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 return decayed || removed;
2763 /* Update task and its cfs_rq load average */
2764 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2766 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2767 u64 now = cfs_rq_clock_task(cfs_rq);
2768 int cpu = cpu_of(rq_of(cfs_rq));
2771 * Track task load average for carrying it to new CPU after migrated, and
2772 * track group sched_entity load average for task_h_load calc in migration
2774 __update_load_avg(now, cpu, &se->avg,
2775 se->on_rq * scale_load_down(se->load.weight),
2776 cfs_rq->curr == se, NULL);
2778 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2779 update_tg_load_avg(cfs_rq, 0);
2781 if (entity_is_task(se))
2782 trace_sched_load_avg_task(task_of(se), &se->avg);
2783 trace_sched_load_avg_cpu(cpu, cfs_rq);
2786 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2788 if (!sched_feat(ATTACH_AGE_LOAD))
2792 * If we got migrated (either between CPUs or between cgroups) we'll
2793 * have aged the average right before clearing @last_update_time.
2795 if (se->avg.last_update_time) {
2796 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2797 &se->avg, 0, 0, NULL);
2800 * XXX: we could have just aged the entire load away if we've been
2801 * absent from the fair class for too long.
2806 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2807 cfs_rq->avg.load_avg += se->avg.load_avg;
2808 cfs_rq->avg.load_sum += se->avg.load_sum;
2809 cfs_rq->avg.util_avg += se->avg.util_avg;
2810 cfs_rq->avg.util_sum += se->avg.util_sum;
2813 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2815 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2816 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2817 cfs_rq->curr == se, NULL);
2819 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2820 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2821 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2822 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2825 /* Add the load generated by se into cfs_rq's load average */
2827 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2829 struct sched_avg *sa = &se->avg;
2830 u64 now = cfs_rq_clock_task(cfs_rq);
2831 int migrated, decayed;
2833 migrated = !sa->last_update_time;
2835 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2836 se->on_rq * scale_load_down(se->load.weight),
2837 cfs_rq->curr == se, NULL);
2840 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2842 cfs_rq->runnable_load_avg += sa->load_avg;
2843 cfs_rq->runnable_load_sum += sa->load_sum;
2846 attach_entity_load_avg(cfs_rq, se);
2848 if (decayed || migrated)
2849 update_tg_load_avg(cfs_rq, 0);
2852 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2854 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2856 update_load_avg(se, 1);
2858 cfs_rq->runnable_load_avg =
2859 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2860 cfs_rq->runnable_load_sum =
2861 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2864 #ifndef CONFIG_64BIT
2865 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2867 u64 last_update_time_copy;
2868 u64 last_update_time;
2871 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2873 last_update_time = cfs_rq->avg.last_update_time;
2874 } while (last_update_time != last_update_time_copy);
2876 return last_update_time;
2879 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2881 return cfs_rq->avg.last_update_time;
2886 * Task first catches up with cfs_rq, and then subtract
2887 * itself from the cfs_rq (task must be off the queue now).
2889 void remove_entity_load_avg(struct sched_entity *se)
2891 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2892 u64 last_update_time;
2895 * Newly created task or never used group entity should not be removed
2896 * from its (source) cfs_rq
2898 if (se->avg.last_update_time == 0)
2901 last_update_time = cfs_rq_last_update_time(cfs_rq);
2903 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2904 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2905 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2909 * Update the rq's load with the elapsed running time before entering
2910 * idle. if the last scheduled task is not a CFS task, idle_enter will
2911 * be the only way to update the runnable statistic.
2913 void idle_enter_fair(struct rq *this_rq)
2918 * Update the rq's load with the elapsed idle time before a task is
2919 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2920 * be the only way to update the runnable statistic.
2922 void idle_exit_fair(struct rq *this_rq)
2926 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2928 return cfs_rq->runnable_load_avg;
2931 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2933 return cfs_rq->avg.load_avg;
2936 static int idle_balance(struct rq *this_rq);
2938 #else /* CONFIG_SMP */
2940 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2942 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2944 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2945 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2948 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2950 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2952 static inline int idle_balance(struct rq *rq)
2957 #endif /* CONFIG_SMP */
2959 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2961 #ifdef CONFIG_SCHEDSTATS
2962 struct task_struct *tsk = NULL;
2964 if (entity_is_task(se))
2967 if (se->statistics.sleep_start) {
2968 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2973 if (unlikely(delta > se->statistics.sleep_max))
2974 se->statistics.sleep_max = delta;
2976 se->statistics.sleep_start = 0;
2977 se->statistics.sum_sleep_runtime += delta;
2980 account_scheduler_latency(tsk, delta >> 10, 1);
2981 trace_sched_stat_sleep(tsk, delta);
2984 if (se->statistics.block_start) {
2985 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2990 if (unlikely(delta > se->statistics.block_max))
2991 se->statistics.block_max = delta;
2993 se->statistics.block_start = 0;
2994 se->statistics.sum_sleep_runtime += delta;
2997 if (tsk->in_iowait) {
2998 se->statistics.iowait_sum += delta;
2999 se->statistics.iowait_count++;
3000 trace_sched_stat_iowait(tsk, delta);
3003 trace_sched_stat_blocked(tsk, delta);
3004 trace_sched_blocked_reason(tsk);
3007 * Blocking time is in units of nanosecs, so shift by
3008 * 20 to get a milliseconds-range estimation of the
3009 * amount of time that the task spent sleeping:
3011 if (unlikely(prof_on == SLEEP_PROFILING)) {
3012 profile_hits(SLEEP_PROFILING,
3013 (void *)get_wchan(tsk),
3016 account_scheduler_latency(tsk, delta >> 10, 0);
3022 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3024 #ifdef CONFIG_SCHED_DEBUG
3025 s64 d = se->vruntime - cfs_rq->min_vruntime;
3030 if (d > 3*sysctl_sched_latency)
3031 schedstat_inc(cfs_rq, nr_spread_over);
3036 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3038 u64 vruntime = cfs_rq->min_vruntime;
3041 * The 'current' period is already promised to the current tasks,
3042 * however the extra weight of the new task will slow them down a
3043 * little, place the new task so that it fits in the slot that
3044 * stays open at the end.
3046 if (initial && sched_feat(START_DEBIT))
3047 vruntime += sched_vslice(cfs_rq, se);
3049 /* sleeps up to a single latency don't count. */
3051 unsigned long thresh = sysctl_sched_latency;
3054 * Halve their sleep time's effect, to allow
3055 * for a gentler effect of sleepers:
3057 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3063 /* ensure we never gain time by being placed backwards. */
3064 se->vruntime = max_vruntime(se->vruntime, vruntime);
3067 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3070 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3073 * Update the normalized vruntime before updating min_vruntime
3074 * through calling update_curr().
3076 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3077 se->vruntime += cfs_rq->min_vruntime;
3080 * Update run-time statistics of the 'current'.
3082 update_curr(cfs_rq);
3083 enqueue_entity_load_avg(cfs_rq, se);
3084 account_entity_enqueue(cfs_rq, se);
3085 update_cfs_shares(cfs_rq);
3087 if (flags & ENQUEUE_WAKEUP) {
3088 place_entity(cfs_rq, se, 0);
3089 enqueue_sleeper(cfs_rq, se);
3092 update_stats_enqueue(cfs_rq, se);
3093 check_spread(cfs_rq, se);
3094 if (se != cfs_rq->curr)
3095 __enqueue_entity(cfs_rq, se);
3098 if (cfs_rq->nr_running == 1) {
3099 list_add_leaf_cfs_rq(cfs_rq);
3100 check_enqueue_throttle(cfs_rq);
3104 static void __clear_buddies_last(struct sched_entity *se)
3106 for_each_sched_entity(se) {
3107 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3108 if (cfs_rq->last != se)
3111 cfs_rq->last = NULL;
3115 static void __clear_buddies_next(struct sched_entity *se)
3117 for_each_sched_entity(se) {
3118 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3119 if (cfs_rq->next != se)
3122 cfs_rq->next = NULL;
3126 static void __clear_buddies_skip(struct sched_entity *se)
3128 for_each_sched_entity(se) {
3129 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3130 if (cfs_rq->skip != se)
3133 cfs_rq->skip = NULL;
3137 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3139 if (cfs_rq->last == se)
3140 __clear_buddies_last(se);
3142 if (cfs_rq->next == se)
3143 __clear_buddies_next(se);
3145 if (cfs_rq->skip == se)
3146 __clear_buddies_skip(se);
3149 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3152 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3155 * Update run-time statistics of the 'current'.
3157 update_curr(cfs_rq);
3158 dequeue_entity_load_avg(cfs_rq, se);
3160 update_stats_dequeue(cfs_rq, se);
3161 if (flags & DEQUEUE_SLEEP) {
3162 #ifdef CONFIG_SCHEDSTATS
3163 if (entity_is_task(se)) {
3164 struct task_struct *tsk = task_of(se);
3166 if (tsk->state & TASK_INTERRUPTIBLE)
3167 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3168 if (tsk->state & TASK_UNINTERRUPTIBLE)
3169 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3174 clear_buddies(cfs_rq, se);
3176 if (se != cfs_rq->curr)
3177 __dequeue_entity(cfs_rq, se);
3179 account_entity_dequeue(cfs_rq, se);
3182 * Normalize the entity after updating the min_vruntime because the
3183 * update can refer to the ->curr item and we need to reflect this
3184 * movement in our normalized position.
3186 if (!(flags & DEQUEUE_SLEEP))
3187 se->vruntime -= cfs_rq->min_vruntime;
3189 /* return excess runtime on last dequeue */
3190 return_cfs_rq_runtime(cfs_rq);
3192 update_min_vruntime(cfs_rq);
3193 update_cfs_shares(cfs_rq);
3197 * Preempt the current task with a newly woken task if needed:
3200 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3202 unsigned long ideal_runtime, delta_exec;
3203 struct sched_entity *se;
3206 ideal_runtime = sched_slice(cfs_rq, curr);
3207 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3208 if (delta_exec > ideal_runtime) {
3209 resched_curr(rq_of(cfs_rq));
3211 * The current task ran long enough, ensure it doesn't get
3212 * re-elected due to buddy favours.
3214 clear_buddies(cfs_rq, curr);
3219 * Ensure that a task that missed wakeup preemption by a
3220 * narrow margin doesn't have to wait for a full slice.
3221 * This also mitigates buddy induced latencies under load.
3223 if (delta_exec < sysctl_sched_min_granularity)
3226 se = __pick_first_entity(cfs_rq);
3227 delta = curr->vruntime - se->vruntime;
3232 if (delta > ideal_runtime)
3233 resched_curr(rq_of(cfs_rq));
3237 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3239 /* 'current' is not kept within the tree. */
3242 * Any task has to be enqueued before it get to execute on
3243 * a CPU. So account for the time it spent waiting on the
3246 update_stats_wait_end(cfs_rq, se);
3247 __dequeue_entity(cfs_rq, se);
3248 update_load_avg(se, 1);
3251 update_stats_curr_start(cfs_rq, se);
3253 #ifdef CONFIG_SCHEDSTATS
3255 * Track our maximum slice length, if the CPU's load is at
3256 * least twice that of our own weight (i.e. dont track it
3257 * when there are only lesser-weight tasks around):
3259 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3260 se->statistics.slice_max = max(se->statistics.slice_max,
3261 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3264 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3268 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3271 * Pick the next process, keeping these things in mind, in this order:
3272 * 1) keep things fair between processes/task groups
3273 * 2) pick the "next" process, since someone really wants that to run
3274 * 3) pick the "last" process, for cache locality
3275 * 4) do not run the "skip" process, if something else is available
3277 static struct sched_entity *
3278 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3280 struct sched_entity *left = __pick_first_entity(cfs_rq);
3281 struct sched_entity *se;
3284 * If curr is set we have to see if its left of the leftmost entity
3285 * still in the tree, provided there was anything in the tree at all.
3287 if (!left || (curr && entity_before(curr, left)))
3290 se = left; /* ideally we run the leftmost entity */
3293 * Avoid running the skip buddy, if running something else can
3294 * be done without getting too unfair.
3296 if (cfs_rq->skip == se) {
3297 struct sched_entity *second;
3300 second = __pick_first_entity(cfs_rq);
3302 second = __pick_next_entity(se);
3303 if (!second || (curr && entity_before(curr, second)))
3307 if (second && wakeup_preempt_entity(second, left) < 1)
3312 * Prefer last buddy, try to return the CPU to a preempted task.
3314 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3318 * Someone really wants this to run. If it's not unfair, run it.
3320 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3323 clear_buddies(cfs_rq, se);
3328 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3330 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3333 * If still on the runqueue then deactivate_task()
3334 * was not called and update_curr() has to be done:
3337 update_curr(cfs_rq);
3339 /* throttle cfs_rqs exceeding runtime */
3340 check_cfs_rq_runtime(cfs_rq);
3342 check_spread(cfs_rq, prev);
3344 update_stats_wait_start(cfs_rq, prev);
3345 /* Put 'current' back into the tree. */
3346 __enqueue_entity(cfs_rq, prev);
3347 /* in !on_rq case, update occurred at dequeue */
3348 update_load_avg(prev, 0);
3350 cfs_rq->curr = NULL;
3354 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3357 * Update run-time statistics of the 'current'.
3359 update_curr(cfs_rq);
3362 * Ensure that runnable average is periodically updated.
3364 update_load_avg(curr, 1);
3365 update_cfs_shares(cfs_rq);
3367 #ifdef CONFIG_SCHED_HRTICK
3369 * queued ticks are scheduled to match the slice, so don't bother
3370 * validating it and just reschedule.
3373 resched_curr(rq_of(cfs_rq));
3377 * don't let the period tick interfere with the hrtick preemption
3379 if (!sched_feat(DOUBLE_TICK) &&
3380 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3384 if (cfs_rq->nr_running > 1)
3385 check_preempt_tick(cfs_rq, curr);
3389 /**************************************************
3390 * CFS bandwidth control machinery
3393 #ifdef CONFIG_CFS_BANDWIDTH
3395 #ifdef HAVE_JUMP_LABEL
3396 static struct static_key __cfs_bandwidth_used;
3398 static inline bool cfs_bandwidth_used(void)
3400 return static_key_false(&__cfs_bandwidth_used);
3403 void cfs_bandwidth_usage_inc(void)
3405 static_key_slow_inc(&__cfs_bandwidth_used);
3408 void cfs_bandwidth_usage_dec(void)
3410 static_key_slow_dec(&__cfs_bandwidth_used);
3412 #else /* HAVE_JUMP_LABEL */
3413 static bool cfs_bandwidth_used(void)
3418 void cfs_bandwidth_usage_inc(void) {}
3419 void cfs_bandwidth_usage_dec(void) {}
3420 #endif /* HAVE_JUMP_LABEL */
3423 * default period for cfs group bandwidth.
3424 * default: 0.1s, units: nanoseconds
3426 static inline u64 default_cfs_period(void)
3428 return 100000000ULL;
3431 static inline u64 sched_cfs_bandwidth_slice(void)
3433 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3437 * Replenish runtime according to assigned quota and update expiration time.
3438 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3439 * additional synchronization around rq->lock.
3441 * requires cfs_b->lock
3443 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3447 if (cfs_b->quota == RUNTIME_INF)
3450 now = sched_clock_cpu(smp_processor_id());
3451 cfs_b->runtime = cfs_b->quota;
3452 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3455 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3457 return &tg->cfs_bandwidth;
3460 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3461 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3463 if (unlikely(cfs_rq->throttle_count))
3464 return cfs_rq->throttled_clock_task;
3466 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3469 /* returns 0 on failure to allocate runtime */
3470 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3472 struct task_group *tg = cfs_rq->tg;
3473 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3474 u64 amount = 0, min_amount, expires;
3476 /* note: this is a positive sum as runtime_remaining <= 0 */
3477 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3479 raw_spin_lock(&cfs_b->lock);
3480 if (cfs_b->quota == RUNTIME_INF)
3481 amount = min_amount;
3483 start_cfs_bandwidth(cfs_b);
3485 if (cfs_b->runtime > 0) {
3486 amount = min(cfs_b->runtime, min_amount);
3487 cfs_b->runtime -= amount;
3491 expires = cfs_b->runtime_expires;
3492 raw_spin_unlock(&cfs_b->lock);
3494 cfs_rq->runtime_remaining += amount;
3496 * we may have advanced our local expiration to account for allowed
3497 * spread between our sched_clock and the one on which runtime was
3500 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3501 cfs_rq->runtime_expires = expires;
3503 return cfs_rq->runtime_remaining > 0;
3507 * Note: This depends on the synchronization provided by sched_clock and the
3508 * fact that rq->clock snapshots this value.
3510 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3512 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3514 /* if the deadline is ahead of our clock, nothing to do */
3515 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3518 if (cfs_rq->runtime_remaining < 0)
3522 * If the local deadline has passed we have to consider the
3523 * possibility that our sched_clock is 'fast' and the global deadline
3524 * has not truly expired.
3526 * Fortunately we can check determine whether this the case by checking
3527 * whether the global deadline has advanced. It is valid to compare
3528 * cfs_b->runtime_expires without any locks since we only care about
3529 * exact equality, so a partial write will still work.
3532 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3533 /* extend local deadline, drift is bounded above by 2 ticks */
3534 cfs_rq->runtime_expires += TICK_NSEC;
3536 /* global deadline is ahead, expiration has passed */
3537 cfs_rq->runtime_remaining = 0;
3541 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3543 /* dock delta_exec before expiring quota (as it could span periods) */
3544 cfs_rq->runtime_remaining -= delta_exec;
3545 expire_cfs_rq_runtime(cfs_rq);
3547 if (likely(cfs_rq->runtime_remaining > 0))
3551 * if we're unable to extend our runtime we resched so that the active
3552 * hierarchy can be throttled
3554 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3555 resched_curr(rq_of(cfs_rq));
3558 static __always_inline
3559 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3561 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3564 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3567 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3569 return cfs_bandwidth_used() && cfs_rq->throttled;
3572 /* check whether cfs_rq, or any parent, is throttled */
3573 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3575 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3579 * Ensure that neither of the group entities corresponding to src_cpu or
3580 * dest_cpu are members of a throttled hierarchy when performing group
3581 * load-balance operations.
3583 static inline int throttled_lb_pair(struct task_group *tg,
3584 int src_cpu, int dest_cpu)
3586 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3588 src_cfs_rq = tg->cfs_rq[src_cpu];
3589 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3591 return throttled_hierarchy(src_cfs_rq) ||
3592 throttled_hierarchy(dest_cfs_rq);
3595 /* updated child weight may affect parent so we have to do this bottom up */
3596 static int tg_unthrottle_up(struct task_group *tg, void *data)
3598 struct rq *rq = data;
3599 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3601 cfs_rq->throttle_count--;
3603 if (!cfs_rq->throttle_count) {
3604 /* adjust cfs_rq_clock_task() */
3605 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3606 cfs_rq->throttled_clock_task;
3613 static int tg_throttle_down(struct task_group *tg, void *data)
3615 struct rq *rq = data;
3616 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3618 /* group is entering throttled state, stop time */
3619 if (!cfs_rq->throttle_count)
3620 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3621 cfs_rq->throttle_count++;
3626 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3628 struct rq *rq = rq_of(cfs_rq);
3629 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3630 struct sched_entity *se;
3631 long task_delta, dequeue = 1;
3634 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3636 /* freeze hierarchy runnable averages while throttled */
3638 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3641 task_delta = cfs_rq->h_nr_running;
3642 for_each_sched_entity(se) {
3643 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3644 /* throttled entity or throttle-on-deactivate */
3649 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3650 qcfs_rq->h_nr_running -= task_delta;
3652 if (qcfs_rq->load.weight)
3657 sub_nr_running(rq, task_delta);
3659 cfs_rq->throttled = 1;
3660 cfs_rq->throttled_clock = rq_clock(rq);
3661 raw_spin_lock(&cfs_b->lock);
3662 empty = list_empty(&cfs_b->throttled_cfs_rq);
3665 * Add to the _head_ of the list, so that an already-started
3666 * distribute_cfs_runtime will not see us
3668 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3671 * If we're the first throttled task, make sure the bandwidth
3675 start_cfs_bandwidth(cfs_b);
3677 raw_spin_unlock(&cfs_b->lock);
3680 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3682 struct rq *rq = rq_of(cfs_rq);
3683 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3684 struct sched_entity *se;
3688 se = cfs_rq->tg->se[cpu_of(rq)];
3690 cfs_rq->throttled = 0;
3692 update_rq_clock(rq);
3694 raw_spin_lock(&cfs_b->lock);
3695 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3696 list_del_rcu(&cfs_rq->throttled_list);
3697 raw_spin_unlock(&cfs_b->lock);
3699 /* update hierarchical throttle state */
3700 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3702 if (!cfs_rq->load.weight)
3705 task_delta = cfs_rq->h_nr_running;
3706 for_each_sched_entity(se) {
3710 cfs_rq = cfs_rq_of(se);
3712 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3713 cfs_rq->h_nr_running += task_delta;
3715 if (cfs_rq_throttled(cfs_rq))
3720 add_nr_running(rq, task_delta);
3722 /* determine whether we need to wake up potentially idle cpu */
3723 if (rq->curr == rq->idle && rq->cfs.nr_running)
3727 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3728 u64 remaining, u64 expires)
3730 struct cfs_rq *cfs_rq;
3732 u64 starting_runtime = remaining;
3735 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3737 struct rq *rq = rq_of(cfs_rq);
3739 raw_spin_lock(&rq->lock);
3740 if (!cfs_rq_throttled(cfs_rq))
3743 runtime = -cfs_rq->runtime_remaining + 1;
3744 if (runtime > remaining)
3745 runtime = remaining;
3746 remaining -= runtime;
3748 cfs_rq->runtime_remaining += runtime;
3749 cfs_rq->runtime_expires = expires;
3751 /* we check whether we're throttled above */
3752 if (cfs_rq->runtime_remaining > 0)
3753 unthrottle_cfs_rq(cfs_rq);
3756 raw_spin_unlock(&rq->lock);
3763 return starting_runtime - remaining;
3767 * Responsible for refilling a task_group's bandwidth and unthrottling its
3768 * cfs_rqs as appropriate. If there has been no activity within the last
3769 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3770 * used to track this state.
3772 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3774 u64 runtime, runtime_expires;
3777 /* no need to continue the timer with no bandwidth constraint */
3778 if (cfs_b->quota == RUNTIME_INF)
3779 goto out_deactivate;
3781 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3782 cfs_b->nr_periods += overrun;
3785 * idle depends on !throttled (for the case of a large deficit), and if
3786 * we're going inactive then everything else can be deferred
3788 if (cfs_b->idle && !throttled)
3789 goto out_deactivate;
3791 __refill_cfs_bandwidth_runtime(cfs_b);
3794 /* mark as potentially idle for the upcoming period */
3799 /* account preceding periods in which throttling occurred */
3800 cfs_b->nr_throttled += overrun;
3802 runtime_expires = cfs_b->runtime_expires;
3805 * This check is repeated as we are holding onto the new bandwidth while
3806 * we unthrottle. This can potentially race with an unthrottled group
3807 * trying to acquire new bandwidth from the global pool. This can result
3808 * in us over-using our runtime if it is all used during this loop, but
3809 * only by limited amounts in that extreme case.
3811 while (throttled && cfs_b->runtime > 0) {
3812 runtime = cfs_b->runtime;
3813 raw_spin_unlock(&cfs_b->lock);
3814 /* we can't nest cfs_b->lock while distributing bandwidth */
3815 runtime = distribute_cfs_runtime(cfs_b, runtime,
3817 raw_spin_lock(&cfs_b->lock);
3819 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3821 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3825 * While we are ensured activity in the period following an
3826 * unthrottle, this also covers the case in which the new bandwidth is
3827 * insufficient to cover the existing bandwidth deficit. (Forcing the
3828 * timer to remain active while there are any throttled entities.)
3838 /* a cfs_rq won't donate quota below this amount */
3839 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3840 /* minimum remaining period time to redistribute slack quota */
3841 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3842 /* how long we wait to gather additional slack before distributing */
3843 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3846 * Are we near the end of the current quota period?
3848 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3849 * hrtimer base being cleared by hrtimer_start. In the case of
3850 * migrate_hrtimers, base is never cleared, so we are fine.
3852 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3854 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3857 /* if the call-back is running a quota refresh is already occurring */
3858 if (hrtimer_callback_running(refresh_timer))
3861 /* is a quota refresh about to occur? */
3862 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3863 if (remaining < min_expire)
3869 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3871 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3873 /* if there's a quota refresh soon don't bother with slack */
3874 if (runtime_refresh_within(cfs_b, min_left))
3877 hrtimer_start(&cfs_b->slack_timer,
3878 ns_to_ktime(cfs_bandwidth_slack_period),
3882 /* we know any runtime found here is valid as update_curr() precedes return */
3883 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3885 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3886 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3888 if (slack_runtime <= 0)
3891 raw_spin_lock(&cfs_b->lock);
3892 if (cfs_b->quota != RUNTIME_INF &&
3893 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3894 cfs_b->runtime += slack_runtime;
3896 /* we are under rq->lock, defer unthrottling using a timer */
3897 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3898 !list_empty(&cfs_b->throttled_cfs_rq))
3899 start_cfs_slack_bandwidth(cfs_b);
3901 raw_spin_unlock(&cfs_b->lock);
3903 /* even if it's not valid for return we don't want to try again */
3904 cfs_rq->runtime_remaining -= slack_runtime;
3907 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3909 if (!cfs_bandwidth_used())
3912 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3915 __return_cfs_rq_runtime(cfs_rq);
3919 * This is done with a timer (instead of inline with bandwidth return) since
3920 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3922 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3924 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3927 /* confirm we're still not at a refresh boundary */
3928 raw_spin_lock(&cfs_b->lock);
3929 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3930 raw_spin_unlock(&cfs_b->lock);
3934 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3935 runtime = cfs_b->runtime;
3937 expires = cfs_b->runtime_expires;
3938 raw_spin_unlock(&cfs_b->lock);
3943 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3945 raw_spin_lock(&cfs_b->lock);
3946 if (expires == cfs_b->runtime_expires)
3947 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3948 raw_spin_unlock(&cfs_b->lock);
3952 * When a group wakes up we want to make sure that its quota is not already
3953 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3954 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3956 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3958 if (!cfs_bandwidth_used())
3961 /* an active group must be handled by the update_curr()->put() path */
3962 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3965 /* ensure the group is not already throttled */
3966 if (cfs_rq_throttled(cfs_rq))
3969 /* update runtime allocation */
3970 account_cfs_rq_runtime(cfs_rq, 0);
3971 if (cfs_rq->runtime_remaining <= 0)
3972 throttle_cfs_rq(cfs_rq);
3975 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3976 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3978 if (!cfs_bandwidth_used())
3981 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3985 * it's possible for a throttled entity to be forced into a running
3986 * state (e.g. set_curr_task), in this case we're finished.
3988 if (cfs_rq_throttled(cfs_rq))
3991 throttle_cfs_rq(cfs_rq);
3995 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3997 struct cfs_bandwidth *cfs_b =
3998 container_of(timer, struct cfs_bandwidth, slack_timer);
4000 do_sched_cfs_slack_timer(cfs_b);
4002 return HRTIMER_NORESTART;
4005 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4007 struct cfs_bandwidth *cfs_b =
4008 container_of(timer, struct cfs_bandwidth, period_timer);
4012 raw_spin_lock(&cfs_b->lock);
4014 overrun = hrtimer_forward_now(timer, cfs_b->period);
4018 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4021 cfs_b->period_active = 0;
4022 raw_spin_unlock(&cfs_b->lock);
4024 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4027 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4029 raw_spin_lock_init(&cfs_b->lock);
4031 cfs_b->quota = RUNTIME_INF;
4032 cfs_b->period = ns_to_ktime(default_cfs_period());
4034 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4035 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4036 cfs_b->period_timer.function = sched_cfs_period_timer;
4037 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4038 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4041 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4043 cfs_rq->runtime_enabled = 0;
4044 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4047 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4049 lockdep_assert_held(&cfs_b->lock);
4051 if (!cfs_b->period_active) {
4052 cfs_b->period_active = 1;
4053 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4054 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4058 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4060 /* init_cfs_bandwidth() was not called */
4061 if (!cfs_b->throttled_cfs_rq.next)
4064 hrtimer_cancel(&cfs_b->period_timer);
4065 hrtimer_cancel(&cfs_b->slack_timer);
4068 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4070 struct cfs_rq *cfs_rq;
4072 for_each_leaf_cfs_rq(rq, cfs_rq) {
4073 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4075 raw_spin_lock(&cfs_b->lock);
4076 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4077 raw_spin_unlock(&cfs_b->lock);
4081 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4083 struct cfs_rq *cfs_rq;
4085 for_each_leaf_cfs_rq(rq, cfs_rq) {
4086 if (!cfs_rq->runtime_enabled)
4090 * clock_task is not advancing so we just need to make sure
4091 * there's some valid quota amount
4093 cfs_rq->runtime_remaining = 1;
4095 * Offline rq is schedulable till cpu is completely disabled
4096 * in take_cpu_down(), so we prevent new cfs throttling here.
4098 cfs_rq->runtime_enabled = 0;
4100 if (cfs_rq_throttled(cfs_rq))
4101 unthrottle_cfs_rq(cfs_rq);
4105 #else /* CONFIG_CFS_BANDWIDTH */
4106 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4108 return rq_clock_task(rq_of(cfs_rq));
4111 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4112 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4113 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4114 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4116 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4121 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4126 static inline int throttled_lb_pair(struct task_group *tg,
4127 int src_cpu, int dest_cpu)
4132 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4134 #ifdef CONFIG_FAIR_GROUP_SCHED
4135 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4138 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4142 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4143 static inline void update_runtime_enabled(struct rq *rq) {}
4144 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4146 #endif /* CONFIG_CFS_BANDWIDTH */
4148 /**************************************************
4149 * CFS operations on tasks:
4152 #ifdef CONFIG_SCHED_HRTICK
4153 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4155 struct sched_entity *se = &p->se;
4156 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4158 WARN_ON(task_rq(p) != rq);
4160 if (cfs_rq->nr_running > 1) {
4161 u64 slice = sched_slice(cfs_rq, se);
4162 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4163 s64 delta = slice - ran;
4170 hrtick_start(rq, delta);
4175 * called from enqueue/dequeue and updates the hrtick when the
4176 * current task is from our class and nr_running is low enough
4179 static void hrtick_update(struct rq *rq)
4181 struct task_struct *curr = rq->curr;
4183 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4186 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4187 hrtick_start_fair(rq, curr);
4189 #else /* !CONFIG_SCHED_HRTICK */
4191 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4195 static inline void hrtick_update(struct rq *rq)
4201 static bool cpu_overutilized(int cpu);
4202 static inline unsigned long boosted_cpu_util(int cpu);
4204 #define boosted_cpu_util(cpu) cpu_util(cpu)
4208 static void update_capacity_of(int cpu)
4210 unsigned long req_cap;
4215 /* Convert scale-invariant capacity to cpu. */
4216 req_cap = boosted_cpu_util(cpu);
4217 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4218 set_cfs_cpu_capacity(cpu, true, req_cap);
4223 * The enqueue_task method is called before nr_running is
4224 * increased. Here we update the fair scheduling stats and
4225 * then put the task into the rbtree:
4228 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4230 struct cfs_rq *cfs_rq;
4231 struct sched_entity *se = &p->se;
4233 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4234 int task_wakeup = flags & ENQUEUE_WAKEUP;
4237 for_each_sched_entity(se) {
4240 cfs_rq = cfs_rq_of(se);
4241 enqueue_entity(cfs_rq, se, flags);
4244 * end evaluation on encountering a throttled cfs_rq
4246 * note: in the case of encountering a throttled cfs_rq we will
4247 * post the final h_nr_running increment below.
4249 if (cfs_rq_throttled(cfs_rq))
4251 cfs_rq->h_nr_running++;
4252 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4254 flags = ENQUEUE_WAKEUP;
4257 for_each_sched_entity(se) {
4258 cfs_rq = cfs_rq_of(se);
4259 cfs_rq->h_nr_running++;
4260 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4262 if (cfs_rq_throttled(cfs_rq))
4265 update_load_avg(se, 1);
4266 update_cfs_shares(cfs_rq);
4270 add_nr_running(rq, 1);
4275 * Update SchedTune accounting.
4277 * We do it before updating the CPU capacity to ensure the
4278 * boost value of the current task is accounted for in the
4279 * selection of the OPP.
4281 * We do it also in the case where we enqueue a throttled task;
4282 * we could argue that a throttled task should not boost a CPU,
4284 * a) properly implementing CPU boosting considering throttled
4285 * tasks will increase a lot the complexity of the solution
4286 * b) it's not easy to quantify the benefits introduced by
4287 * such a more complex solution.
4288 * Thus, for the time being we go for the simple solution and boost
4289 * also for throttled RQs.
4291 schedtune_enqueue_task(p, cpu_of(rq));
4294 walt_inc_cumulative_runnable_avg(rq, p);
4295 if (!task_new && !rq->rd->overutilized &&
4296 cpu_overutilized(rq->cpu)) {
4297 rq->rd->overutilized = true;
4298 trace_sched_overutilized(true);
4302 * We want to potentially trigger a freq switch
4303 * request only for tasks that are waking up; this is
4304 * because we get here also during load balancing, but
4305 * in these cases it seems wise to trigger as single
4306 * request after load balancing is done.
4308 if (task_new || task_wakeup)
4309 update_capacity_of(cpu_of(rq));
4312 #endif /* CONFIG_SMP */
4316 static void set_next_buddy(struct sched_entity *se);
4319 * The dequeue_task method is called before nr_running is
4320 * decreased. We remove the task from the rbtree and
4321 * update the fair scheduling stats:
4323 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4325 struct cfs_rq *cfs_rq;
4326 struct sched_entity *se = &p->se;
4327 int task_sleep = flags & DEQUEUE_SLEEP;
4329 for_each_sched_entity(se) {
4330 cfs_rq = cfs_rq_of(se);
4331 dequeue_entity(cfs_rq, se, flags);
4334 * end evaluation on encountering a throttled cfs_rq
4336 * note: in the case of encountering a throttled cfs_rq we will
4337 * post the final h_nr_running decrement below.
4339 if (cfs_rq_throttled(cfs_rq))
4341 cfs_rq->h_nr_running--;
4342 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4344 /* Don't dequeue parent if it has other entities besides us */
4345 if (cfs_rq->load.weight) {
4347 * Bias pick_next to pick a task from this cfs_rq, as
4348 * p is sleeping when it is within its sched_slice.
4350 if (task_sleep && parent_entity(se))
4351 set_next_buddy(parent_entity(se));
4353 /* avoid re-evaluating load for this entity */
4354 se = parent_entity(se);
4357 flags |= DEQUEUE_SLEEP;
4360 for_each_sched_entity(se) {
4361 cfs_rq = cfs_rq_of(se);
4362 cfs_rq->h_nr_running--;
4363 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4365 if (cfs_rq_throttled(cfs_rq))
4368 update_load_avg(se, 1);
4369 update_cfs_shares(cfs_rq);
4373 sub_nr_running(rq, 1);
4378 * Update SchedTune accounting
4380 * We do it before updating the CPU capacity to ensure the
4381 * boost value of the current task is accounted for in the
4382 * selection of the OPP.
4384 schedtune_dequeue_task(p, cpu_of(rq));
4387 walt_dec_cumulative_runnable_avg(rq, p);
4390 * We want to potentially trigger a freq switch
4391 * request only for tasks that are going to sleep;
4392 * this is because we get here also during load
4393 * balancing, but in these cases it seems wise to
4394 * trigger as single request after load balancing is
4398 if (rq->cfs.nr_running)
4399 update_capacity_of(cpu_of(rq));
4400 else if (sched_freq())
4401 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4405 #endif /* CONFIG_SMP */
4413 * per rq 'load' arrray crap; XXX kill this.
4417 * The exact cpuload at various idx values, calculated at every tick would be
4418 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4420 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4421 * on nth tick when cpu may be busy, then we have:
4422 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4423 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4425 * decay_load_missed() below does efficient calculation of
4426 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4427 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4429 * The calculation is approximated on a 128 point scale.
4430 * degrade_zero_ticks is the number of ticks after which load at any
4431 * particular idx is approximated to be zero.
4432 * degrade_factor is a precomputed table, a row for each load idx.
4433 * Each column corresponds to degradation factor for a power of two ticks,
4434 * based on 128 point scale.
4436 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4437 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4439 * With this power of 2 load factors, we can degrade the load n times
4440 * by looking at 1 bits in n and doing as many mult/shift instead of
4441 * n mult/shifts needed by the exact degradation.
4443 #define DEGRADE_SHIFT 7
4444 static const unsigned char
4445 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4446 static const unsigned char
4447 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4448 {0, 0, 0, 0, 0, 0, 0, 0},
4449 {64, 32, 8, 0, 0, 0, 0, 0},
4450 {96, 72, 40, 12, 1, 0, 0},
4451 {112, 98, 75, 43, 15, 1, 0},
4452 {120, 112, 98, 76, 45, 16, 2} };
4455 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4456 * would be when CPU is idle and so we just decay the old load without
4457 * adding any new load.
4459 static unsigned long
4460 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4464 if (!missed_updates)
4467 if (missed_updates >= degrade_zero_ticks[idx])
4471 return load >> missed_updates;
4473 while (missed_updates) {
4474 if (missed_updates % 2)
4475 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4477 missed_updates >>= 1;
4484 * Update rq->cpu_load[] statistics. This function is usually called every
4485 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4486 * every tick. We fix it up based on jiffies.
4488 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4489 unsigned long pending_updates)
4493 this_rq->nr_load_updates++;
4495 /* Update our load: */
4496 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4497 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4498 unsigned long old_load, new_load;
4500 /* scale is effectively 1 << i now, and >> i divides by scale */
4502 old_load = this_rq->cpu_load[i];
4503 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4504 new_load = this_load;
4506 * Round up the averaging division if load is increasing. This
4507 * prevents us from getting stuck on 9 if the load is 10, for
4510 if (new_load > old_load)
4511 new_load += scale - 1;
4513 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4516 sched_avg_update(this_rq);
4519 /* Used instead of source_load when we know the type == 0 */
4520 static unsigned long weighted_cpuload(const int cpu)
4522 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4525 #ifdef CONFIG_NO_HZ_COMMON
4527 * There is no sane way to deal with nohz on smp when using jiffies because the
4528 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4529 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4531 * Therefore we cannot use the delta approach from the regular tick since that
4532 * would seriously skew the load calculation. However we'll make do for those
4533 * updates happening while idle (nohz_idle_balance) or coming out of idle
4534 * (tick_nohz_idle_exit).
4536 * This means we might still be one tick off for nohz periods.
4540 * Called from nohz_idle_balance() to update the load ratings before doing the
4543 static void update_idle_cpu_load(struct rq *this_rq)
4545 unsigned long curr_jiffies = READ_ONCE(jiffies);
4546 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4547 unsigned long pending_updates;
4550 * bail if there's load or we're actually up-to-date.
4552 if (load || curr_jiffies == this_rq->last_load_update_tick)
4555 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4556 this_rq->last_load_update_tick = curr_jiffies;
4558 __update_cpu_load(this_rq, load, pending_updates);
4562 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4564 void update_cpu_load_nohz(void)
4566 struct rq *this_rq = this_rq();
4567 unsigned long curr_jiffies = READ_ONCE(jiffies);
4568 unsigned long pending_updates;
4570 if (curr_jiffies == this_rq->last_load_update_tick)
4573 raw_spin_lock(&this_rq->lock);
4574 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4575 if (pending_updates) {
4576 this_rq->last_load_update_tick = curr_jiffies;
4578 * We were idle, this means load 0, the current load might be
4579 * !0 due to remote wakeups and the sort.
4581 __update_cpu_load(this_rq, 0, pending_updates);
4583 raw_spin_unlock(&this_rq->lock);
4585 #endif /* CONFIG_NO_HZ */
4588 * Called from scheduler_tick()
4590 void update_cpu_load_active(struct rq *this_rq)
4592 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4594 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4596 this_rq->last_load_update_tick = jiffies;
4597 __update_cpu_load(this_rq, load, 1);
4601 * Return a low guess at the load of a migration-source cpu weighted
4602 * according to the scheduling class and "nice" value.
4604 * We want to under-estimate the load of migration sources, to
4605 * balance conservatively.
4607 static unsigned long source_load(int cpu, int type)
4609 struct rq *rq = cpu_rq(cpu);
4610 unsigned long total = weighted_cpuload(cpu);
4612 if (type == 0 || !sched_feat(LB_BIAS))
4615 return min(rq->cpu_load[type-1], total);
4619 * Return a high guess at the load of a migration-target cpu weighted
4620 * according to the scheduling class and "nice" value.
4622 static unsigned long target_load(int cpu, int type)
4624 struct rq *rq = cpu_rq(cpu);
4625 unsigned long total = weighted_cpuload(cpu);
4627 if (type == 0 || !sched_feat(LB_BIAS))
4630 return max(rq->cpu_load[type-1], total);
4634 static unsigned long cpu_avg_load_per_task(int cpu)
4636 struct rq *rq = cpu_rq(cpu);
4637 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4638 unsigned long load_avg = weighted_cpuload(cpu);
4641 return load_avg / nr_running;
4646 static void record_wakee(struct task_struct *p)
4649 * Rough decay (wiping) for cost saving, don't worry
4650 * about the boundary, really active task won't care
4653 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4654 current->wakee_flips >>= 1;
4655 current->wakee_flip_decay_ts = jiffies;
4658 if (current->last_wakee != p) {
4659 current->last_wakee = p;
4660 current->wakee_flips++;
4664 static void task_waking_fair(struct task_struct *p)
4666 struct sched_entity *se = &p->se;
4667 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4670 #ifndef CONFIG_64BIT
4671 u64 min_vruntime_copy;
4674 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4676 min_vruntime = cfs_rq->min_vruntime;
4677 } while (min_vruntime != min_vruntime_copy);
4679 min_vruntime = cfs_rq->min_vruntime;
4682 se->vruntime -= min_vruntime;
4686 #ifdef CONFIG_FAIR_GROUP_SCHED
4688 * effective_load() calculates the load change as seen from the root_task_group
4690 * Adding load to a group doesn't make a group heavier, but can cause movement
4691 * of group shares between cpus. Assuming the shares were perfectly aligned one
4692 * can calculate the shift in shares.
4694 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4695 * on this @cpu and results in a total addition (subtraction) of @wg to the
4696 * total group weight.
4698 * Given a runqueue weight distribution (rw_i) we can compute a shares
4699 * distribution (s_i) using:
4701 * s_i = rw_i / \Sum rw_j (1)
4703 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4704 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4705 * shares distribution (s_i):
4707 * rw_i = { 2, 4, 1, 0 }
4708 * s_i = { 2/7, 4/7, 1/7, 0 }
4710 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4711 * task used to run on and the CPU the waker is running on), we need to
4712 * compute the effect of waking a task on either CPU and, in case of a sync
4713 * wakeup, compute the effect of the current task going to sleep.
4715 * So for a change of @wl to the local @cpu with an overall group weight change
4716 * of @wl we can compute the new shares distribution (s'_i) using:
4718 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4720 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4721 * differences in waking a task to CPU 0. The additional task changes the
4722 * weight and shares distributions like:
4724 * rw'_i = { 3, 4, 1, 0 }
4725 * s'_i = { 3/8, 4/8, 1/8, 0 }
4727 * We can then compute the difference in effective weight by using:
4729 * dw_i = S * (s'_i - s_i) (3)
4731 * Where 'S' is the group weight as seen by its parent.
4733 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4734 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4735 * 4/7) times the weight of the group.
4737 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4739 struct sched_entity *se = tg->se[cpu];
4741 if (!tg->parent) /* the trivial, non-cgroup case */
4744 for_each_sched_entity(se) {
4745 struct cfs_rq *cfs_rq = se->my_q;
4746 long W, w = cfs_rq_load_avg(cfs_rq);
4751 * W = @wg + \Sum rw_j
4753 W = wg + atomic_long_read(&tg->load_avg);
4755 /* Ensure \Sum rw_j >= rw_i */
4756 W -= cfs_rq->tg_load_avg_contrib;
4765 * wl = S * s'_i; see (2)
4768 wl = (w * (long)tg->shares) / W;
4773 * Per the above, wl is the new se->load.weight value; since
4774 * those are clipped to [MIN_SHARES, ...) do so now. See
4775 * calc_cfs_shares().
4777 if (wl < MIN_SHARES)
4781 * wl = dw_i = S * (s'_i - s_i); see (3)
4783 wl -= se->avg.load_avg;
4786 * Recursively apply this logic to all parent groups to compute
4787 * the final effective load change on the root group. Since
4788 * only the @tg group gets extra weight, all parent groups can
4789 * only redistribute existing shares. @wl is the shift in shares
4790 * resulting from this level per the above.
4799 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4807 * Returns the current capacity of cpu after applying both
4808 * cpu and freq scaling.
4810 unsigned long capacity_curr_of(int cpu)
4812 return cpu_rq(cpu)->cpu_capacity_orig *
4813 arch_scale_freq_capacity(NULL, cpu)
4814 >> SCHED_CAPACITY_SHIFT;
4817 static inline bool energy_aware(void)
4819 return sched_feat(ENERGY_AWARE);
4823 struct sched_group *sg_top;
4824 struct sched_group *sg_cap;
4831 struct task_struct *task;
4846 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4847 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4848 * energy calculations. Using the scale-invariant util returned by
4849 * cpu_util() and approximating scale-invariant util by:
4851 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4853 * the normalized util can be found using the specific capacity.
4855 * capacity = capacity_orig * curr_freq/max_freq
4857 * norm_util = running_time/time ~ util/capacity
4859 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4861 int util = __cpu_util(cpu, delta);
4863 if (util >= capacity)
4864 return SCHED_CAPACITY_SCALE;
4866 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4869 static int calc_util_delta(struct energy_env *eenv, int cpu)
4871 if (cpu == eenv->src_cpu)
4872 return -eenv->util_delta;
4873 if (cpu == eenv->dst_cpu)
4874 return eenv->util_delta;
4879 unsigned long group_max_util(struct energy_env *eenv)
4882 unsigned long max_util = 0;
4884 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4885 delta = calc_util_delta(eenv, i);
4886 max_util = max(max_util, __cpu_util(i, delta));
4893 * group_norm_util() returns the approximated group util relative to it's
4894 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4895 * energy calculations. Since task executions may or may not overlap in time in
4896 * the group the true normalized util is between max(cpu_norm_util(i)) and
4897 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4898 * latter is used as the estimate as it leads to a more pessimistic energy
4899 * estimate (more busy).
4902 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4905 unsigned long util_sum = 0;
4906 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4908 for_each_cpu(i, sched_group_cpus(sg)) {
4909 delta = calc_util_delta(eenv, i);
4910 util_sum += __cpu_norm_util(i, capacity, delta);
4913 if (util_sum > SCHED_CAPACITY_SCALE)
4914 return SCHED_CAPACITY_SCALE;
4918 static int find_new_capacity(struct energy_env *eenv,
4919 const struct sched_group_energy const *sge)
4922 unsigned long util = group_max_util(eenv);
4924 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4925 if (sge->cap_states[idx].cap >= util)
4929 eenv->cap_idx = idx;
4934 static int group_idle_state(struct sched_group *sg)
4936 int i, state = INT_MAX;
4938 /* Find the shallowest idle state in the sched group. */
4939 for_each_cpu(i, sched_group_cpus(sg))
4940 state = min(state, idle_get_state_idx(cpu_rq(i)));
4942 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
4949 * sched_group_energy(): Computes the absolute energy consumption of cpus
4950 * belonging to the sched_group including shared resources shared only by
4951 * members of the group. Iterates over all cpus in the hierarchy below the
4952 * sched_group starting from the bottom working it's way up before going to
4953 * the next cpu until all cpus are covered at all levels. The current
4954 * implementation is likely to gather the same util statistics multiple times.
4955 * This can probably be done in a faster but more complex way.
4956 * Note: sched_group_energy() may fail when racing with sched_domain updates.
4958 static int sched_group_energy(struct energy_env *eenv)
4960 struct sched_domain *sd;
4961 int cpu, total_energy = 0;
4962 struct cpumask visit_cpus;
4963 struct sched_group *sg;
4965 WARN_ON(!eenv->sg_top->sge);
4967 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
4969 while (!cpumask_empty(&visit_cpus)) {
4970 struct sched_group *sg_shared_cap = NULL;
4972 cpu = cpumask_first(&visit_cpus);
4975 * Is the group utilization affected by cpus outside this
4978 sd = rcu_dereference(per_cpu(sd_scs, cpu));
4982 * We most probably raced with hotplug; returning a
4983 * wrong energy estimation is better than entering an
4989 sg_shared_cap = sd->parent->groups;
4991 for_each_domain(cpu, sd) {
4994 /* Has this sched_domain already been visited? */
4995 if (sd->child && group_first_cpu(sg) != cpu)
4999 unsigned long group_util;
5000 int sg_busy_energy, sg_idle_energy;
5001 int cap_idx, idle_idx;
5003 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5004 eenv->sg_cap = sg_shared_cap;
5008 cap_idx = find_new_capacity(eenv, sg->sge);
5010 if (sg->group_weight == 1) {
5011 /* Remove capacity of src CPU (before task move) */
5012 if (eenv->util_delta == 0 &&
5013 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5014 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5015 eenv->cap.delta -= eenv->cap.before;
5017 /* Add capacity of dst CPU (after task move) */
5018 if (eenv->util_delta != 0 &&
5019 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5020 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5021 eenv->cap.delta += eenv->cap.after;
5025 idle_idx = group_idle_state(sg);
5026 group_util = group_norm_util(eenv, sg);
5027 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5028 >> SCHED_CAPACITY_SHIFT;
5029 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5030 * sg->sge->idle_states[idle_idx].power)
5031 >> SCHED_CAPACITY_SHIFT;
5033 total_energy += sg_busy_energy + sg_idle_energy;
5036 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5038 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5041 } while (sg = sg->next, sg != sd->groups);
5044 cpumask_clear_cpu(cpu, &visit_cpus);
5048 eenv->energy = total_energy;
5052 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5054 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5058 * energy_diff(): Estimate the energy impact of changing the utilization
5059 * distribution. eenv specifies the change: utilisation amount, source, and
5060 * destination cpu. Source or destination cpu may be -1 in which case the
5061 * utilization is removed from or added to the system (e.g. task wake-up). If
5062 * both are specified, the utilization is migrated.
5064 static inline int __energy_diff(struct energy_env *eenv)
5066 struct sched_domain *sd;
5067 struct sched_group *sg;
5068 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5070 struct energy_env eenv_before = {
5072 .src_cpu = eenv->src_cpu,
5073 .dst_cpu = eenv->dst_cpu,
5074 .nrg = { 0, 0, 0, 0},
5078 if (eenv->src_cpu == eenv->dst_cpu)
5081 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5082 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5085 return 0; /* Error */
5090 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5091 eenv_before.sg_top = eenv->sg_top = sg;
5093 if (sched_group_energy(&eenv_before))
5094 return 0; /* Invalid result abort */
5095 energy_before += eenv_before.energy;
5097 /* Keep track of SRC cpu (before) capacity */
5098 eenv->cap.before = eenv_before.cap.before;
5099 eenv->cap.delta = eenv_before.cap.delta;
5101 if (sched_group_energy(eenv))
5102 return 0; /* Invalid result abort */
5103 energy_after += eenv->energy;
5105 } while (sg = sg->next, sg != sd->groups);
5107 eenv->nrg.before = energy_before;
5108 eenv->nrg.after = energy_after;
5109 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5112 trace_sched_energy_diff(eenv->task,
5113 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5114 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5115 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5116 eenv->nrg.delta, eenv->payoff);
5118 return eenv->nrg.diff;
5121 #ifdef CONFIG_SCHED_TUNE
5123 struct target_nrg schedtune_target_nrg;
5126 * System energy normalization
5127 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5128 * corresponding to the specified energy variation.
5131 normalize_energy(int energy_diff)
5134 #ifdef CONFIG_SCHED_DEBUG
5137 /* Check for boundaries */
5138 max_delta = schedtune_target_nrg.max_power;
5139 max_delta -= schedtune_target_nrg.min_power;
5140 WARN_ON(abs(energy_diff) >= max_delta);
5143 /* Do scaling using positive numbers to increase the range */
5144 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5146 /* Scale by energy magnitude */
5147 normalized_nrg <<= SCHED_LOAD_SHIFT;
5149 /* Normalize on max energy for target platform */
5150 normalized_nrg = reciprocal_divide(
5151 normalized_nrg, schedtune_target_nrg.rdiv);
5153 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5157 energy_diff(struct energy_env *eenv)
5159 int boost = schedtune_task_boost(eenv->task);
5162 /* Conpute "absolute" energy diff */
5163 __energy_diff(eenv);
5165 /* Return energy diff when boost margin is 0 */
5167 return eenv->nrg.diff;
5169 /* Compute normalized energy diff */
5170 nrg_delta = normalize_energy(eenv->nrg.diff);
5171 eenv->nrg.delta = nrg_delta;
5173 eenv->payoff = schedtune_accept_deltas(
5179 * When SchedTune is enabled, the energy_diff() function will return
5180 * the computed energy payoff value. Since the energy_diff() return
5181 * value is expected to be negative by its callers, this evaluation
5182 * function return a negative value each time the evaluation return a
5183 * positive payoff, which is the condition for the acceptance of
5184 * a scheduling decision
5186 return -eenv->payoff;
5188 #else /* CONFIG_SCHED_TUNE */
5189 #define energy_diff(eenv) __energy_diff(eenv)
5193 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5194 * A waker of many should wake a different task than the one last awakened
5195 * at a frequency roughly N times higher than one of its wakees. In order
5196 * to determine whether we should let the load spread vs consolodating to
5197 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5198 * partner, and a factor of lls_size higher frequency in the other. With
5199 * both conditions met, we can be relatively sure that the relationship is
5200 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5201 * being client/server, worker/dispatcher, interrupt source or whatever is
5202 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5204 static int wake_wide(struct task_struct *p)
5206 unsigned int master = current->wakee_flips;
5207 unsigned int slave = p->wakee_flips;
5208 int factor = this_cpu_read(sd_llc_size);
5211 swap(master, slave);
5212 if (slave < factor || master < slave * factor)
5217 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5219 s64 this_load, load;
5220 s64 this_eff_load, prev_eff_load;
5221 int idx, this_cpu, prev_cpu;
5222 struct task_group *tg;
5223 unsigned long weight;
5227 this_cpu = smp_processor_id();
5228 prev_cpu = task_cpu(p);
5229 load = source_load(prev_cpu, idx);
5230 this_load = target_load(this_cpu, idx);
5233 * If sync wakeup then subtract the (maximum possible)
5234 * effect of the currently running task from the load
5235 * of the current CPU:
5238 tg = task_group(current);
5239 weight = current->se.avg.load_avg;
5241 this_load += effective_load(tg, this_cpu, -weight, -weight);
5242 load += effective_load(tg, prev_cpu, 0, -weight);
5246 weight = p->se.avg.load_avg;
5249 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5250 * due to the sync cause above having dropped this_load to 0, we'll
5251 * always have an imbalance, but there's really nothing you can do
5252 * about that, so that's good too.
5254 * Otherwise check if either cpus are near enough in load to allow this
5255 * task to be woken on this_cpu.
5257 this_eff_load = 100;
5258 this_eff_load *= capacity_of(prev_cpu);
5260 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5261 prev_eff_load *= capacity_of(this_cpu);
5263 if (this_load > 0) {
5264 this_eff_load *= this_load +
5265 effective_load(tg, this_cpu, weight, weight);
5267 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5270 balanced = this_eff_load <= prev_eff_load;
5272 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5277 schedstat_inc(sd, ttwu_move_affine);
5278 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5283 static inline unsigned long task_util(struct task_struct *p)
5285 #ifdef CONFIG_SCHED_WALT
5286 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5287 unsigned long demand = p->ravg.demand;
5288 return (demand << 10) / walt_ravg_window;
5291 return p->se.avg.util_avg;
5294 unsigned int capacity_margin = 1280; /* ~20% margin */
5296 static inline unsigned long boosted_task_util(struct task_struct *task);
5298 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5300 unsigned long capacity = capacity_of(cpu);
5302 util += boosted_task_util(p);
5304 return (capacity * 1024) > (util * capacity_margin);
5307 static inline bool task_fits_max(struct task_struct *p, int cpu)
5309 unsigned long capacity = capacity_of(cpu);
5310 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5312 if (capacity == max_capacity)
5315 if (capacity * capacity_margin > max_capacity * 1024)
5318 return __task_fits(p, cpu, 0);
5321 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5323 return __task_fits(p, cpu, cpu_util(cpu));
5326 static bool cpu_overutilized(int cpu)
5328 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5331 #ifdef CONFIG_SCHED_TUNE
5334 schedtune_margin(unsigned long signal, long boost)
5336 long long margin = 0;
5339 * Signal proportional compensation (SPC)
5341 * The Boost (B) value is used to compute a Margin (M) which is
5342 * proportional to the complement of the original Signal (S):
5343 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5344 * M = B * S, if B is negative
5345 * The obtained M could be used by the caller to "boost" S.
5348 margin = SCHED_LOAD_SCALE - signal;
5351 margin = -signal * boost;
5353 * Fast integer division by constant:
5354 * Constant : (C) = 100
5355 * Precision : 0.1% (P) = 0.1
5356 * Reference : C * 100 / P (R) = 100000
5359 * Shift bits : ceil(log(R,2)) (S) = 17
5360 * Mult const : round(2^S/C) (M) = 1311
5373 schedtune_cpu_margin(unsigned long util, int cpu)
5375 int boost = schedtune_cpu_boost(cpu);
5380 return schedtune_margin(util, boost);
5384 schedtune_task_margin(struct task_struct *task)
5386 int boost = schedtune_task_boost(task);
5393 util = task_util(task);
5394 margin = schedtune_margin(util, boost);
5399 #else /* CONFIG_SCHED_TUNE */
5402 schedtune_cpu_margin(unsigned long util, int cpu)
5408 schedtune_task_margin(struct task_struct *task)
5413 #endif /* CONFIG_SCHED_TUNE */
5415 static inline unsigned long
5416 boosted_cpu_util(int cpu)
5418 unsigned long util = cpu_util(cpu);
5419 long margin = schedtune_cpu_margin(util, cpu);
5421 trace_sched_boost_cpu(cpu, util, margin);
5423 return util + margin;
5426 static inline unsigned long
5427 boosted_task_util(struct task_struct *task)
5429 unsigned long util = task_util(task);
5430 long margin = schedtune_task_margin(task);
5432 trace_sched_boost_task(task, util, margin);
5434 return util + margin;
5438 * find_idlest_group finds and returns the least busy CPU group within the
5441 static struct sched_group *
5442 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5443 int this_cpu, int sd_flag)
5445 struct sched_group *idlest = NULL, *group = sd->groups;
5446 struct sched_group *fit_group = NULL, *spare_group = NULL;
5447 unsigned long min_load = ULONG_MAX, this_load = 0;
5448 unsigned long fit_capacity = ULONG_MAX;
5449 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5450 int load_idx = sd->forkexec_idx;
5451 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5453 if (sd_flag & SD_BALANCE_WAKE)
5454 load_idx = sd->wake_idx;
5457 unsigned long load, avg_load, spare_capacity;
5461 /* Skip over this group if it has no CPUs allowed */
5462 if (!cpumask_intersects(sched_group_cpus(group),
5463 tsk_cpus_allowed(p)))
5466 local_group = cpumask_test_cpu(this_cpu,
5467 sched_group_cpus(group));
5469 /* Tally up the load of all CPUs in the group */
5472 for_each_cpu(i, sched_group_cpus(group)) {
5473 /* Bias balancing toward cpus of our domain */
5475 load = source_load(i, load_idx);
5477 load = target_load(i, load_idx);
5482 * Look for most energy-efficient group that can fit
5483 * that can fit the task.
5485 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5486 fit_capacity = capacity_of(i);
5491 * Look for group which has most spare capacity on a
5494 spare_capacity = capacity_of(i) - cpu_util(i);
5495 if (spare_capacity > max_spare_capacity) {
5496 max_spare_capacity = spare_capacity;
5497 spare_group = group;
5501 /* Adjust by relative CPU capacity of the group */
5502 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5505 this_load = avg_load;
5506 } else if (avg_load < min_load) {
5507 min_load = avg_load;
5510 } while (group = group->next, group != sd->groups);
5518 if (!idlest || 100*this_load < imbalance*min_load)
5524 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5527 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5529 unsigned long load, min_load = ULONG_MAX;
5530 unsigned int min_exit_latency = UINT_MAX;
5531 u64 latest_idle_timestamp = 0;
5532 int least_loaded_cpu = this_cpu;
5533 int shallowest_idle_cpu = -1;
5536 /* Traverse only the allowed CPUs */
5537 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5538 if (task_fits_spare(p, i)) {
5539 struct rq *rq = cpu_rq(i);
5540 struct cpuidle_state *idle = idle_get_state(rq);
5541 if (idle && idle->exit_latency < min_exit_latency) {
5543 * We give priority to a CPU whose idle state
5544 * has the smallest exit latency irrespective
5545 * of any idle timestamp.
5547 min_exit_latency = idle->exit_latency;
5548 latest_idle_timestamp = rq->idle_stamp;
5549 shallowest_idle_cpu = i;
5550 } else if (idle_cpu(i) &&
5551 (!idle || idle->exit_latency == min_exit_latency) &&
5552 rq->idle_stamp > latest_idle_timestamp) {
5554 * If equal or no active idle state, then
5555 * the most recently idled CPU might have
5558 latest_idle_timestamp = rq->idle_stamp;
5559 shallowest_idle_cpu = i;
5560 } else if (shallowest_idle_cpu == -1) {
5562 * If we haven't found an idle CPU yet
5563 * pick a non-idle one that can fit the task as
5566 shallowest_idle_cpu = i;
5568 } else if (shallowest_idle_cpu == -1) {
5569 load = weighted_cpuload(i);
5570 if (load < min_load || (load == min_load && i == this_cpu)) {
5572 least_loaded_cpu = i;
5577 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5581 * Try and locate an idle CPU in the sched_domain.
5583 static int select_idle_sibling(struct task_struct *p, int target)
5585 struct sched_domain *sd;
5586 struct sched_group *sg;
5587 int i = task_cpu(p);
5589 int best_idle_cstate = -1;
5590 int best_idle_capacity = INT_MAX;
5592 if (!sysctl_sched_cstate_aware) {
5593 if (idle_cpu(target))
5597 * If the prevous cpu is cache affine and idle, don't be stupid.
5599 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5604 * Otherwise, iterate the domains and find an elegible idle cpu.
5606 sd = rcu_dereference(per_cpu(sd_llc, target));
5607 for_each_lower_domain(sd) {
5610 if (!cpumask_intersects(sched_group_cpus(sg),
5611 tsk_cpus_allowed(p)))
5614 if (sysctl_sched_cstate_aware) {
5615 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5616 struct rq *rq = cpu_rq(i);
5617 int idle_idx = idle_get_state_idx(rq);
5618 unsigned long new_usage = boosted_task_util(p);
5619 unsigned long capacity_orig = capacity_orig_of(i);
5620 if (new_usage > capacity_orig || !idle_cpu(i))
5623 if (i == target && new_usage <= capacity_curr_of(target))
5626 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5628 best_idle_cstate = idle_idx;
5629 best_idle_capacity = capacity_orig;
5633 for_each_cpu(i, sched_group_cpus(sg)) {
5634 if (i == target || !idle_cpu(i))
5638 target = cpumask_first_and(sched_group_cpus(sg),
5639 tsk_cpus_allowed(p));
5644 } while (sg != sd->groups);
5653 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5656 int target_cpu = -1;
5657 int target_util = 0;
5658 int backup_capacity = 0;
5659 int best_idle_cpu = -1;
5660 int best_idle_cstate = INT_MAX;
5661 int backup_cpu = -1;
5662 unsigned long task_util_boosted, new_util;
5664 task_util_boosted = boosted_task_util(p);
5665 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5671 * Iterate from higher cpus for boosted tasks.
5673 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5675 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5679 * p's blocked utilization is still accounted for on prev_cpu
5680 * so prev_cpu will receive a negative bias due to the double
5681 * accounting. However, the blocked utilization may be zero.
5683 new_util = cpu_util(i) + task_util_boosted;
5686 * Ensure minimum capacity to grant the required boost.
5687 * The target CPU can be already at a capacity level higher
5688 * than the one required to boost the task.
5690 if (new_util > capacity_orig_of(i))
5693 #ifdef CONFIG_SCHED_WALT
5694 if (walt_cpu_high_irqload(i))
5698 * Unconditionally favoring tasks that prefer idle cpus to
5701 if (idle_cpu(i) && prefer_idle) {
5702 if (best_idle_cpu < 0)
5707 cur_capacity = capacity_curr_of(i);
5709 idle_idx = idle_get_state_idx(rq);
5711 if (new_util < cur_capacity) {
5712 if (cpu_rq(i)->nr_running) {
5714 /* Find a target cpu with highest
5717 if (target_util == 0 ||
5718 target_util < new_util) {
5720 target_util = new_util;
5723 /* Find a target cpu with lowest
5726 if (target_util == 0 ||
5727 target_util > new_util) {
5729 target_util = new_util;
5732 } else if (!prefer_idle) {
5733 if (best_idle_cpu < 0 ||
5734 (sysctl_sched_cstate_aware &&
5735 best_idle_cstate > idle_idx)) {
5736 best_idle_cstate = idle_idx;
5740 } else if (backup_capacity == 0 ||
5741 backup_capacity > cur_capacity) {
5742 // Find a backup cpu with least capacity.
5743 backup_capacity = cur_capacity;
5748 if (prefer_idle && best_idle_cpu >= 0)
5749 target_cpu = best_idle_cpu;
5750 else if (target_cpu < 0)
5751 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5756 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5758 struct sched_domain *sd;
5759 struct sched_group *sg, *sg_target;
5760 int target_max_cap = INT_MAX;
5761 int target_cpu = task_cpu(p);
5762 unsigned long task_util_boosted, new_util;
5765 if (sysctl_sched_sync_hint_enable && sync) {
5766 int cpu = smp_processor_id();
5767 cpumask_t search_cpus;
5768 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5769 if (cpumask_test_cpu(cpu, &search_cpus))
5773 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5781 if (sysctl_sched_is_big_little) {
5784 * Find group with sufficient capacity. We only get here if no cpu is
5785 * overutilized. We may end up overutilizing a cpu by adding the task,
5786 * but that should not be any worse than select_idle_sibling().
5787 * load_balance() should sort it out later as we get above the tipping
5791 /* Assuming all cpus are the same in group */
5792 int max_cap_cpu = group_first_cpu(sg);
5795 * Assume smaller max capacity means more energy-efficient.
5796 * Ideally we should query the energy model for the right
5797 * answer but it easily ends up in an exhaustive search.
5799 if (capacity_of(max_cap_cpu) < target_max_cap &&
5800 task_fits_max(p, max_cap_cpu)) {
5802 target_max_cap = capacity_of(max_cap_cpu);
5804 } while (sg = sg->next, sg != sd->groups);
5806 task_util_boosted = boosted_task_util(p);
5807 /* Find cpu with sufficient capacity */
5808 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5810 * p's blocked utilization is still accounted for on prev_cpu
5811 * so prev_cpu will receive a negative bias due to the double
5812 * accounting. However, the blocked utilization may be zero.
5814 new_util = cpu_util(i) + task_util_boosted;
5817 * Ensure minimum capacity to grant the required boost.
5818 * The target CPU can be already at a capacity level higher
5819 * than the one required to boost the task.
5821 if (new_util > capacity_orig_of(i))
5824 if (new_util < capacity_curr_of(i)) {
5826 if (cpu_rq(i)->nr_running)
5830 /* cpu has capacity at higher OPP, keep it as fallback */
5831 if (target_cpu == task_cpu(p))
5836 * Find a cpu with sufficient capacity
5838 #ifdef CONFIG_CGROUP_SCHEDTUNE
5839 bool boosted = schedtune_task_boost(p) > 0;
5840 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5843 bool prefer_idle = 0;
5845 int tmp_target = find_best_target(p, boosted, prefer_idle);
5846 if (tmp_target >= 0) {
5847 target_cpu = tmp_target;
5848 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5853 if (target_cpu != task_cpu(p)) {
5854 struct energy_env eenv = {
5855 .util_delta = task_util(p),
5856 .src_cpu = task_cpu(p),
5857 .dst_cpu = target_cpu,
5861 /* Not enough spare capacity on previous cpu */
5862 if (cpu_overutilized(task_cpu(p)))
5865 if (energy_diff(&eenv) >= 0)
5873 * select_task_rq_fair: Select target runqueue for the waking task in domains
5874 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5875 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5877 * Balances load by selecting the idlest cpu in the idlest group, or under
5878 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5880 * Returns the target cpu number.
5882 * preempt must be disabled.
5885 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5887 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5888 int cpu = smp_processor_id();
5889 int new_cpu = prev_cpu;
5890 int want_affine = 0;
5891 int sync = wake_flags & WF_SYNC;
5893 if (sd_flag & SD_BALANCE_WAKE)
5894 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5895 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5899 for_each_domain(cpu, tmp) {
5900 if (!(tmp->flags & SD_LOAD_BALANCE))
5904 * If both cpu and prev_cpu are part of this domain,
5905 * cpu is a valid SD_WAKE_AFFINE target.
5907 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5908 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5913 if (tmp->flags & sd_flag)
5915 else if (!want_affine)
5920 sd = NULL; /* Prefer wake_affine over balance flags */
5921 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5926 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5927 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5928 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5929 new_cpu = select_idle_sibling(p, new_cpu);
5932 struct sched_group *group;
5935 if (!(sd->flags & sd_flag)) {
5940 group = find_idlest_group(sd, p, cpu, sd_flag);
5946 new_cpu = find_idlest_cpu(group, p, cpu);
5947 if (new_cpu == -1 || new_cpu == cpu) {
5948 /* Now try balancing at a lower domain level of cpu */
5953 /* Now try balancing at a lower domain level of new_cpu */
5955 weight = sd->span_weight;
5957 for_each_domain(cpu, tmp) {
5958 if (weight <= tmp->span_weight)
5960 if (tmp->flags & sd_flag)
5963 /* while loop will break here if sd == NULL */
5971 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5972 * cfs_rq_of(p) references at time of call are still valid and identify the
5973 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5974 * other assumptions, including the state of rq->lock, should be made.
5976 static void migrate_task_rq_fair(struct task_struct *p)
5979 * We are supposed to update the task to "current" time, then its up to date
5980 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5981 * what current time is, so simply throw away the out-of-date time. This
5982 * will result in the wakee task is less decayed, but giving the wakee more
5983 * load sounds not bad.
5985 remove_entity_load_avg(&p->se);
5987 /* Tell new CPU we are migrated */
5988 p->se.avg.last_update_time = 0;
5990 /* We have migrated, no longer consider this task hot */
5991 p->se.exec_start = 0;
5994 static void task_dead_fair(struct task_struct *p)
5996 remove_entity_load_avg(&p->se);
5999 #define task_fits_max(p, cpu) true
6000 #endif /* CONFIG_SMP */
6002 static unsigned long
6003 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6005 unsigned long gran = sysctl_sched_wakeup_granularity;
6008 * Since its curr running now, convert the gran from real-time
6009 * to virtual-time in his units.
6011 * By using 'se' instead of 'curr' we penalize light tasks, so
6012 * they get preempted easier. That is, if 'se' < 'curr' then
6013 * the resulting gran will be larger, therefore penalizing the
6014 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6015 * be smaller, again penalizing the lighter task.
6017 * This is especially important for buddies when the leftmost
6018 * task is higher priority than the buddy.
6020 return calc_delta_fair(gran, se);
6024 * Should 'se' preempt 'curr'.
6038 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6040 s64 gran, vdiff = curr->vruntime - se->vruntime;
6045 gran = wakeup_gran(curr, se);
6052 static void set_last_buddy(struct sched_entity *se)
6054 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6057 for_each_sched_entity(se)
6058 cfs_rq_of(se)->last = se;
6061 static void set_next_buddy(struct sched_entity *se)
6063 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6066 for_each_sched_entity(se)
6067 cfs_rq_of(se)->next = se;
6070 static void set_skip_buddy(struct sched_entity *se)
6072 for_each_sched_entity(se)
6073 cfs_rq_of(se)->skip = se;
6077 * Preempt the current task with a newly woken task if needed:
6079 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6081 struct task_struct *curr = rq->curr;
6082 struct sched_entity *se = &curr->se, *pse = &p->se;
6083 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6084 int scale = cfs_rq->nr_running >= sched_nr_latency;
6085 int next_buddy_marked = 0;
6087 if (unlikely(se == pse))
6091 * This is possible from callers such as attach_tasks(), in which we
6092 * unconditionally check_prempt_curr() after an enqueue (which may have
6093 * lead to a throttle). This both saves work and prevents false
6094 * next-buddy nomination below.
6096 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6099 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6100 set_next_buddy(pse);
6101 next_buddy_marked = 1;
6105 * We can come here with TIF_NEED_RESCHED already set from new task
6108 * Note: this also catches the edge-case of curr being in a throttled
6109 * group (e.g. via set_curr_task), since update_curr() (in the
6110 * enqueue of curr) will have resulted in resched being set. This
6111 * prevents us from potentially nominating it as a false LAST_BUDDY
6114 if (test_tsk_need_resched(curr))
6117 /* Idle tasks are by definition preempted by non-idle tasks. */
6118 if (unlikely(curr->policy == SCHED_IDLE) &&
6119 likely(p->policy != SCHED_IDLE))
6123 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6124 * is driven by the tick):
6126 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6129 find_matching_se(&se, &pse);
6130 update_curr(cfs_rq_of(se));
6132 if (wakeup_preempt_entity(se, pse) == 1) {
6134 * Bias pick_next to pick the sched entity that is
6135 * triggering this preemption.
6137 if (!next_buddy_marked)
6138 set_next_buddy(pse);
6147 * Only set the backward buddy when the current task is still
6148 * on the rq. This can happen when a wakeup gets interleaved
6149 * with schedule on the ->pre_schedule() or idle_balance()
6150 * point, either of which can * drop the rq lock.
6152 * Also, during early boot the idle thread is in the fair class,
6153 * for obvious reasons its a bad idea to schedule back to it.
6155 if (unlikely(!se->on_rq || curr == rq->idle))
6158 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6162 static struct task_struct *
6163 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6165 struct cfs_rq *cfs_rq = &rq->cfs;
6166 struct sched_entity *se;
6167 struct task_struct *p;
6171 #ifdef CONFIG_FAIR_GROUP_SCHED
6172 if (!cfs_rq->nr_running)
6175 if (prev->sched_class != &fair_sched_class)
6179 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6180 * likely that a next task is from the same cgroup as the current.
6182 * Therefore attempt to avoid putting and setting the entire cgroup
6183 * hierarchy, only change the part that actually changes.
6187 struct sched_entity *curr = cfs_rq->curr;
6190 * Since we got here without doing put_prev_entity() we also
6191 * have to consider cfs_rq->curr. If it is still a runnable
6192 * entity, update_curr() will update its vruntime, otherwise
6193 * forget we've ever seen it.
6197 update_curr(cfs_rq);
6202 * This call to check_cfs_rq_runtime() will do the
6203 * throttle and dequeue its entity in the parent(s).
6204 * Therefore the 'simple' nr_running test will indeed
6207 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6211 se = pick_next_entity(cfs_rq, curr);
6212 cfs_rq = group_cfs_rq(se);
6218 * Since we haven't yet done put_prev_entity and if the selected task
6219 * is a different task than we started out with, try and touch the
6220 * least amount of cfs_rqs.
6223 struct sched_entity *pse = &prev->se;
6225 while (!(cfs_rq = is_same_group(se, pse))) {
6226 int se_depth = se->depth;
6227 int pse_depth = pse->depth;
6229 if (se_depth <= pse_depth) {
6230 put_prev_entity(cfs_rq_of(pse), pse);
6231 pse = parent_entity(pse);
6233 if (se_depth >= pse_depth) {
6234 set_next_entity(cfs_rq_of(se), se);
6235 se = parent_entity(se);
6239 put_prev_entity(cfs_rq, pse);
6240 set_next_entity(cfs_rq, se);
6243 if (hrtick_enabled(rq))
6244 hrtick_start_fair(rq, p);
6246 rq->misfit_task = !task_fits_max(p, rq->cpu);
6253 if (!cfs_rq->nr_running)
6256 put_prev_task(rq, prev);
6259 se = pick_next_entity(cfs_rq, NULL);
6260 set_next_entity(cfs_rq, se);
6261 cfs_rq = group_cfs_rq(se);
6266 if (hrtick_enabled(rq))
6267 hrtick_start_fair(rq, p);
6269 rq->misfit_task = !task_fits_max(p, rq->cpu);
6274 rq->misfit_task = 0;
6276 * This is OK, because current is on_cpu, which avoids it being picked
6277 * for load-balance and preemption/IRQs are still disabled avoiding
6278 * further scheduler activity on it and we're being very careful to
6279 * re-start the picking loop.
6281 lockdep_unpin_lock(&rq->lock);
6282 new_tasks = idle_balance(rq);
6283 lockdep_pin_lock(&rq->lock);
6285 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6286 * possible for any higher priority task to appear. In that case we
6287 * must re-start the pick_next_entity() loop.
6299 * Account for a descheduled task:
6301 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6303 struct sched_entity *se = &prev->se;
6304 struct cfs_rq *cfs_rq;
6306 for_each_sched_entity(se) {
6307 cfs_rq = cfs_rq_of(se);
6308 put_prev_entity(cfs_rq, se);
6313 * sched_yield() is very simple
6315 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6317 static void yield_task_fair(struct rq *rq)
6319 struct task_struct *curr = rq->curr;
6320 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6321 struct sched_entity *se = &curr->se;
6324 * Are we the only task in the tree?
6326 if (unlikely(rq->nr_running == 1))
6329 clear_buddies(cfs_rq, se);
6331 if (curr->policy != SCHED_BATCH) {
6332 update_rq_clock(rq);
6334 * Update run-time statistics of the 'current'.
6336 update_curr(cfs_rq);
6338 * Tell update_rq_clock() that we've just updated,
6339 * so we don't do microscopic update in schedule()
6340 * and double the fastpath cost.
6342 rq_clock_skip_update(rq, true);
6348 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6350 struct sched_entity *se = &p->se;
6352 /* throttled hierarchies are not runnable */
6353 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6356 /* Tell the scheduler that we'd really like pse to run next. */
6359 yield_task_fair(rq);
6365 /**************************************************
6366 * Fair scheduling class load-balancing methods.
6370 * The purpose of load-balancing is to achieve the same basic fairness the
6371 * per-cpu scheduler provides, namely provide a proportional amount of compute
6372 * time to each task. This is expressed in the following equation:
6374 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6376 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6377 * W_i,0 is defined as:
6379 * W_i,0 = \Sum_j w_i,j (2)
6381 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6382 * is derived from the nice value as per prio_to_weight[].
6384 * The weight average is an exponential decay average of the instantaneous
6387 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6389 * C_i is the compute capacity of cpu i, typically it is the
6390 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6391 * can also include other factors [XXX].
6393 * To achieve this balance we define a measure of imbalance which follows
6394 * directly from (1):
6396 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6398 * We them move tasks around to minimize the imbalance. In the continuous
6399 * function space it is obvious this converges, in the discrete case we get
6400 * a few fun cases generally called infeasible weight scenarios.
6403 * - infeasible weights;
6404 * - local vs global optima in the discrete case. ]
6409 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6410 * for all i,j solution, we create a tree of cpus that follows the hardware
6411 * topology where each level pairs two lower groups (or better). This results
6412 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6413 * tree to only the first of the previous level and we decrease the frequency
6414 * of load-balance at each level inv. proportional to the number of cpus in
6420 * \Sum { --- * --- * 2^i } = O(n) (5)
6422 * `- size of each group
6423 * | | `- number of cpus doing load-balance
6425 * `- sum over all levels
6427 * Coupled with a limit on how many tasks we can migrate every balance pass,
6428 * this makes (5) the runtime complexity of the balancer.
6430 * An important property here is that each CPU is still (indirectly) connected
6431 * to every other cpu in at most O(log n) steps:
6433 * The adjacency matrix of the resulting graph is given by:
6436 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6439 * And you'll find that:
6441 * A^(log_2 n)_i,j != 0 for all i,j (7)
6443 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6444 * The task movement gives a factor of O(m), giving a convergence complexity
6447 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6452 * In order to avoid CPUs going idle while there's still work to do, new idle
6453 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6454 * tree itself instead of relying on other CPUs to bring it work.
6456 * This adds some complexity to both (5) and (8) but it reduces the total idle
6464 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6467 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6472 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6474 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6476 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6479 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6480 * rewrite all of this once again.]
6483 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6485 enum fbq_type { regular, remote, all };
6494 #define LBF_ALL_PINNED 0x01
6495 #define LBF_NEED_BREAK 0x02
6496 #define LBF_DST_PINNED 0x04
6497 #define LBF_SOME_PINNED 0x08
6500 struct sched_domain *sd;
6508 struct cpumask *dst_grpmask;
6510 enum cpu_idle_type idle;
6512 unsigned int src_grp_nr_running;
6513 /* The set of CPUs under consideration for load-balancing */
6514 struct cpumask *cpus;
6519 unsigned int loop_break;
6520 unsigned int loop_max;
6522 enum fbq_type fbq_type;
6523 enum group_type busiest_group_type;
6524 struct list_head tasks;
6528 * Is this task likely cache-hot:
6530 static int task_hot(struct task_struct *p, struct lb_env *env)
6534 lockdep_assert_held(&env->src_rq->lock);
6536 if (p->sched_class != &fair_sched_class)
6539 if (unlikely(p->policy == SCHED_IDLE))
6543 * Buddy candidates are cache hot:
6545 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6546 (&p->se == cfs_rq_of(&p->se)->next ||
6547 &p->se == cfs_rq_of(&p->se)->last))
6550 if (sysctl_sched_migration_cost == -1)
6552 if (sysctl_sched_migration_cost == 0)
6555 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6557 return delta < (s64)sysctl_sched_migration_cost;
6560 #ifdef CONFIG_NUMA_BALANCING
6562 * Returns 1, if task migration degrades locality
6563 * Returns 0, if task migration improves locality i.e migration preferred.
6564 * Returns -1, if task migration is not affected by locality.
6566 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6568 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6569 unsigned long src_faults, dst_faults;
6570 int src_nid, dst_nid;
6572 if (!static_branch_likely(&sched_numa_balancing))
6575 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6578 src_nid = cpu_to_node(env->src_cpu);
6579 dst_nid = cpu_to_node(env->dst_cpu);
6581 if (src_nid == dst_nid)
6584 /* Migrating away from the preferred node is always bad. */
6585 if (src_nid == p->numa_preferred_nid) {
6586 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6592 /* Encourage migration to the preferred node. */
6593 if (dst_nid == p->numa_preferred_nid)
6597 src_faults = group_faults(p, src_nid);
6598 dst_faults = group_faults(p, dst_nid);
6600 src_faults = task_faults(p, src_nid);
6601 dst_faults = task_faults(p, dst_nid);
6604 return dst_faults < src_faults;
6608 static inline int migrate_degrades_locality(struct task_struct *p,
6616 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6619 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6623 lockdep_assert_held(&env->src_rq->lock);
6626 * We do not migrate tasks that are:
6627 * 1) throttled_lb_pair, or
6628 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6629 * 3) running (obviously), or
6630 * 4) are cache-hot on their current CPU.
6632 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6635 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6638 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6640 env->flags |= LBF_SOME_PINNED;
6643 * Remember if this task can be migrated to any other cpu in
6644 * our sched_group. We may want to revisit it if we couldn't
6645 * meet load balance goals by pulling other tasks on src_cpu.
6647 * Also avoid computing new_dst_cpu if we have already computed
6648 * one in current iteration.
6650 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6653 /* Prevent to re-select dst_cpu via env's cpus */
6654 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6655 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6656 env->flags |= LBF_DST_PINNED;
6657 env->new_dst_cpu = cpu;
6665 /* Record that we found atleast one task that could run on dst_cpu */
6666 env->flags &= ~LBF_ALL_PINNED;
6668 if (task_running(env->src_rq, p)) {
6669 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6674 * Aggressive migration if:
6675 * 1) destination numa is preferred
6676 * 2) task is cache cold, or
6677 * 3) too many balance attempts have failed.
6679 tsk_cache_hot = migrate_degrades_locality(p, env);
6680 if (tsk_cache_hot == -1)
6681 tsk_cache_hot = task_hot(p, env);
6683 if (tsk_cache_hot <= 0 ||
6684 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6685 if (tsk_cache_hot == 1) {
6686 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6687 schedstat_inc(p, se.statistics.nr_forced_migrations);
6692 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6697 * detach_task() -- detach the task for the migration specified in env
6699 static void detach_task(struct task_struct *p, struct lb_env *env)
6701 lockdep_assert_held(&env->src_rq->lock);
6703 deactivate_task(env->src_rq, p, 0);
6704 p->on_rq = TASK_ON_RQ_MIGRATING;
6705 double_lock_balance(env->src_rq, env->dst_rq);
6706 set_task_cpu(p, env->dst_cpu);
6707 double_unlock_balance(env->src_rq, env->dst_rq);
6711 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6712 * part of active balancing operations within "domain".
6714 * Returns a task if successful and NULL otherwise.
6716 static struct task_struct *detach_one_task(struct lb_env *env)
6718 struct task_struct *p, *n;
6720 lockdep_assert_held(&env->src_rq->lock);
6722 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6723 if (!can_migrate_task(p, env))
6726 detach_task(p, env);
6729 * Right now, this is only the second place where
6730 * lb_gained[env->idle] is updated (other is detach_tasks)
6731 * so we can safely collect stats here rather than
6732 * inside detach_tasks().
6734 schedstat_inc(env->sd, lb_gained[env->idle]);
6740 static const unsigned int sched_nr_migrate_break = 32;
6743 * detach_tasks() -- tries to detach up to imbalance weighted load from
6744 * busiest_rq, as part of a balancing operation within domain "sd".
6746 * Returns number of detached tasks if successful and 0 otherwise.
6748 static int detach_tasks(struct lb_env *env)
6750 struct list_head *tasks = &env->src_rq->cfs_tasks;
6751 struct task_struct *p;
6755 lockdep_assert_held(&env->src_rq->lock);
6757 if (env->imbalance <= 0)
6760 while (!list_empty(tasks)) {
6762 * We don't want to steal all, otherwise we may be treated likewise,
6763 * which could at worst lead to a livelock crash.
6765 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6768 p = list_first_entry(tasks, struct task_struct, se.group_node);
6771 /* We've more or less seen every task there is, call it quits */
6772 if (env->loop > env->loop_max)
6775 /* take a breather every nr_migrate tasks */
6776 if (env->loop > env->loop_break) {
6777 env->loop_break += sched_nr_migrate_break;
6778 env->flags |= LBF_NEED_BREAK;
6782 if (!can_migrate_task(p, env))
6785 load = task_h_load(p);
6787 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6790 if ((load / 2) > env->imbalance)
6793 detach_task(p, env);
6794 list_add(&p->se.group_node, &env->tasks);
6797 env->imbalance -= load;
6799 #ifdef CONFIG_PREEMPT
6801 * NEWIDLE balancing is a source of latency, so preemptible
6802 * kernels will stop after the first task is detached to minimize
6803 * the critical section.
6805 if (env->idle == CPU_NEWLY_IDLE)
6810 * We only want to steal up to the prescribed amount of
6813 if (env->imbalance <= 0)
6818 list_move_tail(&p->se.group_node, tasks);
6822 * Right now, this is one of only two places we collect this stat
6823 * so we can safely collect detach_one_task() stats here rather
6824 * than inside detach_one_task().
6826 schedstat_add(env->sd, lb_gained[env->idle], detached);
6832 * attach_task() -- attach the task detached by detach_task() to its new rq.
6834 static void attach_task(struct rq *rq, struct task_struct *p)
6836 lockdep_assert_held(&rq->lock);
6838 BUG_ON(task_rq(p) != rq);
6839 p->on_rq = TASK_ON_RQ_QUEUED;
6840 activate_task(rq, p, 0);
6841 check_preempt_curr(rq, p, 0);
6845 * attach_one_task() -- attaches the task returned from detach_one_task() to
6848 static void attach_one_task(struct rq *rq, struct task_struct *p)
6850 raw_spin_lock(&rq->lock);
6853 * We want to potentially raise target_cpu's OPP.
6855 update_capacity_of(cpu_of(rq));
6856 raw_spin_unlock(&rq->lock);
6860 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6863 static void attach_tasks(struct lb_env *env)
6865 struct list_head *tasks = &env->tasks;
6866 struct task_struct *p;
6868 raw_spin_lock(&env->dst_rq->lock);
6870 while (!list_empty(tasks)) {
6871 p = list_first_entry(tasks, struct task_struct, se.group_node);
6872 list_del_init(&p->se.group_node);
6874 attach_task(env->dst_rq, p);
6878 * We want to potentially raise env.dst_cpu's OPP.
6880 update_capacity_of(env->dst_cpu);
6882 raw_spin_unlock(&env->dst_rq->lock);
6885 #ifdef CONFIG_FAIR_GROUP_SCHED
6886 static void update_blocked_averages(int cpu)
6888 struct rq *rq = cpu_rq(cpu);
6889 struct cfs_rq *cfs_rq;
6890 unsigned long flags;
6892 raw_spin_lock_irqsave(&rq->lock, flags);
6893 update_rq_clock(rq);
6896 * Iterates the task_group tree in a bottom up fashion, see
6897 * list_add_leaf_cfs_rq() for details.
6899 for_each_leaf_cfs_rq(rq, cfs_rq) {
6900 /* throttled entities do not contribute to load */
6901 if (throttled_hierarchy(cfs_rq))
6904 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
6905 update_tg_load_avg(cfs_rq, 0);
6907 raw_spin_unlock_irqrestore(&rq->lock, flags);
6911 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6912 * This needs to be done in a top-down fashion because the load of a child
6913 * group is a fraction of its parents load.
6915 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6917 struct rq *rq = rq_of(cfs_rq);
6918 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6919 unsigned long now = jiffies;
6922 if (cfs_rq->last_h_load_update == now)
6925 cfs_rq->h_load_next = NULL;
6926 for_each_sched_entity(se) {
6927 cfs_rq = cfs_rq_of(se);
6928 cfs_rq->h_load_next = se;
6929 if (cfs_rq->last_h_load_update == now)
6934 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6935 cfs_rq->last_h_load_update = now;
6938 while ((se = cfs_rq->h_load_next) != NULL) {
6939 load = cfs_rq->h_load;
6940 load = div64_ul(load * se->avg.load_avg,
6941 cfs_rq_load_avg(cfs_rq) + 1);
6942 cfs_rq = group_cfs_rq(se);
6943 cfs_rq->h_load = load;
6944 cfs_rq->last_h_load_update = now;
6948 static unsigned long task_h_load(struct task_struct *p)
6950 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6952 update_cfs_rq_h_load(cfs_rq);
6953 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6954 cfs_rq_load_avg(cfs_rq) + 1);
6957 static inline void update_blocked_averages(int cpu)
6959 struct rq *rq = cpu_rq(cpu);
6960 struct cfs_rq *cfs_rq = &rq->cfs;
6961 unsigned long flags;
6963 raw_spin_lock_irqsave(&rq->lock, flags);
6964 update_rq_clock(rq);
6965 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
6966 raw_spin_unlock_irqrestore(&rq->lock, flags);
6969 static unsigned long task_h_load(struct task_struct *p)
6971 return p->se.avg.load_avg;
6975 /********** Helpers for find_busiest_group ************************/
6978 * sg_lb_stats - stats of a sched_group required for load_balancing
6980 struct sg_lb_stats {
6981 unsigned long avg_load; /*Avg load across the CPUs of the group */
6982 unsigned long group_load; /* Total load over the CPUs of the group */
6983 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6984 unsigned long load_per_task;
6985 unsigned long group_capacity;
6986 unsigned long group_util; /* Total utilization of the group */
6987 unsigned int sum_nr_running; /* Nr tasks running in the group */
6988 unsigned int idle_cpus;
6989 unsigned int group_weight;
6990 enum group_type group_type;
6991 int group_no_capacity;
6992 int group_misfit_task; /* A cpu has a task too big for its capacity */
6993 #ifdef CONFIG_NUMA_BALANCING
6994 unsigned int nr_numa_running;
6995 unsigned int nr_preferred_running;
7000 * sd_lb_stats - Structure to store the statistics of a sched_domain
7001 * during load balancing.
7003 struct sd_lb_stats {
7004 struct sched_group *busiest; /* Busiest group in this sd */
7005 struct sched_group *local; /* Local group in this sd */
7006 unsigned long total_load; /* Total load of all groups in sd */
7007 unsigned long total_capacity; /* Total capacity of all groups in sd */
7008 unsigned long avg_load; /* Average load across all groups in sd */
7010 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7011 struct sg_lb_stats local_stat; /* Statistics of the local group */
7014 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7017 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7018 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7019 * We must however clear busiest_stat::avg_load because
7020 * update_sd_pick_busiest() reads this before assignment.
7022 *sds = (struct sd_lb_stats){
7026 .total_capacity = 0UL,
7029 .sum_nr_running = 0,
7030 .group_type = group_other,
7036 * get_sd_load_idx - Obtain the load index for a given sched domain.
7037 * @sd: The sched_domain whose load_idx is to be obtained.
7038 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7040 * Return: The load index.
7042 static inline int get_sd_load_idx(struct sched_domain *sd,
7043 enum cpu_idle_type idle)
7049 load_idx = sd->busy_idx;
7052 case CPU_NEWLY_IDLE:
7053 load_idx = sd->newidle_idx;
7056 load_idx = sd->idle_idx;
7063 static unsigned long scale_rt_capacity(int cpu)
7065 struct rq *rq = cpu_rq(cpu);
7066 u64 total, used, age_stamp, avg;
7070 * Since we're reading these variables without serialization make sure
7071 * we read them once before doing sanity checks on them.
7073 age_stamp = READ_ONCE(rq->age_stamp);
7074 avg = READ_ONCE(rq->rt_avg);
7075 delta = __rq_clock_broken(rq) - age_stamp;
7077 if (unlikely(delta < 0))
7080 total = sched_avg_period() + delta;
7082 used = div_u64(avg, total);
7085 * deadline bandwidth is defined at system level so we must
7086 * weight this bandwidth with the max capacity of the system.
7087 * As a reminder, avg_bw is 20bits width and
7088 * scale_cpu_capacity is 10 bits width
7090 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7092 if (likely(used < SCHED_CAPACITY_SCALE))
7093 return SCHED_CAPACITY_SCALE - used;
7098 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7100 raw_spin_lock_init(&mcc->lock);
7105 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7107 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7108 struct sched_group *sdg = sd->groups;
7109 struct max_cpu_capacity *mcc;
7110 unsigned long max_capacity;
7112 unsigned long flags;
7114 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7116 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7118 raw_spin_lock_irqsave(&mcc->lock, flags);
7119 max_capacity = mcc->val;
7120 max_cap_cpu = mcc->cpu;
7122 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7123 (max_capacity < capacity)) {
7124 mcc->val = capacity;
7126 #ifdef CONFIG_SCHED_DEBUG
7127 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7128 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7133 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7135 skip_unlock: __attribute__ ((unused));
7136 capacity *= scale_rt_capacity(cpu);
7137 capacity >>= SCHED_CAPACITY_SHIFT;
7142 cpu_rq(cpu)->cpu_capacity = capacity;
7143 sdg->sgc->capacity = capacity;
7144 sdg->sgc->max_capacity = capacity;
7147 void update_group_capacity(struct sched_domain *sd, int cpu)
7149 struct sched_domain *child = sd->child;
7150 struct sched_group *group, *sdg = sd->groups;
7151 unsigned long capacity, max_capacity;
7152 unsigned long interval;
7154 interval = msecs_to_jiffies(sd->balance_interval);
7155 interval = clamp(interval, 1UL, max_load_balance_interval);
7156 sdg->sgc->next_update = jiffies + interval;
7159 update_cpu_capacity(sd, cpu);
7166 if (child->flags & SD_OVERLAP) {
7168 * SD_OVERLAP domains cannot assume that child groups
7169 * span the current group.
7172 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7173 struct sched_group_capacity *sgc;
7174 struct rq *rq = cpu_rq(cpu);
7177 * build_sched_domains() -> init_sched_groups_capacity()
7178 * gets here before we've attached the domains to the
7181 * Use capacity_of(), which is set irrespective of domains
7182 * in update_cpu_capacity().
7184 * This avoids capacity from being 0 and
7185 * causing divide-by-zero issues on boot.
7187 if (unlikely(!rq->sd)) {
7188 capacity += capacity_of(cpu);
7190 sgc = rq->sd->groups->sgc;
7191 capacity += sgc->capacity;
7194 max_capacity = max(capacity, max_capacity);
7198 * !SD_OVERLAP domains can assume that child groups
7199 * span the current group.
7202 group = child->groups;
7204 struct sched_group_capacity *sgc = group->sgc;
7206 capacity += sgc->capacity;
7207 max_capacity = max(sgc->max_capacity, max_capacity);
7208 group = group->next;
7209 } while (group != child->groups);
7212 sdg->sgc->capacity = capacity;
7213 sdg->sgc->max_capacity = max_capacity;
7217 * Check whether the capacity of the rq has been noticeably reduced by side
7218 * activity. The imbalance_pct is used for the threshold.
7219 * Return true is the capacity is reduced
7222 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7224 return ((rq->cpu_capacity * sd->imbalance_pct) <
7225 (rq->cpu_capacity_orig * 100));
7229 * Group imbalance indicates (and tries to solve) the problem where balancing
7230 * groups is inadequate due to tsk_cpus_allowed() constraints.
7232 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7233 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7236 * { 0 1 2 3 } { 4 5 6 7 }
7239 * If we were to balance group-wise we'd place two tasks in the first group and
7240 * two tasks in the second group. Clearly this is undesired as it will overload
7241 * cpu 3 and leave one of the cpus in the second group unused.
7243 * The current solution to this issue is detecting the skew in the first group
7244 * by noticing the lower domain failed to reach balance and had difficulty
7245 * moving tasks due to affinity constraints.
7247 * When this is so detected; this group becomes a candidate for busiest; see
7248 * update_sd_pick_busiest(). And calculate_imbalance() and
7249 * find_busiest_group() avoid some of the usual balance conditions to allow it
7250 * to create an effective group imbalance.
7252 * This is a somewhat tricky proposition since the next run might not find the
7253 * group imbalance and decide the groups need to be balanced again. A most
7254 * subtle and fragile situation.
7257 static inline int sg_imbalanced(struct sched_group *group)
7259 return group->sgc->imbalance;
7263 * group_has_capacity returns true if the group has spare capacity that could
7264 * be used by some tasks.
7265 * We consider that a group has spare capacity if the * number of task is
7266 * smaller than the number of CPUs or if the utilization is lower than the
7267 * available capacity for CFS tasks.
7268 * For the latter, we use a threshold to stabilize the state, to take into
7269 * account the variance of the tasks' load and to return true if the available
7270 * capacity in meaningful for the load balancer.
7271 * As an example, an available capacity of 1% can appear but it doesn't make
7272 * any benefit for the load balance.
7275 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7277 if (sgs->sum_nr_running < sgs->group_weight)
7280 if ((sgs->group_capacity * 100) >
7281 (sgs->group_util * env->sd->imbalance_pct))
7288 * group_is_overloaded returns true if the group has more tasks than it can
7290 * group_is_overloaded is not equals to !group_has_capacity because a group
7291 * with the exact right number of tasks, has no more spare capacity but is not
7292 * overloaded so both group_has_capacity and group_is_overloaded return
7296 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7298 if (sgs->sum_nr_running <= sgs->group_weight)
7301 if ((sgs->group_capacity * 100) <
7302 (sgs->group_util * env->sd->imbalance_pct))
7310 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7311 * per-cpu capacity than sched_group ref.
7314 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7316 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7317 ref->sgc->max_capacity;
7321 group_type group_classify(struct sched_group *group,
7322 struct sg_lb_stats *sgs)
7324 if (sgs->group_no_capacity)
7325 return group_overloaded;
7327 if (sg_imbalanced(group))
7328 return group_imbalanced;
7330 if (sgs->group_misfit_task)
7331 return group_misfit_task;
7337 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7338 * @env: The load balancing environment.
7339 * @group: sched_group whose statistics are to be updated.
7340 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7341 * @local_group: Does group contain this_cpu.
7342 * @sgs: variable to hold the statistics for this group.
7343 * @overload: Indicate more than one runnable task for any CPU.
7344 * @overutilized: Indicate overutilization for any CPU.
7346 static inline void update_sg_lb_stats(struct lb_env *env,
7347 struct sched_group *group, int load_idx,
7348 int local_group, struct sg_lb_stats *sgs,
7349 bool *overload, bool *overutilized)
7354 memset(sgs, 0, sizeof(*sgs));
7356 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7357 struct rq *rq = cpu_rq(i);
7359 /* Bias balancing toward cpus of our domain */
7361 load = target_load(i, load_idx);
7363 load = source_load(i, load_idx);
7365 sgs->group_load += load;
7366 sgs->group_util += cpu_util(i);
7367 sgs->sum_nr_running += rq->cfs.h_nr_running;
7369 nr_running = rq->nr_running;
7373 #ifdef CONFIG_NUMA_BALANCING
7374 sgs->nr_numa_running += rq->nr_numa_running;
7375 sgs->nr_preferred_running += rq->nr_preferred_running;
7377 sgs->sum_weighted_load += weighted_cpuload(i);
7379 * No need to call idle_cpu() if nr_running is not 0
7381 if (!nr_running && idle_cpu(i))
7384 if (cpu_overutilized(i)) {
7385 *overutilized = true;
7386 if (!sgs->group_misfit_task && rq->misfit_task)
7387 sgs->group_misfit_task = capacity_of(i);
7391 /* Adjust by relative CPU capacity of the group */
7392 sgs->group_capacity = group->sgc->capacity;
7393 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7395 if (sgs->sum_nr_running)
7396 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7398 sgs->group_weight = group->group_weight;
7400 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7401 sgs->group_type = group_classify(group, sgs);
7405 * update_sd_pick_busiest - return 1 on busiest group
7406 * @env: The load balancing environment.
7407 * @sds: sched_domain statistics
7408 * @sg: sched_group candidate to be checked for being the busiest
7409 * @sgs: sched_group statistics
7411 * Determine if @sg is a busier group than the previously selected
7414 * Return: %true if @sg is a busier group than the previously selected
7415 * busiest group. %false otherwise.
7417 static bool update_sd_pick_busiest(struct lb_env *env,
7418 struct sd_lb_stats *sds,
7419 struct sched_group *sg,
7420 struct sg_lb_stats *sgs)
7422 struct sg_lb_stats *busiest = &sds->busiest_stat;
7424 if (sgs->group_type > busiest->group_type)
7427 if (sgs->group_type < busiest->group_type)
7431 * Candidate sg doesn't face any serious load-balance problems
7432 * so don't pick it if the local sg is already filled up.
7434 if (sgs->group_type == group_other &&
7435 !group_has_capacity(env, &sds->local_stat))
7438 if (sgs->avg_load <= busiest->avg_load)
7442 * Candiate sg has no more than one task per cpu and has higher
7443 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7445 if (sgs->sum_nr_running <= sgs->group_weight &&
7446 group_smaller_cpu_capacity(sds->local, sg))
7449 /* This is the busiest node in its class. */
7450 if (!(env->sd->flags & SD_ASYM_PACKING))
7454 * ASYM_PACKING needs to move all the work to the lowest
7455 * numbered CPUs in the group, therefore mark all groups
7456 * higher than ourself as busy.
7458 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7462 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7469 #ifdef CONFIG_NUMA_BALANCING
7470 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7472 if (sgs->sum_nr_running > sgs->nr_numa_running)
7474 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7479 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7481 if (rq->nr_running > rq->nr_numa_running)
7483 if (rq->nr_running > rq->nr_preferred_running)
7488 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7493 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7497 #endif /* CONFIG_NUMA_BALANCING */
7500 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7501 * @env: The load balancing environment.
7502 * @sds: variable to hold the statistics for this sched_domain.
7504 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7506 struct sched_domain *child = env->sd->child;
7507 struct sched_group *sg = env->sd->groups;
7508 struct sg_lb_stats tmp_sgs;
7509 int load_idx, prefer_sibling = 0;
7510 bool overload = false, overutilized = false;
7512 if (child && child->flags & SD_PREFER_SIBLING)
7515 load_idx = get_sd_load_idx(env->sd, env->idle);
7518 struct sg_lb_stats *sgs = &tmp_sgs;
7521 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7524 sgs = &sds->local_stat;
7526 if (env->idle != CPU_NEWLY_IDLE ||
7527 time_after_eq(jiffies, sg->sgc->next_update))
7528 update_group_capacity(env->sd, env->dst_cpu);
7531 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7532 &overload, &overutilized);
7538 * In case the child domain prefers tasks go to siblings
7539 * first, lower the sg capacity so that we'll try
7540 * and move all the excess tasks away. We lower the capacity
7541 * of a group only if the local group has the capacity to fit
7542 * these excess tasks. The extra check prevents the case where
7543 * you always pull from the heaviest group when it is already
7544 * under-utilized (possible with a large weight task outweighs
7545 * the tasks on the system).
7547 if (prefer_sibling && sds->local &&
7548 group_has_capacity(env, &sds->local_stat) &&
7549 (sgs->sum_nr_running > 1)) {
7550 sgs->group_no_capacity = 1;
7551 sgs->group_type = group_classify(sg, sgs);
7555 * Ignore task groups with misfit tasks if local group has no
7556 * capacity or if per-cpu capacity isn't higher.
7558 if (sgs->group_type == group_misfit_task &&
7559 (!group_has_capacity(env, &sds->local_stat) ||
7560 !group_smaller_cpu_capacity(sg, sds->local)))
7561 sgs->group_type = group_other;
7563 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7565 sds->busiest_stat = *sgs;
7569 /* Now, start updating sd_lb_stats */
7570 sds->total_load += sgs->group_load;
7571 sds->total_capacity += sgs->group_capacity;
7574 } while (sg != env->sd->groups);
7576 if (env->sd->flags & SD_NUMA)
7577 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7579 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7581 if (!env->sd->parent) {
7582 /* update overload indicator if we are at root domain */
7583 if (env->dst_rq->rd->overload != overload)
7584 env->dst_rq->rd->overload = overload;
7586 /* Update over-utilization (tipping point, U >= 0) indicator */
7587 if (env->dst_rq->rd->overutilized != overutilized) {
7588 env->dst_rq->rd->overutilized = overutilized;
7589 trace_sched_overutilized(overutilized);
7592 if (!env->dst_rq->rd->overutilized && overutilized) {
7593 env->dst_rq->rd->overutilized = true;
7594 trace_sched_overutilized(true);
7601 * check_asym_packing - Check to see if the group is packed into the
7604 * This is primarily intended to used at the sibling level. Some
7605 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7606 * case of POWER7, it can move to lower SMT modes only when higher
7607 * threads are idle. When in lower SMT modes, the threads will
7608 * perform better since they share less core resources. Hence when we
7609 * have idle threads, we want them to be the higher ones.
7611 * This packing function is run on idle threads. It checks to see if
7612 * the busiest CPU in this domain (core in the P7 case) has a higher
7613 * CPU number than the packing function is being run on. Here we are
7614 * assuming lower CPU number will be equivalent to lower a SMT thread
7617 * Return: 1 when packing is required and a task should be moved to
7618 * this CPU. The amount of the imbalance is returned in *imbalance.
7620 * @env: The load balancing environment.
7621 * @sds: Statistics of the sched_domain which is to be packed
7623 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7627 if (!(env->sd->flags & SD_ASYM_PACKING))
7633 busiest_cpu = group_first_cpu(sds->busiest);
7634 if (env->dst_cpu > busiest_cpu)
7637 env->imbalance = DIV_ROUND_CLOSEST(
7638 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7639 SCHED_CAPACITY_SCALE);
7645 * fix_small_imbalance - Calculate the minor imbalance that exists
7646 * amongst the groups of a sched_domain, during
7648 * @env: The load balancing environment.
7649 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7652 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7654 unsigned long tmp, capa_now = 0, capa_move = 0;
7655 unsigned int imbn = 2;
7656 unsigned long scaled_busy_load_per_task;
7657 struct sg_lb_stats *local, *busiest;
7659 local = &sds->local_stat;
7660 busiest = &sds->busiest_stat;
7662 if (!local->sum_nr_running)
7663 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7664 else if (busiest->load_per_task > local->load_per_task)
7667 scaled_busy_load_per_task =
7668 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7669 busiest->group_capacity;
7671 if (busiest->avg_load + scaled_busy_load_per_task >=
7672 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7673 env->imbalance = busiest->load_per_task;
7678 * OK, we don't have enough imbalance to justify moving tasks,
7679 * however we may be able to increase total CPU capacity used by
7683 capa_now += busiest->group_capacity *
7684 min(busiest->load_per_task, busiest->avg_load);
7685 capa_now += local->group_capacity *
7686 min(local->load_per_task, local->avg_load);
7687 capa_now /= SCHED_CAPACITY_SCALE;
7689 /* Amount of load we'd subtract */
7690 if (busiest->avg_load > scaled_busy_load_per_task) {
7691 capa_move += busiest->group_capacity *
7692 min(busiest->load_per_task,
7693 busiest->avg_load - scaled_busy_load_per_task);
7696 /* Amount of load we'd add */
7697 if (busiest->avg_load * busiest->group_capacity <
7698 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7699 tmp = (busiest->avg_load * busiest->group_capacity) /
7700 local->group_capacity;
7702 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7703 local->group_capacity;
7705 capa_move += local->group_capacity *
7706 min(local->load_per_task, local->avg_load + tmp);
7707 capa_move /= SCHED_CAPACITY_SCALE;
7709 /* Move if we gain throughput */
7710 if (capa_move > capa_now)
7711 env->imbalance = busiest->load_per_task;
7715 * calculate_imbalance - Calculate the amount of imbalance present within the
7716 * groups of a given sched_domain during load balance.
7717 * @env: load balance environment
7718 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7720 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7722 unsigned long max_pull, load_above_capacity = ~0UL;
7723 struct sg_lb_stats *local, *busiest;
7725 local = &sds->local_stat;
7726 busiest = &sds->busiest_stat;
7728 if (busiest->group_type == group_imbalanced) {
7730 * In the group_imb case we cannot rely on group-wide averages
7731 * to ensure cpu-load equilibrium, look at wider averages. XXX
7733 busiest->load_per_task =
7734 min(busiest->load_per_task, sds->avg_load);
7738 * In the presence of smp nice balancing, certain scenarios can have
7739 * max load less than avg load(as we skip the groups at or below
7740 * its cpu_capacity, while calculating max_load..)
7742 if (busiest->avg_load <= sds->avg_load ||
7743 local->avg_load >= sds->avg_load) {
7744 /* Misfitting tasks should be migrated in any case */
7745 if (busiest->group_type == group_misfit_task) {
7746 env->imbalance = busiest->group_misfit_task;
7751 * Busiest group is overloaded, local is not, use the spare
7752 * cycles to maximize throughput
7754 if (busiest->group_type == group_overloaded &&
7755 local->group_type <= group_misfit_task) {
7756 env->imbalance = busiest->load_per_task;
7761 return fix_small_imbalance(env, sds);
7765 * If there aren't any idle cpus, avoid creating some.
7767 if (busiest->group_type == group_overloaded &&
7768 local->group_type == group_overloaded) {
7769 load_above_capacity = busiest->sum_nr_running *
7771 if (load_above_capacity > busiest->group_capacity)
7772 load_above_capacity -= busiest->group_capacity;
7774 load_above_capacity = ~0UL;
7778 * We're trying to get all the cpus to the average_load, so we don't
7779 * want to push ourselves above the average load, nor do we wish to
7780 * reduce the max loaded cpu below the average load. At the same time,
7781 * we also don't want to reduce the group load below the group capacity
7782 * (so that we can implement power-savings policies etc). Thus we look
7783 * for the minimum possible imbalance.
7785 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7787 /* How much load to actually move to equalise the imbalance */
7788 env->imbalance = min(
7789 max_pull * busiest->group_capacity,
7790 (sds->avg_load - local->avg_load) * local->group_capacity
7791 ) / SCHED_CAPACITY_SCALE;
7793 /* Boost imbalance to allow misfit task to be balanced. */
7794 if (busiest->group_type == group_misfit_task)
7795 env->imbalance = max_t(long, env->imbalance,
7796 busiest->group_misfit_task);
7799 * if *imbalance is less than the average load per runnable task
7800 * there is no guarantee that any tasks will be moved so we'll have
7801 * a think about bumping its value to force at least one task to be
7804 if (env->imbalance < busiest->load_per_task)
7805 return fix_small_imbalance(env, sds);
7808 /******* find_busiest_group() helpers end here *********************/
7811 * find_busiest_group - Returns the busiest group within the sched_domain
7812 * if there is an imbalance. If there isn't an imbalance, and
7813 * the user has opted for power-savings, it returns a group whose
7814 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7815 * such a group exists.
7817 * Also calculates the amount of weighted load which should be moved
7818 * to restore balance.
7820 * @env: The load balancing environment.
7822 * Return: - The busiest group if imbalance exists.
7823 * - If no imbalance and user has opted for power-savings balance,
7824 * return the least loaded group whose CPUs can be
7825 * put to idle by rebalancing its tasks onto our group.
7827 static struct sched_group *find_busiest_group(struct lb_env *env)
7829 struct sg_lb_stats *local, *busiest;
7830 struct sd_lb_stats sds;
7832 init_sd_lb_stats(&sds);
7835 * Compute the various statistics relavent for load balancing at
7838 update_sd_lb_stats(env, &sds);
7840 if (energy_aware() && !env->dst_rq->rd->overutilized)
7843 local = &sds.local_stat;
7844 busiest = &sds.busiest_stat;
7846 /* ASYM feature bypasses nice load balance check */
7847 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7848 check_asym_packing(env, &sds))
7851 /* There is no busy sibling group to pull tasks from */
7852 if (!sds.busiest || busiest->sum_nr_running == 0)
7855 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7856 / sds.total_capacity;
7859 * If the busiest group is imbalanced the below checks don't
7860 * work because they assume all things are equal, which typically
7861 * isn't true due to cpus_allowed constraints and the like.
7863 if (busiest->group_type == group_imbalanced)
7866 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7867 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7868 busiest->group_no_capacity)
7871 /* Misfitting tasks should be dealt with regardless of the avg load */
7872 if (busiest->group_type == group_misfit_task) {
7877 * If the local group is busier than the selected busiest group
7878 * don't try and pull any tasks.
7880 if (local->avg_load >= busiest->avg_load)
7884 * Don't pull any tasks if this group is already above the domain
7887 if (local->avg_load >= sds.avg_load)
7890 if (env->idle == CPU_IDLE) {
7892 * This cpu is idle. If the busiest group is not overloaded
7893 * and there is no imbalance between this and busiest group
7894 * wrt idle cpus, it is balanced. The imbalance becomes
7895 * significant if the diff is greater than 1 otherwise we
7896 * might end up to just move the imbalance on another group
7898 if ((busiest->group_type != group_overloaded) &&
7899 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7900 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7904 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7905 * imbalance_pct to be conservative.
7907 if (100 * busiest->avg_load <=
7908 env->sd->imbalance_pct * local->avg_load)
7913 env->busiest_group_type = busiest->group_type;
7914 /* Looks like there is an imbalance. Compute it */
7915 calculate_imbalance(env, &sds);
7924 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7926 static struct rq *find_busiest_queue(struct lb_env *env,
7927 struct sched_group *group)
7929 struct rq *busiest = NULL, *rq;
7930 unsigned long busiest_load = 0, busiest_capacity = 1;
7933 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7934 unsigned long capacity, wl;
7938 rt = fbq_classify_rq(rq);
7941 * We classify groups/runqueues into three groups:
7942 * - regular: there are !numa tasks
7943 * - remote: there are numa tasks that run on the 'wrong' node
7944 * - all: there is no distinction
7946 * In order to avoid migrating ideally placed numa tasks,
7947 * ignore those when there's better options.
7949 * If we ignore the actual busiest queue to migrate another
7950 * task, the next balance pass can still reduce the busiest
7951 * queue by moving tasks around inside the node.
7953 * If we cannot move enough load due to this classification
7954 * the next pass will adjust the group classification and
7955 * allow migration of more tasks.
7957 * Both cases only affect the total convergence complexity.
7959 if (rt > env->fbq_type)
7962 capacity = capacity_of(i);
7964 wl = weighted_cpuload(i);
7967 * When comparing with imbalance, use weighted_cpuload()
7968 * which is not scaled with the cpu capacity.
7971 if (rq->nr_running == 1 && wl > env->imbalance &&
7972 !check_cpu_capacity(rq, env->sd) &&
7973 env->busiest_group_type != group_misfit_task)
7977 * For the load comparisons with the other cpu's, consider
7978 * the weighted_cpuload() scaled with the cpu capacity, so
7979 * that the load can be moved away from the cpu that is
7980 * potentially running at a lower capacity.
7982 * Thus we're looking for max(wl_i / capacity_i), crosswise
7983 * multiplication to rid ourselves of the division works out
7984 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7985 * our previous maximum.
7987 if (wl * busiest_capacity > busiest_load * capacity) {
7989 busiest_capacity = capacity;
7998 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7999 * so long as it is large enough.
8001 #define MAX_PINNED_INTERVAL 512
8003 /* Working cpumask for load_balance and load_balance_newidle. */
8004 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8006 static int need_active_balance(struct lb_env *env)
8008 struct sched_domain *sd = env->sd;
8010 if (env->idle == CPU_NEWLY_IDLE) {
8013 * ASYM_PACKING needs to force migrate tasks from busy but
8014 * higher numbered CPUs in order to pack all tasks in the
8015 * lowest numbered CPUs.
8017 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8022 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8023 * It's worth migrating the task if the src_cpu's capacity is reduced
8024 * because of other sched_class or IRQs if more capacity stays
8025 * available on dst_cpu.
8027 if ((env->idle != CPU_NOT_IDLE) &&
8028 (env->src_rq->cfs.h_nr_running == 1)) {
8029 if ((check_cpu_capacity(env->src_rq, sd)) &&
8030 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8034 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8035 env->src_rq->cfs.h_nr_running == 1 &&
8036 cpu_overutilized(env->src_cpu) &&
8037 !cpu_overutilized(env->dst_cpu)) {
8041 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8044 static int active_load_balance_cpu_stop(void *data);
8046 static int should_we_balance(struct lb_env *env)
8048 struct sched_group *sg = env->sd->groups;
8049 struct cpumask *sg_cpus, *sg_mask;
8050 int cpu, balance_cpu = -1;
8053 * In the newly idle case, we will allow all the cpu's
8054 * to do the newly idle load balance.
8056 if (env->idle == CPU_NEWLY_IDLE)
8059 sg_cpus = sched_group_cpus(sg);
8060 sg_mask = sched_group_mask(sg);
8061 /* Try to find first idle cpu */
8062 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8063 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8070 if (balance_cpu == -1)
8071 balance_cpu = group_balance_cpu(sg);
8074 * First idle cpu or the first cpu(busiest) in this sched group
8075 * is eligible for doing load balancing at this and above domains.
8077 return balance_cpu == env->dst_cpu;
8081 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8082 * tasks if there is an imbalance.
8084 static int load_balance(int this_cpu, struct rq *this_rq,
8085 struct sched_domain *sd, enum cpu_idle_type idle,
8086 int *continue_balancing)
8088 int ld_moved, cur_ld_moved, active_balance = 0;
8089 struct sched_domain *sd_parent = sd->parent;
8090 struct sched_group *group;
8092 unsigned long flags;
8093 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8095 struct lb_env env = {
8097 .dst_cpu = this_cpu,
8099 .dst_grpmask = sched_group_cpus(sd->groups),
8101 .loop_break = sched_nr_migrate_break,
8104 .tasks = LIST_HEAD_INIT(env.tasks),
8108 * For NEWLY_IDLE load_balancing, we don't need to consider
8109 * other cpus in our group
8111 if (idle == CPU_NEWLY_IDLE)
8112 env.dst_grpmask = NULL;
8114 cpumask_copy(cpus, cpu_active_mask);
8116 schedstat_inc(sd, lb_count[idle]);
8119 if (!should_we_balance(&env)) {
8120 *continue_balancing = 0;
8124 group = find_busiest_group(&env);
8126 schedstat_inc(sd, lb_nobusyg[idle]);
8130 busiest = find_busiest_queue(&env, group);
8132 schedstat_inc(sd, lb_nobusyq[idle]);
8136 BUG_ON(busiest == env.dst_rq);
8138 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8140 env.src_cpu = busiest->cpu;
8141 env.src_rq = busiest;
8144 if (busiest->nr_running > 1) {
8146 * Attempt to move tasks. If find_busiest_group has found
8147 * an imbalance but busiest->nr_running <= 1, the group is
8148 * still unbalanced. ld_moved simply stays zero, so it is
8149 * correctly treated as an imbalance.
8151 env.flags |= LBF_ALL_PINNED;
8152 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8155 raw_spin_lock_irqsave(&busiest->lock, flags);
8158 * cur_ld_moved - load moved in current iteration
8159 * ld_moved - cumulative load moved across iterations
8161 cur_ld_moved = detach_tasks(&env);
8163 * We want to potentially lower env.src_cpu's OPP.
8166 update_capacity_of(env.src_cpu);
8169 * We've detached some tasks from busiest_rq. Every
8170 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8171 * unlock busiest->lock, and we are able to be sure
8172 * that nobody can manipulate the tasks in parallel.
8173 * See task_rq_lock() family for the details.
8176 raw_spin_unlock(&busiest->lock);
8180 ld_moved += cur_ld_moved;
8183 local_irq_restore(flags);
8185 if (env.flags & LBF_NEED_BREAK) {
8186 env.flags &= ~LBF_NEED_BREAK;
8191 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8192 * us and move them to an alternate dst_cpu in our sched_group
8193 * where they can run. The upper limit on how many times we
8194 * iterate on same src_cpu is dependent on number of cpus in our
8197 * This changes load balance semantics a bit on who can move
8198 * load to a given_cpu. In addition to the given_cpu itself
8199 * (or a ilb_cpu acting on its behalf where given_cpu is
8200 * nohz-idle), we now have balance_cpu in a position to move
8201 * load to given_cpu. In rare situations, this may cause
8202 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8203 * _independently_ and at _same_ time to move some load to
8204 * given_cpu) causing exceess load to be moved to given_cpu.
8205 * This however should not happen so much in practice and
8206 * moreover subsequent load balance cycles should correct the
8207 * excess load moved.
8209 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8211 /* Prevent to re-select dst_cpu via env's cpus */
8212 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8214 env.dst_rq = cpu_rq(env.new_dst_cpu);
8215 env.dst_cpu = env.new_dst_cpu;
8216 env.flags &= ~LBF_DST_PINNED;
8218 env.loop_break = sched_nr_migrate_break;
8221 * Go back to "more_balance" rather than "redo" since we
8222 * need to continue with same src_cpu.
8228 * We failed to reach balance because of affinity.
8231 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8233 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8234 *group_imbalance = 1;
8237 /* All tasks on this runqueue were pinned by CPU affinity */
8238 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8239 cpumask_clear_cpu(cpu_of(busiest), cpus);
8240 if (!cpumask_empty(cpus)) {
8242 env.loop_break = sched_nr_migrate_break;
8245 goto out_all_pinned;
8250 schedstat_inc(sd, lb_failed[idle]);
8252 * Increment the failure counter only on periodic balance.
8253 * We do not want newidle balance, which can be very
8254 * frequent, pollute the failure counter causing
8255 * excessive cache_hot migrations and active balances.
8257 if (idle != CPU_NEWLY_IDLE)
8258 if (env.src_grp_nr_running > 1)
8259 sd->nr_balance_failed++;
8261 if (need_active_balance(&env)) {
8262 raw_spin_lock_irqsave(&busiest->lock, flags);
8264 /* don't kick the active_load_balance_cpu_stop,
8265 * if the curr task on busiest cpu can't be
8268 if (!cpumask_test_cpu(this_cpu,
8269 tsk_cpus_allowed(busiest->curr))) {
8270 raw_spin_unlock_irqrestore(&busiest->lock,
8272 env.flags |= LBF_ALL_PINNED;
8273 goto out_one_pinned;
8277 * ->active_balance synchronizes accesses to
8278 * ->active_balance_work. Once set, it's cleared
8279 * only after active load balance is finished.
8281 if (!busiest->active_balance) {
8282 busiest->active_balance = 1;
8283 busiest->push_cpu = this_cpu;
8286 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8288 if (active_balance) {
8289 stop_one_cpu_nowait(cpu_of(busiest),
8290 active_load_balance_cpu_stop, busiest,
8291 &busiest->active_balance_work);
8295 * We've kicked active balancing, reset the failure
8298 sd->nr_balance_failed = sd->cache_nice_tries+1;
8301 sd->nr_balance_failed = 0;
8303 if (likely(!active_balance)) {
8304 /* We were unbalanced, so reset the balancing interval */
8305 sd->balance_interval = sd->min_interval;
8308 * If we've begun active balancing, start to back off. This
8309 * case may not be covered by the all_pinned logic if there
8310 * is only 1 task on the busy runqueue (because we don't call
8313 if (sd->balance_interval < sd->max_interval)
8314 sd->balance_interval *= 2;
8321 * We reach balance although we may have faced some affinity
8322 * constraints. Clear the imbalance flag if it was set.
8325 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8327 if (*group_imbalance)
8328 *group_imbalance = 0;
8333 * We reach balance because all tasks are pinned at this level so
8334 * we can't migrate them. Let the imbalance flag set so parent level
8335 * can try to migrate them.
8337 schedstat_inc(sd, lb_balanced[idle]);
8339 sd->nr_balance_failed = 0;
8342 /* tune up the balancing interval */
8343 if (((env.flags & LBF_ALL_PINNED) &&
8344 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8345 (sd->balance_interval < sd->max_interval))
8346 sd->balance_interval *= 2;
8353 static inline unsigned long
8354 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8356 unsigned long interval = sd->balance_interval;
8359 interval *= sd->busy_factor;
8361 /* scale ms to jiffies */
8362 interval = msecs_to_jiffies(interval);
8363 interval = clamp(interval, 1UL, max_load_balance_interval);
8369 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8371 unsigned long interval, next;
8373 interval = get_sd_balance_interval(sd, cpu_busy);
8374 next = sd->last_balance + interval;
8376 if (time_after(*next_balance, next))
8377 *next_balance = next;
8381 * idle_balance is called by schedule() if this_cpu is about to become
8382 * idle. Attempts to pull tasks from other CPUs.
8384 static int idle_balance(struct rq *this_rq)
8386 unsigned long next_balance = jiffies + HZ;
8387 int this_cpu = this_rq->cpu;
8388 struct sched_domain *sd;
8389 int pulled_task = 0;
8391 long removed_util=0;
8393 idle_enter_fair(this_rq);
8396 * We must set idle_stamp _before_ calling idle_balance(), such that we
8397 * measure the duration of idle_balance() as idle time.
8399 this_rq->idle_stamp = rq_clock(this_rq);
8401 if (!energy_aware() &&
8402 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8403 !this_rq->rd->overload)) {
8405 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8407 update_next_balance(sd, 0, &next_balance);
8413 raw_spin_unlock(&this_rq->lock);
8416 * If removed_util_avg is !0 we most probably migrated some task away
8417 * from this_cpu. In this case we might be willing to trigger an OPP
8418 * update, but we want to do so if we don't find anybody else to pull
8419 * here (we will trigger an OPP update with the pulled task's enqueue
8422 * Record removed_util before calling update_blocked_averages, and use
8423 * it below (before returning) to see if an OPP update is required.
8425 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8426 update_blocked_averages(this_cpu);
8428 for_each_domain(this_cpu, sd) {
8429 int continue_balancing = 1;
8430 u64 t0, domain_cost;
8432 if (!(sd->flags & SD_LOAD_BALANCE))
8435 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8436 update_next_balance(sd, 0, &next_balance);
8440 if (sd->flags & SD_BALANCE_NEWIDLE) {
8441 t0 = sched_clock_cpu(this_cpu);
8443 pulled_task = load_balance(this_cpu, this_rq,
8445 &continue_balancing);
8447 domain_cost = sched_clock_cpu(this_cpu) - t0;
8448 if (domain_cost > sd->max_newidle_lb_cost)
8449 sd->max_newidle_lb_cost = domain_cost;
8451 curr_cost += domain_cost;
8454 update_next_balance(sd, 0, &next_balance);
8457 * Stop searching for tasks to pull if there are
8458 * now runnable tasks on this rq.
8460 if (pulled_task || this_rq->nr_running > 0)
8465 raw_spin_lock(&this_rq->lock);
8467 if (curr_cost > this_rq->max_idle_balance_cost)
8468 this_rq->max_idle_balance_cost = curr_cost;
8471 * While browsing the domains, we released the rq lock, a task could
8472 * have been enqueued in the meantime. Since we're not going idle,
8473 * pretend we pulled a task.
8475 if (this_rq->cfs.h_nr_running && !pulled_task)
8479 /* Move the next balance forward */
8480 if (time_after(this_rq->next_balance, next_balance))
8481 this_rq->next_balance = next_balance;
8483 /* Is there a task of a high priority class? */
8484 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8488 idle_exit_fair(this_rq);
8489 this_rq->idle_stamp = 0;
8490 } else if (removed_util) {
8492 * No task pulled and someone has been migrated away.
8493 * Good case to trigger an OPP update.
8495 update_capacity_of(this_cpu);
8502 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8503 * running tasks off the busiest CPU onto idle CPUs. It requires at
8504 * least 1 task to be running on each physical CPU where possible, and
8505 * avoids physical / logical imbalances.
8507 static int active_load_balance_cpu_stop(void *data)
8509 struct rq *busiest_rq = data;
8510 int busiest_cpu = cpu_of(busiest_rq);
8511 int target_cpu = busiest_rq->push_cpu;
8512 struct rq *target_rq = cpu_rq(target_cpu);
8513 struct sched_domain *sd;
8514 struct task_struct *p = NULL;
8516 raw_spin_lock_irq(&busiest_rq->lock);
8518 /* make sure the requested cpu hasn't gone down in the meantime */
8519 if (unlikely(busiest_cpu != smp_processor_id() ||
8520 !busiest_rq->active_balance))
8523 /* Is there any task to move? */
8524 if (busiest_rq->nr_running <= 1)
8528 * This condition is "impossible", if it occurs
8529 * we need to fix it. Originally reported by
8530 * Bjorn Helgaas on a 128-cpu setup.
8532 BUG_ON(busiest_rq == target_rq);
8534 /* Search for an sd spanning us and the target CPU. */
8536 for_each_domain(target_cpu, sd) {
8537 if ((sd->flags & SD_LOAD_BALANCE) &&
8538 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8543 struct lb_env env = {
8545 .dst_cpu = target_cpu,
8546 .dst_rq = target_rq,
8547 .src_cpu = busiest_rq->cpu,
8548 .src_rq = busiest_rq,
8552 schedstat_inc(sd, alb_count);
8554 p = detach_one_task(&env);
8556 schedstat_inc(sd, alb_pushed);
8558 * We want to potentially lower env.src_cpu's OPP.
8560 update_capacity_of(env.src_cpu);
8563 schedstat_inc(sd, alb_failed);
8567 busiest_rq->active_balance = 0;
8568 raw_spin_unlock(&busiest_rq->lock);
8571 attach_one_task(target_rq, p);
8578 static inline int on_null_domain(struct rq *rq)
8580 return unlikely(!rcu_dereference_sched(rq->sd));
8583 #ifdef CONFIG_NO_HZ_COMMON
8585 * idle load balancing details
8586 * - When one of the busy CPUs notice that there may be an idle rebalancing
8587 * needed, they will kick the idle load balancer, which then does idle
8588 * load balancing for all the idle CPUs.
8591 cpumask_var_t idle_cpus_mask;
8593 unsigned long next_balance; /* in jiffy units */
8594 } nohz ____cacheline_aligned;
8596 static inline int find_new_ilb(void)
8598 int ilb = cpumask_first(nohz.idle_cpus_mask);
8600 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8607 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8608 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8609 * CPU (if there is one).
8611 static void nohz_balancer_kick(void)
8615 nohz.next_balance++;
8617 ilb_cpu = find_new_ilb();
8619 if (ilb_cpu >= nr_cpu_ids)
8622 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8625 * Use smp_send_reschedule() instead of resched_cpu().
8626 * This way we generate a sched IPI on the target cpu which
8627 * is idle. And the softirq performing nohz idle load balance
8628 * will be run before returning from the IPI.
8630 smp_send_reschedule(ilb_cpu);
8634 static inline void nohz_balance_exit_idle(int cpu)
8636 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8638 * Completely isolated CPUs don't ever set, so we must test.
8640 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8641 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8642 atomic_dec(&nohz.nr_cpus);
8644 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8648 static inline void set_cpu_sd_state_busy(void)
8650 struct sched_domain *sd;
8651 int cpu = smp_processor_id();
8654 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8656 if (!sd || !sd->nohz_idle)
8660 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8665 void set_cpu_sd_state_idle(void)
8667 struct sched_domain *sd;
8668 int cpu = smp_processor_id();
8671 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8673 if (!sd || sd->nohz_idle)
8677 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8683 * This routine will record that the cpu is going idle with tick stopped.
8684 * This info will be used in performing idle load balancing in the future.
8686 void nohz_balance_enter_idle(int cpu)
8689 * If this cpu is going down, then nothing needs to be done.
8691 if (!cpu_active(cpu))
8694 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8698 * If we're a completely isolated CPU, we don't play.
8700 if (on_null_domain(cpu_rq(cpu)))
8703 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8704 atomic_inc(&nohz.nr_cpus);
8705 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8708 static int sched_ilb_notifier(struct notifier_block *nfb,
8709 unsigned long action, void *hcpu)
8711 switch (action & ~CPU_TASKS_FROZEN) {
8713 nohz_balance_exit_idle(smp_processor_id());
8721 static DEFINE_SPINLOCK(balancing);
8724 * Scale the max load_balance interval with the number of CPUs in the system.
8725 * This trades load-balance latency on larger machines for less cross talk.
8727 void update_max_interval(void)
8729 max_load_balance_interval = HZ*num_online_cpus()/10;
8733 * It checks each scheduling domain to see if it is due to be balanced,
8734 * and initiates a balancing operation if so.
8736 * Balancing parameters are set up in init_sched_domains.
8738 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8740 int continue_balancing = 1;
8742 unsigned long interval;
8743 struct sched_domain *sd;
8744 /* Earliest time when we have to do rebalance again */
8745 unsigned long next_balance = jiffies + 60*HZ;
8746 int update_next_balance = 0;
8747 int need_serialize, need_decay = 0;
8750 update_blocked_averages(cpu);
8753 for_each_domain(cpu, sd) {
8755 * Decay the newidle max times here because this is a regular
8756 * visit to all the domains. Decay ~1% per second.
8758 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8759 sd->max_newidle_lb_cost =
8760 (sd->max_newidle_lb_cost * 253) / 256;
8761 sd->next_decay_max_lb_cost = jiffies + HZ;
8764 max_cost += sd->max_newidle_lb_cost;
8766 if (!(sd->flags & SD_LOAD_BALANCE))
8770 * Stop the load balance at this level. There is another
8771 * CPU in our sched group which is doing load balancing more
8774 if (!continue_balancing) {
8780 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8782 need_serialize = sd->flags & SD_SERIALIZE;
8783 if (need_serialize) {
8784 if (!spin_trylock(&balancing))
8788 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8789 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8791 * The LBF_DST_PINNED logic could have changed
8792 * env->dst_cpu, so we can't know our idle
8793 * state even if we migrated tasks. Update it.
8795 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8797 sd->last_balance = jiffies;
8798 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8801 spin_unlock(&balancing);
8803 if (time_after(next_balance, sd->last_balance + interval)) {
8804 next_balance = sd->last_balance + interval;
8805 update_next_balance = 1;
8810 * Ensure the rq-wide value also decays but keep it at a
8811 * reasonable floor to avoid funnies with rq->avg_idle.
8813 rq->max_idle_balance_cost =
8814 max((u64)sysctl_sched_migration_cost, max_cost);
8819 * next_balance will be updated only when there is a need.
8820 * When the cpu is attached to null domain for ex, it will not be
8823 if (likely(update_next_balance)) {
8824 rq->next_balance = next_balance;
8826 #ifdef CONFIG_NO_HZ_COMMON
8828 * If this CPU has been elected to perform the nohz idle
8829 * balance. Other idle CPUs have already rebalanced with
8830 * nohz_idle_balance() and nohz.next_balance has been
8831 * updated accordingly. This CPU is now running the idle load
8832 * balance for itself and we need to update the
8833 * nohz.next_balance accordingly.
8835 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8836 nohz.next_balance = rq->next_balance;
8841 #ifdef CONFIG_NO_HZ_COMMON
8843 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8844 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8846 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8848 int this_cpu = this_rq->cpu;
8851 /* Earliest time when we have to do rebalance again */
8852 unsigned long next_balance = jiffies + 60*HZ;
8853 int update_next_balance = 0;
8855 if (idle != CPU_IDLE ||
8856 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8859 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8860 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8864 * If this cpu gets work to do, stop the load balancing
8865 * work being done for other cpus. Next load
8866 * balancing owner will pick it up.
8871 rq = cpu_rq(balance_cpu);
8874 * If time for next balance is due,
8877 if (time_after_eq(jiffies, rq->next_balance)) {
8878 raw_spin_lock_irq(&rq->lock);
8879 update_rq_clock(rq);
8880 update_idle_cpu_load(rq);
8881 raw_spin_unlock_irq(&rq->lock);
8882 rebalance_domains(rq, CPU_IDLE);
8885 if (time_after(next_balance, rq->next_balance)) {
8886 next_balance = rq->next_balance;
8887 update_next_balance = 1;
8892 * next_balance will be updated only when there is a need.
8893 * When the CPU is attached to null domain for ex, it will not be
8896 if (likely(update_next_balance))
8897 nohz.next_balance = next_balance;
8899 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8903 * Current heuristic for kicking the idle load balancer in the presence
8904 * of an idle cpu in the system.
8905 * - This rq has more than one task.
8906 * - This rq has at least one CFS task and the capacity of the CPU is
8907 * significantly reduced because of RT tasks or IRQs.
8908 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8909 * multiple busy cpu.
8910 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8911 * domain span are idle.
8913 static inline bool nohz_kick_needed(struct rq *rq)
8915 unsigned long now = jiffies;
8916 struct sched_domain *sd;
8917 struct sched_group_capacity *sgc;
8918 int nr_busy, cpu = rq->cpu;
8921 if (unlikely(rq->idle_balance))
8925 * We may be recently in ticked or tickless idle mode. At the first
8926 * busy tick after returning from idle, we will update the busy stats.
8928 set_cpu_sd_state_busy();
8929 nohz_balance_exit_idle(cpu);
8932 * None are in tickless mode and hence no need for NOHZ idle load
8935 if (likely(!atomic_read(&nohz.nr_cpus)))
8938 if (time_before(now, nohz.next_balance))
8941 if (rq->nr_running >= 2 &&
8942 (!energy_aware() || cpu_overutilized(cpu)))
8946 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8947 if (sd && !energy_aware()) {
8948 sgc = sd->groups->sgc;
8949 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8958 sd = rcu_dereference(rq->sd);
8960 if ((rq->cfs.h_nr_running >= 1) &&
8961 check_cpu_capacity(rq, sd)) {
8967 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8968 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8969 sched_domain_span(sd)) < cpu)) {
8979 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8983 * run_rebalance_domains is triggered when needed from the scheduler tick.
8984 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8986 static void run_rebalance_domains(struct softirq_action *h)
8988 struct rq *this_rq = this_rq();
8989 enum cpu_idle_type idle = this_rq->idle_balance ?
8990 CPU_IDLE : CPU_NOT_IDLE;
8993 * If this cpu has a pending nohz_balance_kick, then do the
8994 * balancing on behalf of the other idle cpus whose ticks are
8995 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8996 * give the idle cpus a chance to load balance. Else we may
8997 * load balance only within the local sched_domain hierarchy
8998 * and abort nohz_idle_balance altogether if we pull some load.
9000 nohz_idle_balance(this_rq, idle);
9001 rebalance_domains(this_rq, idle);
9005 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9007 void trigger_load_balance(struct rq *rq)
9009 /* Don't need to rebalance while attached to NULL domain */
9010 if (unlikely(on_null_domain(rq)))
9013 if (time_after_eq(jiffies, rq->next_balance))
9014 raise_softirq(SCHED_SOFTIRQ);
9015 #ifdef CONFIG_NO_HZ_COMMON
9016 if (nohz_kick_needed(rq))
9017 nohz_balancer_kick();
9021 static void rq_online_fair(struct rq *rq)
9025 update_runtime_enabled(rq);
9028 static void rq_offline_fair(struct rq *rq)
9032 /* Ensure any throttled groups are reachable by pick_next_task */
9033 unthrottle_offline_cfs_rqs(rq);
9036 #endif /* CONFIG_SMP */
9039 * scheduler tick hitting a task of our scheduling class:
9041 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9043 struct cfs_rq *cfs_rq;
9044 struct sched_entity *se = &curr->se;
9046 for_each_sched_entity(se) {
9047 cfs_rq = cfs_rq_of(se);
9048 entity_tick(cfs_rq, se, queued);
9051 if (static_branch_unlikely(&sched_numa_balancing))
9052 task_tick_numa(rq, curr);
9055 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9056 rq->rd->overutilized = true;
9057 trace_sched_overutilized(true);
9060 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9066 * called on fork with the child task as argument from the parent's context
9067 * - child not yet on the tasklist
9068 * - preemption disabled
9070 static void task_fork_fair(struct task_struct *p)
9072 struct cfs_rq *cfs_rq;
9073 struct sched_entity *se = &p->se, *curr;
9074 int this_cpu = smp_processor_id();
9075 struct rq *rq = this_rq();
9076 unsigned long flags;
9078 raw_spin_lock_irqsave(&rq->lock, flags);
9080 update_rq_clock(rq);
9082 cfs_rq = task_cfs_rq(current);
9083 curr = cfs_rq->curr;
9086 * Not only the cpu but also the task_group of the parent might have
9087 * been changed after parent->se.parent,cfs_rq were copied to
9088 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9089 * of child point to valid ones.
9092 __set_task_cpu(p, this_cpu);
9095 update_curr(cfs_rq);
9098 se->vruntime = curr->vruntime;
9099 place_entity(cfs_rq, se, 1);
9101 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9103 * Upon rescheduling, sched_class::put_prev_task() will place
9104 * 'current' within the tree based on its new key value.
9106 swap(curr->vruntime, se->vruntime);
9110 se->vruntime -= cfs_rq->min_vruntime;
9112 raw_spin_unlock_irqrestore(&rq->lock, flags);
9116 * Priority of the task has changed. Check to see if we preempt
9120 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9122 if (!task_on_rq_queued(p))
9126 * Reschedule if we are currently running on this runqueue and
9127 * our priority decreased, or if we are not currently running on
9128 * this runqueue and our priority is higher than the current's
9130 if (rq->curr == p) {
9131 if (p->prio > oldprio)
9134 check_preempt_curr(rq, p, 0);
9137 static inline bool vruntime_normalized(struct task_struct *p)
9139 struct sched_entity *se = &p->se;
9142 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9143 * the dequeue_entity(.flags=0) will already have normalized the
9150 * When !on_rq, vruntime of the task has usually NOT been normalized.
9151 * But there are some cases where it has already been normalized:
9153 * - A forked child which is waiting for being woken up by
9154 * wake_up_new_task().
9155 * - A task which has been woken up by try_to_wake_up() and
9156 * waiting for actually being woken up by sched_ttwu_pending().
9158 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9164 static void detach_task_cfs_rq(struct task_struct *p)
9166 struct sched_entity *se = &p->se;
9167 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9169 if (!vruntime_normalized(p)) {
9171 * Fix up our vruntime so that the current sleep doesn't
9172 * cause 'unlimited' sleep bonus.
9174 place_entity(cfs_rq, se, 0);
9175 se->vruntime -= cfs_rq->min_vruntime;
9178 /* Catch up with the cfs_rq and remove our load when we leave */
9179 detach_entity_load_avg(cfs_rq, se);
9182 static void attach_task_cfs_rq(struct task_struct *p)
9184 struct sched_entity *se = &p->se;
9185 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9187 #ifdef CONFIG_FAIR_GROUP_SCHED
9189 * Since the real-depth could have been changed (only FAIR
9190 * class maintain depth value), reset depth properly.
9192 se->depth = se->parent ? se->parent->depth + 1 : 0;
9195 /* Synchronize task with its cfs_rq */
9196 attach_entity_load_avg(cfs_rq, se);
9198 if (!vruntime_normalized(p))
9199 se->vruntime += cfs_rq->min_vruntime;
9202 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9204 detach_task_cfs_rq(p);
9207 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9209 attach_task_cfs_rq(p);
9211 if (task_on_rq_queued(p)) {
9213 * We were most likely switched from sched_rt, so
9214 * kick off the schedule if running, otherwise just see
9215 * if we can still preempt the current task.
9220 check_preempt_curr(rq, p, 0);
9224 /* Account for a task changing its policy or group.
9226 * This routine is mostly called to set cfs_rq->curr field when a task
9227 * migrates between groups/classes.
9229 static void set_curr_task_fair(struct rq *rq)
9231 struct sched_entity *se = &rq->curr->se;
9233 for_each_sched_entity(se) {
9234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9236 set_next_entity(cfs_rq, se);
9237 /* ensure bandwidth has been allocated on our new cfs_rq */
9238 account_cfs_rq_runtime(cfs_rq, 0);
9242 void init_cfs_rq(struct cfs_rq *cfs_rq)
9244 cfs_rq->tasks_timeline = RB_ROOT;
9245 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9246 #ifndef CONFIG_64BIT
9247 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9250 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9251 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9255 #ifdef CONFIG_FAIR_GROUP_SCHED
9256 static void task_move_group_fair(struct task_struct *p)
9258 detach_task_cfs_rq(p);
9259 set_task_rq(p, task_cpu(p));
9262 /* Tell se's cfs_rq has been changed -- migrated */
9263 p->se.avg.last_update_time = 0;
9265 attach_task_cfs_rq(p);
9268 void free_fair_sched_group(struct task_group *tg)
9272 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9274 for_each_possible_cpu(i) {
9276 kfree(tg->cfs_rq[i]);
9279 remove_entity_load_avg(tg->se[i]);
9288 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9290 struct cfs_rq *cfs_rq;
9291 struct sched_entity *se;
9294 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9297 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9301 tg->shares = NICE_0_LOAD;
9303 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9305 for_each_possible_cpu(i) {
9306 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9307 GFP_KERNEL, cpu_to_node(i));
9311 se = kzalloc_node(sizeof(struct sched_entity),
9312 GFP_KERNEL, cpu_to_node(i));
9316 init_cfs_rq(cfs_rq);
9317 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9318 init_entity_runnable_average(se);
9329 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9331 struct rq *rq = cpu_rq(cpu);
9332 unsigned long flags;
9335 * Only empty task groups can be destroyed; so we can speculatively
9336 * check on_list without danger of it being re-added.
9338 if (!tg->cfs_rq[cpu]->on_list)
9341 raw_spin_lock_irqsave(&rq->lock, flags);
9342 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9343 raw_spin_unlock_irqrestore(&rq->lock, flags);
9346 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9347 struct sched_entity *se, int cpu,
9348 struct sched_entity *parent)
9350 struct rq *rq = cpu_rq(cpu);
9354 init_cfs_rq_runtime(cfs_rq);
9356 tg->cfs_rq[cpu] = cfs_rq;
9359 /* se could be NULL for root_task_group */
9364 se->cfs_rq = &rq->cfs;
9367 se->cfs_rq = parent->my_q;
9368 se->depth = parent->depth + 1;
9372 /* guarantee group entities always have weight */
9373 update_load_set(&se->load, NICE_0_LOAD);
9374 se->parent = parent;
9377 static DEFINE_MUTEX(shares_mutex);
9379 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9382 unsigned long flags;
9385 * We can't change the weight of the root cgroup.
9390 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9392 mutex_lock(&shares_mutex);
9393 if (tg->shares == shares)
9396 tg->shares = shares;
9397 for_each_possible_cpu(i) {
9398 struct rq *rq = cpu_rq(i);
9399 struct sched_entity *se;
9402 /* Propagate contribution to hierarchy */
9403 raw_spin_lock_irqsave(&rq->lock, flags);
9405 /* Possible calls to update_curr() need rq clock */
9406 update_rq_clock(rq);
9407 for_each_sched_entity(se)
9408 update_cfs_shares(group_cfs_rq(se));
9409 raw_spin_unlock_irqrestore(&rq->lock, flags);
9413 mutex_unlock(&shares_mutex);
9416 #else /* CONFIG_FAIR_GROUP_SCHED */
9418 void free_fair_sched_group(struct task_group *tg) { }
9420 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9425 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9427 #endif /* CONFIG_FAIR_GROUP_SCHED */
9430 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9432 struct sched_entity *se = &task->se;
9433 unsigned int rr_interval = 0;
9436 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9439 if (rq->cfs.load.weight)
9440 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9446 * All the scheduling class methods:
9448 const struct sched_class fair_sched_class = {
9449 .next = &idle_sched_class,
9450 .enqueue_task = enqueue_task_fair,
9451 .dequeue_task = dequeue_task_fair,
9452 .yield_task = yield_task_fair,
9453 .yield_to_task = yield_to_task_fair,
9455 .check_preempt_curr = check_preempt_wakeup,
9457 .pick_next_task = pick_next_task_fair,
9458 .put_prev_task = put_prev_task_fair,
9461 .select_task_rq = select_task_rq_fair,
9462 .migrate_task_rq = migrate_task_rq_fair,
9464 .rq_online = rq_online_fair,
9465 .rq_offline = rq_offline_fair,
9467 .task_waking = task_waking_fair,
9468 .task_dead = task_dead_fair,
9469 .set_cpus_allowed = set_cpus_allowed_common,
9472 .set_curr_task = set_curr_task_fair,
9473 .task_tick = task_tick_fair,
9474 .task_fork = task_fork_fair,
9476 .prio_changed = prio_changed_fair,
9477 .switched_from = switched_from_fair,
9478 .switched_to = switched_to_fair,
9480 .get_rr_interval = get_rr_interval_fair,
9482 .update_curr = update_curr_fair,
9484 #ifdef CONFIG_FAIR_GROUP_SCHED
9485 .task_move_group = task_move_group_fair,
9489 #ifdef CONFIG_SCHED_DEBUG
9490 void print_cfs_stats(struct seq_file *m, int cpu)
9492 struct cfs_rq *cfs_rq;
9495 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9496 print_cfs_rq(m, cpu, cfs_rq);
9500 #ifdef CONFIG_NUMA_BALANCING
9501 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9504 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9506 for_each_online_node(node) {
9507 if (p->numa_faults) {
9508 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9509 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9511 if (p->numa_group) {
9512 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9513 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9515 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9518 #endif /* CONFIG_NUMA_BALANCING */
9519 #endif /* CONFIG_SCHED_DEBUG */
9521 __init void init_sched_fair_class(void)
9524 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9526 #ifdef CONFIG_NO_HZ_COMMON
9527 nohz.next_balance = jiffies;
9528 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9529 cpu_notifier(sched_ilb_notifier, 0);