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_sync_hint_enable = 1;
57 unsigned int sysctl_sched_initial_task_util = 0;
58 unsigned int sysctl_sched_cstate_aware = 1;
60 #ifdef CONFIG_SCHED_WALT
61 unsigned int sysctl_sched_use_walt_cpu_util = 1;
62 unsigned int sysctl_sched_use_walt_task_util = 1;
63 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
67 * The initial- and re-scaling of tunables is configurable
68 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
72 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
73 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75 enum sched_tunable_scaling sysctl_sched_tunable_scaling
76 = SCHED_TUNABLESCALING_LOG;
79 * Minimal preemption granularity for CPU-bound tasks:
80 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_min_granularity = 750000ULL;
83 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
86 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
88 static unsigned int sched_nr_latency = 8;
91 * After fork, child runs first. If set to 0 (default) then
92 * parent will (try to) run first.
94 unsigned int sysctl_sched_child_runs_first __read_mostly;
97 * SCHED_OTHER wake-up granularity.
98 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
100 * This option delays the preemption effects of decoupled workloads
101 * and reduces their over-scheduling. Synchronous workloads will still
102 * have immediate wakeup/sleep latencies.
104 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
105 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
107 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
110 * The exponential sliding window over which load is averaged for shares
114 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
116 #ifdef CONFIG_CFS_BANDWIDTH
118 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
119 * each time a cfs_rq requests quota.
121 * Note: in the case that the slice exceeds the runtime remaining (either due
122 * to consumption or the quota being specified to be smaller than the slice)
123 * we will always only issue the remaining available time.
125 * default: 5 msec, units: microseconds
127 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 * The margin used when comparing utilization with CPU capacity:
132 * util * margin < capacity * 1024
134 unsigned int capacity_margin = 1280; /* ~20% */
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
172 case SCHED_TUNABLESCALING_LINEAR:
175 case SCHED_TUNABLESCALING_LOG:
177 factor = 1 + ilog2(cpus);
184 static void update_sysctl(void)
186 unsigned int factor = get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity);
191 SET_SYSCTL(sched_latency);
192 SET_SYSCTL(sched_wakeup_granularity);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
208 if (likely(lw->inv_weight))
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
235 u64 fact = scale_load_down(weight);
236 int shift = WMULT_SHIFT;
238 __update_inv_weight(lw);
240 if (unlikely(fact >> 32)) {
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
255 return mul_u64_u32_shr(delta_exec, fact, shift);
259 const struct sched_class fair_sched_class;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct *task_of(struct sched_entity *se)
278 #ifdef CONFIG_SCHED_DEBUG
279 WARN_ON_ONCE(!entity_is_task(se));
281 return container_of(se, struct task_struct, se);
284 /* Walk up scheduling entities hierarchy */
285 #define for_each_sched_entity(se) \
286 for (; se; se = se->parent)
288 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
293 /* runqueue on which this entity is (to be) queued */
294 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
299 /* runqueue "owned" by this group */
300 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
305 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
307 if (!cfs_rq->on_list) {
309 * Ensure we either appear before our parent (if already
310 * enqueued) or force our parent to appear after us when it is
311 * enqueued. The fact that we always enqueue bottom-up
312 * reduces this to two cases.
314 if (cfs_rq->tg->parent &&
315 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
316 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
317 &rq_of(cfs_rq)->leaf_cfs_rq_list);
319 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
320 &rq_of(cfs_rq)->leaf_cfs_rq_list);
327 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
329 if (cfs_rq->on_list) {
330 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
335 /* Iterate thr' all leaf cfs_rq's on a runqueue */
336 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
337 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
339 /* Do the two (enqueued) entities belong to the same group ? */
340 static inline struct cfs_rq *
341 is_same_group(struct sched_entity *se, struct sched_entity *pse)
343 if (se->cfs_rq == pse->cfs_rq)
349 static inline struct sched_entity *parent_entity(struct sched_entity *se)
355 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
357 int se_depth, pse_depth;
360 * preemption test can be made between sibling entities who are in the
361 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
362 * both tasks until we find their ancestors who are siblings of common
366 /* First walk up until both entities are at same depth */
367 se_depth = (*se)->depth;
368 pse_depth = (*pse)->depth;
370 while (se_depth > pse_depth) {
372 *se = parent_entity(*se);
375 while (pse_depth > se_depth) {
377 *pse = parent_entity(*pse);
380 while (!is_same_group(*se, *pse)) {
381 *se = parent_entity(*se);
382 *pse = parent_entity(*pse);
386 #else /* !CONFIG_FAIR_GROUP_SCHED */
388 static inline struct task_struct *task_of(struct sched_entity *se)
390 return container_of(se, struct task_struct, se);
393 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
395 return container_of(cfs_rq, struct rq, cfs);
398 #define entity_is_task(se) 1
400 #define for_each_sched_entity(se) \
401 for (; se; se = NULL)
403 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
405 return &task_rq(p)->cfs;
408 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
410 struct task_struct *p = task_of(se);
411 struct rq *rq = task_rq(p);
416 /* runqueue "owned" by this group */
417 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
422 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
426 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
430 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
431 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 unsigned int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
617 if (unlikely(se->load.weight != NICE_0_LOAD))
618 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
624 * The idea is to set a period in which each task runs once.
626 * When there are too many tasks (sched_nr_latency) we have to stretch
627 * this period because otherwise the slices get too small.
629 * p = (nr <= nl) ? l : l*nr/nl
631 static u64 __sched_period(unsigned long nr_running)
633 if (unlikely(nr_running > sched_nr_latency))
634 return nr_running * sysctl_sched_min_granularity;
636 return sysctl_sched_latency;
640 * We calculate the wall-time slice from the period by taking a part
641 * proportional to the weight.
645 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
649 for_each_sched_entity(se) {
650 struct load_weight *load;
651 struct load_weight lw;
653 cfs_rq = cfs_rq_of(se);
654 load = &cfs_rq->load;
656 if (unlikely(!se->on_rq)) {
659 update_load_add(&lw, se->load.weight);
662 slice = __calc_delta(slice, se->load.weight, load);
668 * We calculate the vruntime slice of a to-be-inserted task.
672 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 return calc_delta_fair(sched_slice(cfs_rq, se), se);
678 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
679 static unsigned long task_h_load(struct task_struct *p);
682 * We choose a half-life close to 1 scheduling period.
683 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
684 * dependent on this value.
686 #define LOAD_AVG_PERIOD 32
687 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
688 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
690 /* Give new sched_entity start runnable values to heavy its load in infant time */
691 void init_entity_runnable_average(struct sched_entity *se)
693 struct sched_avg *sa = &se->avg;
695 sa->last_update_time = 0;
697 * sched_avg's period_contrib should be strictly less then 1024, so
698 * we give it 1023 to make sure it is almost a period (1024us), and
699 * will definitely be update (after enqueue).
701 sa->period_contrib = 1023;
702 sa->load_avg = scale_load_down(se->load.weight);
703 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
705 * In previous Android versions, we used to have:
706 * sa->util_avg = sched_freq() ?
707 * sysctl_sched_initial_task_util :
708 * scale_load_down(SCHED_LOAD_SCALE);
709 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
710 * However, that functionality has been moved to enqueue.
711 * It is unclear if we should restore this in enqueue.
714 * At this point, util_avg won't be used in select_task_rq_fair anyway
718 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
722 * With new tasks being created, their initial util_avgs are extrapolated
723 * based on the cfs_rq's current util_avg:
725 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
727 * However, in many cases, the above util_avg does not give a desired
728 * value. Moreover, the sum of the util_avgs may be divergent, such
729 * as when the series is a harmonic series.
731 * To solve this problem, we also cap the util_avg of successive tasks to
732 * only 1/2 of the left utilization budget:
734 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
736 * where n denotes the nth task.
738 * For example, a simplest series from the beginning would be like:
740 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
741 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
743 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
744 * if util_avg > util_avg_cap.
746 void post_init_entity_util_avg(struct sched_entity *se)
748 struct cfs_rq *cfs_rq = cfs_rq_of(se);
749 struct sched_avg *sa = &se->avg;
750 long cap = (long)(scale_load_down(SCHED_LOAD_SCALE) - cfs_rq->avg.util_avg) / 2;
753 if (cfs_rq->avg.util_avg != 0) {
754 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
755 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
757 if (sa->util_avg > cap)
763 * If we wish to restore tuning via setting initial util,
764 * this is where we should do it.
766 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
770 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
771 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
773 void init_entity_runnable_average(struct sched_entity *se)
776 void post_init_entity_util_avg(struct sched_entity *se)
782 * Update the current task's runtime statistics.
784 static void update_curr(struct cfs_rq *cfs_rq)
786 struct sched_entity *curr = cfs_rq->curr;
787 u64 now = rq_clock_task(rq_of(cfs_rq));
793 delta_exec = now - curr->exec_start;
794 if (unlikely((s64)delta_exec <= 0))
797 curr->exec_start = now;
799 schedstat_set(curr->statistics.exec_max,
800 max(delta_exec, curr->statistics.exec_max));
802 curr->sum_exec_runtime += delta_exec;
803 schedstat_add(cfs_rq, exec_clock, delta_exec);
805 curr->vruntime += calc_delta_fair(delta_exec, curr);
806 update_min_vruntime(cfs_rq);
808 if (entity_is_task(curr)) {
809 struct task_struct *curtask = task_of(curr);
811 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
812 cpuacct_charge(curtask, delta_exec);
813 account_group_exec_runtime(curtask, delta_exec);
816 account_cfs_rq_runtime(cfs_rq, delta_exec);
819 static void update_curr_fair(struct rq *rq)
821 update_curr(cfs_rq_of(&rq->curr->se));
825 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
827 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
831 * Task is being enqueued - update stats:
833 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
836 * Are we enqueueing a waiting task? (for current tasks
837 * a dequeue/enqueue event is a NOP)
839 if (se != cfs_rq->curr)
840 update_stats_wait_start(cfs_rq, se);
844 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
846 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
847 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
848 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
849 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
850 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
851 #ifdef CONFIG_SCHEDSTATS
852 if (entity_is_task(se)) {
853 trace_sched_stat_wait(task_of(se),
854 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
857 schedstat_set(se->statistics.wait_start, 0);
861 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
864 * Mark the end of the wait period if dequeueing a
867 if (se != cfs_rq->curr)
868 update_stats_wait_end(cfs_rq, se);
872 * We are picking a new current task - update its stats:
875 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
878 * We are starting a new run period:
880 se->exec_start = rq_clock_task(rq_of(cfs_rq));
883 /**************************************************
884 * Scheduling class queueing methods:
887 #ifdef CONFIG_NUMA_BALANCING
889 * Approximate time to scan a full NUMA task in ms. The task scan period is
890 * calculated based on the tasks virtual memory size and
891 * numa_balancing_scan_size.
893 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
894 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
896 /* Portion of address space to scan in MB */
897 unsigned int sysctl_numa_balancing_scan_size = 256;
899 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
900 unsigned int sysctl_numa_balancing_scan_delay = 1000;
902 static unsigned int task_nr_scan_windows(struct task_struct *p)
904 unsigned long rss = 0;
905 unsigned long nr_scan_pages;
908 * Calculations based on RSS as non-present and empty pages are skipped
909 * by the PTE scanner and NUMA hinting faults should be trapped based
912 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
913 rss = get_mm_rss(p->mm);
917 rss = round_up(rss, nr_scan_pages);
918 return rss / nr_scan_pages;
921 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
922 #define MAX_SCAN_WINDOW 2560
924 static unsigned int task_scan_min(struct task_struct *p)
926 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
927 unsigned int scan, floor;
928 unsigned int windows = 1;
930 if (scan_size < MAX_SCAN_WINDOW)
931 windows = MAX_SCAN_WINDOW / scan_size;
932 floor = 1000 / windows;
934 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
935 return max_t(unsigned int, floor, scan);
938 static unsigned int task_scan_max(struct task_struct *p)
940 unsigned int smin = task_scan_min(p);
943 /* Watch for min being lower than max due to floor calculations */
944 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
945 return max(smin, smax);
948 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
950 rq->nr_numa_running += (p->numa_preferred_nid != -1);
951 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
954 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
956 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
957 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
963 spinlock_t lock; /* nr_tasks, tasks */
968 nodemask_t active_nodes;
969 unsigned long total_faults;
971 * Faults_cpu is used to decide whether memory should move
972 * towards the CPU. As a consequence, these stats are weighted
973 * more by CPU use than by memory faults.
975 unsigned long *faults_cpu;
976 unsigned long faults[0];
979 /* Shared or private faults. */
980 #define NR_NUMA_HINT_FAULT_TYPES 2
982 /* Memory and CPU locality */
983 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
985 /* Averaged statistics, and temporary buffers. */
986 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
988 pid_t task_numa_group_id(struct task_struct *p)
990 return p->numa_group ? p->numa_group->gid : 0;
994 * The averaged statistics, shared & private, memory & cpu,
995 * occupy the first half of the array. The second half of the
996 * array is for current counters, which are averaged into the
997 * first set by task_numa_placement.
999 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1001 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1004 static inline unsigned long task_faults(struct task_struct *p, int nid)
1006 if (!p->numa_faults)
1009 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1010 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1013 static inline unsigned long group_faults(struct task_struct *p, int nid)
1018 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1019 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1022 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1024 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1025 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1028 /* Handle placement on systems where not all nodes are directly connected. */
1029 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1030 int maxdist, bool task)
1032 unsigned long score = 0;
1036 * All nodes are directly connected, and the same distance
1037 * from each other. No need for fancy placement algorithms.
1039 if (sched_numa_topology_type == NUMA_DIRECT)
1043 * This code is called for each node, introducing N^2 complexity,
1044 * which should be ok given the number of nodes rarely exceeds 8.
1046 for_each_online_node(node) {
1047 unsigned long faults;
1048 int dist = node_distance(nid, node);
1051 * The furthest away nodes in the system are not interesting
1052 * for placement; nid was already counted.
1054 if (dist == sched_max_numa_distance || node == nid)
1058 * On systems with a backplane NUMA topology, compare groups
1059 * of nodes, and move tasks towards the group with the most
1060 * memory accesses. When comparing two nodes at distance
1061 * "hoplimit", only nodes closer by than "hoplimit" are part
1062 * of each group. Skip other nodes.
1064 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1068 /* Add up the faults from nearby nodes. */
1070 faults = task_faults(p, node);
1072 faults = group_faults(p, node);
1075 * On systems with a glueless mesh NUMA topology, there are
1076 * no fixed "groups of nodes". Instead, nodes that are not
1077 * directly connected bounce traffic through intermediate
1078 * nodes; a numa_group can occupy any set of nodes.
1079 * The further away a node is, the less the faults count.
1080 * This seems to result in good task placement.
1082 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1083 faults *= (sched_max_numa_distance - dist);
1084 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1094 * These return the fraction of accesses done by a particular task, or
1095 * task group, on a particular numa node. The group weight is given a
1096 * larger multiplier, in order to group tasks together that are almost
1097 * evenly spread out between numa nodes.
1099 static inline unsigned long task_weight(struct task_struct *p, int nid,
1102 unsigned long faults, total_faults;
1104 if (!p->numa_faults)
1107 total_faults = p->total_numa_faults;
1112 faults = task_faults(p, nid);
1113 faults += score_nearby_nodes(p, nid, dist, true);
1115 return 1000 * faults / total_faults;
1118 static inline unsigned long group_weight(struct task_struct *p, int nid,
1121 unsigned long faults, total_faults;
1126 total_faults = p->numa_group->total_faults;
1131 faults = group_faults(p, nid);
1132 faults += score_nearby_nodes(p, nid, dist, false);
1134 return 1000 * faults / total_faults;
1137 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1138 int src_nid, int dst_cpu)
1140 struct numa_group *ng = p->numa_group;
1141 int dst_nid = cpu_to_node(dst_cpu);
1142 int last_cpupid, this_cpupid;
1144 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1147 * Multi-stage node selection is used in conjunction with a periodic
1148 * migration fault to build a temporal task<->page relation. By using
1149 * a two-stage filter we remove short/unlikely relations.
1151 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1152 * a task's usage of a particular page (n_p) per total usage of this
1153 * page (n_t) (in a given time-span) to a probability.
1155 * Our periodic faults will sample this probability and getting the
1156 * same result twice in a row, given these samples are fully
1157 * independent, is then given by P(n)^2, provided our sample period
1158 * is sufficiently short compared to the usage pattern.
1160 * This quadric squishes small probabilities, making it less likely we
1161 * act on an unlikely task<->page relation.
1163 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1164 if (!cpupid_pid_unset(last_cpupid) &&
1165 cpupid_to_nid(last_cpupid) != dst_nid)
1168 /* Always allow migrate on private faults */
1169 if (cpupid_match_pid(p, last_cpupid))
1172 /* A shared fault, but p->numa_group has not been set up yet. */
1177 * Do not migrate if the destination is not a node that
1178 * is actively used by this numa group.
1180 if (!node_isset(dst_nid, ng->active_nodes))
1184 * Source is a node that is not actively used by this
1185 * numa group, while the destination is. Migrate.
1187 if (!node_isset(src_nid, ng->active_nodes))
1191 * Both source and destination are nodes in active
1192 * use by this numa group. Maximize memory bandwidth
1193 * by migrating from more heavily used groups, to less
1194 * heavily used ones, spreading the load around.
1195 * Use a 1/4 hysteresis to avoid spurious page movement.
1197 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1200 static unsigned long weighted_cpuload(const int cpu);
1201 static unsigned long source_load(int cpu, int type);
1202 static unsigned long target_load(int cpu, int type);
1203 static unsigned long capacity_of(int cpu);
1204 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1206 /* Cached statistics for all CPUs within a node */
1208 unsigned long nr_running;
1211 /* Total compute capacity of CPUs on a node */
1212 unsigned long compute_capacity;
1214 /* Approximate capacity in terms of runnable tasks on a node */
1215 unsigned long task_capacity;
1216 int has_free_capacity;
1220 * XXX borrowed from update_sg_lb_stats
1222 static void update_numa_stats(struct numa_stats *ns, int nid)
1224 int smt, cpu, cpus = 0;
1225 unsigned long capacity;
1227 memset(ns, 0, sizeof(*ns));
1228 for_each_cpu(cpu, cpumask_of_node(nid)) {
1229 struct rq *rq = cpu_rq(cpu);
1231 ns->nr_running += rq->nr_running;
1232 ns->load += weighted_cpuload(cpu);
1233 ns->compute_capacity += capacity_of(cpu);
1239 * If we raced with hotplug and there are no CPUs left in our mask
1240 * the @ns structure is NULL'ed and task_numa_compare() will
1241 * not find this node attractive.
1243 * We'll either bail at !has_free_capacity, or we'll detect a huge
1244 * imbalance and bail there.
1249 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1250 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1251 capacity = cpus / smt; /* cores */
1253 ns->task_capacity = min_t(unsigned, capacity,
1254 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1255 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1258 struct task_numa_env {
1259 struct task_struct *p;
1261 int src_cpu, src_nid;
1262 int dst_cpu, dst_nid;
1264 struct numa_stats src_stats, dst_stats;
1269 struct task_struct *best_task;
1274 static void task_numa_assign(struct task_numa_env *env,
1275 struct task_struct *p, long imp)
1278 put_task_struct(env->best_task);
1281 env->best_imp = imp;
1282 env->best_cpu = env->dst_cpu;
1285 static bool load_too_imbalanced(long src_load, long dst_load,
1286 struct task_numa_env *env)
1289 long orig_src_load, orig_dst_load;
1290 long src_capacity, dst_capacity;
1293 * The load is corrected for the CPU capacity available on each node.
1296 * ------------ vs ---------
1297 * src_capacity dst_capacity
1299 src_capacity = env->src_stats.compute_capacity;
1300 dst_capacity = env->dst_stats.compute_capacity;
1302 /* We care about the slope of the imbalance, not the direction. */
1303 if (dst_load < src_load)
1304 swap(dst_load, src_load);
1306 /* Is the difference below the threshold? */
1307 imb = dst_load * src_capacity * 100 -
1308 src_load * dst_capacity * env->imbalance_pct;
1313 * The imbalance is above the allowed threshold.
1314 * Compare it with the old imbalance.
1316 orig_src_load = env->src_stats.load;
1317 orig_dst_load = env->dst_stats.load;
1319 if (orig_dst_load < orig_src_load)
1320 swap(orig_dst_load, orig_src_load);
1322 old_imb = orig_dst_load * src_capacity * 100 -
1323 orig_src_load * dst_capacity * env->imbalance_pct;
1325 /* Would this change make things worse? */
1326 return (imb > old_imb);
1330 * This checks if the overall compute and NUMA accesses of the system would
1331 * be improved if the source tasks was migrated to the target dst_cpu taking
1332 * into account that it might be best if task running on the dst_cpu should
1333 * be exchanged with the source task
1335 static void task_numa_compare(struct task_numa_env *env,
1336 long taskimp, long groupimp)
1338 struct rq *src_rq = cpu_rq(env->src_cpu);
1339 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1340 struct task_struct *cur;
1341 long src_load, dst_load;
1343 long imp = env->p->numa_group ? groupimp : taskimp;
1345 int dist = env->dist;
1346 bool assigned = false;
1350 raw_spin_lock_irq(&dst_rq->lock);
1353 * No need to move the exiting task or idle task.
1355 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1359 * The task_struct must be protected here to protect the
1360 * p->numa_faults access in the task_weight since the
1361 * numa_faults could already be freed in the following path:
1362 * finish_task_switch()
1363 * --> put_task_struct()
1364 * --> __put_task_struct()
1365 * --> task_numa_free()
1367 get_task_struct(cur);
1370 raw_spin_unlock_irq(&dst_rq->lock);
1373 * Because we have preemption enabled we can get migrated around and
1374 * end try selecting ourselves (current == env->p) as a swap candidate.
1380 * "imp" is the fault differential for the source task between the
1381 * source and destination node. Calculate the total differential for
1382 * the source task and potential destination task. The more negative
1383 * the value is, the more rmeote accesses that would be expected to
1384 * be incurred if the tasks were swapped.
1387 /* Skip this swap candidate if cannot move to the source cpu */
1388 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1392 * If dst and source tasks are in the same NUMA group, or not
1393 * in any group then look only at task weights.
1395 if (cur->numa_group == env->p->numa_group) {
1396 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1397 task_weight(cur, env->dst_nid, dist);
1399 * Add some hysteresis to prevent swapping the
1400 * tasks within a group over tiny differences.
1402 if (cur->numa_group)
1406 * Compare the group weights. If a task is all by
1407 * itself (not part of a group), use the task weight
1410 if (cur->numa_group)
1411 imp += group_weight(cur, env->src_nid, dist) -
1412 group_weight(cur, env->dst_nid, dist);
1414 imp += task_weight(cur, env->src_nid, dist) -
1415 task_weight(cur, env->dst_nid, dist);
1419 if (imp <= env->best_imp && moveimp <= env->best_imp)
1423 /* Is there capacity at our destination? */
1424 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1425 !env->dst_stats.has_free_capacity)
1431 /* Balance doesn't matter much if we're running a task per cpu */
1432 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1433 dst_rq->nr_running == 1)
1437 * In the overloaded case, try and keep the load balanced.
1440 load = task_h_load(env->p);
1441 dst_load = env->dst_stats.load + load;
1442 src_load = env->src_stats.load - load;
1444 if (moveimp > imp && moveimp > env->best_imp) {
1446 * If the improvement from just moving env->p direction is
1447 * better than swapping tasks around, check if a move is
1448 * possible. Store a slightly smaller score than moveimp,
1449 * so an actually idle CPU will win.
1451 if (!load_too_imbalanced(src_load, dst_load, env)) {
1453 put_task_struct(cur);
1459 if (imp <= env->best_imp)
1463 load = task_h_load(cur);
1468 if (load_too_imbalanced(src_load, dst_load, env))
1472 * One idle CPU per node is evaluated for a task numa move.
1473 * Call select_idle_sibling to maybe find a better one.
1476 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1481 task_numa_assign(env, cur, imp);
1485 * The dst_rq->curr isn't assigned. The protection for task_struct is
1488 if (cur && !assigned)
1489 put_task_struct(cur);
1492 static void task_numa_find_cpu(struct task_numa_env *env,
1493 long taskimp, long groupimp)
1497 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1498 /* Skip this CPU if the source task cannot migrate */
1499 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1503 task_numa_compare(env, taskimp, groupimp);
1507 /* Only move tasks to a NUMA node less busy than the current node. */
1508 static bool numa_has_capacity(struct task_numa_env *env)
1510 struct numa_stats *src = &env->src_stats;
1511 struct numa_stats *dst = &env->dst_stats;
1513 if (src->has_free_capacity && !dst->has_free_capacity)
1517 * Only consider a task move if the source has a higher load
1518 * than the destination, corrected for CPU capacity on each node.
1520 * src->load dst->load
1521 * --------------------- vs ---------------------
1522 * src->compute_capacity dst->compute_capacity
1524 if (src->load * dst->compute_capacity * env->imbalance_pct >
1526 dst->load * src->compute_capacity * 100)
1532 static int task_numa_migrate(struct task_struct *p)
1534 struct task_numa_env env = {
1537 .src_cpu = task_cpu(p),
1538 .src_nid = task_node(p),
1540 .imbalance_pct = 112,
1546 struct sched_domain *sd;
1547 unsigned long taskweight, groupweight;
1549 long taskimp, groupimp;
1552 * Pick the lowest SD_NUMA domain, as that would have the smallest
1553 * imbalance and would be the first to start moving tasks about.
1555 * And we want to avoid any moving of tasks about, as that would create
1556 * random movement of tasks -- counter the numa conditions we're trying
1560 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1562 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1566 * Cpusets can break the scheduler domain tree into smaller
1567 * balance domains, some of which do not cross NUMA boundaries.
1568 * Tasks that are "trapped" in such domains cannot be migrated
1569 * elsewhere, so there is no point in (re)trying.
1571 if (unlikely(!sd)) {
1572 p->numa_preferred_nid = task_node(p);
1576 env.dst_nid = p->numa_preferred_nid;
1577 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1578 taskweight = task_weight(p, env.src_nid, dist);
1579 groupweight = group_weight(p, env.src_nid, dist);
1580 update_numa_stats(&env.src_stats, env.src_nid);
1581 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1582 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1583 update_numa_stats(&env.dst_stats, env.dst_nid);
1585 /* Try to find a spot on the preferred nid. */
1586 if (numa_has_capacity(&env))
1587 task_numa_find_cpu(&env, taskimp, groupimp);
1590 * Look at other nodes in these cases:
1591 * - there is no space available on the preferred_nid
1592 * - the task is part of a numa_group that is interleaved across
1593 * multiple NUMA nodes; in order to better consolidate the group,
1594 * we need to check other locations.
1596 if (env.best_cpu == -1 || (p->numa_group &&
1597 nodes_weight(p->numa_group->active_nodes) > 1)) {
1598 for_each_online_node(nid) {
1599 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1602 dist = node_distance(env.src_nid, env.dst_nid);
1603 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1605 taskweight = task_weight(p, env.src_nid, dist);
1606 groupweight = group_weight(p, env.src_nid, dist);
1609 /* Only consider nodes where both task and groups benefit */
1610 taskimp = task_weight(p, nid, dist) - taskweight;
1611 groupimp = group_weight(p, nid, dist) - groupweight;
1612 if (taskimp < 0 && groupimp < 0)
1617 update_numa_stats(&env.dst_stats, env.dst_nid);
1618 if (numa_has_capacity(&env))
1619 task_numa_find_cpu(&env, taskimp, groupimp);
1624 * If the task is part of a workload that spans multiple NUMA nodes,
1625 * and is migrating into one of the workload's active nodes, remember
1626 * this node as the task's preferred numa node, so the workload can
1628 * A task that migrated to a second choice node will be better off
1629 * trying for a better one later. Do not set the preferred node here.
1631 if (p->numa_group) {
1632 if (env.best_cpu == -1)
1637 if (node_isset(nid, p->numa_group->active_nodes))
1638 sched_setnuma(p, env.dst_nid);
1641 /* No better CPU than the current one was found. */
1642 if (env.best_cpu == -1)
1646 * Reset the scan period if the task is being rescheduled on an
1647 * alternative node to recheck if the tasks is now properly placed.
1649 p->numa_scan_period = task_scan_min(p);
1651 if (env.best_task == NULL) {
1652 ret = migrate_task_to(p, env.best_cpu);
1654 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1658 ret = migrate_swap(p, env.best_task);
1660 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1661 put_task_struct(env.best_task);
1665 /* Attempt to migrate a task to a CPU on the preferred node. */
1666 static void numa_migrate_preferred(struct task_struct *p)
1668 unsigned long interval = HZ;
1670 /* This task has no NUMA fault statistics yet */
1671 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1674 /* Periodically retry migrating the task to the preferred node */
1675 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1676 p->numa_migrate_retry = jiffies + interval;
1678 /* Success if task is already running on preferred CPU */
1679 if (task_node(p) == p->numa_preferred_nid)
1682 /* Otherwise, try migrate to a CPU on the preferred node */
1683 task_numa_migrate(p);
1687 * Find the nodes on which the workload is actively running. We do this by
1688 * tracking the nodes from which NUMA hinting faults are triggered. This can
1689 * be different from the set of nodes where the workload's memory is currently
1692 * The bitmask is used to make smarter decisions on when to do NUMA page
1693 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1694 * are added when they cause over 6/16 of the maximum number of faults, but
1695 * only removed when they drop below 3/16.
1697 static void update_numa_active_node_mask(struct numa_group *numa_group)
1699 unsigned long faults, max_faults = 0;
1702 for_each_online_node(nid) {
1703 faults = group_faults_cpu(numa_group, nid);
1704 if (faults > max_faults)
1705 max_faults = faults;
1708 for_each_online_node(nid) {
1709 faults = group_faults_cpu(numa_group, nid);
1710 if (!node_isset(nid, numa_group->active_nodes)) {
1711 if (faults > max_faults * 6 / 16)
1712 node_set(nid, numa_group->active_nodes);
1713 } else if (faults < max_faults * 3 / 16)
1714 node_clear(nid, numa_group->active_nodes);
1719 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1720 * increments. The more local the fault statistics are, the higher the scan
1721 * period will be for the next scan window. If local/(local+remote) ratio is
1722 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1723 * the scan period will decrease. Aim for 70% local accesses.
1725 #define NUMA_PERIOD_SLOTS 10
1726 #define NUMA_PERIOD_THRESHOLD 7
1729 * Increase the scan period (slow down scanning) if the majority of
1730 * our memory is already on our local node, or if the majority of
1731 * the page accesses are shared with other processes.
1732 * Otherwise, decrease the scan period.
1734 static void update_task_scan_period(struct task_struct *p,
1735 unsigned long shared, unsigned long private)
1737 unsigned int period_slot;
1741 unsigned long remote = p->numa_faults_locality[0];
1742 unsigned long local = p->numa_faults_locality[1];
1745 * If there were no record hinting faults then either the task is
1746 * completely idle or all activity is areas that are not of interest
1747 * to automatic numa balancing. Related to that, if there were failed
1748 * migration then it implies we are migrating too quickly or the local
1749 * node is overloaded. In either case, scan slower
1751 if (local + shared == 0 || p->numa_faults_locality[2]) {
1752 p->numa_scan_period = min(p->numa_scan_period_max,
1753 p->numa_scan_period << 1);
1755 p->mm->numa_next_scan = jiffies +
1756 msecs_to_jiffies(p->numa_scan_period);
1762 * Prepare to scale scan period relative to the current period.
1763 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1764 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1765 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1767 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1768 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1769 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1770 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1773 diff = slot * period_slot;
1775 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1778 * Scale scan rate increases based on sharing. There is an
1779 * inverse relationship between the degree of sharing and
1780 * the adjustment made to the scanning period. Broadly
1781 * speaking the intent is that there is little point
1782 * scanning faster if shared accesses dominate as it may
1783 * simply bounce migrations uselessly
1785 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1786 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1789 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1790 task_scan_min(p), task_scan_max(p));
1791 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1795 * Get the fraction of time the task has been running since the last
1796 * NUMA placement cycle. The scheduler keeps similar statistics, but
1797 * decays those on a 32ms period, which is orders of magnitude off
1798 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1799 * stats only if the task is so new there are no NUMA statistics yet.
1801 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1803 u64 runtime, delta, now;
1804 /* Use the start of this time slice to avoid calculations. */
1805 now = p->se.exec_start;
1806 runtime = p->se.sum_exec_runtime;
1808 if (p->last_task_numa_placement) {
1809 delta = runtime - p->last_sum_exec_runtime;
1810 *period = now - p->last_task_numa_placement;
1812 delta = p->se.avg.load_sum / p->se.load.weight;
1813 *period = LOAD_AVG_MAX;
1816 p->last_sum_exec_runtime = runtime;
1817 p->last_task_numa_placement = now;
1823 * Determine the preferred nid for a task in a numa_group. This needs to
1824 * be done in a way that produces consistent results with group_weight,
1825 * otherwise workloads might not converge.
1827 static int preferred_group_nid(struct task_struct *p, int nid)
1832 /* Direct connections between all NUMA nodes. */
1833 if (sched_numa_topology_type == NUMA_DIRECT)
1837 * On a system with glueless mesh NUMA topology, group_weight
1838 * scores nodes according to the number of NUMA hinting faults on
1839 * both the node itself, and on nearby nodes.
1841 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1842 unsigned long score, max_score = 0;
1843 int node, max_node = nid;
1845 dist = sched_max_numa_distance;
1847 for_each_online_node(node) {
1848 score = group_weight(p, node, dist);
1849 if (score > max_score) {
1858 * Finding the preferred nid in a system with NUMA backplane
1859 * interconnect topology is more involved. The goal is to locate
1860 * tasks from numa_groups near each other in the system, and
1861 * untangle workloads from different sides of the system. This requires
1862 * searching down the hierarchy of node groups, recursively searching
1863 * inside the highest scoring group of nodes. The nodemask tricks
1864 * keep the complexity of the search down.
1866 nodes = node_online_map;
1867 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1868 unsigned long max_faults = 0;
1869 nodemask_t max_group = NODE_MASK_NONE;
1872 /* Are there nodes at this distance from each other? */
1873 if (!find_numa_distance(dist))
1876 for_each_node_mask(a, nodes) {
1877 unsigned long faults = 0;
1878 nodemask_t this_group;
1879 nodes_clear(this_group);
1881 /* Sum group's NUMA faults; includes a==b case. */
1882 for_each_node_mask(b, nodes) {
1883 if (node_distance(a, b) < dist) {
1884 faults += group_faults(p, b);
1885 node_set(b, this_group);
1886 node_clear(b, nodes);
1890 /* Remember the top group. */
1891 if (faults > max_faults) {
1892 max_faults = faults;
1893 max_group = this_group;
1895 * subtle: at the smallest distance there is
1896 * just one node left in each "group", the
1897 * winner is the preferred nid.
1902 /* Next round, evaluate the nodes within max_group. */
1910 static void task_numa_placement(struct task_struct *p)
1912 int seq, nid, max_nid = -1, max_group_nid = -1;
1913 unsigned long max_faults = 0, max_group_faults = 0;
1914 unsigned long fault_types[2] = { 0, 0 };
1915 unsigned long total_faults;
1916 u64 runtime, period;
1917 spinlock_t *group_lock = NULL;
1920 * The p->mm->numa_scan_seq field gets updated without
1921 * exclusive access. Use READ_ONCE() here to ensure
1922 * that the field is read in a single access:
1924 seq = READ_ONCE(p->mm->numa_scan_seq);
1925 if (p->numa_scan_seq == seq)
1927 p->numa_scan_seq = seq;
1928 p->numa_scan_period_max = task_scan_max(p);
1930 total_faults = p->numa_faults_locality[0] +
1931 p->numa_faults_locality[1];
1932 runtime = numa_get_avg_runtime(p, &period);
1934 /* If the task is part of a group prevent parallel updates to group stats */
1935 if (p->numa_group) {
1936 group_lock = &p->numa_group->lock;
1937 spin_lock_irq(group_lock);
1940 /* Find the node with the highest number of faults */
1941 for_each_online_node(nid) {
1942 /* Keep track of the offsets in numa_faults array */
1943 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1944 unsigned long faults = 0, group_faults = 0;
1947 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1948 long diff, f_diff, f_weight;
1950 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1951 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1952 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1953 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1955 /* Decay existing window, copy faults since last scan */
1956 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1957 fault_types[priv] += p->numa_faults[membuf_idx];
1958 p->numa_faults[membuf_idx] = 0;
1961 * Normalize the faults_from, so all tasks in a group
1962 * count according to CPU use, instead of by the raw
1963 * number of faults. Tasks with little runtime have
1964 * little over-all impact on throughput, and thus their
1965 * faults are less important.
1967 f_weight = div64_u64(runtime << 16, period + 1);
1968 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1970 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1971 p->numa_faults[cpubuf_idx] = 0;
1973 p->numa_faults[mem_idx] += diff;
1974 p->numa_faults[cpu_idx] += f_diff;
1975 faults += p->numa_faults[mem_idx];
1976 p->total_numa_faults += diff;
1977 if (p->numa_group) {
1979 * safe because we can only change our own group
1981 * mem_idx represents the offset for a given
1982 * nid and priv in a specific region because it
1983 * is at the beginning of the numa_faults array.
1985 p->numa_group->faults[mem_idx] += diff;
1986 p->numa_group->faults_cpu[mem_idx] += f_diff;
1987 p->numa_group->total_faults += diff;
1988 group_faults += p->numa_group->faults[mem_idx];
1992 if (faults > max_faults) {
1993 max_faults = faults;
1997 if (group_faults > max_group_faults) {
1998 max_group_faults = group_faults;
1999 max_group_nid = nid;
2003 update_task_scan_period(p, fault_types[0], fault_types[1]);
2005 if (p->numa_group) {
2006 update_numa_active_node_mask(p->numa_group);
2007 spin_unlock_irq(group_lock);
2008 max_nid = preferred_group_nid(p, max_group_nid);
2012 /* Set the new preferred node */
2013 if (max_nid != p->numa_preferred_nid)
2014 sched_setnuma(p, max_nid);
2016 if (task_node(p) != p->numa_preferred_nid)
2017 numa_migrate_preferred(p);
2021 static inline int get_numa_group(struct numa_group *grp)
2023 return atomic_inc_not_zero(&grp->refcount);
2026 static inline void put_numa_group(struct numa_group *grp)
2028 if (atomic_dec_and_test(&grp->refcount))
2029 kfree_rcu(grp, rcu);
2032 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2035 struct numa_group *grp, *my_grp;
2036 struct task_struct *tsk;
2038 int cpu = cpupid_to_cpu(cpupid);
2041 if (unlikely(!p->numa_group)) {
2042 unsigned int size = sizeof(struct numa_group) +
2043 4*nr_node_ids*sizeof(unsigned long);
2045 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2049 atomic_set(&grp->refcount, 1);
2050 spin_lock_init(&grp->lock);
2052 /* Second half of the array tracks nids where faults happen */
2053 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2056 node_set(task_node(current), grp->active_nodes);
2058 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2059 grp->faults[i] = p->numa_faults[i];
2061 grp->total_faults = p->total_numa_faults;
2064 rcu_assign_pointer(p->numa_group, grp);
2068 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2070 if (!cpupid_match_pid(tsk, cpupid))
2073 grp = rcu_dereference(tsk->numa_group);
2077 my_grp = p->numa_group;
2082 * Only join the other group if its bigger; if we're the bigger group,
2083 * the other task will join us.
2085 if (my_grp->nr_tasks > grp->nr_tasks)
2089 * Tie-break on the grp address.
2091 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2094 /* Always join threads in the same process. */
2095 if (tsk->mm == current->mm)
2098 /* Simple filter to avoid false positives due to PID collisions */
2099 if (flags & TNF_SHARED)
2102 /* Update priv based on whether false sharing was detected */
2105 if (join && !get_numa_group(grp))
2113 BUG_ON(irqs_disabled());
2114 double_lock_irq(&my_grp->lock, &grp->lock);
2116 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2117 my_grp->faults[i] -= p->numa_faults[i];
2118 grp->faults[i] += p->numa_faults[i];
2120 my_grp->total_faults -= p->total_numa_faults;
2121 grp->total_faults += p->total_numa_faults;
2126 spin_unlock(&my_grp->lock);
2127 spin_unlock_irq(&grp->lock);
2129 rcu_assign_pointer(p->numa_group, grp);
2131 put_numa_group(my_grp);
2139 void task_numa_free(struct task_struct *p)
2141 struct numa_group *grp = p->numa_group;
2142 void *numa_faults = p->numa_faults;
2143 unsigned long flags;
2147 spin_lock_irqsave(&grp->lock, flags);
2148 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2149 grp->faults[i] -= p->numa_faults[i];
2150 grp->total_faults -= p->total_numa_faults;
2153 spin_unlock_irqrestore(&grp->lock, flags);
2154 RCU_INIT_POINTER(p->numa_group, NULL);
2155 put_numa_group(grp);
2158 p->numa_faults = NULL;
2163 * Got a PROT_NONE fault for a page on @node.
2165 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2167 struct task_struct *p = current;
2168 bool migrated = flags & TNF_MIGRATED;
2169 int cpu_node = task_node(current);
2170 int local = !!(flags & TNF_FAULT_LOCAL);
2173 if (!static_branch_likely(&sched_numa_balancing))
2176 /* for example, ksmd faulting in a user's mm */
2180 /* Allocate buffer to track faults on a per-node basis */
2181 if (unlikely(!p->numa_faults)) {
2182 int size = sizeof(*p->numa_faults) *
2183 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2185 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2186 if (!p->numa_faults)
2189 p->total_numa_faults = 0;
2190 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2194 * First accesses are treated as private, otherwise consider accesses
2195 * to be private if the accessing pid has not changed
2197 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2200 priv = cpupid_match_pid(p, last_cpupid);
2201 if (!priv && !(flags & TNF_NO_GROUP))
2202 task_numa_group(p, last_cpupid, flags, &priv);
2206 * If a workload spans multiple NUMA nodes, a shared fault that
2207 * occurs wholly within the set of nodes that the workload is
2208 * actively using should be counted as local. This allows the
2209 * scan rate to slow down when a workload has settled down.
2211 if (!priv && !local && p->numa_group &&
2212 node_isset(cpu_node, p->numa_group->active_nodes) &&
2213 node_isset(mem_node, p->numa_group->active_nodes))
2216 task_numa_placement(p);
2219 * Retry task to preferred node migration periodically, in case it
2220 * case it previously failed, or the scheduler moved us.
2222 if (time_after(jiffies, p->numa_migrate_retry))
2223 numa_migrate_preferred(p);
2226 p->numa_pages_migrated += pages;
2227 if (flags & TNF_MIGRATE_FAIL)
2228 p->numa_faults_locality[2] += pages;
2230 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2231 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2232 p->numa_faults_locality[local] += pages;
2235 static void reset_ptenuma_scan(struct task_struct *p)
2238 * We only did a read acquisition of the mmap sem, so
2239 * p->mm->numa_scan_seq is written to without exclusive access
2240 * and the update is not guaranteed to be atomic. That's not
2241 * much of an issue though, since this is just used for
2242 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2243 * expensive, to avoid any form of compiler optimizations:
2245 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2246 p->mm->numa_scan_offset = 0;
2250 * The expensive part of numa migration is done from task_work context.
2251 * Triggered from task_tick_numa().
2253 void task_numa_work(struct callback_head *work)
2255 unsigned long migrate, next_scan, now = jiffies;
2256 struct task_struct *p = current;
2257 struct mm_struct *mm = p->mm;
2258 struct vm_area_struct *vma;
2259 unsigned long start, end;
2260 unsigned long nr_pte_updates = 0;
2261 long pages, virtpages;
2263 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2265 work->next = work; /* protect against double add */
2267 * Who cares about NUMA placement when they're dying.
2269 * NOTE: make sure not to dereference p->mm before this check,
2270 * exit_task_work() happens _after_ exit_mm() so we could be called
2271 * without p->mm even though we still had it when we enqueued this
2274 if (p->flags & PF_EXITING)
2277 if (!mm->numa_next_scan) {
2278 mm->numa_next_scan = now +
2279 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2283 * Enforce maximal scan/migration frequency..
2285 migrate = mm->numa_next_scan;
2286 if (time_before(now, migrate))
2289 if (p->numa_scan_period == 0) {
2290 p->numa_scan_period_max = task_scan_max(p);
2291 p->numa_scan_period = task_scan_min(p);
2294 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2295 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2299 * Delay this task enough that another task of this mm will likely win
2300 * the next time around.
2302 p->node_stamp += 2 * TICK_NSEC;
2304 start = mm->numa_scan_offset;
2305 pages = sysctl_numa_balancing_scan_size;
2306 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2307 virtpages = pages * 8; /* Scan up to this much virtual space */
2312 down_read(&mm->mmap_sem);
2313 vma = find_vma(mm, start);
2315 reset_ptenuma_scan(p);
2319 for (; vma; vma = vma->vm_next) {
2320 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2321 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2326 * Shared library pages mapped by multiple processes are not
2327 * migrated as it is expected they are cache replicated. Avoid
2328 * hinting faults in read-only file-backed mappings or the vdso
2329 * as migrating the pages will be of marginal benefit.
2332 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2336 * Skip inaccessible VMAs to avoid any confusion between
2337 * PROT_NONE and NUMA hinting ptes
2339 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2343 start = max(start, vma->vm_start);
2344 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2345 end = min(end, vma->vm_end);
2346 nr_pte_updates = change_prot_numa(vma, start, end);
2349 * Try to scan sysctl_numa_balancing_size worth of
2350 * hpages that have at least one present PTE that
2351 * is not already pte-numa. If the VMA contains
2352 * areas that are unused or already full of prot_numa
2353 * PTEs, scan up to virtpages, to skip through those
2357 pages -= (end - start) >> PAGE_SHIFT;
2358 virtpages -= (end - start) >> PAGE_SHIFT;
2361 if (pages <= 0 || virtpages <= 0)
2365 } while (end != vma->vm_end);
2370 * It is possible to reach the end of the VMA list but the last few
2371 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2372 * would find the !migratable VMA on the next scan but not reset the
2373 * scanner to the start so check it now.
2376 mm->numa_scan_offset = start;
2378 reset_ptenuma_scan(p);
2379 up_read(&mm->mmap_sem);
2383 * Drive the periodic memory faults..
2385 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2387 struct callback_head *work = &curr->numa_work;
2391 * We don't care about NUMA placement if we don't have memory.
2393 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2397 * Using runtime rather than walltime has the dual advantage that
2398 * we (mostly) drive the selection from busy threads and that the
2399 * task needs to have done some actual work before we bother with
2402 now = curr->se.sum_exec_runtime;
2403 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2405 if (now > curr->node_stamp + period) {
2406 if (!curr->node_stamp)
2407 curr->numa_scan_period = task_scan_min(curr);
2408 curr->node_stamp += period;
2410 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2411 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2412 task_work_add(curr, work, true);
2417 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2421 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2425 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2428 #endif /* CONFIG_NUMA_BALANCING */
2431 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2433 update_load_add(&cfs_rq->load, se->load.weight);
2434 if (!parent_entity(se))
2435 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2437 if (entity_is_task(se)) {
2438 struct rq *rq = rq_of(cfs_rq);
2440 account_numa_enqueue(rq, task_of(se));
2441 list_add(&se->group_node, &rq->cfs_tasks);
2444 cfs_rq->nr_running++;
2448 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2450 update_load_sub(&cfs_rq->load, se->load.weight);
2451 if (!parent_entity(se))
2452 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2453 if (entity_is_task(se)) {
2454 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2455 list_del_init(&se->group_node);
2457 cfs_rq->nr_running--;
2460 #ifdef CONFIG_FAIR_GROUP_SCHED
2462 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2467 * Use this CPU's real-time load instead of the last load contribution
2468 * as the updating of the contribution is delayed, and we will use the
2469 * the real-time load to calc the share. See update_tg_load_avg().
2471 tg_weight = atomic_long_read(&tg->load_avg);
2472 tg_weight -= cfs_rq->tg_load_avg_contrib;
2473 tg_weight += cfs_rq->load.weight;
2478 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2480 long tg_weight, load, shares;
2482 tg_weight = calc_tg_weight(tg, cfs_rq);
2483 load = cfs_rq->load.weight;
2485 shares = (tg->shares * load);
2487 shares /= tg_weight;
2489 if (shares < MIN_SHARES)
2490 shares = MIN_SHARES;
2491 if (shares > tg->shares)
2492 shares = tg->shares;
2496 # else /* CONFIG_SMP */
2497 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2501 # endif /* CONFIG_SMP */
2502 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2503 unsigned long weight)
2506 /* commit outstanding execution time */
2507 if (cfs_rq->curr == se)
2508 update_curr(cfs_rq);
2509 account_entity_dequeue(cfs_rq, se);
2512 update_load_set(&se->load, weight);
2515 account_entity_enqueue(cfs_rq, se);
2518 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2520 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2522 struct task_group *tg;
2523 struct sched_entity *se;
2527 se = tg->se[cpu_of(rq_of(cfs_rq))];
2528 if (!se || throttled_hierarchy(cfs_rq))
2531 if (likely(se->load.weight == tg->shares))
2534 shares = calc_cfs_shares(cfs_rq, tg);
2536 reweight_entity(cfs_rq_of(se), se, shares);
2538 #else /* CONFIG_FAIR_GROUP_SCHED */
2539 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2542 #endif /* CONFIG_FAIR_GROUP_SCHED */
2545 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2546 static const u32 runnable_avg_yN_inv[] = {
2547 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2548 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2549 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2550 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2551 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2552 0x85aac367, 0x82cd8698,
2556 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2557 * over-estimates when re-combining.
2559 static const u32 runnable_avg_yN_sum[] = {
2560 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2561 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2562 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2567 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2569 static __always_inline u64 decay_load(u64 val, u64 n)
2571 unsigned int local_n;
2575 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2578 /* after bounds checking we can collapse to 32-bit */
2582 * As y^PERIOD = 1/2, we can combine
2583 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2584 * With a look-up table which covers y^n (n<PERIOD)
2586 * To achieve constant time decay_load.
2588 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2589 val >>= local_n / LOAD_AVG_PERIOD;
2590 local_n %= LOAD_AVG_PERIOD;
2593 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2598 * For updates fully spanning n periods, the contribution to runnable
2599 * average will be: \Sum 1024*y^n
2601 * We can compute this reasonably efficiently by combining:
2602 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2604 static u32 __compute_runnable_contrib(u64 n)
2608 if (likely(n <= LOAD_AVG_PERIOD))
2609 return runnable_avg_yN_sum[n];
2610 else if (unlikely(n >= LOAD_AVG_MAX_N))
2611 return LOAD_AVG_MAX;
2613 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2615 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2616 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2618 n -= LOAD_AVG_PERIOD;
2619 } while (n > LOAD_AVG_PERIOD);
2621 contrib = decay_load(contrib, n);
2622 return contrib + runnable_avg_yN_sum[n];
2625 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2626 #error "load tracking assumes 2^10 as unit"
2629 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2632 * We can represent the historical contribution to runnable average as the
2633 * coefficients of a geometric series. To do this we sub-divide our runnable
2634 * history into segments of approximately 1ms (1024us); label the segment that
2635 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2637 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2639 * (now) (~1ms ago) (~2ms ago)
2641 * Let u_i denote the fraction of p_i that the entity was runnable.
2643 * We then designate the fractions u_i as our co-efficients, yielding the
2644 * following representation of historical load:
2645 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2647 * We choose y based on the with of a reasonably scheduling period, fixing:
2650 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2651 * approximately half as much as the contribution to load within the last ms
2654 * When a period "rolls over" and we have new u_0`, multiplying the previous
2655 * sum again by y is sufficient to update:
2656 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2657 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2659 static __always_inline int
2660 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2661 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2663 u64 delta, scaled_delta, periods;
2665 unsigned int delta_w, scaled_delta_w, decayed = 0;
2666 unsigned long scale_freq, scale_cpu;
2668 delta = now - sa->last_update_time;
2670 * This should only happen when time goes backwards, which it
2671 * unfortunately does during sched clock init when we swap over to TSC.
2673 if ((s64)delta < 0) {
2674 sa->last_update_time = now;
2679 * Use 1024ns as the unit of measurement since it's a reasonable
2680 * approximation of 1us and fast to compute.
2685 sa->last_update_time = now;
2687 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2688 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2689 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2691 /* delta_w is the amount already accumulated against our next period */
2692 delta_w = sa->period_contrib;
2693 if (delta + delta_w >= 1024) {
2696 /* how much left for next period will start over, we don't know yet */
2697 sa->period_contrib = 0;
2700 * Now that we know we're crossing a period boundary, figure
2701 * out how much from delta we need to complete the current
2702 * period and accrue it.
2704 delta_w = 1024 - delta_w;
2705 scaled_delta_w = cap_scale(delta_w, scale_freq);
2707 sa->load_sum += weight * scaled_delta_w;
2709 cfs_rq->runnable_load_sum +=
2710 weight * scaled_delta_w;
2714 sa->util_sum += scaled_delta_w * scale_cpu;
2718 /* Figure out how many additional periods this update spans */
2719 periods = delta / 1024;
2722 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2724 cfs_rq->runnable_load_sum =
2725 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2727 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2729 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2730 contrib = __compute_runnable_contrib(periods);
2731 contrib = cap_scale(contrib, scale_freq);
2733 sa->load_sum += weight * contrib;
2735 cfs_rq->runnable_load_sum += weight * contrib;
2738 sa->util_sum += contrib * scale_cpu;
2741 /* Remainder of delta accrued against u_0` */
2742 scaled_delta = cap_scale(delta, scale_freq);
2744 sa->load_sum += weight * scaled_delta;
2746 cfs_rq->runnable_load_sum += weight * scaled_delta;
2749 sa->util_sum += scaled_delta * scale_cpu;
2751 sa->period_contrib += delta;
2754 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2756 cfs_rq->runnable_load_avg =
2757 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2759 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2765 #ifdef CONFIG_FAIR_GROUP_SCHED
2767 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2768 * and effective_load (which is not done because it is too costly).
2770 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2772 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2774 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2775 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2776 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2780 #else /* CONFIG_FAIR_GROUP_SCHED */
2781 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2782 #endif /* CONFIG_FAIR_GROUP_SCHED */
2784 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2786 if (&this_rq()->cfs == cfs_rq) {
2788 * There are a few boundary cases this might miss but it should
2789 * get called often enough that that should (hopefully) not be
2790 * a real problem -- added to that it only calls on the local
2791 * CPU, so if we enqueue remotely we'll miss an update, but
2792 * the next tick/schedule should update.
2794 * It will not get called when we go idle, because the idle
2795 * thread is a different class (!fair), nor will the utilization
2796 * number include things like RT tasks.
2798 * As is, the util number is not freq-invariant (we'd have to
2799 * implement arch_scale_freq_capacity() for that).
2803 cpufreq_update_util(rq_of(cfs_rq), 0);
2807 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2810 * Unsigned subtract and clamp on underflow.
2812 * Explicitly do a load-store to ensure the intermediate value never hits
2813 * memory. This allows lockless observations without ever seeing the negative
2816 #define sub_positive(_ptr, _val) do { \
2817 typeof(_ptr) ptr = (_ptr); \
2818 typeof(*ptr) val = (_val); \
2819 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2823 WRITE_ONCE(*ptr, res); \
2826 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2827 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq,
2830 struct sched_avg *sa = &cfs_rq->avg;
2831 int decayed, removed = 0, removed_util = 0;
2833 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2834 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2835 sub_positive(&sa->load_avg, r);
2836 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2840 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2841 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2842 sub_positive(&sa->util_avg, r);
2843 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2847 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2848 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2850 #ifndef CONFIG_64BIT
2852 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2855 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2856 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2857 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2859 if (update_freq && (decayed || removed_util))
2860 cfs_rq_util_change(cfs_rq);
2862 return decayed || removed;
2865 /* Update task and its cfs_rq load average */
2866 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2868 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2869 u64 now = cfs_rq_clock_task(cfs_rq);
2870 int cpu = cpu_of(rq_of(cfs_rq));
2873 * Track task load average for carrying it to new CPU after migrated, and
2874 * track group sched_entity load average for task_h_load calc in migration
2876 __update_load_avg(now, cpu, &se->avg,
2877 se->on_rq * scale_load_down(se->load.weight),
2878 cfs_rq->curr == se, NULL);
2880 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2881 update_tg_load_avg(cfs_rq, 0);
2883 if (entity_is_task(se))
2884 trace_sched_load_avg_task(task_of(se), &se->avg);
2887 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2889 if (!sched_feat(ATTACH_AGE_LOAD))
2893 * If we got migrated (either between CPUs or between cgroups) we'll
2894 * have aged the average right before clearing @last_update_time.
2896 if (se->avg.last_update_time) {
2897 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2898 &se->avg, 0, 0, NULL);
2901 * XXX: we could have just aged the entire load away if we've been
2902 * absent from the fair class for too long.
2907 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2908 cfs_rq->avg.load_avg += se->avg.load_avg;
2909 cfs_rq->avg.load_sum += se->avg.load_sum;
2910 cfs_rq->avg.util_avg += se->avg.util_avg;
2911 cfs_rq->avg.util_sum += se->avg.util_sum;
2913 cfs_rq_util_change(cfs_rq);
2916 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2918 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2919 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2920 cfs_rq->curr == se, NULL);
2922 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2923 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2924 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2925 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2927 cfs_rq_util_change(cfs_rq);
2930 /* Add the load generated by se into cfs_rq's load average */
2932 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2934 struct sched_avg *sa = &se->avg;
2935 u64 now = cfs_rq_clock_task(cfs_rq);
2936 int migrated, decayed;
2938 migrated = !sa->last_update_time;
2940 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2941 se->on_rq * scale_load_down(se->load.weight),
2942 cfs_rq->curr == se, NULL);
2945 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
2947 cfs_rq->runnable_load_avg += sa->load_avg;
2948 cfs_rq->runnable_load_sum += sa->load_sum;
2951 attach_entity_load_avg(cfs_rq, se);
2953 if (decayed || migrated)
2954 update_tg_load_avg(cfs_rq, 0);
2957 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2959 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2961 update_load_avg(se, 1);
2963 cfs_rq->runnable_load_avg =
2964 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2965 cfs_rq->runnable_load_sum =
2966 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2969 #ifndef CONFIG_64BIT
2970 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2972 u64 last_update_time_copy;
2973 u64 last_update_time;
2976 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2978 last_update_time = cfs_rq->avg.last_update_time;
2979 } while (last_update_time != last_update_time_copy);
2981 return last_update_time;
2984 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2986 return cfs_rq->avg.last_update_time;
2991 * Synchronize entity load avg of dequeued entity without locking
2994 void sync_entity_load_avg(struct sched_entity *se)
2996 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2997 u64 last_update_time;
2999 last_update_time = cfs_rq_last_update_time(cfs_rq);
3000 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3004 * Task first catches up with cfs_rq, and then subtract
3005 * itself from the cfs_rq (task must be off the queue now).
3007 void remove_entity_load_avg(struct sched_entity *se)
3009 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3012 * Newly created task or never used group entity should not be removed
3013 * from its (source) cfs_rq
3015 if (se->avg.last_update_time == 0)
3018 sync_entity_load_avg(se);
3019 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3020 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3024 * Update the rq's load with the elapsed running time before entering
3025 * idle. if the last scheduled task is not a CFS task, idle_enter will
3026 * be the only way to update the runnable statistic.
3028 void idle_enter_fair(struct rq *this_rq)
3033 * Update the rq's load with the elapsed idle time before a task is
3034 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3035 * be the only way to update the runnable statistic.
3037 void idle_exit_fair(struct rq *this_rq)
3041 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3043 return cfs_rq->runnable_load_avg;
3046 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3048 return cfs_rq->avg.load_avg;
3051 static int idle_balance(struct rq *this_rq);
3053 #else /* CONFIG_SMP */
3055 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3057 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3061 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3063 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3064 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3067 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3069 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3071 static inline int idle_balance(struct rq *rq)
3076 #endif /* CONFIG_SMP */
3078 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3080 #ifdef CONFIG_SCHEDSTATS
3081 struct task_struct *tsk = NULL;
3083 if (entity_is_task(se))
3086 if (se->statistics.sleep_start) {
3087 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3092 if (unlikely(delta > se->statistics.sleep_max))
3093 se->statistics.sleep_max = delta;
3095 se->statistics.sleep_start = 0;
3096 se->statistics.sum_sleep_runtime += delta;
3099 account_scheduler_latency(tsk, delta >> 10, 1);
3100 trace_sched_stat_sleep(tsk, delta);
3103 if (se->statistics.block_start) {
3104 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3109 if (unlikely(delta > se->statistics.block_max))
3110 se->statistics.block_max = delta;
3112 se->statistics.block_start = 0;
3113 se->statistics.sum_sleep_runtime += delta;
3116 if (tsk->in_iowait) {
3117 se->statistics.iowait_sum += delta;
3118 se->statistics.iowait_count++;
3119 trace_sched_stat_iowait(tsk, delta);
3122 trace_sched_stat_blocked(tsk, delta);
3123 trace_sched_blocked_reason(tsk);
3126 * Blocking time is in units of nanosecs, so shift by
3127 * 20 to get a milliseconds-range estimation of the
3128 * amount of time that the task spent sleeping:
3130 if (unlikely(prof_on == SLEEP_PROFILING)) {
3131 profile_hits(SLEEP_PROFILING,
3132 (void *)get_wchan(tsk),
3135 account_scheduler_latency(tsk, delta >> 10, 0);
3141 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3143 #ifdef CONFIG_SCHED_DEBUG
3144 s64 d = se->vruntime - cfs_rq->min_vruntime;
3149 if (d > 3*sysctl_sched_latency)
3150 schedstat_inc(cfs_rq, nr_spread_over);
3155 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3157 u64 vruntime = cfs_rq->min_vruntime;
3160 * The 'current' period is already promised to the current tasks,
3161 * however the extra weight of the new task will slow them down a
3162 * little, place the new task so that it fits in the slot that
3163 * stays open at the end.
3165 if (initial && sched_feat(START_DEBIT))
3166 vruntime += sched_vslice(cfs_rq, se);
3168 /* sleeps up to a single latency don't count. */
3170 unsigned long thresh = sysctl_sched_latency;
3173 * Halve their sleep time's effect, to allow
3174 * for a gentler effect of sleepers:
3176 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3182 /* ensure we never gain time by being placed backwards. */
3183 se->vruntime = max_vruntime(se->vruntime, vruntime);
3186 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3189 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3192 * Update the normalized vruntime before updating min_vruntime
3193 * through calling update_curr().
3195 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3196 se->vruntime += cfs_rq->min_vruntime;
3199 * Update run-time statistics of the 'current'.
3201 update_curr(cfs_rq);
3202 enqueue_entity_load_avg(cfs_rq, se);
3203 account_entity_enqueue(cfs_rq, se);
3204 update_cfs_shares(cfs_rq);
3206 if (flags & ENQUEUE_WAKEUP) {
3207 place_entity(cfs_rq, se, 0);
3208 enqueue_sleeper(cfs_rq, se);
3211 update_stats_enqueue(cfs_rq, se);
3212 check_spread(cfs_rq, se);
3213 if (se != cfs_rq->curr)
3214 __enqueue_entity(cfs_rq, se);
3217 if (cfs_rq->nr_running == 1) {
3218 list_add_leaf_cfs_rq(cfs_rq);
3219 check_enqueue_throttle(cfs_rq);
3223 static void __clear_buddies_last(struct sched_entity *se)
3225 for_each_sched_entity(se) {
3226 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3227 if (cfs_rq->last != se)
3230 cfs_rq->last = NULL;
3234 static void __clear_buddies_next(struct sched_entity *se)
3236 for_each_sched_entity(se) {
3237 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3238 if (cfs_rq->next != se)
3241 cfs_rq->next = NULL;
3245 static void __clear_buddies_skip(struct sched_entity *se)
3247 for_each_sched_entity(se) {
3248 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3249 if (cfs_rq->skip != se)
3252 cfs_rq->skip = NULL;
3256 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3258 if (cfs_rq->last == se)
3259 __clear_buddies_last(se);
3261 if (cfs_rq->next == se)
3262 __clear_buddies_next(se);
3264 if (cfs_rq->skip == se)
3265 __clear_buddies_skip(se);
3268 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3271 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3274 * Update run-time statistics of the 'current'.
3276 update_curr(cfs_rq);
3277 dequeue_entity_load_avg(cfs_rq, se);
3279 update_stats_dequeue(cfs_rq, se);
3280 if (flags & DEQUEUE_SLEEP) {
3281 #ifdef CONFIG_SCHEDSTATS
3282 if (entity_is_task(se)) {
3283 struct task_struct *tsk = task_of(se);
3285 if (tsk->state & TASK_INTERRUPTIBLE)
3286 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3287 if (tsk->state & TASK_UNINTERRUPTIBLE)
3288 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3293 clear_buddies(cfs_rq, se);
3295 if (se != cfs_rq->curr)
3296 __dequeue_entity(cfs_rq, se);
3298 account_entity_dequeue(cfs_rq, se);
3301 * Normalize the entity after updating the min_vruntime because the
3302 * update can refer to the ->curr item and we need to reflect this
3303 * movement in our normalized position.
3305 if (!(flags & DEQUEUE_SLEEP))
3306 se->vruntime -= cfs_rq->min_vruntime;
3308 /* return excess runtime on last dequeue */
3309 return_cfs_rq_runtime(cfs_rq);
3311 update_min_vruntime(cfs_rq);
3312 update_cfs_shares(cfs_rq);
3316 * Preempt the current task with a newly woken task if needed:
3319 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3321 unsigned long ideal_runtime, delta_exec;
3322 struct sched_entity *se;
3325 ideal_runtime = sched_slice(cfs_rq, curr);
3326 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3327 if (delta_exec > ideal_runtime) {
3328 resched_curr(rq_of(cfs_rq));
3330 * The current task ran long enough, ensure it doesn't get
3331 * re-elected due to buddy favours.
3333 clear_buddies(cfs_rq, curr);
3338 * Ensure that a task that missed wakeup preemption by a
3339 * narrow margin doesn't have to wait for a full slice.
3340 * This also mitigates buddy induced latencies under load.
3342 if (delta_exec < sysctl_sched_min_granularity)
3345 se = __pick_first_entity(cfs_rq);
3346 delta = curr->vruntime - se->vruntime;
3351 if (delta > ideal_runtime)
3352 resched_curr(rq_of(cfs_rq));
3356 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3358 /* 'current' is not kept within the tree. */
3361 * Any task has to be enqueued before it get to execute on
3362 * a CPU. So account for the time it spent waiting on the
3365 update_stats_wait_end(cfs_rq, se);
3366 __dequeue_entity(cfs_rq, se);
3367 update_load_avg(se, 1);
3370 update_stats_curr_start(cfs_rq, se);
3372 #ifdef CONFIG_SCHEDSTATS
3374 * Track our maximum slice length, if the CPU's load is at
3375 * least twice that of our own weight (i.e. dont track it
3376 * when there are only lesser-weight tasks around):
3378 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3379 se->statistics.slice_max = max(se->statistics.slice_max,
3380 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3383 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3387 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3390 * Pick the next process, keeping these things in mind, in this order:
3391 * 1) keep things fair between processes/task groups
3392 * 2) pick the "next" process, since someone really wants that to run
3393 * 3) pick the "last" process, for cache locality
3394 * 4) do not run the "skip" process, if something else is available
3396 static struct sched_entity *
3397 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3399 struct sched_entity *left = __pick_first_entity(cfs_rq);
3400 struct sched_entity *se;
3403 * If curr is set we have to see if its left of the leftmost entity
3404 * still in the tree, provided there was anything in the tree at all.
3406 if (!left || (curr && entity_before(curr, left)))
3409 se = left; /* ideally we run the leftmost entity */
3412 * Avoid running the skip buddy, if running something else can
3413 * be done without getting too unfair.
3415 if (cfs_rq->skip == se) {
3416 struct sched_entity *second;
3419 second = __pick_first_entity(cfs_rq);
3421 second = __pick_next_entity(se);
3422 if (!second || (curr && entity_before(curr, second)))
3426 if (second && wakeup_preempt_entity(second, left) < 1)
3431 * Prefer last buddy, try to return the CPU to a preempted task.
3433 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3437 * Someone really wants this to run. If it's not unfair, run it.
3439 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3442 clear_buddies(cfs_rq, se);
3447 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3449 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3452 * If still on the runqueue then deactivate_task()
3453 * was not called and update_curr() has to be done:
3456 update_curr(cfs_rq);
3458 /* throttle cfs_rqs exceeding runtime */
3459 check_cfs_rq_runtime(cfs_rq);
3461 check_spread(cfs_rq, prev);
3463 update_stats_wait_start(cfs_rq, prev);
3464 /* Put 'current' back into the tree. */
3465 __enqueue_entity(cfs_rq, prev);
3466 /* in !on_rq case, update occurred at dequeue */
3467 update_load_avg(prev, 0);
3469 cfs_rq->curr = NULL;
3473 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3476 * Update run-time statistics of the 'current'.
3478 update_curr(cfs_rq);
3481 * Ensure that runnable average is periodically updated.
3483 update_load_avg(curr, 1);
3484 update_cfs_shares(cfs_rq);
3486 #ifdef CONFIG_SCHED_HRTICK
3488 * queued ticks are scheduled to match the slice, so don't bother
3489 * validating it and just reschedule.
3492 resched_curr(rq_of(cfs_rq));
3496 * don't let the period tick interfere with the hrtick preemption
3498 if (!sched_feat(DOUBLE_TICK) &&
3499 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3503 if (cfs_rq->nr_running > 1)
3504 check_preempt_tick(cfs_rq, curr);
3508 /**************************************************
3509 * CFS bandwidth control machinery
3512 #ifdef CONFIG_CFS_BANDWIDTH
3514 #ifdef HAVE_JUMP_LABEL
3515 static struct static_key __cfs_bandwidth_used;
3517 static inline bool cfs_bandwidth_used(void)
3519 return static_key_false(&__cfs_bandwidth_used);
3522 void cfs_bandwidth_usage_inc(void)
3524 static_key_slow_inc(&__cfs_bandwidth_used);
3527 void cfs_bandwidth_usage_dec(void)
3529 static_key_slow_dec(&__cfs_bandwidth_used);
3531 #else /* HAVE_JUMP_LABEL */
3532 static bool cfs_bandwidth_used(void)
3537 void cfs_bandwidth_usage_inc(void) {}
3538 void cfs_bandwidth_usage_dec(void) {}
3539 #endif /* HAVE_JUMP_LABEL */
3542 * default period for cfs group bandwidth.
3543 * default: 0.1s, units: nanoseconds
3545 static inline u64 default_cfs_period(void)
3547 return 100000000ULL;
3550 static inline u64 sched_cfs_bandwidth_slice(void)
3552 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3556 * Replenish runtime according to assigned quota and update expiration time.
3557 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3558 * additional synchronization around rq->lock.
3560 * requires cfs_b->lock
3562 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3566 if (cfs_b->quota == RUNTIME_INF)
3569 now = sched_clock_cpu(smp_processor_id());
3570 cfs_b->runtime = cfs_b->quota;
3571 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3574 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3576 return &tg->cfs_bandwidth;
3579 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3580 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3582 if (unlikely(cfs_rq->throttle_count))
3583 return cfs_rq->throttled_clock_task;
3585 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3588 /* returns 0 on failure to allocate runtime */
3589 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3591 struct task_group *tg = cfs_rq->tg;
3592 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3593 u64 amount = 0, min_amount, expires;
3595 /* note: this is a positive sum as runtime_remaining <= 0 */
3596 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3598 raw_spin_lock(&cfs_b->lock);
3599 if (cfs_b->quota == RUNTIME_INF)
3600 amount = min_amount;
3602 start_cfs_bandwidth(cfs_b);
3604 if (cfs_b->runtime > 0) {
3605 amount = min(cfs_b->runtime, min_amount);
3606 cfs_b->runtime -= amount;
3610 expires = cfs_b->runtime_expires;
3611 raw_spin_unlock(&cfs_b->lock);
3613 cfs_rq->runtime_remaining += amount;
3615 * we may have advanced our local expiration to account for allowed
3616 * spread between our sched_clock and the one on which runtime was
3619 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3620 cfs_rq->runtime_expires = expires;
3622 return cfs_rq->runtime_remaining > 0;
3626 * Note: This depends on the synchronization provided by sched_clock and the
3627 * fact that rq->clock snapshots this value.
3629 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3631 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3633 /* if the deadline is ahead of our clock, nothing to do */
3634 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3637 if (cfs_rq->runtime_remaining < 0)
3641 * If the local deadline has passed we have to consider the
3642 * possibility that our sched_clock is 'fast' and the global deadline
3643 * has not truly expired.
3645 * Fortunately we can check determine whether this the case by checking
3646 * whether the global deadline has advanced. It is valid to compare
3647 * cfs_b->runtime_expires without any locks since we only care about
3648 * exact equality, so a partial write will still work.
3651 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3652 /* extend local deadline, drift is bounded above by 2 ticks */
3653 cfs_rq->runtime_expires += TICK_NSEC;
3655 /* global deadline is ahead, expiration has passed */
3656 cfs_rq->runtime_remaining = 0;
3660 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3662 /* dock delta_exec before expiring quota (as it could span periods) */
3663 cfs_rq->runtime_remaining -= delta_exec;
3664 expire_cfs_rq_runtime(cfs_rq);
3666 if (likely(cfs_rq->runtime_remaining > 0))
3670 * if we're unable to extend our runtime we resched so that the active
3671 * hierarchy can be throttled
3673 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3674 resched_curr(rq_of(cfs_rq));
3677 static __always_inline
3678 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3680 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3683 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3686 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3688 return cfs_bandwidth_used() && cfs_rq->throttled;
3691 /* check whether cfs_rq, or any parent, is throttled */
3692 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3694 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3698 * Ensure that neither of the group entities corresponding to src_cpu or
3699 * dest_cpu are members of a throttled hierarchy when performing group
3700 * load-balance operations.
3702 static inline int throttled_lb_pair(struct task_group *tg,
3703 int src_cpu, int dest_cpu)
3705 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3707 src_cfs_rq = tg->cfs_rq[src_cpu];
3708 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3710 return throttled_hierarchy(src_cfs_rq) ||
3711 throttled_hierarchy(dest_cfs_rq);
3714 /* updated child weight may affect parent so we have to do this bottom up */
3715 static int tg_unthrottle_up(struct task_group *tg, void *data)
3717 struct rq *rq = data;
3718 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3720 cfs_rq->throttle_count--;
3722 if (!cfs_rq->throttle_count) {
3723 /* adjust cfs_rq_clock_task() */
3724 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3725 cfs_rq->throttled_clock_task;
3732 static int tg_throttle_down(struct task_group *tg, void *data)
3734 struct rq *rq = data;
3735 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3737 /* group is entering throttled state, stop time */
3738 if (!cfs_rq->throttle_count)
3739 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3740 cfs_rq->throttle_count++;
3745 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3747 struct rq *rq = rq_of(cfs_rq);
3748 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3749 struct sched_entity *se;
3750 long task_delta, dequeue = 1;
3753 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3755 /* freeze hierarchy runnable averages while throttled */
3757 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3760 task_delta = cfs_rq->h_nr_running;
3761 for_each_sched_entity(se) {
3762 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3763 /* throttled entity or throttle-on-deactivate */
3768 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3769 qcfs_rq->h_nr_running -= task_delta;
3771 if (qcfs_rq->load.weight)
3776 sub_nr_running(rq, task_delta);
3778 cfs_rq->throttled = 1;
3779 cfs_rq->throttled_clock = rq_clock(rq);
3780 raw_spin_lock(&cfs_b->lock);
3781 empty = list_empty(&cfs_b->throttled_cfs_rq);
3784 * Add to the _head_ of the list, so that an already-started
3785 * distribute_cfs_runtime will not see us
3787 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3790 * If we're the first throttled task, make sure the bandwidth
3794 start_cfs_bandwidth(cfs_b);
3796 raw_spin_unlock(&cfs_b->lock);
3799 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3801 struct rq *rq = rq_of(cfs_rq);
3802 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3803 struct sched_entity *se;
3807 se = cfs_rq->tg->se[cpu_of(rq)];
3809 cfs_rq->throttled = 0;
3811 update_rq_clock(rq);
3813 raw_spin_lock(&cfs_b->lock);
3814 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3815 list_del_rcu(&cfs_rq->throttled_list);
3816 raw_spin_unlock(&cfs_b->lock);
3818 /* update hierarchical throttle state */
3819 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3821 if (!cfs_rq->load.weight)
3824 task_delta = cfs_rq->h_nr_running;
3825 for_each_sched_entity(se) {
3829 cfs_rq = cfs_rq_of(se);
3831 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3832 cfs_rq->h_nr_running += task_delta;
3834 if (cfs_rq_throttled(cfs_rq))
3839 add_nr_running(rq, task_delta);
3841 /* determine whether we need to wake up potentially idle cpu */
3842 if (rq->curr == rq->idle && rq->cfs.nr_running)
3846 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3847 u64 remaining, u64 expires)
3849 struct cfs_rq *cfs_rq;
3851 u64 starting_runtime = remaining;
3854 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3856 struct rq *rq = rq_of(cfs_rq);
3858 raw_spin_lock(&rq->lock);
3859 if (!cfs_rq_throttled(cfs_rq))
3862 runtime = -cfs_rq->runtime_remaining + 1;
3863 if (runtime > remaining)
3864 runtime = remaining;
3865 remaining -= runtime;
3867 cfs_rq->runtime_remaining += runtime;
3868 cfs_rq->runtime_expires = expires;
3870 /* we check whether we're throttled above */
3871 if (cfs_rq->runtime_remaining > 0)
3872 unthrottle_cfs_rq(cfs_rq);
3875 raw_spin_unlock(&rq->lock);
3882 return starting_runtime - remaining;
3886 * Responsible for refilling a task_group's bandwidth and unthrottling its
3887 * cfs_rqs as appropriate. If there has been no activity within the last
3888 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3889 * used to track this state.
3891 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3893 u64 runtime, runtime_expires;
3896 /* no need to continue the timer with no bandwidth constraint */
3897 if (cfs_b->quota == RUNTIME_INF)
3898 goto out_deactivate;
3900 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3901 cfs_b->nr_periods += overrun;
3904 * idle depends on !throttled (for the case of a large deficit), and if
3905 * we're going inactive then everything else can be deferred
3907 if (cfs_b->idle && !throttled)
3908 goto out_deactivate;
3910 __refill_cfs_bandwidth_runtime(cfs_b);
3913 /* mark as potentially idle for the upcoming period */
3918 /* account preceding periods in which throttling occurred */
3919 cfs_b->nr_throttled += overrun;
3921 runtime_expires = cfs_b->runtime_expires;
3924 * This check is repeated as we are holding onto the new bandwidth while
3925 * we unthrottle. This can potentially race with an unthrottled group
3926 * trying to acquire new bandwidth from the global pool. This can result
3927 * in us over-using our runtime if it is all used during this loop, but
3928 * only by limited amounts in that extreme case.
3930 while (throttled && cfs_b->runtime > 0) {
3931 runtime = cfs_b->runtime;
3932 raw_spin_unlock(&cfs_b->lock);
3933 /* we can't nest cfs_b->lock while distributing bandwidth */
3934 runtime = distribute_cfs_runtime(cfs_b, runtime,
3936 raw_spin_lock(&cfs_b->lock);
3938 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3940 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3944 * While we are ensured activity in the period following an
3945 * unthrottle, this also covers the case in which the new bandwidth is
3946 * insufficient to cover the existing bandwidth deficit. (Forcing the
3947 * timer to remain active while there are any throttled entities.)
3957 /* a cfs_rq won't donate quota below this amount */
3958 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3959 /* minimum remaining period time to redistribute slack quota */
3960 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3961 /* how long we wait to gather additional slack before distributing */
3962 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3965 * Are we near the end of the current quota period?
3967 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3968 * hrtimer base being cleared by hrtimer_start. In the case of
3969 * migrate_hrtimers, base is never cleared, so we are fine.
3971 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3973 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3976 /* if the call-back is running a quota refresh is already occurring */
3977 if (hrtimer_callback_running(refresh_timer))
3980 /* is a quota refresh about to occur? */
3981 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3982 if (remaining < min_expire)
3988 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3990 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3992 /* if there's a quota refresh soon don't bother with slack */
3993 if (runtime_refresh_within(cfs_b, min_left))
3996 hrtimer_start(&cfs_b->slack_timer,
3997 ns_to_ktime(cfs_bandwidth_slack_period),
4001 /* we know any runtime found here is valid as update_curr() precedes return */
4002 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4004 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4005 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4007 if (slack_runtime <= 0)
4010 raw_spin_lock(&cfs_b->lock);
4011 if (cfs_b->quota != RUNTIME_INF &&
4012 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4013 cfs_b->runtime += slack_runtime;
4015 /* we are under rq->lock, defer unthrottling using a timer */
4016 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4017 !list_empty(&cfs_b->throttled_cfs_rq))
4018 start_cfs_slack_bandwidth(cfs_b);
4020 raw_spin_unlock(&cfs_b->lock);
4022 /* even if it's not valid for return we don't want to try again */
4023 cfs_rq->runtime_remaining -= slack_runtime;
4026 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4028 if (!cfs_bandwidth_used())
4031 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4034 __return_cfs_rq_runtime(cfs_rq);
4038 * This is done with a timer (instead of inline with bandwidth return) since
4039 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4041 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4043 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4046 /* confirm we're still not at a refresh boundary */
4047 raw_spin_lock(&cfs_b->lock);
4048 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4049 raw_spin_unlock(&cfs_b->lock);
4053 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4054 runtime = cfs_b->runtime;
4056 expires = cfs_b->runtime_expires;
4057 raw_spin_unlock(&cfs_b->lock);
4062 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4064 raw_spin_lock(&cfs_b->lock);
4065 if (expires == cfs_b->runtime_expires)
4066 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4067 raw_spin_unlock(&cfs_b->lock);
4071 * When a group wakes up we want to make sure that its quota is not already
4072 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4073 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4075 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4077 if (!cfs_bandwidth_used())
4080 /* Synchronize hierarchical throttle counter: */
4081 if (unlikely(!cfs_rq->throttle_uptodate)) {
4082 struct rq *rq = rq_of(cfs_rq);
4083 struct cfs_rq *pcfs_rq;
4084 struct task_group *tg;
4086 cfs_rq->throttle_uptodate = 1;
4088 /* Get closest up-to-date node, because leaves go first: */
4089 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4090 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4091 if (pcfs_rq->throttle_uptodate)
4095 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4096 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4100 /* an active group must be handled by the update_curr()->put() path */
4101 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4104 /* ensure the group is not already throttled */
4105 if (cfs_rq_throttled(cfs_rq))
4108 /* update runtime allocation */
4109 account_cfs_rq_runtime(cfs_rq, 0);
4110 if (cfs_rq->runtime_remaining <= 0)
4111 throttle_cfs_rq(cfs_rq);
4114 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4115 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4117 if (!cfs_bandwidth_used())
4120 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4124 * it's possible for a throttled entity to be forced into a running
4125 * state (e.g. set_curr_task), in this case we're finished.
4127 if (cfs_rq_throttled(cfs_rq))
4130 throttle_cfs_rq(cfs_rq);
4134 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4136 struct cfs_bandwidth *cfs_b =
4137 container_of(timer, struct cfs_bandwidth, slack_timer);
4139 do_sched_cfs_slack_timer(cfs_b);
4141 return HRTIMER_NORESTART;
4144 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4146 struct cfs_bandwidth *cfs_b =
4147 container_of(timer, struct cfs_bandwidth, period_timer);
4151 raw_spin_lock(&cfs_b->lock);
4153 overrun = hrtimer_forward_now(timer, cfs_b->period);
4157 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4160 cfs_b->period_active = 0;
4161 raw_spin_unlock(&cfs_b->lock);
4163 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4166 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4168 raw_spin_lock_init(&cfs_b->lock);
4170 cfs_b->quota = RUNTIME_INF;
4171 cfs_b->period = ns_to_ktime(default_cfs_period());
4173 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4174 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4175 cfs_b->period_timer.function = sched_cfs_period_timer;
4176 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4177 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4180 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4182 cfs_rq->runtime_enabled = 0;
4183 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4186 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4188 lockdep_assert_held(&cfs_b->lock);
4190 if (!cfs_b->period_active) {
4191 cfs_b->period_active = 1;
4192 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4193 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4197 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4199 /* init_cfs_bandwidth() was not called */
4200 if (!cfs_b->throttled_cfs_rq.next)
4203 hrtimer_cancel(&cfs_b->period_timer);
4204 hrtimer_cancel(&cfs_b->slack_timer);
4207 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4209 struct cfs_rq *cfs_rq;
4211 for_each_leaf_cfs_rq(rq, cfs_rq) {
4212 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4214 raw_spin_lock(&cfs_b->lock);
4215 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4216 raw_spin_unlock(&cfs_b->lock);
4220 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4222 struct cfs_rq *cfs_rq;
4224 for_each_leaf_cfs_rq(rq, cfs_rq) {
4225 if (!cfs_rq->runtime_enabled)
4229 * clock_task is not advancing so we just need to make sure
4230 * there's some valid quota amount
4232 cfs_rq->runtime_remaining = 1;
4234 * Offline rq is schedulable till cpu is completely disabled
4235 * in take_cpu_down(), so we prevent new cfs throttling here.
4237 cfs_rq->runtime_enabled = 0;
4239 if (cfs_rq_throttled(cfs_rq))
4240 unthrottle_cfs_rq(cfs_rq);
4244 #else /* CONFIG_CFS_BANDWIDTH */
4245 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4247 return rq_clock_task(rq_of(cfs_rq));
4250 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4251 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4252 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4253 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4255 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4260 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4265 static inline int throttled_lb_pair(struct task_group *tg,
4266 int src_cpu, int dest_cpu)
4271 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4273 #ifdef CONFIG_FAIR_GROUP_SCHED
4274 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4277 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4281 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4282 static inline void update_runtime_enabled(struct rq *rq) {}
4283 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4285 #endif /* CONFIG_CFS_BANDWIDTH */
4287 /**************************************************
4288 * CFS operations on tasks:
4291 #ifdef CONFIG_SCHED_HRTICK
4292 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4294 struct sched_entity *se = &p->se;
4295 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4297 WARN_ON(task_rq(p) != rq);
4299 if (cfs_rq->nr_running > 1) {
4300 u64 slice = sched_slice(cfs_rq, se);
4301 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4302 s64 delta = slice - ran;
4309 hrtick_start(rq, delta);
4314 * called from enqueue/dequeue and updates the hrtick when the
4315 * current task is from our class and nr_running is low enough
4318 static void hrtick_update(struct rq *rq)
4320 struct task_struct *curr = rq->curr;
4322 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4325 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4326 hrtick_start_fair(rq, curr);
4328 #else /* !CONFIG_SCHED_HRTICK */
4330 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4334 static inline void hrtick_update(struct rq *rq)
4340 static bool cpu_overutilized(int cpu);
4341 unsigned long boosted_cpu_util(int cpu);
4343 #define boosted_cpu_util(cpu) cpu_util(cpu)
4347 static void update_capacity_of(int cpu)
4349 unsigned long req_cap;
4354 /* Convert scale-invariant capacity to cpu. */
4355 req_cap = boosted_cpu_util(cpu);
4356 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4357 set_cfs_cpu_capacity(cpu, true, req_cap);
4362 * The enqueue_task method is called before nr_running is
4363 * increased. Here we update the fair scheduling stats and
4364 * then put the task into the rbtree:
4367 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4369 struct cfs_rq *cfs_rq;
4370 struct sched_entity *se = &p->se;
4372 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4373 int task_wakeup = flags & ENQUEUE_WAKEUP;
4377 * If in_iowait is set, the code below may not trigger any cpufreq
4378 * utilization updates, so do it here explicitly with the IOWAIT flag
4382 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4384 for_each_sched_entity(se) {
4387 cfs_rq = cfs_rq_of(se);
4388 enqueue_entity(cfs_rq, se, flags);
4391 * end evaluation on encountering a throttled cfs_rq
4393 * note: in the case of encountering a throttled cfs_rq we will
4394 * post the final h_nr_running increment below.
4396 if (cfs_rq_throttled(cfs_rq))
4398 cfs_rq->h_nr_running++;
4399 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4401 flags = ENQUEUE_WAKEUP;
4404 for_each_sched_entity(se) {
4405 cfs_rq = cfs_rq_of(se);
4406 cfs_rq->h_nr_running++;
4407 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4409 if (cfs_rq_throttled(cfs_rq))
4412 update_load_avg(se, 1);
4413 update_cfs_shares(cfs_rq);
4417 add_nr_running(rq, 1);
4422 * Update SchedTune accounting.
4424 * We do it before updating the CPU capacity to ensure the
4425 * boost value of the current task is accounted for in the
4426 * selection of the OPP.
4428 * We do it also in the case where we enqueue a throttled task;
4429 * we could argue that a throttled task should not boost a CPU,
4431 * a) properly implementing CPU boosting considering throttled
4432 * tasks will increase a lot the complexity of the solution
4433 * b) it's not easy to quantify the benefits introduced by
4434 * such a more complex solution.
4435 * Thus, for the time being we go for the simple solution and boost
4436 * also for throttled RQs.
4438 schedtune_enqueue_task(p, cpu_of(rq));
4441 walt_inc_cumulative_runnable_avg(rq, p);
4442 if (!task_new && !rq->rd->overutilized &&
4443 cpu_overutilized(rq->cpu)) {
4444 rq->rd->overutilized = true;
4445 trace_sched_overutilized(true);
4449 * We want to potentially trigger a freq switch
4450 * request only for tasks that are waking up; this is
4451 * because we get here also during load balancing, but
4452 * in these cases it seems wise to trigger as single
4453 * request after load balancing is done.
4455 if (task_new || task_wakeup)
4456 update_capacity_of(cpu_of(rq));
4459 #endif /* CONFIG_SMP */
4463 static void set_next_buddy(struct sched_entity *se);
4466 * The dequeue_task method is called before nr_running is
4467 * decreased. We remove the task from the rbtree and
4468 * update the fair scheduling stats:
4470 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4472 struct cfs_rq *cfs_rq;
4473 struct sched_entity *se = &p->se;
4474 int task_sleep = flags & DEQUEUE_SLEEP;
4476 for_each_sched_entity(se) {
4477 cfs_rq = cfs_rq_of(se);
4478 dequeue_entity(cfs_rq, se, flags);
4481 * end evaluation on encountering a throttled cfs_rq
4483 * note: in the case of encountering a throttled cfs_rq we will
4484 * post the final h_nr_running decrement below.
4486 if (cfs_rq_throttled(cfs_rq))
4488 cfs_rq->h_nr_running--;
4489 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4491 /* Don't dequeue parent if it has other entities besides us */
4492 if (cfs_rq->load.weight) {
4493 /* Avoid re-evaluating load for this entity: */
4494 se = parent_entity(se);
4496 * Bias pick_next to pick a task from this cfs_rq, as
4497 * p is sleeping when it is within its sched_slice.
4499 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4503 flags |= DEQUEUE_SLEEP;
4506 for_each_sched_entity(se) {
4507 cfs_rq = cfs_rq_of(se);
4508 cfs_rq->h_nr_running--;
4509 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4511 if (cfs_rq_throttled(cfs_rq))
4514 update_load_avg(se, 1);
4515 update_cfs_shares(cfs_rq);
4519 sub_nr_running(rq, 1);
4524 * Update SchedTune accounting
4526 * We do it before updating the CPU capacity to ensure the
4527 * boost value of the current task is accounted for in the
4528 * selection of the OPP.
4530 schedtune_dequeue_task(p, cpu_of(rq));
4533 walt_dec_cumulative_runnable_avg(rq, p);
4536 * We want to potentially trigger a freq switch
4537 * request only for tasks that are going to sleep;
4538 * this is because we get here also during load
4539 * balancing, but in these cases it seems wise to
4540 * trigger as single request after load balancing is
4544 if (rq->cfs.nr_running)
4545 update_capacity_of(cpu_of(rq));
4546 else if (sched_freq())
4547 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4551 #endif /* CONFIG_SMP */
4559 * per rq 'load' arrray crap; XXX kill this.
4563 * The exact cpuload at various idx values, calculated at every tick would be
4564 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4566 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4567 * on nth tick when cpu may be busy, then we have:
4568 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4569 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4571 * decay_load_missed() below does efficient calculation of
4572 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4573 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4575 * The calculation is approximated on a 128 point scale.
4576 * degrade_zero_ticks is the number of ticks after which load at any
4577 * particular idx is approximated to be zero.
4578 * degrade_factor is a precomputed table, a row for each load idx.
4579 * Each column corresponds to degradation factor for a power of two ticks,
4580 * based on 128 point scale.
4582 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4583 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4585 * With this power of 2 load factors, we can degrade the load n times
4586 * by looking at 1 bits in n and doing as many mult/shift instead of
4587 * n mult/shifts needed by the exact degradation.
4589 #define DEGRADE_SHIFT 7
4590 static const unsigned char
4591 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4592 static const unsigned char
4593 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4594 {0, 0, 0, 0, 0, 0, 0, 0},
4595 {64, 32, 8, 0, 0, 0, 0, 0},
4596 {96, 72, 40, 12, 1, 0, 0},
4597 {112, 98, 75, 43, 15, 1, 0},
4598 {120, 112, 98, 76, 45, 16, 2} };
4601 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4602 * would be when CPU is idle and so we just decay the old load without
4603 * adding any new load.
4605 static unsigned long
4606 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4610 if (!missed_updates)
4613 if (missed_updates >= degrade_zero_ticks[idx])
4617 return load >> missed_updates;
4619 while (missed_updates) {
4620 if (missed_updates % 2)
4621 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4623 missed_updates >>= 1;
4630 * Update rq->cpu_load[] statistics. This function is usually called every
4631 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4632 * every tick. We fix it up based on jiffies.
4634 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4635 unsigned long pending_updates)
4639 this_rq->nr_load_updates++;
4641 /* Update our load: */
4642 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4643 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4644 unsigned long old_load, new_load;
4646 /* scale is effectively 1 << i now, and >> i divides by scale */
4648 old_load = this_rq->cpu_load[i];
4649 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4650 new_load = this_load;
4652 * Round up the averaging division if load is increasing. This
4653 * prevents us from getting stuck on 9 if the load is 10, for
4656 if (new_load > old_load)
4657 new_load += scale - 1;
4659 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4662 sched_avg_update(this_rq);
4665 /* Used instead of source_load when we know the type == 0 */
4666 static unsigned long weighted_cpuload(const int cpu)
4668 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4671 #ifdef CONFIG_NO_HZ_COMMON
4673 * There is no sane way to deal with nohz on smp when using jiffies because the
4674 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4675 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4677 * Therefore we cannot use the delta approach from the regular tick since that
4678 * would seriously skew the load calculation. However we'll make do for those
4679 * updates happening while idle (nohz_idle_balance) or coming out of idle
4680 * (tick_nohz_idle_exit).
4682 * This means we might still be one tick off for nohz periods.
4686 * Called from nohz_idle_balance() to update the load ratings before doing the
4689 static void update_idle_cpu_load(struct rq *this_rq)
4691 unsigned long curr_jiffies = READ_ONCE(jiffies);
4692 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4693 unsigned long pending_updates;
4696 * bail if there's load or we're actually up-to-date.
4698 if (load || curr_jiffies == this_rq->last_load_update_tick)
4701 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4702 this_rq->last_load_update_tick = curr_jiffies;
4704 __update_cpu_load(this_rq, load, pending_updates);
4708 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4710 void update_cpu_load_nohz(void)
4712 struct rq *this_rq = this_rq();
4713 unsigned long curr_jiffies = READ_ONCE(jiffies);
4714 unsigned long pending_updates;
4716 if (curr_jiffies == this_rq->last_load_update_tick)
4719 raw_spin_lock(&this_rq->lock);
4720 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4721 if (pending_updates) {
4722 this_rq->last_load_update_tick = curr_jiffies;
4724 * We were idle, this means load 0, the current load might be
4725 * !0 due to remote wakeups and the sort.
4727 __update_cpu_load(this_rq, 0, pending_updates);
4729 raw_spin_unlock(&this_rq->lock);
4731 #endif /* CONFIG_NO_HZ */
4734 * Called from scheduler_tick()
4736 void update_cpu_load_active(struct rq *this_rq)
4738 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4740 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4742 this_rq->last_load_update_tick = jiffies;
4743 __update_cpu_load(this_rq, load, 1);
4747 * Return a low guess at the load of a migration-source cpu weighted
4748 * according to the scheduling class and "nice" value.
4750 * We want to under-estimate the load of migration sources, to
4751 * balance conservatively.
4753 static unsigned long source_load(int cpu, int type)
4755 struct rq *rq = cpu_rq(cpu);
4756 unsigned long total = weighted_cpuload(cpu);
4758 if (type == 0 || !sched_feat(LB_BIAS))
4761 return min(rq->cpu_load[type-1], total);
4765 * Return a high guess at the load of a migration-target cpu weighted
4766 * according to the scheduling class and "nice" value.
4768 static unsigned long target_load(int cpu, int type)
4770 struct rq *rq = cpu_rq(cpu);
4771 unsigned long total = weighted_cpuload(cpu);
4773 if (type == 0 || !sched_feat(LB_BIAS))
4776 return max(rq->cpu_load[type-1], total);
4780 static unsigned long cpu_avg_load_per_task(int cpu)
4782 struct rq *rq = cpu_rq(cpu);
4783 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4784 unsigned long load_avg = weighted_cpuload(cpu);
4787 return load_avg / nr_running;
4792 static void record_wakee(struct task_struct *p)
4795 * Rough decay (wiping) for cost saving, don't worry
4796 * about the boundary, really active task won't care
4799 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4800 current->wakee_flips >>= 1;
4801 current->wakee_flip_decay_ts = jiffies;
4804 if (current->last_wakee != p) {
4805 current->last_wakee = p;
4806 current->wakee_flips++;
4810 static void task_waking_fair(struct task_struct *p)
4812 struct sched_entity *se = &p->se;
4813 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4816 #ifndef CONFIG_64BIT
4817 u64 min_vruntime_copy;
4820 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4822 min_vruntime = cfs_rq->min_vruntime;
4823 } while (min_vruntime != min_vruntime_copy);
4825 min_vruntime = cfs_rq->min_vruntime;
4828 se->vruntime -= min_vruntime;
4832 #ifdef CONFIG_FAIR_GROUP_SCHED
4834 * effective_load() calculates the load change as seen from the root_task_group
4836 * Adding load to a group doesn't make a group heavier, but can cause movement
4837 * of group shares between cpus. Assuming the shares were perfectly aligned one
4838 * can calculate the shift in shares.
4840 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4841 * on this @cpu and results in a total addition (subtraction) of @wg to the
4842 * total group weight.
4844 * Given a runqueue weight distribution (rw_i) we can compute a shares
4845 * distribution (s_i) using:
4847 * s_i = rw_i / \Sum rw_j (1)
4849 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4850 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4851 * shares distribution (s_i):
4853 * rw_i = { 2, 4, 1, 0 }
4854 * s_i = { 2/7, 4/7, 1/7, 0 }
4856 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4857 * task used to run on and the CPU the waker is running on), we need to
4858 * compute the effect of waking a task on either CPU and, in case of a sync
4859 * wakeup, compute the effect of the current task going to sleep.
4861 * So for a change of @wl to the local @cpu with an overall group weight change
4862 * of @wl we can compute the new shares distribution (s'_i) using:
4864 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4866 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4867 * differences in waking a task to CPU 0. The additional task changes the
4868 * weight and shares distributions like:
4870 * rw'_i = { 3, 4, 1, 0 }
4871 * s'_i = { 3/8, 4/8, 1/8, 0 }
4873 * We can then compute the difference in effective weight by using:
4875 * dw_i = S * (s'_i - s_i) (3)
4877 * Where 'S' is the group weight as seen by its parent.
4879 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4880 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4881 * 4/7) times the weight of the group.
4883 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4885 struct sched_entity *se = tg->se[cpu];
4887 if (!tg->parent) /* the trivial, non-cgroup case */
4890 for_each_sched_entity(se) {
4891 struct cfs_rq *cfs_rq = se->my_q;
4892 long W, w = cfs_rq_load_avg(cfs_rq);
4897 * W = @wg + \Sum rw_j
4899 W = wg + atomic_long_read(&tg->load_avg);
4901 /* Ensure \Sum rw_j >= rw_i */
4902 W -= cfs_rq->tg_load_avg_contrib;
4911 * wl = S * s'_i; see (2)
4914 wl = (w * (long)tg->shares) / W;
4919 * Per the above, wl is the new se->load.weight value; since
4920 * those are clipped to [MIN_SHARES, ...) do so now. See
4921 * calc_cfs_shares().
4923 if (wl < MIN_SHARES)
4927 * wl = dw_i = S * (s'_i - s_i); see (3)
4929 wl -= se->avg.load_avg;
4932 * Recursively apply this logic to all parent groups to compute
4933 * the final effective load change on the root group. Since
4934 * only the @tg group gets extra weight, all parent groups can
4935 * only redistribute existing shares. @wl is the shift in shares
4936 * resulting from this level per the above.
4945 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4953 * Returns the current capacity of cpu after applying both
4954 * cpu and freq scaling.
4956 unsigned long capacity_curr_of(int cpu)
4958 return cpu_rq(cpu)->cpu_capacity_orig *
4959 arch_scale_freq_capacity(NULL, cpu)
4960 >> SCHED_CAPACITY_SHIFT;
4963 static inline bool energy_aware(void)
4965 return sched_feat(ENERGY_AWARE);
4969 struct sched_group *sg_top;
4970 struct sched_group *sg_cap;
4977 struct task_struct *task;
4992 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4993 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4994 * energy calculations. Using the scale-invariant util returned by
4995 * cpu_util() and approximating scale-invariant util by:
4997 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4999 * the normalized util can be found using the specific capacity.
5001 * capacity = capacity_orig * curr_freq/max_freq
5003 * norm_util = running_time/time ~ util/capacity
5005 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5007 int util = __cpu_util(cpu, delta);
5009 if (util >= capacity)
5010 return SCHED_CAPACITY_SCALE;
5012 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5015 static int calc_util_delta(struct energy_env *eenv, int cpu)
5017 if (cpu == eenv->src_cpu)
5018 return -eenv->util_delta;
5019 if (cpu == eenv->dst_cpu)
5020 return eenv->util_delta;
5025 unsigned long group_max_util(struct energy_env *eenv)
5028 unsigned long max_util = 0;
5030 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5031 delta = calc_util_delta(eenv, i);
5032 max_util = max(max_util, __cpu_util(i, delta));
5039 * group_norm_util() returns the approximated group util relative to it's
5040 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5041 * energy calculations. Since task executions may or may not overlap in time in
5042 * the group the true normalized util is between max(cpu_norm_util(i)) and
5043 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5044 * latter is used as the estimate as it leads to a more pessimistic energy
5045 * estimate (more busy).
5048 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5051 unsigned long util_sum = 0;
5052 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5054 for_each_cpu(i, sched_group_cpus(sg)) {
5055 delta = calc_util_delta(eenv, i);
5056 util_sum += __cpu_norm_util(i, capacity, delta);
5059 if (util_sum > SCHED_CAPACITY_SCALE)
5060 return SCHED_CAPACITY_SCALE;
5064 static int find_new_capacity(struct energy_env *eenv,
5065 const struct sched_group_energy * const sge)
5068 unsigned long util = group_max_util(eenv);
5070 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5071 if (sge->cap_states[idx].cap >= util)
5075 eenv->cap_idx = idx;
5080 static int group_idle_state(struct sched_group *sg)
5082 int i, state = INT_MAX;
5084 /* Find the shallowest idle state in the sched group. */
5085 for_each_cpu(i, sched_group_cpus(sg))
5086 state = min(state, idle_get_state_idx(cpu_rq(i)));
5088 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5095 * sched_group_energy(): Computes the absolute energy consumption of cpus
5096 * belonging to the sched_group including shared resources shared only by
5097 * members of the group. Iterates over all cpus in the hierarchy below the
5098 * sched_group starting from the bottom working it's way up before going to
5099 * the next cpu until all cpus are covered at all levels. The current
5100 * implementation is likely to gather the same util statistics multiple times.
5101 * This can probably be done in a faster but more complex way.
5102 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5104 static int sched_group_energy(struct energy_env *eenv)
5106 struct sched_domain *sd;
5107 int cpu, total_energy = 0;
5108 struct cpumask visit_cpus;
5109 struct sched_group *sg;
5111 WARN_ON(!eenv->sg_top->sge);
5113 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5115 while (!cpumask_empty(&visit_cpus)) {
5116 struct sched_group *sg_shared_cap = NULL;
5118 cpu = cpumask_first(&visit_cpus);
5121 * Is the group utilization affected by cpus outside this
5124 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5128 * We most probably raced with hotplug; returning a
5129 * wrong energy estimation is better than entering an
5135 sg_shared_cap = sd->parent->groups;
5137 for_each_domain(cpu, sd) {
5140 /* Has this sched_domain already been visited? */
5141 if (sd->child && group_first_cpu(sg) != cpu)
5145 unsigned long group_util;
5146 int sg_busy_energy, sg_idle_energy;
5147 int cap_idx, idle_idx;
5149 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5150 eenv->sg_cap = sg_shared_cap;
5154 cap_idx = find_new_capacity(eenv, sg->sge);
5156 if (sg->group_weight == 1) {
5157 /* Remove capacity of src CPU (before task move) */
5158 if (eenv->util_delta == 0 &&
5159 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5160 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5161 eenv->cap.delta -= eenv->cap.before;
5163 /* Add capacity of dst CPU (after task move) */
5164 if (eenv->util_delta != 0 &&
5165 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5166 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5167 eenv->cap.delta += eenv->cap.after;
5171 idle_idx = group_idle_state(sg);
5172 group_util = group_norm_util(eenv, sg);
5173 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5174 >> SCHED_CAPACITY_SHIFT;
5175 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5176 * sg->sge->idle_states[idle_idx].power)
5177 >> SCHED_CAPACITY_SHIFT;
5179 total_energy += sg_busy_energy + sg_idle_energy;
5182 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5184 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5187 } while (sg = sg->next, sg != sd->groups);
5190 cpumask_clear_cpu(cpu, &visit_cpus);
5194 eenv->energy = total_energy;
5198 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5200 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5204 * energy_diff(): Estimate the energy impact of changing the utilization
5205 * distribution. eenv specifies the change: utilisation amount, source, and
5206 * destination cpu. Source or destination cpu may be -1 in which case the
5207 * utilization is removed from or added to the system (e.g. task wake-up). If
5208 * both are specified, the utilization is migrated.
5210 static inline int __energy_diff(struct energy_env *eenv)
5212 struct sched_domain *sd;
5213 struct sched_group *sg;
5214 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5217 struct energy_env eenv_before = {
5219 .src_cpu = eenv->src_cpu,
5220 .dst_cpu = eenv->dst_cpu,
5221 .nrg = { 0, 0, 0, 0},
5225 if (eenv->src_cpu == eenv->dst_cpu)
5228 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5229 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5232 return 0; /* Error */
5237 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5238 eenv_before.sg_top = eenv->sg_top = sg;
5240 if (sched_group_energy(&eenv_before))
5241 return 0; /* Invalid result abort */
5242 energy_before += eenv_before.energy;
5244 /* Keep track of SRC cpu (before) capacity */
5245 eenv->cap.before = eenv_before.cap.before;
5246 eenv->cap.delta = eenv_before.cap.delta;
5248 if (sched_group_energy(eenv))
5249 return 0; /* Invalid result abort */
5250 energy_after += eenv->energy;
5252 } while (sg = sg->next, sg != sd->groups);
5254 eenv->nrg.before = energy_before;
5255 eenv->nrg.after = energy_after;
5256 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5259 trace_sched_energy_diff(eenv->task,
5260 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5261 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5262 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5263 eenv->nrg.delta, eenv->payoff);
5266 * Dead-zone margin preventing too many migrations.
5269 margin = eenv->nrg.before >> 6; /* ~1.56% */
5271 diff = eenv->nrg.after - eenv->nrg.before;
5273 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5275 return eenv->nrg.diff;
5278 #ifdef CONFIG_SCHED_TUNE
5280 struct target_nrg schedtune_target_nrg;
5283 * System energy normalization
5284 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5285 * corresponding to the specified energy variation.
5288 normalize_energy(int energy_diff)
5291 #ifdef CONFIG_SCHED_DEBUG
5294 /* Check for boundaries */
5295 max_delta = schedtune_target_nrg.max_power;
5296 max_delta -= schedtune_target_nrg.min_power;
5297 WARN_ON(abs(energy_diff) >= max_delta);
5300 /* Do scaling using positive numbers to increase the range */
5301 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5303 /* Scale by energy magnitude */
5304 normalized_nrg <<= SCHED_LOAD_SHIFT;
5306 /* Normalize on max energy for target platform */
5307 normalized_nrg = reciprocal_divide(
5308 normalized_nrg, schedtune_target_nrg.rdiv);
5310 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5314 energy_diff(struct energy_env *eenv)
5316 int boost = schedtune_task_boost(eenv->task);
5319 /* Conpute "absolute" energy diff */
5320 __energy_diff(eenv);
5322 /* Return energy diff when boost margin is 0 */
5324 return eenv->nrg.diff;
5326 /* Compute normalized energy diff */
5327 nrg_delta = normalize_energy(eenv->nrg.diff);
5328 eenv->nrg.delta = nrg_delta;
5330 eenv->payoff = schedtune_accept_deltas(
5336 * When SchedTune is enabled, the energy_diff() function will return
5337 * the computed energy payoff value. Since the energy_diff() return
5338 * value is expected to be negative by its callers, this evaluation
5339 * function return a negative value each time the evaluation return a
5340 * positive payoff, which is the condition for the acceptance of
5341 * a scheduling decision
5343 return -eenv->payoff;
5345 #else /* CONFIG_SCHED_TUNE */
5346 #define energy_diff(eenv) __energy_diff(eenv)
5350 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5351 * A waker of many should wake a different task than the one last awakened
5352 * at a frequency roughly N times higher than one of its wakees. In order
5353 * to determine whether we should let the load spread vs consolodating to
5354 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5355 * partner, and a factor of lls_size higher frequency in the other. With
5356 * both conditions met, we can be relatively sure that the relationship is
5357 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5358 * being client/server, worker/dispatcher, interrupt source or whatever is
5359 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5361 static int wake_wide(struct task_struct *p)
5363 unsigned int master = current->wakee_flips;
5364 unsigned int slave = p->wakee_flips;
5365 int factor = this_cpu_read(sd_llc_size);
5368 swap(master, slave);
5369 if (slave < factor || master < slave * factor)
5374 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5375 int prev_cpu, int sync)
5377 s64 this_load, load;
5378 s64 this_eff_load, prev_eff_load;
5380 struct task_group *tg;
5381 unsigned long weight;
5385 this_cpu = smp_processor_id();
5386 load = source_load(prev_cpu, idx);
5387 this_load = target_load(this_cpu, idx);
5390 * If sync wakeup then subtract the (maximum possible)
5391 * effect of the currently running task from the load
5392 * of the current CPU:
5395 tg = task_group(current);
5396 weight = current->se.avg.load_avg;
5398 this_load += effective_load(tg, this_cpu, -weight, -weight);
5399 load += effective_load(tg, prev_cpu, 0, -weight);
5403 weight = p->se.avg.load_avg;
5406 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5407 * due to the sync cause above having dropped this_load to 0, we'll
5408 * always have an imbalance, but there's really nothing you can do
5409 * about that, so that's good too.
5411 * Otherwise check if either cpus are near enough in load to allow this
5412 * task to be woken on this_cpu.
5414 this_eff_load = 100;
5415 this_eff_load *= capacity_of(prev_cpu);
5417 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5418 prev_eff_load *= capacity_of(this_cpu);
5420 if (this_load > 0) {
5421 this_eff_load *= this_load +
5422 effective_load(tg, this_cpu, weight, weight);
5424 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5427 balanced = this_eff_load <= prev_eff_load;
5429 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5434 schedstat_inc(sd, ttwu_move_affine);
5435 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5440 static inline unsigned long task_util(struct task_struct *p)
5442 #ifdef CONFIG_SCHED_WALT
5443 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5444 unsigned long demand = p->ravg.demand;
5445 return (demand << 10) / walt_ravg_window;
5448 return p->se.avg.util_avg;
5451 static inline unsigned long boosted_task_util(struct task_struct *task);
5453 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5455 unsigned long capacity = capacity_of(cpu);
5457 util += boosted_task_util(p);
5459 return (capacity * 1024) > (util * capacity_margin);
5462 static inline bool task_fits_max(struct task_struct *p, int cpu)
5464 unsigned long capacity = capacity_of(cpu);
5465 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5467 if (capacity == max_capacity)
5470 if (capacity * capacity_margin > max_capacity * 1024)
5473 return __task_fits(p, cpu, 0);
5476 static bool cpu_overutilized(int cpu)
5478 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5481 #ifdef CONFIG_SCHED_TUNE
5484 schedtune_margin(unsigned long signal, long boost)
5486 long long margin = 0;
5489 * Signal proportional compensation (SPC)
5491 * The Boost (B) value is used to compute a Margin (M) which is
5492 * proportional to the complement of the original Signal (S):
5493 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5494 * M = B * S, if B is negative
5495 * The obtained M could be used by the caller to "boost" S.
5498 margin = SCHED_LOAD_SCALE - signal;
5501 margin = -signal * boost;
5503 * Fast integer division by constant:
5504 * Constant : (C) = 100
5505 * Precision : 0.1% (P) = 0.1
5506 * Reference : C * 100 / P (R) = 100000
5509 * Shift bits : ceil(log(R,2)) (S) = 17
5510 * Mult const : round(2^S/C) (M) = 1311
5523 schedtune_cpu_margin(unsigned long util, int cpu)
5525 int boost = schedtune_cpu_boost(cpu);
5530 return schedtune_margin(util, boost);
5534 schedtune_task_margin(struct task_struct *task)
5536 int boost = schedtune_task_boost(task);
5543 util = task_util(task);
5544 margin = schedtune_margin(util, boost);
5549 #else /* CONFIG_SCHED_TUNE */
5552 schedtune_cpu_margin(unsigned long util, int cpu)
5558 schedtune_task_margin(struct task_struct *task)
5563 #endif /* CONFIG_SCHED_TUNE */
5566 boosted_cpu_util(int cpu)
5568 unsigned long util = cpu_util(cpu);
5569 long margin = schedtune_cpu_margin(util, cpu);
5571 trace_sched_boost_cpu(cpu, util, margin);
5573 return util + margin;
5576 static inline unsigned long
5577 boosted_task_util(struct task_struct *task)
5579 unsigned long util = task_util(task);
5580 long margin = schedtune_task_margin(task);
5582 trace_sched_boost_task(task, util, margin);
5584 return util + margin;
5587 static int cpu_util_wake(int cpu, struct task_struct *p);
5589 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5591 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5595 * find_idlest_group finds and returns the least busy CPU group within the
5598 static struct sched_group *
5599 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5600 int this_cpu, int sd_flag)
5602 struct sched_group *idlest = NULL, *group = sd->groups;
5603 struct sched_group *most_spare_sg = NULL;
5604 unsigned long min_load = ULONG_MAX, this_load = 0;
5605 unsigned long most_spare = 0, this_spare = 0;
5606 int load_idx = sd->forkexec_idx;
5607 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5609 if (sd_flag & SD_BALANCE_WAKE)
5610 load_idx = sd->wake_idx;
5613 unsigned long load, avg_load, spare_cap, max_spare_cap;
5617 /* Skip over this group if it has no CPUs allowed */
5618 if (!cpumask_intersects(sched_group_cpus(group),
5619 tsk_cpus_allowed(p)))
5622 local_group = cpumask_test_cpu(this_cpu,
5623 sched_group_cpus(group));
5626 * Tally up the load of all CPUs in the group and find
5627 * the group containing the CPU with most spare capacity.
5632 for_each_cpu(i, sched_group_cpus(group)) {
5633 /* Bias balancing toward cpus of our domain */
5635 load = source_load(i, load_idx);
5637 load = target_load(i, load_idx);
5641 spare_cap = capacity_spare_wake(i, p);
5643 if (spare_cap > max_spare_cap)
5644 max_spare_cap = spare_cap;
5647 /* Adjust by relative CPU capacity of the group */
5648 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5651 this_load = avg_load;
5652 this_spare = max_spare_cap;
5654 if (avg_load < min_load) {
5655 min_load = avg_load;
5659 if (most_spare < max_spare_cap) {
5660 most_spare = max_spare_cap;
5661 most_spare_sg = group;
5664 } while (group = group->next, group != sd->groups);
5667 * The cross-over point between using spare capacity or least load
5668 * is too conservative for high utilization tasks on partially
5669 * utilized systems if we require spare_capacity > task_util(p),
5670 * so we allow for some task stuffing by using
5671 * spare_capacity > task_util(p)/2.
5673 if (this_spare > task_util(p) / 2 &&
5674 imbalance*this_spare > 100*most_spare)
5676 else if (most_spare > task_util(p) / 2)
5677 return most_spare_sg;
5679 if (!idlest || 100*this_load < imbalance*min_load)
5685 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5688 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5690 unsigned long load, min_load = ULONG_MAX;
5691 unsigned int min_exit_latency = UINT_MAX;
5692 u64 latest_idle_timestamp = 0;
5693 int least_loaded_cpu = this_cpu;
5694 int shallowest_idle_cpu = -1;
5697 /* Check if we have any choice: */
5698 if (group->group_weight == 1)
5699 return cpumask_first(sched_group_cpus(group));
5701 /* Traverse only the allowed CPUs */
5702 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5704 struct rq *rq = cpu_rq(i);
5705 struct cpuidle_state *idle = idle_get_state(rq);
5706 if (idle && idle->exit_latency < min_exit_latency) {
5708 * We give priority to a CPU whose idle state
5709 * has the smallest exit latency irrespective
5710 * of any idle timestamp.
5712 min_exit_latency = idle->exit_latency;
5713 latest_idle_timestamp = rq->idle_stamp;
5714 shallowest_idle_cpu = i;
5715 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5716 rq->idle_stamp > latest_idle_timestamp) {
5718 * If equal or no active idle state, then
5719 * the most recently idled CPU might have
5722 latest_idle_timestamp = rq->idle_stamp;
5723 shallowest_idle_cpu = i;
5725 } else if (shallowest_idle_cpu == -1) {
5726 load = weighted_cpuload(i);
5727 if (load < min_load || (load == min_load && i == this_cpu)) {
5729 least_loaded_cpu = i;
5734 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5738 * Try and locate an idle CPU in the sched_domain.
5740 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5742 struct sched_domain *sd;
5743 struct sched_group *sg;
5744 int best_idle_cpu = -1;
5745 int best_idle_cstate = INT_MAX;
5746 unsigned long best_idle_capacity = ULONG_MAX;
5748 if (!sysctl_sched_cstate_aware) {
5749 if (idle_cpu(target))
5753 * If the prevous cpu is cache affine and idle, don't be stupid.
5755 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5760 * Otherwise, iterate the domains and find an elegible idle cpu.
5762 sd = rcu_dereference(per_cpu(sd_llc, target));
5763 for_each_lower_domain(sd) {
5767 if (!cpumask_intersects(sched_group_cpus(sg),
5768 tsk_cpus_allowed(p)))
5771 if (sysctl_sched_cstate_aware) {
5772 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5773 int idle_idx = idle_get_state_idx(cpu_rq(i));
5774 unsigned long new_usage = boosted_task_util(p);
5775 unsigned long capacity_orig = capacity_orig_of(i);
5777 if (new_usage > capacity_orig || !idle_cpu(i))
5780 if (i == target && new_usage <= capacity_curr_of(target))
5783 if (idle_idx < best_idle_cstate &&
5784 capacity_orig <= best_idle_capacity) {
5786 best_idle_cstate = idle_idx;
5787 best_idle_capacity = capacity_orig;
5791 for_each_cpu(i, sched_group_cpus(sg)) {
5792 if (i == target || !idle_cpu(i))
5796 target = cpumask_first_and(sched_group_cpus(sg),
5797 tsk_cpus_allowed(p));
5802 } while (sg != sd->groups);
5805 if (best_idle_cpu >= 0)
5806 target = best_idle_cpu;
5812 static int start_cpu(bool boosted)
5814 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5816 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5817 "sched RCU must be held");
5819 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5822 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5824 int target_cpu = -1;
5825 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5826 unsigned long backup_capacity = ULONG_MAX;
5827 int best_idle_cpu = -1;
5828 int best_idle_cstate = INT_MAX;
5829 int backup_cpu = -1;
5830 unsigned long min_util = boosted_task_util(p);
5831 struct sched_domain *sd;
5832 struct sched_group *sg;
5833 int cpu = start_cpu(boosted);
5838 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5848 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5849 unsigned long cur_capacity, new_util;
5855 * p's blocked utilization is still accounted for on prev_cpu
5856 * so prev_cpu will receive a negative bias due to the double
5857 * accounting. However, the blocked utilization may be zero.
5859 new_util = cpu_util(i) + task_util(p);
5862 * Ensure minimum capacity to grant the required boost.
5863 * The target CPU can be already at a capacity level higher
5864 * than the one required to boost the task.
5866 new_util = max(min_util, new_util);
5868 if (new_util > capacity_orig_of(i))
5871 #ifdef CONFIG_SCHED_WALT
5872 if (walt_cpu_high_irqload(i))
5877 * Unconditionally favoring tasks that prefer idle cpus to
5880 if (idle_cpu(i) && prefer_idle)
5883 cur_capacity = capacity_curr_of(i);
5885 if (new_util < cur_capacity) {
5886 if (cpu_rq(i)->nr_running) {
5888 * Find a target cpu with the lowest/highest
5889 * utilization if prefer_idle/!prefer_idle.
5891 if ((prefer_idle && target_util > new_util) ||
5892 (!prefer_idle && target_util < new_util)) {
5893 target_util = new_util;
5896 } else if (!prefer_idle) {
5897 int idle_idx = idle_get_state_idx(cpu_rq(i));
5899 if (best_idle_cpu < 0 ||
5900 (sysctl_sched_cstate_aware &&
5901 best_idle_cstate > idle_idx)) {
5902 best_idle_cstate = idle_idx;
5906 } else if (backup_capacity > cur_capacity) {
5907 /* Find a backup cpu with least capacity. */
5908 backup_capacity = cur_capacity;
5912 } while (sg = sg->next, sg != sd->groups);
5915 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5921 * cpu_util_wake: Compute cpu utilization with any contributions from
5922 * the waking task p removed.
5924 static int cpu_util_wake(int cpu, struct task_struct *p)
5926 unsigned long util, capacity;
5928 /* Task has no contribution or is new */
5929 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5930 return cpu_util(cpu);
5932 capacity = capacity_orig_of(cpu);
5933 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5935 return (util >= capacity) ? capacity : util;
5939 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5940 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5942 * In that case WAKE_AFFINE doesn't make sense and we'll let
5943 * BALANCE_WAKE sort things out.
5945 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5947 long min_cap, max_cap;
5949 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5950 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5952 /* Minimum capacity is close to max, no need to abort wake_affine */
5953 if (max_cap - min_cap < max_cap >> 3)
5956 /* Bring task utilization in sync with prev_cpu */
5957 sync_entity_load_avg(&p->se);
5959 return min_cap * 1024 < task_util(p) * capacity_margin;
5962 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
5964 struct sched_domain *sd;
5965 int target_cpu = prev_cpu, tmp_target;
5966 bool boosted, prefer_idle;
5968 if (sysctl_sched_sync_hint_enable && sync) {
5969 int cpu = smp_processor_id();
5971 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5976 #ifdef CONFIG_CGROUP_SCHEDTUNE
5977 boosted = schedtune_task_boost(p) > 0;
5978 prefer_idle = schedtune_prefer_idle(p) > 0;
5980 boosted = get_sysctl_sched_cfs_boost() > 0;
5984 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
5985 /* Find a cpu with sufficient capacity */
5986 tmp_target = find_best_target(p, boosted, prefer_idle);
5990 if (tmp_target >= 0) {
5991 target_cpu = tmp_target;
5992 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5996 if (target_cpu != prev_cpu) {
5997 struct energy_env eenv = {
5998 .util_delta = task_util(p),
5999 .src_cpu = prev_cpu,
6000 .dst_cpu = target_cpu,
6004 /* Not enough spare capacity on previous cpu */
6005 if (cpu_overutilized(prev_cpu))
6008 if (energy_diff(&eenv) >= 0)
6009 target_cpu = prev_cpu;
6018 * select_task_rq_fair: Select target runqueue for the waking task in domains
6019 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6020 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6022 * Balances load by selecting the idlest cpu in the idlest group, or under
6023 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6025 * Returns the target cpu number.
6027 * preempt must be disabled.
6030 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6032 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6033 int cpu = smp_processor_id();
6034 int new_cpu = prev_cpu;
6035 int want_affine = 0;
6036 int sync = wake_flags & WF_SYNC;
6038 if (sd_flag & SD_BALANCE_WAKE)
6039 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6040 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6042 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6043 return select_energy_cpu_brute(p, prev_cpu, sync);
6046 for_each_domain(cpu, tmp) {
6047 if (!(tmp->flags & SD_LOAD_BALANCE))
6051 * If both cpu and prev_cpu are part of this domain,
6052 * cpu is a valid SD_WAKE_AFFINE target.
6054 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6055 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6060 if (tmp->flags & sd_flag)
6062 else if (!want_affine)
6067 sd = NULL; /* Prefer wake_affine over balance flags */
6068 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6073 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6074 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6077 struct sched_group *group;
6080 if (!(sd->flags & sd_flag)) {
6085 group = find_idlest_group(sd, p, cpu, sd_flag);
6091 new_cpu = find_idlest_cpu(group, p, cpu);
6092 if (new_cpu == -1 || new_cpu == cpu) {
6093 /* Now try balancing at a lower domain level of cpu */
6098 /* Now try balancing at a lower domain level of new_cpu */
6100 weight = sd->span_weight;
6102 for_each_domain(cpu, tmp) {
6103 if (weight <= tmp->span_weight)
6105 if (tmp->flags & sd_flag)
6108 /* while loop will break here if sd == NULL */
6116 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6117 * cfs_rq_of(p) references at time of call are still valid and identify the
6118 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6119 * other assumptions, including the state of rq->lock, should be made.
6121 static void migrate_task_rq_fair(struct task_struct *p)
6124 * We are supposed to update the task to "current" time, then its up to date
6125 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6126 * what current time is, so simply throw away the out-of-date time. This
6127 * will result in the wakee task is less decayed, but giving the wakee more
6128 * load sounds not bad.
6130 remove_entity_load_avg(&p->se);
6132 /* Tell new CPU we are migrated */
6133 p->se.avg.last_update_time = 0;
6135 /* We have migrated, no longer consider this task hot */
6136 p->se.exec_start = 0;
6139 static void task_dead_fair(struct task_struct *p)
6141 remove_entity_load_avg(&p->se);
6144 #define task_fits_max(p, cpu) true
6145 #endif /* CONFIG_SMP */
6147 static unsigned long
6148 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6150 unsigned long gran = sysctl_sched_wakeup_granularity;
6153 * Since its curr running now, convert the gran from real-time
6154 * to virtual-time in his units.
6156 * By using 'se' instead of 'curr' we penalize light tasks, so
6157 * they get preempted easier. That is, if 'se' < 'curr' then
6158 * the resulting gran will be larger, therefore penalizing the
6159 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6160 * be smaller, again penalizing the lighter task.
6162 * This is especially important for buddies when the leftmost
6163 * task is higher priority than the buddy.
6165 return calc_delta_fair(gran, se);
6169 * Should 'se' preempt 'curr'.
6183 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6185 s64 gran, vdiff = curr->vruntime - se->vruntime;
6190 gran = wakeup_gran(curr, se);
6197 static void set_last_buddy(struct sched_entity *se)
6199 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6202 for_each_sched_entity(se)
6203 cfs_rq_of(se)->last = se;
6206 static void set_next_buddy(struct sched_entity *se)
6208 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6211 for_each_sched_entity(se)
6212 cfs_rq_of(se)->next = se;
6215 static void set_skip_buddy(struct sched_entity *se)
6217 for_each_sched_entity(se)
6218 cfs_rq_of(se)->skip = se;
6222 * Preempt the current task with a newly woken task if needed:
6224 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6226 struct task_struct *curr = rq->curr;
6227 struct sched_entity *se = &curr->se, *pse = &p->se;
6228 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6229 int scale = cfs_rq->nr_running >= sched_nr_latency;
6230 int next_buddy_marked = 0;
6232 if (unlikely(se == pse))
6236 * This is possible from callers such as attach_tasks(), in which we
6237 * unconditionally check_prempt_curr() after an enqueue (which may have
6238 * lead to a throttle). This both saves work and prevents false
6239 * next-buddy nomination below.
6241 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6244 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6245 set_next_buddy(pse);
6246 next_buddy_marked = 1;
6250 * We can come here with TIF_NEED_RESCHED already set from new task
6253 * Note: this also catches the edge-case of curr being in a throttled
6254 * group (e.g. via set_curr_task), since update_curr() (in the
6255 * enqueue of curr) will have resulted in resched being set. This
6256 * prevents us from potentially nominating it as a false LAST_BUDDY
6259 if (test_tsk_need_resched(curr))
6262 /* Idle tasks are by definition preempted by non-idle tasks. */
6263 if (unlikely(curr->policy == SCHED_IDLE) &&
6264 likely(p->policy != SCHED_IDLE))
6268 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6269 * is driven by the tick):
6271 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6274 find_matching_se(&se, &pse);
6275 update_curr(cfs_rq_of(se));
6277 if (wakeup_preempt_entity(se, pse) == 1) {
6279 * Bias pick_next to pick the sched entity that is
6280 * triggering this preemption.
6282 if (!next_buddy_marked)
6283 set_next_buddy(pse);
6292 * Only set the backward buddy when the current task is still
6293 * on the rq. This can happen when a wakeup gets interleaved
6294 * with schedule on the ->pre_schedule() or idle_balance()
6295 * point, either of which can * drop the rq lock.
6297 * Also, during early boot the idle thread is in the fair class,
6298 * for obvious reasons its a bad idea to schedule back to it.
6300 if (unlikely(!se->on_rq || curr == rq->idle))
6303 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6307 static struct task_struct *
6308 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6310 struct cfs_rq *cfs_rq = &rq->cfs;
6311 struct sched_entity *se;
6312 struct task_struct *p;
6316 #ifdef CONFIG_FAIR_GROUP_SCHED
6317 if (!cfs_rq->nr_running)
6320 if (prev->sched_class != &fair_sched_class)
6324 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6325 * likely that a next task is from the same cgroup as the current.
6327 * Therefore attempt to avoid putting and setting the entire cgroup
6328 * hierarchy, only change the part that actually changes.
6332 struct sched_entity *curr = cfs_rq->curr;
6335 * Since we got here without doing put_prev_entity() we also
6336 * have to consider cfs_rq->curr. If it is still a runnable
6337 * entity, update_curr() will update its vruntime, otherwise
6338 * forget we've ever seen it.
6342 update_curr(cfs_rq);
6347 * This call to check_cfs_rq_runtime() will do the
6348 * throttle and dequeue its entity in the parent(s).
6349 * Therefore the 'simple' nr_running test will indeed
6352 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6356 se = pick_next_entity(cfs_rq, curr);
6357 cfs_rq = group_cfs_rq(se);
6363 * Since we haven't yet done put_prev_entity and if the selected task
6364 * is a different task than we started out with, try and touch the
6365 * least amount of cfs_rqs.
6368 struct sched_entity *pse = &prev->se;
6370 while (!(cfs_rq = is_same_group(se, pse))) {
6371 int se_depth = se->depth;
6372 int pse_depth = pse->depth;
6374 if (se_depth <= pse_depth) {
6375 put_prev_entity(cfs_rq_of(pse), pse);
6376 pse = parent_entity(pse);
6378 if (se_depth >= pse_depth) {
6379 set_next_entity(cfs_rq_of(se), se);
6380 se = parent_entity(se);
6384 put_prev_entity(cfs_rq, pse);
6385 set_next_entity(cfs_rq, se);
6388 if (hrtick_enabled(rq))
6389 hrtick_start_fair(rq, p);
6391 rq->misfit_task = !task_fits_max(p, rq->cpu);
6398 if (!cfs_rq->nr_running)
6401 put_prev_task(rq, prev);
6404 se = pick_next_entity(cfs_rq, NULL);
6405 set_next_entity(cfs_rq, se);
6406 cfs_rq = group_cfs_rq(se);
6411 if (hrtick_enabled(rq))
6412 hrtick_start_fair(rq, p);
6414 rq->misfit_task = !task_fits_max(p, rq->cpu);
6419 rq->misfit_task = 0;
6421 * This is OK, because current is on_cpu, which avoids it being picked
6422 * for load-balance and preemption/IRQs are still disabled avoiding
6423 * further scheduler activity on it and we're being very careful to
6424 * re-start the picking loop.
6426 lockdep_unpin_lock(&rq->lock);
6427 new_tasks = idle_balance(rq);
6428 lockdep_pin_lock(&rq->lock);
6430 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6431 * possible for any higher priority task to appear. In that case we
6432 * must re-start the pick_next_entity() loop.
6444 * Account for a descheduled task:
6446 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6448 struct sched_entity *se = &prev->se;
6449 struct cfs_rq *cfs_rq;
6451 for_each_sched_entity(se) {
6452 cfs_rq = cfs_rq_of(se);
6453 put_prev_entity(cfs_rq, se);
6458 * sched_yield() is very simple
6460 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6462 static void yield_task_fair(struct rq *rq)
6464 struct task_struct *curr = rq->curr;
6465 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6466 struct sched_entity *se = &curr->se;
6469 * Are we the only task in the tree?
6471 if (unlikely(rq->nr_running == 1))
6474 clear_buddies(cfs_rq, se);
6476 if (curr->policy != SCHED_BATCH) {
6477 update_rq_clock(rq);
6479 * Update run-time statistics of the 'current'.
6481 update_curr(cfs_rq);
6483 * Tell update_rq_clock() that we've just updated,
6484 * so we don't do microscopic update in schedule()
6485 * and double the fastpath cost.
6487 rq_clock_skip_update(rq, true);
6493 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6495 struct sched_entity *se = &p->se;
6497 /* throttled hierarchies are not runnable */
6498 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6501 /* Tell the scheduler that we'd really like pse to run next. */
6504 yield_task_fair(rq);
6510 /**************************************************
6511 * Fair scheduling class load-balancing methods.
6515 * The purpose of load-balancing is to achieve the same basic fairness the
6516 * per-cpu scheduler provides, namely provide a proportional amount of compute
6517 * time to each task. This is expressed in the following equation:
6519 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6521 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6522 * W_i,0 is defined as:
6524 * W_i,0 = \Sum_j w_i,j (2)
6526 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6527 * is derived from the nice value as per prio_to_weight[].
6529 * The weight average is an exponential decay average of the instantaneous
6532 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6534 * C_i is the compute capacity of cpu i, typically it is the
6535 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6536 * can also include other factors [XXX].
6538 * To achieve this balance we define a measure of imbalance which follows
6539 * directly from (1):
6541 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6543 * We them move tasks around to minimize the imbalance. In the continuous
6544 * function space it is obvious this converges, in the discrete case we get
6545 * a few fun cases generally called infeasible weight scenarios.
6548 * - infeasible weights;
6549 * - local vs global optima in the discrete case. ]
6554 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6555 * for all i,j solution, we create a tree of cpus that follows the hardware
6556 * topology where each level pairs two lower groups (or better). This results
6557 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6558 * tree to only the first of the previous level and we decrease the frequency
6559 * of load-balance at each level inv. proportional to the number of cpus in
6565 * \Sum { --- * --- * 2^i } = O(n) (5)
6567 * `- size of each group
6568 * | | `- number of cpus doing load-balance
6570 * `- sum over all levels
6572 * Coupled with a limit on how many tasks we can migrate every balance pass,
6573 * this makes (5) the runtime complexity of the balancer.
6575 * An important property here is that each CPU is still (indirectly) connected
6576 * to every other cpu in at most O(log n) steps:
6578 * The adjacency matrix of the resulting graph is given by:
6581 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6584 * And you'll find that:
6586 * A^(log_2 n)_i,j != 0 for all i,j (7)
6588 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6589 * The task movement gives a factor of O(m), giving a convergence complexity
6592 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6597 * In order to avoid CPUs going idle while there's still work to do, new idle
6598 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6599 * tree itself instead of relying on other CPUs to bring it work.
6601 * This adds some complexity to both (5) and (8) but it reduces the total idle
6609 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6612 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6617 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6619 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6621 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6624 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6625 * rewrite all of this once again.]
6628 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6630 enum fbq_type { regular, remote, all };
6639 #define LBF_ALL_PINNED 0x01
6640 #define LBF_NEED_BREAK 0x02
6641 #define LBF_DST_PINNED 0x04
6642 #define LBF_SOME_PINNED 0x08
6645 struct sched_domain *sd;
6653 struct cpumask *dst_grpmask;
6655 enum cpu_idle_type idle;
6657 unsigned int src_grp_nr_running;
6658 /* The set of CPUs under consideration for load-balancing */
6659 struct cpumask *cpus;
6664 unsigned int loop_break;
6665 unsigned int loop_max;
6667 enum fbq_type fbq_type;
6668 enum group_type busiest_group_type;
6669 struct list_head tasks;
6673 * Is this task likely cache-hot:
6675 static int task_hot(struct task_struct *p, struct lb_env *env)
6679 lockdep_assert_held(&env->src_rq->lock);
6681 if (p->sched_class != &fair_sched_class)
6684 if (unlikely(p->policy == SCHED_IDLE))
6688 * Buddy candidates are cache hot:
6690 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6691 (&p->se == cfs_rq_of(&p->se)->next ||
6692 &p->se == cfs_rq_of(&p->se)->last))
6695 if (sysctl_sched_migration_cost == -1)
6697 if (sysctl_sched_migration_cost == 0)
6700 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6702 return delta < (s64)sysctl_sched_migration_cost;
6705 #ifdef CONFIG_NUMA_BALANCING
6707 * Returns 1, if task migration degrades locality
6708 * Returns 0, if task migration improves locality i.e migration preferred.
6709 * Returns -1, if task migration is not affected by locality.
6711 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6713 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6714 unsigned long src_faults, dst_faults;
6715 int src_nid, dst_nid;
6717 if (!static_branch_likely(&sched_numa_balancing))
6720 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6723 src_nid = cpu_to_node(env->src_cpu);
6724 dst_nid = cpu_to_node(env->dst_cpu);
6726 if (src_nid == dst_nid)
6729 /* Migrating away from the preferred node is always bad. */
6730 if (src_nid == p->numa_preferred_nid) {
6731 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6737 /* Encourage migration to the preferred node. */
6738 if (dst_nid == p->numa_preferred_nid)
6742 src_faults = group_faults(p, src_nid);
6743 dst_faults = group_faults(p, dst_nid);
6745 src_faults = task_faults(p, src_nid);
6746 dst_faults = task_faults(p, dst_nid);
6749 return dst_faults < src_faults;
6753 static inline int migrate_degrades_locality(struct task_struct *p,
6761 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6764 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6768 lockdep_assert_held(&env->src_rq->lock);
6771 * We do not migrate tasks that are:
6772 * 1) throttled_lb_pair, or
6773 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6774 * 3) running (obviously), or
6775 * 4) are cache-hot on their current CPU.
6777 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6780 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6783 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6785 env->flags |= LBF_SOME_PINNED;
6788 * Remember if this task can be migrated to any other cpu in
6789 * our sched_group. We may want to revisit it if we couldn't
6790 * meet load balance goals by pulling other tasks on src_cpu.
6792 * Also avoid computing new_dst_cpu if we have already computed
6793 * one in current iteration.
6795 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6798 /* Prevent to re-select dst_cpu via env's cpus */
6799 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6800 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6801 env->flags |= LBF_DST_PINNED;
6802 env->new_dst_cpu = cpu;
6810 /* Record that we found atleast one task that could run on dst_cpu */
6811 env->flags &= ~LBF_ALL_PINNED;
6813 if (task_running(env->src_rq, p)) {
6814 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6819 * Aggressive migration if:
6820 * 1) destination numa is preferred
6821 * 2) task is cache cold, or
6822 * 3) too many balance attempts have failed.
6824 tsk_cache_hot = migrate_degrades_locality(p, env);
6825 if (tsk_cache_hot == -1)
6826 tsk_cache_hot = task_hot(p, env);
6828 if (tsk_cache_hot <= 0 ||
6829 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6830 if (tsk_cache_hot == 1) {
6831 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6832 schedstat_inc(p, se.statistics.nr_forced_migrations);
6837 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6842 * detach_task() -- detach the task for the migration specified in env
6844 static void detach_task(struct task_struct *p, struct lb_env *env)
6846 lockdep_assert_held(&env->src_rq->lock);
6848 deactivate_task(env->src_rq, p, 0);
6849 p->on_rq = TASK_ON_RQ_MIGRATING;
6850 double_lock_balance(env->src_rq, env->dst_rq);
6851 set_task_cpu(p, env->dst_cpu);
6852 double_unlock_balance(env->src_rq, env->dst_rq);
6856 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6857 * part of active balancing operations within "domain".
6859 * Returns a task if successful and NULL otherwise.
6861 static struct task_struct *detach_one_task(struct lb_env *env)
6863 struct task_struct *p, *n;
6865 lockdep_assert_held(&env->src_rq->lock);
6867 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6868 if (!can_migrate_task(p, env))
6871 detach_task(p, env);
6874 * Right now, this is only the second place where
6875 * lb_gained[env->idle] is updated (other is detach_tasks)
6876 * so we can safely collect stats here rather than
6877 * inside detach_tasks().
6879 schedstat_inc(env->sd, lb_gained[env->idle]);
6885 static const unsigned int sched_nr_migrate_break = 32;
6888 * detach_tasks() -- tries to detach up to imbalance weighted load from
6889 * busiest_rq, as part of a balancing operation within domain "sd".
6891 * Returns number of detached tasks if successful and 0 otherwise.
6893 static int detach_tasks(struct lb_env *env)
6895 struct list_head *tasks = &env->src_rq->cfs_tasks;
6896 struct task_struct *p;
6900 lockdep_assert_held(&env->src_rq->lock);
6902 if (env->imbalance <= 0)
6905 while (!list_empty(tasks)) {
6907 * We don't want to steal all, otherwise we may be treated likewise,
6908 * which could at worst lead to a livelock crash.
6910 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6913 p = list_first_entry(tasks, struct task_struct, se.group_node);
6916 /* We've more or less seen every task there is, call it quits */
6917 if (env->loop > env->loop_max)
6920 /* take a breather every nr_migrate tasks */
6921 if (env->loop > env->loop_break) {
6922 env->loop_break += sched_nr_migrate_break;
6923 env->flags |= LBF_NEED_BREAK;
6927 if (!can_migrate_task(p, env))
6930 load = task_h_load(p);
6932 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6935 if ((load / 2) > env->imbalance)
6938 detach_task(p, env);
6939 list_add(&p->se.group_node, &env->tasks);
6942 env->imbalance -= load;
6944 #ifdef CONFIG_PREEMPT
6946 * NEWIDLE balancing is a source of latency, so preemptible
6947 * kernels will stop after the first task is detached to minimize
6948 * the critical section.
6950 if (env->idle == CPU_NEWLY_IDLE)
6955 * We only want to steal up to the prescribed amount of
6958 if (env->imbalance <= 0)
6963 list_move_tail(&p->se.group_node, tasks);
6967 * Right now, this is one of only two places we collect this stat
6968 * so we can safely collect detach_one_task() stats here rather
6969 * than inside detach_one_task().
6971 schedstat_add(env->sd, lb_gained[env->idle], detached);
6977 * attach_task() -- attach the task detached by detach_task() to its new rq.
6979 static void attach_task(struct rq *rq, struct task_struct *p)
6981 lockdep_assert_held(&rq->lock);
6983 BUG_ON(task_rq(p) != rq);
6984 p->on_rq = TASK_ON_RQ_QUEUED;
6985 activate_task(rq, p, 0);
6986 check_preempt_curr(rq, p, 0);
6990 * attach_one_task() -- attaches the task returned from detach_one_task() to
6993 static void attach_one_task(struct rq *rq, struct task_struct *p)
6995 raw_spin_lock(&rq->lock);
6998 * We want to potentially raise target_cpu's OPP.
7000 update_capacity_of(cpu_of(rq));
7001 raw_spin_unlock(&rq->lock);
7005 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7008 static void attach_tasks(struct lb_env *env)
7010 struct list_head *tasks = &env->tasks;
7011 struct task_struct *p;
7013 raw_spin_lock(&env->dst_rq->lock);
7015 while (!list_empty(tasks)) {
7016 p = list_first_entry(tasks, struct task_struct, se.group_node);
7017 list_del_init(&p->se.group_node);
7019 attach_task(env->dst_rq, p);
7023 * We want to potentially raise env.dst_cpu's OPP.
7025 update_capacity_of(env->dst_cpu);
7027 raw_spin_unlock(&env->dst_rq->lock);
7030 #ifdef CONFIG_FAIR_GROUP_SCHED
7031 static void update_blocked_averages(int cpu)
7033 struct rq *rq = cpu_rq(cpu);
7034 struct cfs_rq *cfs_rq;
7035 unsigned long flags;
7037 raw_spin_lock_irqsave(&rq->lock, flags);
7038 update_rq_clock(rq);
7041 * Iterates the task_group tree in a bottom up fashion, see
7042 * list_add_leaf_cfs_rq() for details.
7044 for_each_leaf_cfs_rq(rq, cfs_rq) {
7045 /* throttled entities do not contribute to load */
7046 if (throttled_hierarchy(cfs_rq))
7049 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7051 update_tg_load_avg(cfs_rq, 0);
7053 raw_spin_unlock_irqrestore(&rq->lock, flags);
7057 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7058 * This needs to be done in a top-down fashion because the load of a child
7059 * group is a fraction of its parents load.
7061 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7063 struct rq *rq = rq_of(cfs_rq);
7064 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7065 unsigned long now = jiffies;
7068 if (cfs_rq->last_h_load_update == now)
7071 cfs_rq->h_load_next = NULL;
7072 for_each_sched_entity(se) {
7073 cfs_rq = cfs_rq_of(se);
7074 cfs_rq->h_load_next = se;
7075 if (cfs_rq->last_h_load_update == now)
7080 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7081 cfs_rq->last_h_load_update = now;
7084 while ((se = cfs_rq->h_load_next) != NULL) {
7085 load = cfs_rq->h_load;
7086 load = div64_ul(load * se->avg.load_avg,
7087 cfs_rq_load_avg(cfs_rq) + 1);
7088 cfs_rq = group_cfs_rq(se);
7089 cfs_rq->h_load = load;
7090 cfs_rq->last_h_load_update = now;
7094 static unsigned long task_h_load(struct task_struct *p)
7096 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7098 update_cfs_rq_h_load(cfs_rq);
7099 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7100 cfs_rq_load_avg(cfs_rq) + 1);
7103 static inline void update_blocked_averages(int cpu)
7105 struct rq *rq = cpu_rq(cpu);
7106 struct cfs_rq *cfs_rq = &rq->cfs;
7107 unsigned long flags;
7109 raw_spin_lock_irqsave(&rq->lock, flags);
7110 update_rq_clock(rq);
7111 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7112 raw_spin_unlock_irqrestore(&rq->lock, flags);
7115 static unsigned long task_h_load(struct task_struct *p)
7117 return p->se.avg.load_avg;
7121 /********** Helpers for find_busiest_group ************************/
7124 * sg_lb_stats - stats of a sched_group required for load_balancing
7126 struct sg_lb_stats {
7127 unsigned long avg_load; /*Avg load across the CPUs of the group */
7128 unsigned long group_load; /* Total load over the CPUs of the group */
7129 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7130 unsigned long load_per_task;
7131 unsigned long group_capacity;
7132 unsigned long group_util; /* Total utilization of the group */
7133 unsigned int sum_nr_running; /* Nr tasks running in the group */
7134 unsigned int idle_cpus;
7135 unsigned int group_weight;
7136 enum group_type group_type;
7137 int group_no_capacity;
7138 int group_misfit_task; /* A cpu has a task too big for its capacity */
7139 #ifdef CONFIG_NUMA_BALANCING
7140 unsigned int nr_numa_running;
7141 unsigned int nr_preferred_running;
7146 * sd_lb_stats - Structure to store the statistics of a sched_domain
7147 * during load balancing.
7149 struct sd_lb_stats {
7150 struct sched_group *busiest; /* Busiest group in this sd */
7151 struct sched_group *local; /* Local group in this sd */
7152 unsigned long total_load; /* Total load of all groups in sd */
7153 unsigned long total_capacity; /* Total capacity of all groups in sd */
7154 unsigned long avg_load; /* Average load across all groups in sd */
7156 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7157 struct sg_lb_stats local_stat; /* Statistics of the local group */
7160 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7163 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7164 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7165 * We must however clear busiest_stat::avg_load because
7166 * update_sd_pick_busiest() reads this before assignment.
7168 *sds = (struct sd_lb_stats){
7172 .total_capacity = 0UL,
7175 .sum_nr_running = 0,
7176 .group_type = group_other,
7182 * get_sd_load_idx - Obtain the load index for a given sched domain.
7183 * @sd: The sched_domain whose load_idx is to be obtained.
7184 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7186 * Return: The load index.
7188 static inline int get_sd_load_idx(struct sched_domain *sd,
7189 enum cpu_idle_type idle)
7195 load_idx = sd->busy_idx;
7198 case CPU_NEWLY_IDLE:
7199 load_idx = sd->newidle_idx;
7202 load_idx = sd->idle_idx;
7209 static unsigned long scale_rt_capacity(int cpu)
7211 struct rq *rq = cpu_rq(cpu);
7212 u64 total, used, age_stamp, avg;
7216 * Since we're reading these variables without serialization make sure
7217 * we read them once before doing sanity checks on them.
7219 age_stamp = READ_ONCE(rq->age_stamp);
7220 avg = READ_ONCE(rq->rt_avg);
7221 delta = __rq_clock_broken(rq) - age_stamp;
7223 if (unlikely(delta < 0))
7226 total = sched_avg_period() + delta;
7228 used = div_u64(avg, total);
7231 * deadline bandwidth is defined at system level so we must
7232 * weight this bandwidth with the max capacity of the system.
7233 * As a reminder, avg_bw is 20bits width and
7234 * scale_cpu_capacity is 10 bits width
7236 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7238 if (likely(used < SCHED_CAPACITY_SCALE))
7239 return SCHED_CAPACITY_SCALE - used;
7244 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7246 raw_spin_lock_init(&mcc->lock);
7251 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7253 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7254 struct sched_group *sdg = sd->groups;
7255 struct max_cpu_capacity *mcc;
7256 unsigned long max_capacity;
7258 unsigned long flags;
7260 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7262 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7264 raw_spin_lock_irqsave(&mcc->lock, flags);
7265 max_capacity = mcc->val;
7266 max_cap_cpu = mcc->cpu;
7268 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7269 (max_capacity < capacity)) {
7270 mcc->val = capacity;
7272 #ifdef CONFIG_SCHED_DEBUG
7273 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7274 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7279 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7281 skip_unlock: __attribute__ ((unused));
7282 capacity *= scale_rt_capacity(cpu);
7283 capacity >>= SCHED_CAPACITY_SHIFT;
7288 cpu_rq(cpu)->cpu_capacity = capacity;
7289 sdg->sgc->capacity = capacity;
7290 sdg->sgc->max_capacity = capacity;
7291 sdg->sgc->min_capacity = capacity;
7294 void update_group_capacity(struct sched_domain *sd, int cpu)
7296 struct sched_domain *child = sd->child;
7297 struct sched_group *group, *sdg = sd->groups;
7298 unsigned long capacity, max_capacity, min_capacity;
7299 unsigned long interval;
7301 interval = msecs_to_jiffies(sd->balance_interval);
7302 interval = clamp(interval, 1UL, max_load_balance_interval);
7303 sdg->sgc->next_update = jiffies + interval;
7306 update_cpu_capacity(sd, cpu);
7312 min_capacity = ULONG_MAX;
7314 if (child->flags & SD_OVERLAP) {
7316 * SD_OVERLAP domains cannot assume that child groups
7317 * span the current group.
7320 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7321 struct sched_group_capacity *sgc;
7322 struct rq *rq = cpu_rq(cpu);
7325 * build_sched_domains() -> init_sched_groups_capacity()
7326 * gets here before we've attached the domains to the
7329 * Use capacity_of(), which is set irrespective of domains
7330 * in update_cpu_capacity().
7332 * This avoids capacity from being 0 and
7333 * causing divide-by-zero issues on boot.
7335 if (unlikely(!rq->sd)) {
7336 capacity += capacity_of(cpu);
7338 sgc = rq->sd->groups->sgc;
7339 capacity += sgc->capacity;
7342 max_capacity = max(capacity, max_capacity);
7343 min_capacity = min(capacity, min_capacity);
7347 * !SD_OVERLAP domains can assume that child groups
7348 * span the current group.
7351 group = child->groups;
7353 struct sched_group_capacity *sgc = group->sgc;
7355 capacity += sgc->capacity;
7356 max_capacity = max(sgc->max_capacity, max_capacity);
7357 min_capacity = min(sgc->min_capacity, min_capacity);
7358 group = group->next;
7359 } while (group != child->groups);
7362 sdg->sgc->capacity = capacity;
7363 sdg->sgc->max_capacity = max_capacity;
7364 sdg->sgc->min_capacity = min_capacity;
7368 * Check whether the capacity of the rq has been noticeably reduced by side
7369 * activity. The imbalance_pct is used for the threshold.
7370 * Return true is the capacity is reduced
7373 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7375 return ((rq->cpu_capacity * sd->imbalance_pct) <
7376 (rq->cpu_capacity_orig * 100));
7380 * Group imbalance indicates (and tries to solve) the problem where balancing
7381 * groups is inadequate due to tsk_cpus_allowed() constraints.
7383 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7384 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7387 * { 0 1 2 3 } { 4 5 6 7 }
7390 * If we were to balance group-wise we'd place two tasks in the first group and
7391 * two tasks in the second group. Clearly this is undesired as it will overload
7392 * cpu 3 and leave one of the cpus in the second group unused.
7394 * The current solution to this issue is detecting the skew in the first group
7395 * by noticing the lower domain failed to reach balance and had difficulty
7396 * moving tasks due to affinity constraints.
7398 * When this is so detected; this group becomes a candidate for busiest; see
7399 * update_sd_pick_busiest(). And calculate_imbalance() and
7400 * find_busiest_group() avoid some of the usual balance conditions to allow it
7401 * to create an effective group imbalance.
7403 * This is a somewhat tricky proposition since the next run might not find the
7404 * group imbalance and decide the groups need to be balanced again. A most
7405 * subtle and fragile situation.
7408 static inline int sg_imbalanced(struct sched_group *group)
7410 return group->sgc->imbalance;
7414 * group_has_capacity returns true if the group has spare capacity that could
7415 * be used by some tasks.
7416 * We consider that a group has spare capacity if the * number of task is
7417 * smaller than the number of CPUs or if the utilization is lower than the
7418 * available capacity for CFS tasks.
7419 * For the latter, we use a threshold to stabilize the state, to take into
7420 * account the variance of the tasks' load and to return true if the available
7421 * capacity in meaningful for the load balancer.
7422 * As an example, an available capacity of 1% can appear but it doesn't make
7423 * any benefit for the load balance.
7426 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7428 if (sgs->sum_nr_running < sgs->group_weight)
7431 if ((sgs->group_capacity * 100) >
7432 (sgs->group_util * env->sd->imbalance_pct))
7439 * group_is_overloaded returns true if the group has more tasks than it can
7441 * group_is_overloaded is not equals to !group_has_capacity because a group
7442 * with the exact right number of tasks, has no more spare capacity but is not
7443 * overloaded so both group_has_capacity and group_is_overloaded return
7447 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7449 if (sgs->sum_nr_running <= sgs->group_weight)
7452 if ((sgs->group_capacity * 100) <
7453 (sgs->group_util * env->sd->imbalance_pct))
7461 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7462 * per-cpu capacity than sched_group ref.
7465 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7467 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7468 ref->sgc->max_capacity;
7472 group_type group_classify(struct sched_group *group,
7473 struct sg_lb_stats *sgs)
7475 if (sgs->group_no_capacity)
7476 return group_overloaded;
7478 if (sg_imbalanced(group))
7479 return group_imbalanced;
7481 if (sgs->group_misfit_task)
7482 return group_misfit_task;
7488 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7489 * @env: The load balancing environment.
7490 * @group: sched_group whose statistics are to be updated.
7491 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7492 * @local_group: Does group contain this_cpu.
7493 * @sgs: variable to hold the statistics for this group.
7494 * @overload: Indicate more than one runnable task for any CPU.
7495 * @overutilized: Indicate overutilization for any CPU.
7497 static inline void update_sg_lb_stats(struct lb_env *env,
7498 struct sched_group *group, int load_idx,
7499 int local_group, struct sg_lb_stats *sgs,
7500 bool *overload, bool *overutilized)
7505 memset(sgs, 0, sizeof(*sgs));
7507 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7508 struct rq *rq = cpu_rq(i);
7510 /* Bias balancing toward cpus of our domain */
7512 load = target_load(i, load_idx);
7514 load = source_load(i, load_idx);
7516 sgs->group_load += load;
7517 sgs->group_util += cpu_util(i);
7518 sgs->sum_nr_running += rq->cfs.h_nr_running;
7520 nr_running = rq->nr_running;
7524 #ifdef CONFIG_NUMA_BALANCING
7525 sgs->nr_numa_running += rq->nr_numa_running;
7526 sgs->nr_preferred_running += rq->nr_preferred_running;
7528 sgs->sum_weighted_load += weighted_cpuload(i);
7530 * No need to call idle_cpu() if nr_running is not 0
7532 if (!nr_running && idle_cpu(i))
7535 if (cpu_overutilized(i)) {
7536 *overutilized = true;
7537 if (!sgs->group_misfit_task && rq->misfit_task)
7538 sgs->group_misfit_task = capacity_of(i);
7542 /* Adjust by relative CPU capacity of the group */
7543 sgs->group_capacity = group->sgc->capacity;
7544 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7546 if (sgs->sum_nr_running)
7547 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7549 sgs->group_weight = group->group_weight;
7551 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7552 sgs->group_type = group_classify(group, sgs);
7556 * update_sd_pick_busiest - return 1 on busiest group
7557 * @env: The load balancing environment.
7558 * @sds: sched_domain statistics
7559 * @sg: sched_group candidate to be checked for being the busiest
7560 * @sgs: sched_group statistics
7562 * Determine if @sg is a busier group than the previously selected
7565 * Return: %true if @sg is a busier group than the previously selected
7566 * busiest group. %false otherwise.
7568 static bool update_sd_pick_busiest(struct lb_env *env,
7569 struct sd_lb_stats *sds,
7570 struct sched_group *sg,
7571 struct sg_lb_stats *sgs)
7573 struct sg_lb_stats *busiest = &sds->busiest_stat;
7575 if (sgs->group_type > busiest->group_type)
7578 if (sgs->group_type < busiest->group_type)
7582 * Candidate sg doesn't face any serious load-balance problems
7583 * so don't pick it if the local sg is already filled up.
7585 if (sgs->group_type == group_other &&
7586 !group_has_capacity(env, &sds->local_stat))
7589 if (sgs->avg_load <= busiest->avg_load)
7592 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7596 * Candidate sg has no more than one task per CPU and
7597 * has higher per-CPU capacity. Migrating tasks to less
7598 * capable CPUs may harm throughput. Maximize throughput,
7599 * power/energy consequences are not considered.
7601 if (sgs->sum_nr_running <= sgs->group_weight &&
7602 group_smaller_cpu_capacity(sds->local, sg))
7606 /* This is the busiest node in its class. */
7607 if (!(env->sd->flags & SD_ASYM_PACKING))
7611 * ASYM_PACKING needs to move all the work to the lowest
7612 * numbered CPUs in the group, therefore mark all groups
7613 * higher than ourself as busy.
7615 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7619 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7626 #ifdef CONFIG_NUMA_BALANCING
7627 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7629 if (sgs->sum_nr_running > sgs->nr_numa_running)
7631 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7636 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7638 if (rq->nr_running > rq->nr_numa_running)
7640 if (rq->nr_running > rq->nr_preferred_running)
7645 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7650 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7654 #endif /* CONFIG_NUMA_BALANCING */
7657 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7658 * @env: The load balancing environment.
7659 * @sds: variable to hold the statistics for this sched_domain.
7661 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7663 struct sched_domain *child = env->sd->child;
7664 struct sched_group *sg = env->sd->groups;
7665 struct sg_lb_stats tmp_sgs;
7666 int load_idx, prefer_sibling = 0;
7667 bool overload = false, overutilized = false;
7669 if (child && child->flags & SD_PREFER_SIBLING)
7672 load_idx = get_sd_load_idx(env->sd, env->idle);
7675 struct sg_lb_stats *sgs = &tmp_sgs;
7678 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7681 sgs = &sds->local_stat;
7683 if (env->idle != CPU_NEWLY_IDLE ||
7684 time_after_eq(jiffies, sg->sgc->next_update))
7685 update_group_capacity(env->sd, env->dst_cpu);
7688 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7689 &overload, &overutilized);
7695 * In case the child domain prefers tasks go to siblings
7696 * first, lower the sg capacity so that we'll try
7697 * and move all the excess tasks away. We lower the capacity
7698 * of a group only if the local group has the capacity to fit
7699 * these excess tasks. The extra check prevents the case where
7700 * you always pull from the heaviest group when it is already
7701 * under-utilized (possible with a large weight task outweighs
7702 * the tasks on the system).
7704 if (prefer_sibling && sds->local &&
7705 group_has_capacity(env, &sds->local_stat) &&
7706 (sgs->sum_nr_running > 1)) {
7707 sgs->group_no_capacity = 1;
7708 sgs->group_type = group_classify(sg, sgs);
7712 * Ignore task groups with misfit tasks if local group has no
7713 * capacity or if per-cpu capacity isn't higher.
7715 if (sgs->group_type == group_misfit_task &&
7716 (!group_has_capacity(env, &sds->local_stat) ||
7717 !group_smaller_cpu_capacity(sg, sds->local)))
7718 sgs->group_type = group_other;
7720 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7722 sds->busiest_stat = *sgs;
7726 /* Now, start updating sd_lb_stats */
7727 sds->total_load += sgs->group_load;
7728 sds->total_capacity += sgs->group_capacity;
7731 } while (sg != env->sd->groups);
7733 if (env->sd->flags & SD_NUMA)
7734 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7736 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7738 if (!env->sd->parent) {
7739 /* update overload indicator if we are at root domain */
7740 if (env->dst_rq->rd->overload != overload)
7741 env->dst_rq->rd->overload = overload;
7743 /* Update over-utilization (tipping point, U >= 0) indicator */
7744 if (env->dst_rq->rd->overutilized != overutilized) {
7745 env->dst_rq->rd->overutilized = overutilized;
7746 trace_sched_overutilized(overutilized);
7749 if (!env->dst_rq->rd->overutilized && overutilized) {
7750 env->dst_rq->rd->overutilized = true;
7751 trace_sched_overutilized(true);
7758 * check_asym_packing - Check to see if the group is packed into the
7761 * This is primarily intended to used at the sibling level. Some
7762 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7763 * case of POWER7, it can move to lower SMT modes only when higher
7764 * threads are idle. When in lower SMT modes, the threads will
7765 * perform better since they share less core resources. Hence when we
7766 * have idle threads, we want them to be the higher ones.
7768 * This packing function is run on idle threads. It checks to see if
7769 * the busiest CPU in this domain (core in the P7 case) has a higher
7770 * CPU number than the packing function is being run on. Here we are
7771 * assuming lower CPU number will be equivalent to lower a SMT thread
7774 * Return: 1 when packing is required and a task should be moved to
7775 * this CPU. The amount of the imbalance is returned in *imbalance.
7777 * @env: The load balancing environment.
7778 * @sds: Statistics of the sched_domain which is to be packed
7780 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7784 if (!(env->sd->flags & SD_ASYM_PACKING))
7790 busiest_cpu = group_first_cpu(sds->busiest);
7791 if (env->dst_cpu > busiest_cpu)
7794 env->imbalance = DIV_ROUND_CLOSEST(
7795 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7796 SCHED_CAPACITY_SCALE);
7802 * fix_small_imbalance - Calculate the minor imbalance that exists
7803 * amongst the groups of a sched_domain, during
7805 * @env: The load balancing environment.
7806 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7809 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7811 unsigned long tmp, capa_now = 0, capa_move = 0;
7812 unsigned int imbn = 2;
7813 unsigned long scaled_busy_load_per_task;
7814 struct sg_lb_stats *local, *busiest;
7816 local = &sds->local_stat;
7817 busiest = &sds->busiest_stat;
7819 if (!local->sum_nr_running)
7820 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7821 else if (busiest->load_per_task > local->load_per_task)
7824 scaled_busy_load_per_task =
7825 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7826 busiest->group_capacity;
7828 if (busiest->avg_load + scaled_busy_load_per_task >=
7829 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7830 env->imbalance = busiest->load_per_task;
7835 * OK, we don't have enough imbalance to justify moving tasks,
7836 * however we may be able to increase total CPU capacity used by
7840 capa_now += busiest->group_capacity *
7841 min(busiest->load_per_task, busiest->avg_load);
7842 capa_now += local->group_capacity *
7843 min(local->load_per_task, local->avg_load);
7844 capa_now /= SCHED_CAPACITY_SCALE;
7846 /* Amount of load we'd subtract */
7847 if (busiest->avg_load > scaled_busy_load_per_task) {
7848 capa_move += busiest->group_capacity *
7849 min(busiest->load_per_task,
7850 busiest->avg_load - scaled_busy_load_per_task);
7853 /* Amount of load we'd add */
7854 if (busiest->avg_load * busiest->group_capacity <
7855 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7856 tmp = (busiest->avg_load * busiest->group_capacity) /
7857 local->group_capacity;
7859 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7860 local->group_capacity;
7862 capa_move += local->group_capacity *
7863 min(local->load_per_task, local->avg_load + tmp);
7864 capa_move /= SCHED_CAPACITY_SCALE;
7866 /* Move if we gain throughput */
7867 if (capa_move > capa_now)
7868 env->imbalance = busiest->load_per_task;
7872 * calculate_imbalance - Calculate the amount of imbalance present within the
7873 * groups of a given sched_domain during load balance.
7874 * @env: load balance environment
7875 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7877 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7879 unsigned long max_pull, load_above_capacity = ~0UL;
7880 struct sg_lb_stats *local, *busiest;
7882 local = &sds->local_stat;
7883 busiest = &sds->busiest_stat;
7885 if (busiest->group_type == group_imbalanced) {
7887 * In the group_imb case we cannot rely on group-wide averages
7888 * to ensure cpu-load equilibrium, look at wider averages. XXX
7890 busiest->load_per_task =
7891 min(busiest->load_per_task, sds->avg_load);
7895 * In the presence of smp nice balancing, certain scenarios can have
7896 * max load less than avg load(as we skip the groups at or below
7897 * its cpu_capacity, while calculating max_load..)
7899 if (busiest->avg_load <= sds->avg_load ||
7900 local->avg_load >= sds->avg_load) {
7901 /* Misfitting tasks should be migrated in any case */
7902 if (busiest->group_type == group_misfit_task) {
7903 env->imbalance = busiest->group_misfit_task;
7908 * Busiest group is overloaded, local is not, use the spare
7909 * cycles to maximize throughput
7911 if (busiest->group_type == group_overloaded &&
7912 local->group_type <= group_misfit_task) {
7913 env->imbalance = busiest->load_per_task;
7918 return fix_small_imbalance(env, sds);
7922 * If there aren't any idle cpus, avoid creating some.
7924 if (busiest->group_type == group_overloaded &&
7925 local->group_type == group_overloaded) {
7926 load_above_capacity = busiest->sum_nr_running *
7928 if (load_above_capacity > busiest->group_capacity)
7929 load_above_capacity -= busiest->group_capacity;
7931 load_above_capacity = ~0UL;
7935 * We're trying to get all the cpus to the average_load, so we don't
7936 * want to push ourselves above the average load, nor do we wish to
7937 * reduce the max loaded cpu below the average load. At the same time,
7938 * we also don't want to reduce the group load below the group capacity
7939 * (so that we can implement power-savings policies etc). Thus we look
7940 * for the minimum possible imbalance.
7942 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7944 /* How much load to actually move to equalise the imbalance */
7945 env->imbalance = min(
7946 max_pull * busiest->group_capacity,
7947 (sds->avg_load - local->avg_load) * local->group_capacity
7948 ) / SCHED_CAPACITY_SCALE;
7950 /* Boost imbalance to allow misfit task to be balanced. */
7951 if (busiest->group_type == group_misfit_task)
7952 env->imbalance = max_t(long, env->imbalance,
7953 busiest->group_misfit_task);
7956 * if *imbalance is less than the average load per runnable task
7957 * there is no guarantee that any tasks will be moved so we'll have
7958 * a think about bumping its value to force at least one task to be
7961 if (env->imbalance < busiest->load_per_task)
7962 return fix_small_imbalance(env, sds);
7965 /******* find_busiest_group() helpers end here *********************/
7968 * find_busiest_group - Returns the busiest group within the sched_domain
7969 * if there is an imbalance. If there isn't an imbalance, and
7970 * the user has opted for power-savings, it returns a group whose
7971 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7972 * such a group exists.
7974 * Also calculates the amount of weighted load which should be moved
7975 * to restore balance.
7977 * @env: The load balancing environment.
7979 * Return: - The busiest group if imbalance exists.
7980 * - If no imbalance and user has opted for power-savings balance,
7981 * return the least loaded group whose CPUs can be
7982 * put to idle by rebalancing its tasks onto our group.
7984 static struct sched_group *find_busiest_group(struct lb_env *env)
7986 struct sg_lb_stats *local, *busiest;
7987 struct sd_lb_stats sds;
7989 init_sd_lb_stats(&sds);
7992 * Compute the various statistics relavent for load balancing at
7995 update_sd_lb_stats(env, &sds);
7997 if (energy_aware() && !env->dst_rq->rd->overutilized)
8000 local = &sds.local_stat;
8001 busiest = &sds.busiest_stat;
8003 /* ASYM feature bypasses nice load balance check */
8004 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8005 check_asym_packing(env, &sds))
8008 /* There is no busy sibling group to pull tasks from */
8009 if (!sds.busiest || busiest->sum_nr_running == 0)
8012 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8013 / sds.total_capacity;
8016 * If the busiest group is imbalanced the below checks don't
8017 * work because they assume all things are equal, which typically
8018 * isn't true due to cpus_allowed constraints and the like.
8020 if (busiest->group_type == group_imbalanced)
8023 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8024 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8025 busiest->group_no_capacity)
8028 /* Misfitting tasks should be dealt with regardless of the avg load */
8029 if (busiest->group_type == group_misfit_task) {
8034 * If the local group is busier than the selected busiest group
8035 * don't try and pull any tasks.
8037 if (local->avg_load >= busiest->avg_load)
8041 * Don't pull any tasks if this group is already above the domain
8044 if (local->avg_load >= sds.avg_load)
8047 if (env->idle == CPU_IDLE) {
8049 * This cpu is idle. If the busiest group is not overloaded
8050 * and there is no imbalance between this and busiest group
8051 * wrt idle cpus, it is balanced. The imbalance becomes
8052 * significant if the diff is greater than 1 otherwise we
8053 * might end up to just move the imbalance on another group
8055 if ((busiest->group_type != group_overloaded) &&
8056 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8057 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8061 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8062 * imbalance_pct to be conservative.
8064 if (100 * busiest->avg_load <=
8065 env->sd->imbalance_pct * local->avg_load)
8070 env->busiest_group_type = busiest->group_type;
8071 /* Looks like there is an imbalance. Compute it */
8072 calculate_imbalance(env, &sds);
8081 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8083 static struct rq *find_busiest_queue(struct lb_env *env,
8084 struct sched_group *group)
8086 struct rq *busiest = NULL, *rq;
8087 unsigned long busiest_load = 0, busiest_capacity = 1;
8090 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8091 unsigned long capacity, wl;
8095 rt = fbq_classify_rq(rq);
8098 * We classify groups/runqueues into three groups:
8099 * - regular: there are !numa tasks
8100 * - remote: there are numa tasks that run on the 'wrong' node
8101 * - all: there is no distinction
8103 * In order to avoid migrating ideally placed numa tasks,
8104 * ignore those when there's better options.
8106 * If we ignore the actual busiest queue to migrate another
8107 * task, the next balance pass can still reduce the busiest
8108 * queue by moving tasks around inside the node.
8110 * If we cannot move enough load due to this classification
8111 * the next pass will adjust the group classification and
8112 * allow migration of more tasks.
8114 * Both cases only affect the total convergence complexity.
8116 if (rt > env->fbq_type)
8119 capacity = capacity_of(i);
8121 wl = weighted_cpuload(i);
8124 * When comparing with imbalance, use weighted_cpuload()
8125 * which is not scaled with the cpu capacity.
8128 if (rq->nr_running == 1 && wl > env->imbalance &&
8129 !check_cpu_capacity(rq, env->sd) &&
8130 env->busiest_group_type != group_misfit_task)
8134 * For the load comparisons with the other cpu's, consider
8135 * the weighted_cpuload() scaled with the cpu capacity, so
8136 * that the load can be moved away from the cpu that is
8137 * potentially running at a lower capacity.
8139 * Thus we're looking for max(wl_i / capacity_i), crosswise
8140 * multiplication to rid ourselves of the division works out
8141 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8142 * our previous maximum.
8144 if (wl * busiest_capacity > busiest_load * capacity) {
8146 busiest_capacity = capacity;
8155 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8156 * so long as it is large enough.
8158 #define MAX_PINNED_INTERVAL 512
8160 /* Working cpumask for load_balance and load_balance_newidle. */
8161 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8163 static int need_active_balance(struct lb_env *env)
8165 struct sched_domain *sd = env->sd;
8167 if (env->idle == CPU_NEWLY_IDLE) {
8170 * ASYM_PACKING needs to force migrate tasks from busy but
8171 * higher numbered CPUs in order to pack all tasks in the
8172 * lowest numbered CPUs.
8174 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8179 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8180 * It's worth migrating the task if the src_cpu's capacity is reduced
8181 * because of other sched_class or IRQs if more capacity stays
8182 * available on dst_cpu.
8184 if ((env->idle != CPU_NOT_IDLE) &&
8185 (env->src_rq->cfs.h_nr_running == 1)) {
8186 if ((check_cpu_capacity(env->src_rq, sd)) &&
8187 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8191 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8192 env->src_rq->cfs.h_nr_running == 1 &&
8193 cpu_overutilized(env->src_cpu) &&
8194 !cpu_overutilized(env->dst_cpu)) {
8198 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8201 static int active_load_balance_cpu_stop(void *data);
8203 static int should_we_balance(struct lb_env *env)
8205 struct sched_group *sg = env->sd->groups;
8206 struct cpumask *sg_cpus, *sg_mask;
8207 int cpu, balance_cpu = -1;
8210 * In the newly idle case, we will allow all the cpu's
8211 * to do the newly idle load balance.
8213 if (env->idle == CPU_NEWLY_IDLE)
8216 sg_cpus = sched_group_cpus(sg);
8217 sg_mask = sched_group_mask(sg);
8218 /* Try to find first idle cpu */
8219 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8220 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8227 if (balance_cpu == -1)
8228 balance_cpu = group_balance_cpu(sg);
8231 * First idle cpu or the first cpu(busiest) in this sched group
8232 * is eligible for doing load balancing at this and above domains.
8234 return balance_cpu == env->dst_cpu;
8238 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8239 * tasks if there is an imbalance.
8241 static int load_balance(int this_cpu, struct rq *this_rq,
8242 struct sched_domain *sd, enum cpu_idle_type idle,
8243 int *continue_balancing)
8245 int ld_moved, cur_ld_moved, active_balance = 0;
8246 struct sched_domain *sd_parent = sd->parent;
8247 struct sched_group *group;
8249 unsigned long flags;
8250 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8252 struct lb_env env = {
8254 .dst_cpu = this_cpu,
8256 .dst_grpmask = sched_group_cpus(sd->groups),
8258 .loop_break = sched_nr_migrate_break,
8261 .tasks = LIST_HEAD_INIT(env.tasks),
8265 * For NEWLY_IDLE load_balancing, we don't need to consider
8266 * other cpus in our group
8268 if (idle == CPU_NEWLY_IDLE)
8269 env.dst_grpmask = NULL;
8271 cpumask_copy(cpus, cpu_active_mask);
8273 schedstat_inc(sd, lb_count[idle]);
8276 if (!should_we_balance(&env)) {
8277 *continue_balancing = 0;
8281 group = find_busiest_group(&env);
8283 schedstat_inc(sd, lb_nobusyg[idle]);
8287 busiest = find_busiest_queue(&env, group);
8289 schedstat_inc(sd, lb_nobusyq[idle]);
8293 BUG_ON(busiest == env.dst_rq);
8295 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8297 env.src_cpu = busiest->cpu;
8298 env.src_rq = busiest;
8301 if (busiest->nr_running > 1) {
8303 * Attempt to move tasks. If find_busiest_group has found
8304 * an imbalance but busiest->nr_running <= 1, the group is
8305 * still unbalanced. ld_moved simply stays zero, so it is
8306 * correctly treated as an imbalance.
8308 env.flags |= LBF_ALL_PINNED;
8309 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8312 raw_spin_lock_irqsave(&busiest->lock, flags);
8315 * cur_ld_moved - load moved in current iteration
8316 * ld_moved - cumulative load moved across iterations
8318 cur_ld_moved = detach_tasks(&env);
8320 * We want to potentially lower env.src_cpu's OPP.
8323 update_capacity_of(env.src_cpu);
8326 * We've detached some tasks from busiest_rq. Every
8327 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8328 * unlock busiest->lock, and we are able to be sure
8329 * that nobody can manipulate the tasks in parallel.
8330 * See task_rq_lock() family for the details.
8333 raw_spin_unlock(&busiest->lock);
8337 ld_moved += cur_ld_moved;
8340 local_irq_restore(flags);
8342 if (env.flags & LBF_NEED_BREAK) {
8343 env.flags &= ~LBF_NEED_BREAK;
8348 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8349 * us and move them to an alternate dst_cpu in our sched_group
8350 * where they can run. The upper limit on how many times we
8351 * iterate on same src_cpu is dependent on number of cpus in our
8354 * This changes load balance semantics a bit on who can move
8355 * load to a given_cpu. In addition to the given_cpu itself
8356 * (or a ilb_cpu acting on its behalf where given_cpu is
8357 * nohz-idle), we now have balance_cpu in a position to move
8358 * load to given_cpu. In rare situations, this may cause
8359 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8360 * _independently_ and at _same_ time to move some load to
8361 * given_cpu) causing exceess load to be moved to given_cpu.
8362 * This however should not happen so much in practice and
8363 * moreover subsequent load balance cycles should correct the
8364 * excess load moved.
8366 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8368 /* Prevent to re-select dst_cpu via env's cpus */
8369 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8371 env.dst_rq = cpu_rq(env.new_dst_cpu);
8372 env.dst_cpu = env.new_dst_cpu;
8373 env.flags &= ~LBF_DST_PINNED;
8375 env.loop_break = sched_nr_migrate_break;
8378 * Go back to "more_balance" rather than "redo" since we
8379 * need to continue with same src_cpu.
8385 * We failed to reach balance because of affinity.
8388 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8390 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8391 *group_imbalance = 1;
8394 /* All tasks on this runqueue were pinned by CPU affinity */
8395 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8396 cpumask_clear_cpu(cpu_of(busiest), cpus);
8397 if (!cpumask_empty(cpus)) {
8399 env.loop_break = sched_nr_migrate_break;
8402 goto out_all_pinned;
8407 schedstat_inc(sd, lb_failed[idle]);
8409 * Increment the failure counter only on periodic balance.
8410 * We do not want newidle balance, which can be very
8411 * frequent, pollute the failure counter causing
8412 * excessive cache_hot migrations and active balances.
8414 if (idle != CPU_NEWLY_IDLE)
8415 if (env.src_grp_nr_running > 1)
8416 sd->nr_balance_failed++;
8418 if (need_active_balance(&env)) {
8419 raw_spin_lock_irqsave(&busiest->lock, flags);
8421 /* don't kick the active_load_balance_cpu_stop,
8422 * if the curr task on busiest cpu can't be
8425 if (!cpumask_test_cpu(this_cpu,
8426 tsk_cpus_allowed(busiest->curr))) {
8427 raw_spin_unlock_irqrestore(&busiest->lock,
8429 env.flags |= LBF_ALL_PINNED;
8430 goto out_one_pinned;
8434 * ->active_balance synchronizes accesses to
8435 * ->active_balance_work. Once set, it's cleared
8436 * only after active load balance is finished.
8438 if (!busiest->active_balance) {
8439 busiest->active_balance = 1;
8440 busiest->push_cpu = this_cpu;
8443 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8445 if (active_balance) {
8446 stop_one_cpu_nowait(cpu_of(busiest),
8447 active_load_balance_cpu_stop, busiest,
8448 &busiest->active_balance_work);
8452 * We've kicked active balancing, reset the failure
8455 sd->nr_balance_failed = sd->cache_nice_tries+1;
8458 sd->nr_balance_failed = 0;
8460 if (likely(!active_balance)) {
8461 /* We were unbalanced, so reset the balancing interval */
8462 sd->balance_interval = sd->min_interval;
8465 * If we've begun active balancing, start to back off. This
8466 * case may not be covered by the all_pinned logic if there
8467 * is only 1 task on the busy runqueue (because we don't call
8470 if (sd->balance_interval < sd->max_interval)
8471 sd->balance_interval *= 2;
8478 * We reach balance although we may have faced some affinity
8479 * constraints. Clear the imbalance flag if it was set.
8482 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8484 if (*group_imbalance)
8485 *group_imbalance = 0;
8490 * We reach balance because all tasks are pinned at this level so
8491 * we can't migrate them. Let the imbalance flag set so parent level
8492 * can try to migrate them.
8494 schedstat_inc(sd, lb_balanced[idle]);
8496 sd->nr_balance_failed = 0;
8499 /* tune up the balancing interval */
8500 if (((env.flags & LBF_ALL_PINNED) &&
8501 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8502 (sd->balance_interval < sd->max_interval))
8503 sd->balance_interval *= 2;
8510 static inline unsigned long
8511 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8513 unsigned long interval = sd->balance_interval;
8516 interval *= sd->busy_factor;
8518 /* scale ms to jiffies */
8519 interval = msecs_to_jiffies(interval);
8520 interval = clamp(interval, 1UL, max_load_balance_interval);
8526 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8528 unsigned long interval, next;
8530 interval = get_sd_balance_interval(sd, cpu_busy);
8531 next = sd->last_balance + interval;
8533 if (time_after(*next_balance, next))
8534 *next_balance = next;
8538 * idle_balance is called by schedule() if this_cpu is about to become
8539 * idle. Attempts to pull tasks from other CPUs.
8541 static int idle_balance(struct rq *this_rq)
8543 unsigned long next_balance = jiffies + HZ;
8544 int this_cpu = this_rq->cpu;
8545 struct sched_domain *sd;
8546 int pulled_task = 0;
8548 long removed_util=0;
8550 idle_enter_fair(this_rq);
8553 * We must set idle_stamp _before_ calling idle_balance(), such that we
8554 * measure the duration of idle_balance() as idle time.
8556 this_rq->idle_stamp = rq_clock(this_rq);
8558 if (!energy_aware() &&
8559 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8560 !this_rq->rd->overload)) {
8562 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8564 update_next_balance(sd, 0, &next_balance);
8570 raw_spin_unlock(&this_rq->lock);
8573 * If removed_util_avg is !0 we most probably migrated some task away
8574 * from this_cpu. In this case we might be willing to trigger an OPP
8575 * update, but we want to do so if we don't find anybody else to pull
8576 * here (we will trigger an OPP update with the pulled task's enqueue
8579 * Record removed_util before calling update_blocked_averages, and use
8580 * it below (before returning) to see if an OPP update is required.
8582 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8583 update_blocked_averages(this_cpu);
8585 for_each_domain(this_cpu, sd) {
8586 int continue_balancing = 1;
8587 u64 t0, domain_cost;
8589 if (!(sd->flags & SD_LOAD_BALANCE))
8592 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8593 update_next_balance(sd, 0, &next_balance);
8597 if (sd->flags & SD_BALANCE_NEWIDLE) {
8598 t0 = sched_clock_cpu(this_cpu);
8600 pulled_task = load_balance(this_cpu, this_rq,
8602 &continue_balancing);
8604 domain_cost = sched_clock_cpu(this_cpu) - t0;
8605 if (domain_cost > sd->max_newidle_lb_cost)
8606 sd->max_newidle_lb_cost = domain_cost;
8608 curr_cost += domain_cost;
8611 update_next_balance(sd, 0, &next_balance);
8614 * Stop searching for tasks to pull if there are
8615 * now runnable tasks on this rq.
8617 if (pulled_task || this_rq->nr_running > 0)
8622 raw_spin_lock(&this_rq->lock);
8624 if (curr_cost > this_rq->max_idle_balance_cost)
8625 this_rq->max_idle_balance_cost = curr_cost;
8628 * While browsing the domains, we released the rq lock, a task could
8629 * have been enqueued in the meantime. Since we're not going idle,
8630 * pretend we pulled a task.
8632 if (this_rq->cfs.h_nr_running && !pulled_task)
8636 /* Move the next balance forward */
8637 if (time_after(this_rq->next_balance, next_balance))
8638 this_rq->next_balance = next_balance;
8640 /* Is there a task of a high priority class? */
8641 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8645 idle_exit_fair(this_rq);
8646 this_rq->idle_stamp = 0;
8647 } else if (removed_util) {
8649 * No task pulled and someone has been migrated away.
8650 * Good case to trigger an OPP update.
8652 update_capacity_of(this_cpu);
8659 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8660 * running tasks off the busiest CPU onto idle CPUs. It requires at
8661 * least 1 task to be running on each physical CPU where possible, and
8662 * avoids physical / logical imbalances.
8664 static int active_load_balance_cpu_stop(void *data)
8666 struct rq *busiest_rq = data;
8667 int busiest_cpu = cpu_of(busiest_rq);
8668 int target_cpu = busiest_rq->push_cpu;
8669 struct rq *target_rq = cpu_rq(target_cpu);
8670 struct sched_domain *sd;
8671 struct task_struct *p = NULL;
8673 raw_spin_lock_irq(&busiest_rq->lock);
8675 /* make sure the requested cpu hasn't gone down in the meantime */
8676 if (unlikely(busiest_cpu != smp_processor_id() ||
8677 !busiest_rq->active_balance))
8680 /* Is there any task to move? */
8681 if (busiest_rq->nr_running <= 1)
8685 * This condition is "impossible", if it occurs
8686 * we need to fix it. Originally reported by
8687 * Bjorn Helgaas on a 128-cpu setup.
8689 BUG_ON(busiest_rq == target_rq);
8691 /* Search for an sd spanning us and the target CPU. */
8693 for_each_domain(target_cpu, sd) {
8694 if ((sd->flags & SD_LOAD_BALANCE) &&
8695 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8700 struct lb_env env = {
8702 .dst_cpu = target_cpu,
8703 .dst_rq = target_rq,
8704 .src_cpu = busiest_rq->cpu,
8705 .src_rq = busiest_rq,
8709 schedstat_inc(sd, alb_count);
8711 p = detach_one_task(&env);
8713 schedstat_inc(sd, alb_pushed);
8715 * We want to potentially lower env.src_cpu's OPP.
8717 update_capacity_of(env.src_cpu);
8720 schedstat_inc(sd, alb_failed);
8724 busiest_rq->active_balance = 0;
8725 raw_spin_unlock(&busiest_rq->lock);
8728 attach_one_task(target_rq, p);
8735 static inline int on_null_domain(struct rq *rq)
8737 return unlikely(!rcu_dereference_sched(rq->sd));
8740 #ifdef CONFIG_NO_HZ_COMMON
8742 * idle load balancing details
8743 * - When one of the busy CPUs notice that there may be an idle rebalancing
8744 * needed, they will kick the idle load balancer, which then does idle
8745 * load balancing for all the idle CPUs.
8748 cpumask_var_t idle_cpus_mask;
8750 unsigned long next_balance; /* in jiffy units */
8751 } nohz ____cacheline_aligned;
8753 static inline int find_new_ilb(void)
8755 int ilb = cpumask_first(nohz.idle_cpus_mask);
8757 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8764 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8765 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8766 * CPU (if there is one).
8768 static void nohz_balancer_kick(void)
8772 nohz.next_balance++;
8774 ilb_cpu = find_new_ilb();
8776 if (ilb_cpu >= nr_cpu_ids)
8779 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8782 * Use smp_send_reschedule() instead of resched_cpu().
8783 * This way we generate a sched IPI on the target cpu which
8784 * is idle. And the softirq performing nohz idle load balance
8785 * will be run before returning from the IPI.
8787 smp_send_reschedule(ilb_cpu);
8791 static inline void nohz_balance_exit_idle(int cpu)
8793 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8795 * Completely isolated CPUs don't ever set, so we must test.
8797 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8798 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8799 atomic_dec(&nohz.nr_cpus);
8801 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8805 static inline void set_cpu_sd_state_busy(void)
8807 struct sched_domain *sd;
8808 int cpu = smp_processor_id();
8811 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8813 if (!sd || !sd->nohz_idle)
8817 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8822 void set_cpu_sd_state_idle(void)
8824 struct sched_domain *sd;
8825 int cpu = smp_processor_id();
8828 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8830 if (!sd || sd->nohz_idle)
8834 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8840 * This routine will record that the cpu is going idle with tick stopped.
8841 * This info will be used in performing idle load balancing in the future.
8843 void nohz_balance_enter_idle(int cpu)
8846 * If this cpu is going down, then nothing needs to be done.
8848 if (!cpu_active(cpu))
8851 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8855 * If we're a completely isolated CPU, we don't play.
8857 if (on_null_domain(cpu_rq(cpu)))
8860 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8861 atomic_inc(&nohz.nr_cpus);
8862 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8865 static int sched_ilb_notifier(struct notifier_block *nfb,
8866 unsigned long action, void *hcpu)
8868 switch (action & ~CPU_TASKS_FROZEN) {
8870 nohz_balance_exit_idle(smp_processor_id());
8878 static DEFINE_SPINLOCK(balancing);
8881 * Scale the max load_balance interval with the number of CPUs in the system.
8882 * This trades load-balance latency on larger machines for less cross talk.
8884 void update_max_interval(void)
8886 max_load_balance_interval = HZ*num_online_cpus()/10;
8890 * It checks each scheduling domain to see if it is due to be balanced,
8891 * and initiates a balancing operation if so.
8893 * Balancing parameters are set up in init_sched_domains.
8895 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8897 int continue_balancing = 1;
8899 unsigned long interval;
8900 struct sched_domain *sd;
8901 /* Earliest time when we have to do rebalance again */
8902 unsigned long next_balance = jiffies + 60*HZ;
8903 int update_next_balance = 0;
8904 int need_serialize, need_decay = 0;
8907 update_blocked_averages(cpu);
8910 for_each_domain(cpu, sd) {
8912 * Decay the newidle max times here because this is a regular
8913 * visit to all the domains. Decay ~1% per second.
8915 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8916 sd->max_newidle_lb_cost =
8917 (sd->max_newidle_lb_cost * 253) / 256;
8918 sd->next_decay_max_lb_cost = jiffies + HZ;
8921 max_cost += sd->max_newidle_lb_cost;
8923 if (!(sd->flags & SD_LOAD_BALANCE))
8927 * Stop the load balance at this level. There is another
8928 * CPU in our sched group which is doing load balancing more
8931 if (!continue_balancing) {
8937 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8939 need_serialize = sd->flags & SD_SERIALIZE;
8940 if (need_serialize) {
8941 if (!spin_trylock(&balancing))
8945 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8946 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8948 * The LBF_DST_PINNED logic could have changed
8949 * env->dst_cpu, so we can't know our idle
8950 * state even if we migrated tasks. Update it.
8952 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8954 sd->last_balance = jiffies;
8955 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8958 spin_unlock(&balancing);
8960 if (time_after(next_balance, sd->last_balance + interval)) {
8961 next_balance = sd->last_balance + interval;
8962 update_next_balance = 1;
8967 * Ensure the rq-wide value also decays but keep it at a
8968 * reasonable floor to avoid funnies with rq->avg_idle.
8970 rq->max_idle_balance_cost =
8971 max((u64)sysctl_sched_migration_cost, max_cost);
8976 * next_balance will be updated only when there is a need.
8977 * When the cpu is attached to null domain for ex, it will not be
8980 if (likely(update_next_balance)) {
8981 rq->next_balance = next_balance;
8983 #ifdef CONFIG_NO_HZ_COMMON
8985 * If this CPU has been elected to perform the nohz idle
8986 * balance. Other idle CPUs have already rebalanced with
8987 * nohz_idle_balance() and nohz.next_balance has been
8988 * updated accordingly. This CPU is now running the idle load
8989 * balance for itself and we need to update the
8990 * nohz.next_balance accordingly.
8992 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8993 nohz.next_balance = rq->next_balance;
8998 #ifdef CONFIG_NO_HZ_COMMON
9000 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9001 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9003 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9005 int this_cpu = this_rq->cpu;
9008 /* Earliest time when we have to do rebalance again */
9009 unsigned long next_balance = jiffies + 60*HZ;
9010 int update_next_balance = 0;
9012 if (idle != CPU_IDLE ||
9013 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9016 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9017 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9021 * If this cpu gets work to do, stop the load balancing
9022 * work being done for other cpus. Next load
9023 * balancing owner will pick it up.
9028 rq = cpu_rq(balance_cpu);
9031 * If time for next balance is due,
9034 if (time_after_eq(jiffies, rq->next_balance)) {
9035 raw_spin_lock_irq(&rq->lock);
9036 update_rq_clock(rq);
9037 update_idle_cpu_load(rq);
9038 raw_spin_unlock_irq(&rq->lock);
9039 rebalance_domains(rq, CPU_IDLE);
9042 if (time_after(next_balance, rq->next_balance)) {
9043 next_balance = rq->next_balance;
9044 update_next_balance = 1;
9049 * next_balance will be updated only when there is a need.
9050 * When the CPU is attached to null domain for ex, it will not be
9053 if (likely(update_next_balance))
9054 nohz.next_balance = next_balance;
9056 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9060 * Current heuristic for kicking the idle load balancer in the presence
9061 * of an idle cpu in the system.
9062 * - This rq has more than one task.
9063 * - This rq has at least one CFS task and the capacity of the CPU is
9064 * significantly reduced because of RT tasks or IRQs.
9065 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9066 * multiple busy cpu.
9067 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9068 * domain span are idle.
9070 static inline bool nohz_kick_needed(struct rq *rq)
9072 unsigned long now = jiffies;
9073 struct sched_domain *sd;
9074 struct sched_group_capacity *sgc;
9075 int nr_busy, cpu = rq->cpu;
9078 if (unlikely(rq->idle_balance))
9082 * We may be recently in ticked or tickless idle mode. At the first
9083 * busy tick after returning from idle, we will update the busy stats.
9085 set_cpu_sd_state_busy();
9086 nohz_balance_exit_idle(cpu);
9089 * None are in tickless mode and hence no need for NOHZ idle load
9092 if (likely(!atomic_read(&nohz.nr_cpus)))
9095 if (time_before(now, nohz.next_balance))
9098 if (rq->nr_running >= 2 &&
9099 (!energy_aware() || cpu_overutilized(cpu)))
9103 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9104 if (sd && !energy_aware()) {
9105 sgc = sd->groups->sgc;
9106 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9115 sd = rcu_dereference(rq->sd);
9117 if ((rq->cfs.h_nr_running >= 1) &&
9118 check_cpu_capacity(rq, sd)) {
9124 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9125 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9126 sched_domain_span(sd)) < cpu)) {
9136 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9140 * run_rebalance_domains is triggered when needed from the scheduler tick.
9141 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9143 static void run_rebalance_domains(struct softirq_action *h)
9145 struct rq *this_rq = this_rq();
9146 enum cpu_idle_type idle = this_rq->idle_balance ?
9147 CPU_IDLE : CPU_NOT_IDLE;
9150 * If this cpu has a pending nohz_balance_kick, then do the
9151 * balancing on behalf of the other idle cpus whose ticks are
9152 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9153 * give the idle cpus a chance to load balance. Else we may
9154 * load balance only within the local sched_domain hierarchy
9155 * and abort nohz_idle_balance altogether if we pull some load.
9157 nohz_idle_balance(this_rq, idle);
9158 rebalance_domains(this_rq, idle);
9162 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9164 void trigger_load_balance(struct rq *rq)
9166 /* Don't need to rebalance while attached to NULL domain */
9167 if (unlikely(on_null_domain(rq)))
9170 if (time_after_eq(jiffies, rq->next_balance))
9171 raise_softirq(SCHED_SOFTIRQ);
9172 #ifdef CONFIG_NO_HZ_COMMON
9173 if (nohz_kick_needed(rq))
9174 nohz_balancer_kick();
9178 static void rq_online_fair(struct rq *rq)
9182 update_runtime_enabled(rq);
9185 static void rq_offline_fair(struct rq *rq)
9189 /* Ensure any throttled groups are reachable by pick_next_task */
9190 unthrottle_offline_cfs_rqs(rq);
9193 #endif /* CONFIG_SMP */
9196 * scheduler tick hitting a task of our scheduling class:
9198 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9200 struct cfs_rq *cfs_rq;
9201 struct sched_entity *se = &curr->se;
9203 for_each_sched_entity(se) {
9204 cfs_rq = cfs_rq_of(se);
9205 entity_tick(cfs_rq, se, queued);
9208 if (static_branch_unlikely(&sched_numa_balancing))
9209 task_tick_numa(rq, curr);
9212 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9213 rq->rd->overutilized = true;
9214 trace_sched_overutilized(true);
9217 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9223 * called on fork with the child task as argument from the parent's context
9224 * - child not yet on the tasklist
9225 * - preemption disabled
9227 static void task_fork_fair(struct task_struct *p)
9229 struct cfs_rq *cfs_rq;
9230 struct sched_entity *se = &p->se, *curr;
9231 int this_cpu = smp_processor_id();
9232 struct rq *rq = this_rq();
9233 unsigned long flags;
9235 raw_spin_lock_irqsave(&rq->lock, flags);
9237 update_rq_clock(rq);
9239 cfs_rq = task_cfs_rq(current);
9240 curr = cfs_rq->curr;
9243 * Not only the cpu but also the task_group of the parent might have
9244 * been changed after parent->se.parent,cfs_rq were copied to
9245 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9246 * of child point to valid ones.
9249 __set_task_cpu(p, this_cpu);
9252 update_curr(cfs_rq);
9255 se->vruntime = curr->vruntime;
9256 place_entity(cfs_rq, se, 1);
9258 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9260 * Upon rescheduling, sched_class::put_prev_task() will place
9261 * 'current' within the tree based on its new key value.
9263 swap(curr->vruntime, se->vruntime);
9267 se->vruntime -= cfs_rq->min_vruntime;
9269 raw_spin_unlock_irqrestore(&rq->lock, flags);
9273 * Priority of the task has changed. Check to see if we preempt
9277 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9279 if (!task_on_rq_queued(p))
9283 * Reschedule if we are currently running on this runqueue and
9284 * our priority decreased, or if we are not currently running on
9285 * this runqueue and our priority is higher than the current's
9287 if (rq->curr == p) {
9288 if (p->prio > oldprio)
9291 check_preempt_curr(rq, p, 0);
9294 static inline bool vruntime_normalized(struct task_struct *p)
9296 struct sched_entity *se = &p->se;
9299 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9300 * the dequeue_entity(.flags=0) will already have normalized the
9307 * When !on_rq, vruntime of the task has usually NOT been normalized.
9308 * But there are some cases where it has already been normalized:
9310 * - A forked child which is waiting for being woken up by
9311 * wake_up_new_task().
9312 * - A task which has been woken up by try_to_wake_up() and
9313 * waiting for actually being woken up by sched_ttwu_pending().
9315 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9321 static void detach_task_cfs_rq(struct task_struct *p)
9323 struct sched_entity *se = &p->se;
9324 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9326 if (!vruntime_normalized(p)) {
9328 * Fix up our vruntime so that the current sleep doesn't
9329 * cause 'unlimited' sleep bonus.
9331 place_entity(cfs_rq, se, 0);
9332 se->vruntime -= cfs_rq->min_vruntime;
9335 /* Catch up with the cfs_rq and remove our load when we leave */
9336 detach_entity_load_avg(cfs_rq, se);
9339 static void attach_task_cfs_rq(struct task_struct *p)
9341 struct sched_entity *se = &p->se;
9342 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9344 #ifdef CONFIG_FAIR_GROUP_SCHED
9346 * Since the real-depth could have been changed (only FAIR
9347 * class maintain depth value), reset depth properly.
9349 se->depth = se->parent ? se->parent->depth + 1 : 0;
9352 /* Synchronize task with its cfs_rq */
9353 attach_entity_load_avg(cfs_rq, se);
9355 if (!vruntime_normalized(p))
9356 se->vruntime += cfs_rq->min_vruntime;
9359 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9361 detach_task_cfs_rq(p);
9364 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9366 attach_task_cfs_rq(p);
9368 if (task_on_rq_queued(p)) {
9370 * We were most likely switched from sched_rt, so
9371 * kick off the schedule if running, otherwise just see
9372 * if we can still preempt the current task.
9377 check_preempt_curr(rq, p, 0);
9381 /* Account for a task changing its policy or group.
9383 * This routine is mostly called to set cfs_rq->curr field when a task
9384 * migrates between groups/classes.
9386 static void set_curr_task_fair(struct rq *rq)
9388 struct sched_entity *se = &rq->curr->se;
9390 for_each_sched_entity(se) {
9391 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9393 set_next_entity(cfs_rq, se);
9394 /* ensure bandwidth has been allocated on our new cfs_rq */
9395 account_cfs_rq_runtime(cfs_rq, 0);
9399 void init_cfs_rq(struct cfs_rq *cfs_rq)
9401 cfs_rq->tasks_timeline = RB_ROOT;
9402 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9403 #ifndef CONFIG_64BIT
9404 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9407 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9408 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9412 #ifdef CONFIG_FAIR_GROUP_SCHED
9413 static void task_move_group_fair(struct task_struct *p)
9415 detach_task_cfs_rq(p);
9416 set_task_rq(p, task_cpu(p));
9419 /* Tell se's cfs_rq has been changed -- migrated */
9420 p->se.avg.last_update_time = 0;
9422 attach_task_cfs_rq(p);
9425 void free_fair_sched_group(struct task_group *tg)
9429 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9431 for_each_possible_cpu(i) {
9433 kfree(tg->cfs_rq[i]);
9436 remove_entity_load_avg(tg->se[i]);
9445 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9447 struct sched_entity *se;
9448 struct cfs_rq *cfs_rq;
9452 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9455 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9459 tg->shares = NICE_0_LOAD;
9461 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9463 for_each_possible_cpu(i) {
9466 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9467 GFP_KERNEL, cpu_to_node(i));
9471 se = kzalloc_node(sizeof(struct sched_entity),
9472 GFP_KERNEL, cpu_to_node(i));
9476 init_cfs_rq(cfs_rq);
9477 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9478 init_entity_runnable_average(se);
9480 raw_spin_lock_irq(&rq->lock);
9481 post_init_entity_util_avg(se);
9482 raw_spin_unlock_irq(&rq->lock);
9493 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9495 struct rq *rq = cpu_rq(cpu);
9496 unsigned long flags;
9499 * Only empty task groups can be destroyed; so we can speculatively
9500 * check on_list without danger of it being re-added.
9502 if (!tg->cfs_rq[cpu]->on_list)
9505 raw_spin_lock_irqsave(&rq->lock, flags);
9506 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9507 raw_spin_unlock_irqrestore(&rq->lock, flags);
9510 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9511 struct sched_entity *se, int cpu,
9512 struct sched_entity *parent)
9514 struct rq *rq = cpu_rq(cpu);
9518 init_cfs_rq_runtime(cfs_rq);
9520 tg->cfs_rq[cpu] = cfs_rq;
9523 /* se could be NULL for root_task_group */
9528 se->cfs_rq = &rq->cfs;
9531 se->cfs_rq = parent->my_q;
9532 se->depth = parent->depth + 1;
9536 /* guarantee group entities always have weight */
9537 update_load_set(&se->load, NICE_0_LOAD);
9538 se->parent = parent;
9541 static DEFINE_MUTEX(shares_mutex);
9543 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9546 unsigned long flags;
9549 * We can't change the weight of the root cgroup.
9554 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9556 mutex_lock(&shares_mutex);
9557 if (tg->shares == shares)
9560 tg->shares = shares;
9561 for_each_possible_cpu(i) {
9562 struct rq *rq = cpu_rq(i);
9563 struct sched_entity *se;
9566 /* Propagate contribution to hierarchy */
9567 raw_spin_lock_irqsave(&rq->lock, flags);
9569 /* Possible calls to update_curr() need rq clock */
9570 update_rq_clock(rq);
9571 for_each_sched_entity(se)
9572 update_cfs_shares(group_cfs_rq(se));
9573 raw_spin_unlock_irqrestore(&rq->lock, flags);
9577 mutex_unlock(&shares_mutex);
9580 #else /* CONFIG_FAIR_GROUP_SCHED */
9582 void free_fair_sched_group(struct task_group *tg) { }
9584 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9589 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9591 #endif /* CONFIG_FAIR_GROUP_SCHED */
9594 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9596 struct sched_entity *se = &task->se;
9597 unsigned int rr_interval = 0;
9600 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9603 if (rq->cfs.load.weight)
9604 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9610 * All the scheduling class methods:
9612 const struct sched_class fair_sched_class = {
9613 .next = &idle_sched_class,
9614 .enqueue_task = enqueue_task_fair,
9615 .dequeue_task = dequeue_task_fair,
9616 .yield_task = yield_task_fair,
9617 .yield_to_task = yield_to_task_fair,
9619 .check_preempt_curr = check_preempt_wakeup,
9621 .pick_next_task = pick_next_task_fair,
9622 .put_prev_task = put_prev_task_fair,
9625 .select_task_rq = select_task_rq_fair,
9626 .migrate_task_rq = migrate_task_rq_fair,
9628 .rq_online = rq_online_fair,
9629 .rq_offline = rq_offline_fair,
9631 .task_waking = task_waking_fair,
9632 .task_dead = task_dead_fair,
9633 .set_cpus_allowed = set_cpus_allowed_common,
9636 .set_curr_task = set_curr_task_fair,
9637 .task_tick = task_tick_fair,
9638 .task_fork = task_fork_fair,
9640 .prio_changed = prio_changed_fair,
9641 .switched_from = switched_from_fair,
9642 .switched_to = switched_to_fair,
9644 .get_rr_interval = get_rr_interval_fair,
9646 .update_curr = update_curr_fair,
9648 #ifdef CONFIG_FAIR_GROUP_SCHED
9649 .task_move_group = task_move_group_fair,
9653 #ifdef CONFIG_SCHED_DEBUG
9654 void print_cfs_stats(struct seq_file *m, int cpu)
9656 struct cfs_rq *cfs_rq;
9659 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9660 print_cfs_rq(m, cpu, cfs_rq);
9664 #ifdef CONFIG_NUMA_BALANCING
9665 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9668 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9670 for_each_online_node(node) {
9671 if (p->numa_faults) {
9672 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9673 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9675 if (p->numa_group) {
9676 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9677 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9679 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9682 #endif /* CONFIG_NUMA_BALANCING */
9683 #endif /* CONFIG_SCHED_DEBUG */
9685 __init void init_sched_fair_class(void)
9688 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9690 #ifdef CONFIG_NO_HZ_COMMON
9691 nohz.next_balance = jiffies;
9692 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9693 cpu_notifier(sched_ilb_notifier, 0);