 static void task_move_group_fair(struct task_struct *p)
9261 detach_task_cfs_rq(p);
9262 set_task_rq(p, task_cpu(p));
9265 /* Tell se's cfs_rq has been changed -- migrated */
9266 p->se.avg.last_update_time = 0;
9268 attach_task_cfs_rq(p);
9271 void free_fair_sched_group(struct task_group *tg)
9275 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9277 for_each_possible_cpu(i) {
9279 kfree(tg->cfs_rq[i]);
9282 remove_entity_load_avg(tg->se[i]);
9291 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9293 struct cfs_rq *cfs_rq;
9294 struct sched_entity *se;
9297 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9300 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9304 tg->shares = NICE_0_LOAD;
9306 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9308 for_each_possible_cpu(i) {
9309 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9310 GFP_KERNEL, cpu_to_node(i));
9314 se = kzalloc_node(sizeof(struct sched_entity),
9315 GFP_KERNEL, cpu_to_node(i));
9319 init_cfs_rq(cfs_rq);
9320 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9321 init_entity_runnable_average(se);
9332 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9334 struct rq *rq = cpu_rq(cpu);
9335 unsigned long flags;
9338 * Only empty task groups can be destroyed; so we can speculatively
9339 * check on_list without danger of it being re-added.
9341 if (!tg->cfs_rq[cpu]->on_list)
9344 raw_spin_lock_irqsave(&rq->lock, flags);
9345 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9346 raw_spin_unlock_irqrestore(&rq->lock, flags);
9349 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9350 struct sched_entity *se, int cpu,
9351 struct sched_entity *parent)
9353 struct rq *rq = cpu_rq(cpu);
9357 init_cfs_rq_runtime(cfs_rq);
9359 tg->cfs_rq[cpu] = cfs_rq;
9362 /* se could be NULL for root_task_group */
9367 se->cfs_rq = &rq->cfs;
9370 se->cfs_rq = parent->my_q;
9371 se->depth = parent->depth + 1;
9375 /* guarantee group entities always have weight */
9376 update_load_set(&se->load, NICE_0_LOAD);
9377 se->parent = parent;
9380 static DEFINE_MUTEX(shares_mutex);
9382 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9385 unsigned long flags;
9388 * We can't change the weight of the root cgroup.
9393 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9395 mutex_lock(&shares_mutex);
9396 if (tg->shares == shares)
9399 tg->shares = shares;
9400 for_each_possible_cpu(i) {
9401 struct rq *rq = cpu_rq(i);
9402 struct sched_entity *se;
9405 /* Propagate contribution to hierarchy */
9406 raw_spin_lock_irqsave(&rq->lock, flags);
9408 /* Possible calls to update_curr() need rq clock */
9409 update_rq_clock(rq);
9410 for_each_sched_entity(se)
9411 update_cfs_shares(group_cfs_rq(se));
9412 raw_spin_unlock_irqrestore(&rq->lock, flags);
9416 mutex_unlock(&shares_mutex);
9419 #else /* CONFIG_FAIR_GROUP_SCHED */
9421 void free_fair_sched_group(struct task_group *tg) { }
9423 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9428 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9430 #endif /* CONFIG_FAIR_GROUP_SCHED */
9433 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9435 struct sched_entity *se = &task->se;
9436 unsigned int rr_interval = 0;
9439 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9442 if (rq->cfs.load.weight)
9443 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9449 * All the scheduling class methods:
9451 const struct sched_class fair_sched_class = {
9452 .next = &idle_sched_class,
9453 .enqueue_task = enqueue_task_fair,
9454 .dequeue_task = dequeue_task_fair,
9455 .yield_task = yield_task_fair,
9456 .yield_to_task = yield_to_task_fair,
9458 .check_preempt_curr = check_preempt_wakeup,
9460 .pick_next_task = pick_next_task_fair,
9461 .put_prev_task = put_prev_task_fair,
9464 .select_task_rq = select_task_rq_fair,
9465 .migrate_task_rq = migrate_task_rq_fair,
9467 .rq_online = rq_online_fair,
9468 .rq_offline = rq_offline_fair,
9470 .task_waking = task_waking_fair,
9471 .task_dead = task_dead_fair,
9472 .set_cpus_allowed = set_cpus_allowed_common,
9475 .set_curr_task = set_curr_task_fair,
9476 .task_tick = task_tick_fair,
9477 .task_fork = task_fork_fair,
9479 .prio_changed = prio_changed_fair,
9480 .switched_from = switched_from_fair,
9481 .switched_to = switched_to_fair,
9483 .get_rr_interval = get_rr_interval_fair,
9485 .update_curr = update_curr_fair,
9487 #ifdef CONFIG_FAIR_GROUP_SCHED
9488 .task_move_group = task_move_group_fair,
9492 #ifdef CONFIG_SCHED_DEBUG
9493 void print_cfs_stats(struct seq_file *m, int cpu)
9495 struct cfs_rq *cfs_rq;
9498 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9499 print_cfs_rq(m, cpu, cfs_rq);
9503 #ifdef CONFIG_NUMA_BALANCING
9504 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9507 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9509 for_each_online_node(node) {
9510 if (p->numa_faults) {
9511 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9512 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9514 if (p->numa_group) {
9515 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9516 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9518 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9521 #endif /* CONFIG_NUMA_BALANCING */
9522 #endif /* CONFIG_SCHED_DEBUG */
9524 __init void init_sched_fair_class(void)
9527 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9529 #ifdef CONFIG_NO_HZ_COMMON
9530 nohz.next_balance = jiffies;
9531 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9532 cpu_notifier(sched_ilb_notifier, 0);