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/sched.h>
24 #include <linux/latencytop.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>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 * The margin used when comparing utilization with CPU capacity:
119 * util * 1024 < capacity * margin
121 unsigned int capacity_margin = 1280; /* ~20% */
123 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
129 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
135 static inline void update_load_set(struct load_weight *lw, unsigned long w)
142 * Increase the granularity value when there are more CPUs,
143 * because with more CPUs the 'effective latency' as visible
144 * to users decreases. But the relationship is not linear,
145 * so pick a second-best guess by going with the log2 of the
148 * This idea comes from the SD scheduler of Con Kolivas:
150 static unsigned int get_update_sysctl_factor(void)
152 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
155 switch (sysctl_sched_tunable_scaling) {
156 case SCHED_TUNABLESCALING_NONE:
159 case SCHED_TUNABLESCALING_LINEAR:
162 case SCHED_TUNABLESCALING_LOG:
164 factor = 1 + ilog2(cpus);
171 static void update_sysctl(void)
173 unsigned int factor = get_update_sysctl_factor();
175 #define SET_SYSCTL(name) \
176 (sysctl_##name = (factor) * normalized_sysctl_##name)
177 SET_SYSCTL(sched_min_granularity);
178 SET_SYSCTL(sched_latency);
179 SET_SYSCTL(sched_wakeup_granularity);
183 void sched_init_granularity(void)
188 #define WMULT_CONST (~0U)
189 #define WMULT_SHIFT 32
191 static void __update_inv_weight(struct load_weight *lw)
195 if (likely(lw->inv_weight))
198 w = scale_load_down(lw->weight);
200 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
202 else if (unlikely(!w))
203 lw->inv_weight = WMULT_CONST;
205 lw->inv_weight = WMULT_CONST / w;
209 * delta_exec * weight / lw.weight
211 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
213 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
214 * we're guaranteed shift stays positive because inv_weight is guaranteed to
215 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
217 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
218 * weight/lw.weight <= 1, and therefore our shift will also be positive.
220 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
222 u64 fact = scale_load_down(weight);
223 int shift = WMULT_SHIFT;
225 __update_inv_weight(lw);
227 if (unlikely(fact >> 32)) {
234 /* hint to use a 32x32->64 mul */
235 fact = (u64)(u32)fact * lw->inv_weight;
242 return mul_u64_u32_shr(delta_exec, fact, shift);
246 const struct sched_class fair_sched_class;
248 /**************************************************************
249 * CFS operations on generic schedulable entities:
252 #ifdef CONFIG_FAIR_GROUP_SCHED
254 /* cpu runqueue to which this cfs_rq is attached */
255 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
260 /* An entity is a task if it doesn't "own" a runqueue */
261 #define entity_is_task(se) (!se->my_q)
263 static inline struct task_struct *task_of(struct sched_entity *se)
265 #ifdef CONFIG_SCHED_DEBUG
266 WARN_ON_ONCE(!entity_is_task(se));
268 return container_of(se, struct task_struct, se);
271 /* Walk up scheduling entities hierarchy */
272 #define for_each_sched_entity(se) \
273 for (; se; se = se->parent)
275 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
280 /* runqueue on which this entity is (to be) queued */
281 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
286 /* runqueue "owned" by this group */
287 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
292 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 if (!cfs_rq->on_list) {
296 * Ensure we either appear before our parent (if already
297 * enqueued) or force our parent to appear after us when it is
298 * enqueued. The fact that we always enqueue bottom-up
299 * reduces this to two cases.
301 if (cfs_rq->tg->parent &&
302 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
303 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
304 &rq_of(cfs_rq)->leaf_cfs_rq_list);
306 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
307 &rq_of(cfs_rq)->leaf_cfs_rq_list);
314 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
316 if (cfs_rq->on_list) {
317 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
322 /* Iterate thr' all leaf cfs_rq's on a runqueue */
323 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
324 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
326 /* Do the two (enqueued) entities belong to the same group ? */
327 static inline struct cfs_rq *
328 is_same_group(struct sched_entity *se, struct sched_entity *pse)
330 if (se->cfs_rq == pse->cfs_rq)
336 static inline struct sched_entity *parent_entity(struct sched_entity *se)
342 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
344 int se_depth, pse_depth;
347 * preemption test can be made between sibling entities who are in the
348 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
349 * both tasks until we find their ancestors who are siblings of common
353 /* First walk up until both entities are at same depth */
354 se_depth = (*se)->depth;
355 pse_depth = (*pse)->depth;
357 while (se_depth > pse_depth) {
359 *se = parent_entity(*se);
362 while (pse_depth > se_depth) {
364 *pse = parent_entity(*pse);
367 while (!is_same_group(*se, *pse)) {
368 *se = parent_entity(*se);
369 *pse = parent_entity(*pse);
373 #else /* !CONFIG_FAIR_GROUP_SCHED */
375 static inline struct task_struct *task_of(struct sched_entity *se)
377 return container_of(se, struct task_struct, se);
380 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
382 return container_of(cfs_rq, struct rq, cfs);
385 #define entity_is_task(se) 1
387 #define for_each_sched_entity(se) \
388 for (; se; se = NULL)
390 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
392 return &task_rq(p)->cfs;
395 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
397 struct task_struct *p = task_of(se);
398 struct rq *rq = task_rq(p);
403 /* runqueue "owned" by this group */
404 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
409 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
413 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
417 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
418 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
420 static inline struct sched_entity *parent_entity(struct sched_entity *se)
426 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
430 #endif /* CONFIG_FAIR_GROUP_SCHED */
432 static __always_inline
433 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
435 /**************************************************************
436 * Scheduling class tree data structure manipulation methods:
439 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
441 s64 delta = (s64)(vruntime - max_vruntime);
443 max_vruntime = vruntime;
448 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
450 s64 delta = (s64)(vruntime - min_vruntime);
452 min_vruntime = vruntime;
457 static inline int entity_before(struct sched_entity *a,
458 struct sched_entity *b)
460 return (s64)(a->vruntime - b->vruntime) < 0;
463 static void update_min_vruntime(struct cfs_rq *cfs_rq)
465 u64 vruntime = cfs_rq->min_vruntime;
468 vruntime = cfs_rq->curr->vruntime;
470 if (cfs_rq->rb_leftmost) {
471 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
476 vruntime = se->vruntime;
478 vruntime = min_vruntime(vruntime, se->vruntime);
481 /* ensure we never gain time by being placed backwards. */
482 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
485 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
490 * Enqueue an entity into the rb-tree:
492 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
494 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
495 struct rb_node *parent = NULL;
496 struct sched_entity *entry;
500 * Find the right place in the rbtree:
504 entry = rb_entry(parent, struct sched_entity, run_node);
506 * We dont care about collisions. Nodes with
507 * the same key stay together.
509 if (entity_before(se, entry)) {
510 link = &parent->rb_left;
512 link = &parent->rb_right;
518 * Maintain a cache of leftmost tree entries (it is frequently
522 cfs_rq->rb_leftmost = &se->run_node;
524 rb_link_node(&se->run_node, parent, link);
525 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
528 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
530 if (cfs_rq->rb_leftmost == &se->run_node) {
531 struct rb_node *next_node;
533 next_node = rb_next(&se->run_node);
534 cfs_rq->rb_leftmost = next_node;
537 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
540 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
542 struct rb_node *left = cfs_rq->rb_leftmost;
547 return rb_entry(left, struct sched_entity, run_node);
550 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
552 struct rb_node *next = rb_next(&se->run_node);
557 return rb_entry(next, struct sched_entity, run_node);
560 #ifdef CONFIG_SCHED_DEBUG
561 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
563 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
568 return rb_entry(last, struct sched_entity, run_node);
571 /**************************************************************
572 * Scheduling class statistics methods:
575 int sched_proc_update_handler(struct ctl_table *table, int write,
576 void __user *buffer, size_t *lenp,
579 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
580 unsigned int factor = get_update_sysctl_factor();
585 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
586 sysctl_sched_min_granularity);
588 #define WRT_SYSCTL(name) \
589 (normalized_sysctl_##name = sysctl_##name / (factor))
590 WRT_SYSCTL(sched_min_granularity);
591 WRT_SYSCTL(sched_latency);
592 WRT_SYSCTL(sched_wakeup_granularity);
602 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
604 if (unlikely(se->load.weight != NICE_0_LOAD))
605 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
611 * The idea is to set a period in which each task runs once.
613 * When there are too many tasks (sched_nr_latency) we have to stretch
614 * this period because otherwise the slices get too small.
616 * p = (nr <= nl) ? l : l*nr/nl
618 static u64 __sched_period(unsigned long nr_running)
620 if (unlikely(nr_running > sched_nr_latency))
621 return nr_running * sysctl_sched_min_granularity;
623 return sysctl_sched_latency;
627 * We calculate the wall-time slice from the period by taking a part
628 * proportional to the weight.
632 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
634 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
636 for_each_sched_entity(se) {
637 struct load_weight *load;
638 struct load_weight lw;
640 cfs_rq = cfs_rq_of(se);
641 load = &cfs_rq->load;
643 if (unlikely(!se->on_rq)) {
646 update_load_add(&lw, se->load.weight);
649 slice = __calc_delta(slice, se->load.weight, load);
655 * We calculate the vruntime slice of a to-be-inserted task.
659 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
661 return calc_delta_fair(sched_slice(cfs_rq, se), se);
665 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
666 static unsigned long task_h_load(struct task_struct *p);
669 * We choose a half-life close to 1 scheduling period.
670 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
671 * dependent on this value.
673 #define LOAD_AVG_PERIOD 32
674 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
675 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
677 /* Give new sched_entity start runnable values to heavy its load in infant time */
678 void init_entity_runnable_average(struct sched_entity *se)
680 struct sched_avg *sa = &se->avg;
682 sa->last_update_time = 0;
684 * sched_avg's period_contrib should be strictly less then 1024, so
685 * we give it 1023 to make sure it is almost a period (1024us), and
686 * will definitely be update (after enqueue).
688 sa->period_contrib = 1023;
689 sa->load_avg = scale_load_down(se->load.weight);
690 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
692 * At this point, util_avg won't be used in select_task_rq_fair anyway
696 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
699 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
700 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
701 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
702 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
705 * With new tasks being created, their initial util_avgs are extrapolated
706 * based on the cfs_rq's current util_avg:
708 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
710 * However, in many cases, the above util_avg does not give a desired
711 * value. Moreover, the sum of the util_avgs may be divergent, such
712 * as when the series is a harmonic series.
714 * To solve this problem, we also cap the util_avg of successive tasks to
715 * only 1/2 of the left utilization budget:
717 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
719 * where n denotes the nth task.
721 * For example, a simplest series from the beginning would be like:
723 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
724 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
726 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
727 * if util_avg > util_avg_cap.
729 void post_init_entity_util_avg(struct sched_entity *se)
731 struct cfs_rq *cfs_rq = cfs_rq_of(se);
732 struct sched_avg *sa = &se->avg;
733 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
734 u64 now = cfs_rq_clock_task(cfs_rq);
737 if (cfs_rq->avg.util_avg != 0) {
738 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
739 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
741 if (sa->util_avg > cap)
746 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
749 if (entity_is_task(se)) {
750 struct task_struct *p = task_of(se);
751 if (p->sched_class != &fair_sched_class) {
753 * For !fair tasks do:
755 update_cfs_rq_load_avg(now, cfs_rq, false);
756 attach_entity_load_avg(cfs_rq, se);
757 switched_from_fair(rq, p);
759 * such that the next switched_to_fair() has the
762 se->avg.last_update_time = now;
767 update_cfs_rq_load_avg(now, cfs_rq, false);
768 attach_entity_load_avg(cfs_rq, se);
769 update_tg_load_avg(cfs_rq, false);
772 #else /* !CONFIG_SMP */
773 void init_entity_runnable_average(struct sched_entity *se)
776 void post_init_entity_util_avg(struct sched_entity *se)
779 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
782 #endif /* CONFIG_SMP */
785 * Update the current task's runtime statistics.
787 static void update_curr(struct cfs_rq *cfs_rq)
789 struct sched_entity *curr = cfs_rq->curr;
790 u64 now = rq_clock_task(rq_of(cfs_rq));
796 delta_exec = now - curr->exec_start;
797 if (unlikely((s64)delta_exec <= 0))
800 curr->exec_start = now;
802 schedstat_set(curr->statistics.exec_max,
803 max(delta_exec, curr->statistics.exec_max));
805 curr->sum_exec_runtime += delta_exec;
806 schedstat_add(cfs_rq->exec_clock, delta_exec);
808 curr->vruntime += calc_delta_fair(delta_exec, curr);
809 update_min_vruntime(cfs_rq);
811 if (entity_is_task(curr)) {
812 struct task_struct *curtask = task_of(curr);
814 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
815 cpuacct_charge(curtask, delta_exec);
816 account_group_exec_runtime(curtask, delta_exec);
819 account_cfs_rq_runtime(cfs_rq, delta_exec);
822 static void update_curr_fair(struct rq *rq)
824 update_curr(cfs_rq_of(&rq->curr->se));
828 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
830 u64 wait_start, prev_wait_start;
832 if (!schedstat_enabled())
835 wait_start = rq_clock(rq_of(cfs_rq));
836 prev_wait_start = schedstat_val(se->statistics.wait_start);
838 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
839 likely(wait_start > prev_wait_start))
840 wait_start -= prev_wait_start;
842 schedstat_set(se->statistics.wait_start, wait_start);
846 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
848 struct task_struct *p;
851 if (!schedstat_enabled())
854 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
856 if (entity_is_task(se)) {
858 if (task_on_rq_migrating(p)) {
860 * Preserve migrating task's wait time so wait_start
861 * time stamp can be adjusted to accumulate wait time
862 * prior to migration.
864 schedstat_set(se->statistics.wait_start, delta);
867 trace_sched_stat_wait(p, delta);
870 schedstat_set(se->statistics.wait_max,
871 max(schedstat_val(se->statistics.wait_max), delta));
872 schedstat_inc(se->statistics.wait_count);
873 schedstat_add(se->statistics.wait_sum, delta);
874 schedstat_set(se->statistics.wait_start, 0);
878 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
880 struct task_struct *tsk = NULL;
881 u64 sleep_start, block_start;
883 if (!schedstat_enabled())
886 sleep_start = schedstat_val(se->statistics.sleep_start);
887 block_start = schedstat_val(se->statistics.block_start);
889 if (entity_is_task(se))
893 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
898 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
899 schedstat_set(se->statistics.sleep_max, delta);
901 schedstat_set(se->statistics.sleep_start, 0);
902 schedstat_add(se->statistics.sum_sleep_runtime, delta);
905 account_scheduler_latency(tsk, delta >> 10, 1);
906 trace_sched_stat_sleep(tsk, delta);
910 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
915 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
916 schedstat_set(se->statistics.block_max, delta);
918 schedstat_set(se->statistics.block_start, 0);
919 schedstat_add(se->statistics.sum_sleep_runtime, delta);
922 if (tsk->in_iowait) {
923 schedstat_add(se->statistics.iowait_sum, delta);
924 schedstat_inc(se->statistics.iowait_count);
925 trace_sched_stat_iowait(tsk, delta);
928 trace_sched_stat_blocked(tsk, delta);
931 * Blocking time is in units of nanosecs, so shift by
932 * 20 to get a milliseconds-range estimation of the
933 * amount of time that the task spent sleeping:
935 if (unlikely(prof_on == SLEEP_PROFILING)) {
936 profile_hits(SLEEP_PROFILING,
937 (void *)get_wchan(tsk),
940 account_scheduler_latency(tsk, delta >> 10, 0);
946 * Task is being enqueued - update stats:
949 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
951 if (!schedstat_enabled())
955 * Are we enqueueing a waiting task? (for current tasks
956 * a dequeue/enqueue event is a NOP)
958 if (se != cfs_rq->curr)
959 update_stats_wait_start(cfs_rq, se);
961 if (flags & ENQUEUE_WAKEUP)
962 update_stats_enqueue_sleeper(cfs_rq, se);
966 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
969 if (!schedstat_enabled())
973 * Mark the end of the wait period if dequeueing a
976 if (se != cfs_rq->curr)
977 update_stats_wait_end(cfs_rq, se);
979 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
980 struct task_struct *tsk = task_of(se);
982 if (tsk->state & TASK_INTERRUPTIBLE)
983 schedstat_set(se->statistics.sleep_start,
984 rq_clock(rq_of(cfs_rq)));
985 if (tsk->state & TASK_UNINTERRUPTIBLE)
986 schedstat_set(se->statistics.block_start,
987 rq_clock(rq_of(cfs_rq)));
992 * We are picking a new current task - update its stats:
995 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
998 * We are starting a new run period:
1000 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1003 /**************************************************
1004 * Scheduling class queueing methods:
1007 #ifdef CONFIG_NUMA_BALANCING
1009 * Approximate time to scan a full NUMA task in ms. The task scan period is
1010 * calculated based on the tasks virtual memory size and
1011 * numa_balancing_scan_size.
1013 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1014 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1016 /* Portion of address space to scan in MB */
1017 unsigned int sysctl_numa_balancing_scan_size = 256;
1019 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1020 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1022 static unsigned int task_nr_scan_windows(struct task_struct *p)
1024 unsigned long rss = 0;
1025 unsigned long nr_scan_pages;
1028 * Calculations based on RSS as non-present and empty pages are skipped
1029 * by the PTE scanner and NUMA hinting faults should be trapped based
1032 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1033 rss = get_mm_rss(p->mm);
1035 rss = nr_scan_pages;
1037 rss = round_up(rss, nr_scan_pages);
1038 return rss / nr_scan_pages;
1041 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1042 #define MAX_SCAN_WINDOW 2560
1044 static unsigned int task_scan_min(struct task_struct *p)
1046 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1047 unsigned int scan, floor;
1048 unsigned int windows = 1;
1050 if (scan_size < MAX_SCAN_WINDOW)
1051 windows = MAX_SCAN_WINDOW / scan_size;
1052 floor = 1000 / windows;
1054 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1055 return max_t(unsigned int, floor, scan);
1058 static unsigned int task_scan_max(struct task_struct *p)
1060 unsigned int smin = task_scan_min(p);
1063 /* Watch for min being lower than max due to floor calculations */
1064 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1065 return max(smin, smax);
1068 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1070 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1071 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1074 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1076 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1077 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1083 spinlock_t lock; /* nr_tasks, tasks */
1088 struct rcu_head rcu;
1089 unsigned long total_faults;
1090 unsigned long max_faults_cpu;
1092 * Faults_cpu is used to decide whether memory should move
1093 * towards the CPU. As a consequence, these stats are weighted
1094 * more by CPU use than by memory faults.
1096 unsigned long *faults_cpu;
1097 unsigned long faults[0];
1100 /* Shared or private faults. */
1101 #define NR_NUMA_HINT_FAULT_TYPES 2
1103 /* Memory and CPU locality */
1104 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1106 /* Averaged statistics, and temporary buffers. */
1107 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1109 pid_t task_numa_group_id(struct task_struct *p)
1111 return p->numa_group ? p->numa_group->gid : 0;
1115 * The averaged statistics, shared & private, memory & cpu,
1116 * occupy the first half of the array. The second half of the
1117 * array is for current counters, which are averaged into the
1118 * first set by task_numa_placement.
1120 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1122 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1125 static inline unsigned long task_faults(struct task_struct *p, int nid)
1127 if (!p->numa_faults)
1130 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1131 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1134 static inline unsigned long group_faults(struct task_struct *p, int nid)
1139 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1140 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1143 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1145 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1146 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1150 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1151 * considered part of a numa group's pseudo-interleaving set. Migrations
1152 * between these nodes are slowed down, to allow things to settle down.
1154 #define ACTIVE_NODE_FRACTION 3
1156 static bool numa_is_active_node(int nid, struct numa_group *ng)
1158 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1161 /* Handle placement on systems where not all nodes are directly connected. */
1162 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1163 int maxdist, bool task)
1165 unsigned long score = 0;
1169 * All nodes are directly connected, and the same distance
1170 * from each other. No need for fancy placement algorithms.
1172 if (sched_numa_topology_type == NUMA_DIRECT)
1176 * This code is called for each node, introducing N^2 complexity,
1177 * which should be ok given the number of nodes rarely exceeds 8.
1179 for_each_online_node(node) {
1180 unsigned long faults;
1181 int dist = node_distance(nid, node);
1184 * The furthest away nodes in the system are not interesting
1185 * for placement; nid was already counted.
1187 if (dist == sched_max_numa_distance || node == nid)
1191 * On systems with a backplane NUMA topology, compare groups
1192 * of nodes, and move tasks towards the group with the most
1193 * memory accesses. When comparing two nodes at distance
1194 * "hoplimit", only nodes closer by than "hoplimit" are part
1195 * of each group. Skip other nodes.
1197 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1201 /* Add up the faults from nearby nodes. */
1203 faults = task_faults(p, node);
1205 faults = group_faults(p, node);
1208 * On systems with a glueless mesh NUMA topology, there are
1209 * no fixed "groups of nodes". Instead, nodes that are not
1210 * directly connected bounce traffic through intermediate
1211 * nodes; a numa_group can occupy any set of nodes.
1212 * The further away a node is, the less the faults count.
1213 * This seems to result in good task placement.
1215 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1216 faults *= (sched_max_numa_distance - dist);
1217 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1227 * These return the fraction of accesses done by a particular task, or
1228 * task group, on a particular numa node. The group weight is given a
1229 * larger multiplier, in order to group tasks together that are almost
1230 * evenly spread out between numa nodes.
1232 static inline unsigned long task_weight(struct task_struct *p, int nid,
1235 unsigned long faults, total_faults;
1237 if (!p->numa_faults)
1240 total_faults = p->total_numa_faults;
1245 faults = task_faults(p, nid);
1246 faults += score_nearby_nodes(p, nid, dist, true);
1248 return 1000 * faults / total_faults;
1251 static inline unsigned long group_weight(struct task_struct *p, int nid,
1254 unsigned long faults, total_faults;
1259 total_faults = p->numa_group->total_faults;
1264 faults = group_faults(p, nid);
1265 faults += score_nearby_nodes(p, nid, dist, false);
1267 return 1000 * faults / total_faults;
1270 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1271 int src_nid, int dst_cpu)
1273 struct numa_group *ng = p->numa_group;
1274 int dst_nid = cpu_to_node(dst_cpu);
1275 int last_cpupid, this_cpupid;
1277 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1280 * Multi-stage node selection is used in conjunction with a periodic
1281 * migration fault to build a temporal task<->page relation. By using
1282 * a two-stage filter we remove short/unlikely relations.
1284 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1285 * a task's usage of a particular page (n_p) per total usage of this
1286 * page (n_t) (in a given time-span) to a probability.
1288 * Our periodic faults will sample this probability and getting the
1289 * same result twice in a row, given these samples are fully
1290 * independent, is then given by P(n)^2, provided our sample period
1291 * is sufficiently short compared to the usage pattern.
1293 * This quadric squishes small probabilities, making it less likely we
1294 * act on an unlikely task<->page relation.
1296 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1297 if (!cpupid_pid_unset(last_cpupid) &&
1298 cpupid_to_nid(last_cpupid) != dst_nid)
1301 /* Always allow migrate on private faults */
1302 if (cpupid_match_pid(p, last_cpupid))
1305 /* A shared fault, but p->numa_group has not been set up yet. */
1310 * Destination node is much more heavily used than the source
1311 * node? Allow migration.
1313 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1314 ACTIVE_NODE_FRACTION)
1318 * Distribute memory according to CPU & memory use on each node,
1319 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1321 * faults_cpu(dst) 3 faults_cpu(src)
1322 * --------------- * - > ---------------
1323 * faults_mem(dst) 4 faults_mem(src)
1325 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1326 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1329 static unsigned long weighted_cpuload(const int cpu);
1330 static unsigned long source_load(int cpu, int type);
1331 static unsigned long target_load(int cpu, int type);
1332 static unsigned long capacity_of(int cpu);
1333 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1335 /* Cached statistics for all CPUs within a node */
1337 unsigned long nr_running;
1340 /* Total compute capacity of CPUs on a node */
1341 unsigned long compute_capacity;
1343 /* Approximate capacity in terms of runnable tasks on a node */
1344 unsigned long task_capacity;
1345 int has_free_capacity;
1349 * XXX borrowed from update_sg_lb_stats
1351 static void update_numa_stats(struct numa_stats *ns, int nid)
1353 int smt, cpu, cpus = 0;
1354 unsigned long capacity;
1356 memset(ns, 0, sizeof(*ns));
1357 for_each_cpu(cpu, cpumask_of_node(nid)) {
1358 struct rq *rq = cpu_rq(cpu);
1360 ns->nr_running += rq->nr_running;
1361 ns->load += weighted_cpuload(cpu);
1362 ns->compute_capacity += capacity_of(cpu);
1368 * If we raced with hotplug and there are no CPUs left in our mask
1369 * the @ns structure is NULL'ed and task_numa_compare() will
1370 * not find this node attractive.
1372 * We'll either bail at !has_free_capacity, or we'll detect a huge
1373 * imbalance and bail there.
1378 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1379 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1380 capacity = cpus / smt; /* cores */
1382 ns->task_capacity = min_t(unsigned, capacity,
1383 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1384 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1387 struct task_numa_env {
1388 struct task_struct *p;
1390 int src_cpu, src_nid;
1391 int dst_cpu, dst_nid;
1393 struct numa_stats src_stats, dst_stats;
1398 struct task_struct *best_task;
1403 static void task_numa_assign(struct task_numa_env *env,
1404 struct task_struct *p, long imp)
1407 put_task_struct(env->best_task);
1412 env->best_imp = imp;
1413 env->best_cpu = env->dst_cpu;
1416 static bool load_too_imbalanced(long src_load, long dst_load,
1417 struct task_numa_env *env)
1420 long orig_src_load, orig_dst_load;
1421 long src_capacity, dst_capacity;
1424 * The load is corrected for the CPU capacity available on each node.
1427 * ------------ vs ---------
1428 * src_capacity dst_capacity
1430 src_capacity = env->src_stats.compute_capacity;
1431 dst_capacity = env->dst_stats.compute_capacity;
1433 /* We care about the slope of the imbalance, not the direction. */
1434 if (dst_load < src_load)
1435 swap(dst_load, src_load);
1437 /* Is the difference below the threshold? */
1438 imb = dst_load * src_capacity * 100 -
1439 src_load * dst_capacity * env->imbalance_pct;
1444 * The imbalance is above the allowed threshold.
1445 * Compare it with the old imbalance.
1447 orig_src_load = env->src_stats.load;
1448 orig_dst_load = env->dst_stats.load;
1450 if (orig_dst_load < orig_src_load)
1451 swap(orig_dst_load, orig_src_load);
1453 old_imb = orig_dst_load * src_capacity * 100 -
1454 orig_src_load * dst_capacity * env->imbalance_pct;
1456 /* Would this change make things worse? */
1457 return (imb > old_imb);
1461 * This checks if the overall compute and NUMA accesses of the system would
1462 * be improved if the source tasks was migrated to the target dst_cpu taking
1463 * into account that it might be best if task running on the dst_cpu should
1464 * be exchanged with the source task
1466 static void task_numa_compare(struct task_numa_env *env,
1467 long taskimp, long groupimp)
1469 struct rq *src_rq = cpu_rq(env->src_cpu);
1470 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1471 struct task_struct *cur;
1472 long src_load, dst_load;
1474 long imp = env->p->numa_group ? groupimp : taskimp;
1476 int dist = env->dist;
1479 cur = task_rcu_dereference(&dst_rq->curr);
1480 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1484 * Because we have preemption enabled we can get migrated around and
1485 * end try selecting ourselves (current == env->p) as a swap candidate.
1491 * "imp" is the fault differential for the source task between the
1492 * source and destination node. Calculate the total differential for
1493 * the source task and potential destination task. The more negative
1494 * the value is, the more rmeote accesses that would be expected to
1495 * be incurred if the tasks were swapped.
1498 /* Skip this swap candidate if cannot move to the source cpu */
1499 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1503 * If dst and source tasks are in the same NUMA group, or not
1504 * in any group then look only at task weights.
1506 if (cur->numa_group == env->p->numa_group) {
1507 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1508 task_weight(cur, env->dst_nid, dist);
1510 * Add some hysteresis to prevent swapping the
1511 * tasks within a group over tiny differences.
1513 if (cur->numa_group)
1517 * Compare the group weights. If a task is all by
1518 * itself (not part of a group), use the task weight
1521 if (cur->numa_group)
1522 imp += group_weight(cur, env->src_nid, dist) -
1523 group_weight(cur, env->dst_nid, dist);
1525 imp += task_weight(cur, env->src_nid, dist) -
1526 task_weight(cur, env->dst_nid, dist);
1530 if (imp <= env->best_imp && moveimp <= env->best_imp)
1534 /* Is there capacity at our destination? */
1535 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1536 !env->dst_stats.has_free_capacity)
1542 /* Balance doesn't matter much if we're running a task per cpu */
1543 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1544 dst_rq->nr_running == 1)
1548 * In the overloaded case, try and keep the load balanced.
1551 load = task_h_load(env->p);
1552 dst_load = env->dst_stats.load + load;
1553 src_load = env->src_stats.load - load;
1555 if (moveimp > imp && moveimp > env->best_imp) {
1557 * If the improvement from just moving env->p direction is
1558 * better than swapping tasks around, check if a move is
1559 * possible. Store a slightly smaller score than moveimp,
1560 * so an actually idle CPU will win.
1562 if (!load_too_imbalanced(src_load, dst_load, env)) {
1569 if (imp <= env->best_imp)
1573 load = task_h_load(cur);
1578 if (load_too_imbalanced(src_load, dst_load, env))
1582 * One idle CPU per node is evaluated for a task numa move.
1583 * Call select_idle_sibling to maybe find a better one.
1586 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1590 task_numa_assign(env, cur, imp);
1595 static void task_numa_find_cpu(struct task_numa_env *env,
1596 long taskimp, long groupimp)
1600 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1601 /* Skip this CPU if the source task cannot migrate */
1602 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1606 task_numa_compare(env, taskimp, groupimp);
1610 /* Only move tasks to a NUMA node less busy than the current node. */
1611 static bool numa_has_capacity(struct task_numa_env *env)
1613 struct numa_stats *src = &env->src_stats;
1614 struct numa_stats *dst = &env->dst_stats;
1616 if (src->has_free_capacity && !dst->has_free_capacity)
1620 * Only consider a task move if the source has a higher load
1621 * than the destination, corrected for CPU capacity on each node.
1623 * src->load dst->load
1624 * --------------------- vs ---------------------
1625 * src->compute_capacity dst->compute_capacity
1627 if (src->load * dst->compute_capacity * env->imbalance_pct >
1629 dst->load * src->compute_capacity * 100)
1635 static int task_numa_migrate(struct task_struct *p)
1637 struct task_numa_env env = {
1640 .src_cpu = task_cpu(p),
1641 .src_nid = task_node(p),
1643 .imbalance_pct = 112,
1649 struct sched_domain *sd;
1650 unsigned long taskweight, groupweight;
1652 long taskimp, groupimp;
1655 * Pick the lowest SD_NUMA domain, as that would have the smallest
1656 * imbalance and would be the first to start moving tasks about.
1658 * And we want to avoid any moving of tasks about, as that would create
1659 * random movement of tasks -- counter the numa conditions we're trying
1663 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1665 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1669 * Cpusets can break the scheduler domain tree into smaller
1670 * balance domains, some of which do not cross NUMA boundaries.
1671 * Tasks that are "trapped" in such domains cannot be migrated
1672 * elsewhere, so there is no point in (re)trying.
1674 if (unlikely(!sd)) {
1675 p->numa_preferred_nid = task_node(p);
1679 env.dst_nid = p->numa_preferred_nid;
1680 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1681 taskweight = task_weight(p, env.src_nid, dist);
1682 groupweight = group_weight(p, env.src_nid, dist);
1683 update_numa_stats(&env.src_stats, env.src_nid);
1684 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1685 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1686 update_numa_stats(&env.dst_stats, env.dst_nid);
1688 /* Try to find a spot on the preferred nid. */
1689 if (numa_has_capacity(&env))
1690 task_numa_find_cpu(&env, taskimp, groupimp);
1693 * Look at other nodes in these cases:
1694 * - there is no space available on the preferred_nid
1695 * - the task is part of a numa_group that is interleaved across
1696 * multiple NUMA nodes; in order to better consolidate the group,
1697 * we need to check other locations.
1699 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1700 for_each_online_node(nid) {
1701 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1704 dist = node_distance(env.src_nid, env.dst_nid);
1705 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1707 taskweight = task_weight(p, env.src_nid, dist);
1708 groupweight = group_weight(p, env.src_nid, dist);
1711 /* Only consider nodes where both task and groups benefit */
1712 taskimp = task_weight(p, nid, dist) - taskweight;
1713 groupimp = group_weight(p, nid, dist) - groupweight;
1714 if (taskimp < 0 && groupimp < 0)
1719 update_numa_stats(&env.dst_stats, env.dst_nid);
1720 if (numa_has_capacity(&env))
1721 task_numa_find_cpu(&env, taskimp, groupimp);
1726 * If the task is part of a workload that spans multiple NUMA nodes,
1727 * and is migrating into one of the workload's active nodes, remember
1728 * this node as the task's preferred numa node, so the workload can
1730 * A task that migrated to a second choice node will be better off
1731 * trying for a better one later. Do not set the preferred node here.
1733 if (p->numa_group) {
1734 struct numa_group *ng = p->numa_group;
1736 if (env.best_cpu == -1)
1741 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1742 sched_setnuma(p, env.dst_nid);
1745 /* No better CPU than the current one was found. */
1746 if (env.best_cpu == -1)
1750 * Reset the scan period if the task is being rescheduled on an
1751 * alternative node to recheck if the tasks is now properly placed.
1753 p->numa_scan_period = task_scan_min(p);
1755 if (env.best_task == NULL) {
1756 ret = migrate_task_to(p, env.best_cpu);
1758 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1762 ret = migrate_swap(p, env.best_task);
1764 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1765 put_task_struct(env.best_task);
1769 /* Attempt to migrate a task to a CPU on the preferred node. */
1770 static void numa_migrate_preferred(struct task_struct *p)
1772 unsigned long interval = HZ;
1774 /* This task has no NUMA fault statistics yet */
1775 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1778 /* Periodically retry migrating the task to the preferred node */
1779 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1780 p->numa_migrate_retry = jiffies + interval;
1782 /* Success if task is already running on preferred CPU */
1783 if (task_node(p) == p->numa_preferred_nid)
1786 /* Otherwise, try migrate to a CPU on the preferred node */
1787 task_numa_migrate(p);
1791 * Find out how many nodes on the workload is actively running on. Do this by
1792 * tracking the nodes from which NUMA hinting faults are triggered. This can
1793 * be different from the set of nodes where the workload's memory is currently
1796 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1798 unsigned long faults, max_faults = 0;
1799 int nid, active_nodes = 0;
1801 for_each_online_node(nid) {
1802 faults = group_faults_cpu(numa_group, nid);
1803 if (faults > max_faults)
1804 max_faults = faults;
1807 for_each_online_node(nid) {
1808 faults = group_faults_cpu(numa_group, nid);
1809 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1813 numa_group->max_faults_cpu = max_faults;
1814 numa_group->active_nodes = active_nodes;
1818 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1819 * increments. The more local the fault statistics are, the higher the scan
1820 * period will be for the next scan window. If local/(local+remote) ratio is
1821 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1822 * the scan period will decrease. Aim for 70% local accesses.
1824 #define NUMA_PERIOD_SLOTS 10
1825 #define NUMA_PERIOD_THRESHOLD 7
1828 * Increase the scan period (slow down scanning) if the majority of
1829 * our memory is already on our local node, or if the majority of
1830 * the page accesses are shared with other processes.
1831 * Otherwise, decrease the scan period.
1833 static void update_task_scan_period(struct task_struct *p,
1834 unsigned long shared, unsigned long private)
1836 unsigned int period_slot;
1840 unsigned long remote = p->numa_faults_locality[0];
1841 unsigned long local = p->numa_faults_locality[1];
1844 * If there were no record hinting faults then either the task is
1845 * completely idle or all activity is areas that are not of interest
1846 * to automatic numa balancing. Related to that, if there were failed
1847 * migration then it implies we are migrating too quickly or the local
1848 * node is overloaded. In either case, scan slower
1850 if (local + shared == 0 || p->numa_faults_locality[2]) {
1851 p->numa_scan_period = min(p->numa_scan_period_max,
1852 p->numa_scan_period << 1);
1854 p->mm->numa_next_scan = jiffies +
1855 msecs_to_jiffies(p->numa_scan_period);
1861 * Prepare to scale scan period relative to the current period.
1862 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1863 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1864 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1866 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1867 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1868 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1869 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1872 diff = slot * period_slot;
1874 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1877 * Scale scan rate increases based on sharing. There is an
1878 * inverse relationship between the degree of sharing and
1879 * the adjustment made to the scanning period. Broadly
1880 * speaking the intent is that there is little point
1881 * scanning faster if shared accesses dominate as it may
1882 * simply bounce migrations uselessly
1884 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1885 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1888 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1889 task_scan_min(p), task_scan_max(p));
1890 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1894 * Get the fraction of time the task has been running since the last
1895 * NUMA placement cycle. The scheduler keeps similar statistics, but
1896 * decays those on a 32ms period, which is orders of magnitude off
1897 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1898 * stats only if the task is so new there are no NUMA statistics yet.
1900 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1902 u64 runtime, delta, now;
1903 /* Use the start of this time slice to avoid calculations. */
1904 now = p->se.exec_start;
1905 runtime = p->se.sum_exec_runtime;
1907 if (p->last_task_numa_placement) {
1908 delta = runtime - p->last_sum_exec_runtime;
1909 *period = now - p->last_task_numa_placement;
1911 delta = p->se.avg.load_sum / p->se.load.weight;
1912 *period = LOAD_AVG_MAX;
1915 p->last_sum_exec_runtime = runtime;
1916 p->last_task_numa_placement = now;
1922 * Determine the preferred nid for a task in a numa_group. This needs to
1923 * be done in a way that produces consistent results with group_weight,
1924 * otherwise workloads might not converge.
1926 static int preferred_group_nid(struct task_struct *p, int nid)
1931 /* Direct connections between all NUMA nodes. */
1932 if (sched_numa_topology_type == NUMA_DIRECT)
1936 * On a system with glueless mesh NUMA topology, group_weight
1937 * scores nodes according to the number of NUMA hinting faults on
1938 * both the node itself, and on nearby nodes.
1940 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1941 unsigned long score, max_score = 0;
1942 int node, max_node = nid;
1944 dist = sched_max_numa_distance;
1946 for_each_online_node(node) {
1947 score = group_weight(p, node, dist);
1948 if (score > max_score) {
1957 * Finding the preferred nid in a system with NUMA backplane
1958 * interconnect topology is more involved. The goal is to locate
1959 * tasks from numa_groups near each other in the system, and
1960 * untangle workloads from different sides of the system. This requires
1961 * searching down the hierarchy of node groups, recursively searching
1962 * inside the highest scoring group of nodes. The nodemask tricks
1963 * keep the complexity of the search down.
1965 nodes = node_online_map;
1966 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1967 unsigned long max_faults = 0;
1968 nodemask_t max_group = NODE_MASK_NONE;
1971 /* Are there nodes at this distance from each other? */
1972 if (!find_numa_distance(dist))
1975 for_each_node_mask(a, nodes) {
1976 unsigned long faults = 0;
1977 nodemask_t this_group;
1978 nodes_clear(this_group);
1980 /* Sum group's NUMA faults; includes a==b case. */
1981 for_each_node_mask(b, nodes) {
1982 if (node_distance(a, b) < dist) {
1983 faults += group_faults(p, b);
1984 node_set(b, this_group);
1985 node_clear(b, nodes);
1989 /* Remember the top group. */
1990 if (faults > max_faults) {
1991 max_faults = faults;
1992 max_group = this_group;
1994 * subtle: at the smallest distance there is
1995 * just one node left in each "group", the
1996 * winner is the preferred nid.
2001 /* Next round, evaluate the nodes within max_group. */
2009 static void task_numa_placement(struct task_struct *p)
2011 int seq, nid, max_nid = -1, max_group_nid = -1;
2012 unsigned long max_faults = 0, max_group_faults = 0;
2013 unsigned long fault_types[2] = { 0, 0 };
2014 unsigned long total_faults;
2015 u64 runtime, period;
2016 spinlock_t *group_lock = NULL;
2019 * The p->mm->numa_scan_seq field gets updated without
2020 * exclusive access. Use READ_ONCE() here to ensure
2021 * that the field is read in a single access:
2023 seq = READ_ONCE(p->mm->numa_scan_seq);
2024 if (p->numa_scan_seq == seq)
2026 p->numa_scan_seq = seq;
2027 p->numa_scan_period_max = task_scan_max(p);
2029 total_faults = p->numa_faults_locality[0] +
2030 p->numa_faults_locality[1];
2031 runtime = numa_get_avg_runtime(p, &period);
2033 /* If the task is part of a group prevent parallel updates to group stats */
2034 if (p->numa_group) {
2035 group_lock = &p->numa_group->lock;
2036 spin_lock_irq(group_lock);
2039 /* Find the node with the highest number of faults */
2040 for_each_online_node(nid) {
2041 /* Keep track of the offsets in numa_faults array */
2042 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2043 unsigned long faults = 0, group_faults = 0;
2046 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2047 long diff, f_diff, f_weight;
2049 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2050 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2051 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2052 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2054 /* Decay existing window, copy faults since last scan */
2055 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2056 fault_types[priv] += p->numa_faults[membuf_idx];
2057 p->numa_faults[membuf_idx] = 0;
2060 * Normalize the faults_from, so all tasks in a group
2061 * count according to CPU use, instead of by the raw
2062 * number of faults. Tasks with little runtime have
2063 * little over-all impact on throughput, and thus their
2064 * faults are less important.
2066 f_weight = div64_u64(runtime << 16, period + 1);
2067 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2069 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2070 p->numa_faults[cpubuf_idx] = 0;
2072 p->numa_faults[mem_idx] += diff;
2073 p->numa_faults[cpu_idx] += f_diff;
2074 faults += p->numa_faults[mem_idx];
2075 p->total_numa_faults += diff;
2076 if (p->numa_group) {
2078 * safe because we can only change our own group
2080 * mem_idx represents the offset for a given
2081 * nid and priv in a specific region because it
2082 * is at the beginning of the numa_faults array.
2084 p->numa_group->faults[mem_idx] += diff;
2085 p->numa_group->faults_cpu[mem_idx] += f_diff;
2086 p->numa_group->total_faults += diff;
2087 group_faults += p->numa_group->faults[mem_idx];
2091 if (faults > max_faults) {
2092 max_faults = faults;
2096 if (group_faults > max_group_faults) {
2097 max_group_faults = group_faults;
2098 max_group_nid = nid;
2102 update_task_scan_period(p, fault_types[0], fault_types[1]);
2104 if (p->numa_group) {
2105 numa_group_count_active_nodes(p->numa_group);
2106 spin_unlock_irq(group_lock);
2107 max_nid = preferred_group_nid(p, max_group_nid);
2111 /* Set the new preferred node */
2112 if (max_nid != p->numa_preferred_nid)
2113 sched_setnuma(p, max_nid);
2115 if (task_node(p) != p->numa_preferred_nid)
2116 numa_migrate_preferred(p);
2120 static inline int get_numa_group(struct numa_group *grp)
2122 return atomic_inc_not_zero(&grp->refcount);
2125 static inline void put_numa_group(struct numa_group *grp)
2127 if (atomic_dec_and_test(&grp->refcount))
2128 kfree_rcu(grp, rcu);
2131 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2134 struct numa_group *grp, *my_grp;
2135 struct task_struct *tsk;
2137 int cpu = cpupid_to_cpu(cpupid);
2140 if (unlikely(!p->numa_group)) {
2141 unsigned int size = sizeof(struct numa_group) +
2142 4*nr_node_ids*sizeof(unsigned long);
2144 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2148 atomic_set(&grp->refcount, 1);
2149 grp->active_nodes = 1;
2150 grp->max_faults_cpu = 0;
2151 spin_lock_init(&grp->lock);
2153 /* Second half of the array tracks nids where faults happen */
2154 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2157 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2158 grp->faults[i] = p->numa_faults[i];
2160 grp->total_faults = p->total_numa_faults;
2163 rcu_assign_pointer(p->numa_group, grp);
2167 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2169 if (!cpupid_match_pid(tsk, cpupid))
2172 grp = rcu_dereference(tsk->numa_group);
2176 my_grp = p->numa_group;
2181 * Only join the other group if its bigger; if we're the bigger group,
2182 * the other task will join us.
2184 if (my_grp->nr_tasks > grp->nr_tasks)
2188 * Tie-break on the grp address.
2190 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2193 /* Always join threads in the same process. */
2194 if (tsk->mm == current->mm)
2197 /* Simple filter to avoid false positives due to PID collisions */
2198 if (flags & TNF_SHARED)
2201 /* Update priv based on whether false sharing was detected */
2204 if (join && !get_numa_group(grp))
2212 BUG_ON(irqs_disabled());
2213 double_lock_irq(&my_grp->lock, &grp->lock);
2215 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2216 my_grp->faults[i] -= p->numa_faults[i];
2217 grp->faults[i] += p->numa_faults[i];
2219 my_grp->total_faults -= p->total_numa_faults;
2220 grp->total_faults += p->total_numa_faults;
2225 spin_unlock(&my_grp->lock);
2226 spin_unlock_irq(&grp->lock);
2228 rcu_assign_pointer(p->numa_group, grp);
2230 put_numa_group(my_grp);
2238 void task_numa_free(struct task_struct *p)
2240 struct numa_group *grp = p->numa_group;
2241 void *numa_faults = p->numa_faults;
2242 unsigned long flags;
2246 spin_lock_irqsave(&grp->lock, flags);
2247 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2248 grp->faults[i] -= p->numa_faults[i];
2249 grp->total_faults -= p->total_numa_faults;
2252 spin_unlock_irqrestore(&grp->lock, flags);
2253 RCU_INIT_POINTER(p->numa_group, NULL);
2254 put_numa_group(grp);
2257 p->numa_faults = NULL;
2262 * Got a PROT_NONE fault for a page on @node.
2264 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2266 struct task_struct *p = current;
2267 bool migrated = flags & TNF_MIGRATED;
2268 int cpu_node = task_node(current);
2269 int local = !!(flags & TNF_FAULT_LOCAL);
2270 struct numa_group *ng;
2273 if (!static_branch_likely(&sched_numa_balancing))
2276 /* for example, ksmd faulting in a user's mm */
2280 /* Allocate buffer to track faults on a per-node basis */
2281 if (unlikely(!p->numa_faults)) {
2282 int size = sizeof(*p->numa_faults) *
2283 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2285 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2286 if (!p->numa_faults)
2289 p->total_numa_faults = 0;
2290 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2294 * First accesses are treated as private, otherwise consider accesses
2295 * to be private if the accessing pid has not changed
2297 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2300 priv = cpupid_match_pid(p, last_cpupid);
2301 if (!priv && !(flags & TNF_NO_GROUP))
2302 task_numa_group(p, last_cpupid, flags, &priv);
2306 * If a workload spans multiple NUMA nodes, a shared fault that
2307 * occurs wholly within the set of nodes that the workload is
2308 * actively using should be counted as local. This allows the
2309 * scan rate to slow down when a workload has settled down.
2312 if (!priv && !local && ng && ng->active_nodes > 1 &&
2313 numa_is_active_node(cpu_node, ng) &&
2314 numa_is_active_node(mem_node, ng))
2317 task_numa_placement(p);
2320 * Retry task to preferred node migration periodically, in case it
2321 * case it previously failed, or the scheduler moved us.
2323 if (time_after(jiffies, p->numa_migrate_retry))
2324 numa_migrate_preferred(p);
2327 p->numa_pages_migrated += pages;
2328 if (flags & TNF_MIGRATE_FAIL)
2329 p->numa_faults_locality[2] += pages;
2331 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2332 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2333 p->numa_faults_locality[local] += pages;
2336 static void reset_ptenuma_scan(struct task_struct *p)
2339 * We only did a read acquisition of the mmap sem, so
2340 * p->mm->numa_scan_seq is written to without exclusive access
2341 * and the update is not guaranteed to be atomic. That's not
2342 * much of an issue though, since this is just used for
2343 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2344 * expensive, to avoid any form of compiler optimizations:
2346 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2347 p->mm->numa_scan_offset = 0;
2351 * The expensive part of numa migration is done from task_work context.
2352 * Triggered from task_tick_numa().
2354 void task_numa_work(struct callback_head *work)
2356 unsigned long migrate, next_scan, now = jiffies;
2357 struct task_struct *p = current;
2358 struct mm_struct *mm = p->mm;
2359 u64 runtime = p->se.sum_exec_runtime;
2360 struct vm_area_struct *vma;
2361 unsigned long start, end;
2362 unsigned long nr_pte_updates = 0;
2363 long pages, virtpages;
2365 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2367 work->next = work; /* protect against double add */
2369 * Who cares about NUMA placement when they're dying.
2371 * NOTE: make sure not to dereference p->mm before this check,
2372 * exit_task_work() happens _after_ exit_mm() so we could be called
2373 * without p->mm even though we still had it when we enqueued this
2376 if (p->flags & PF_EXITING)
2379 if (!mm->numa_next_scan) {
2380 mm->numa_next_scan = now +
2381 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2385 * Enforce maximal scan/migration frequency..
2387 migrate = mm->numa_next_scan;
2388 if (time_before(now, migrate))
2391 if (p->numa_scan_period == 0) {
2392 p->numa_scan_period_max = task_scan_max(p);
2393 p->numa_scan_period = task_scan_min(p);
2396 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2397 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2401 * Delay this task enough that another task of this mm will likely win
2402 * the next time around.
2404 p->node_stamp += 2 * TICK_NSEC;
2406 start = mm->numa_scan_offset;
2407 pages = sysctl_numa_balancing_scan_size;
2408 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2409 virtpages = pages * 8; /* Scan up to this much virtual space */
2414 down_read(&mm->mmap_sem);
2415 vma = find_vma(mm, start);
2417 reset_ptenuma_scan(p);
2421 for (; vma; vma = vma->vm_next) {
2422 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2423 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2428 * Shared library pages mapped by multiple processes are not
2429 * migrated as it is expected they are cache replicated. Avoid
2430 * hinting faults in read-only file-backed mappings or the vdso
2431 * as migrating the pages will be of marginal benefit.
2434 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2438 * Skip inaccessible VMAs to avoid any confusion between
2439 * PROT_NONE and NUMA hinting ptes
2441 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2445 start = max(start, vma->vm_start);
2446 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2447 end = min(end, vma->vm_end);
2448 nr_pte_updates = change_prot_numa(vma, start, end);
2451 * Try to scan sysctl_numa_balancing_size worth of
2452 * hpages that have at least one present PTE that
2453 * is not already pte-numa. If the VMA contains
2454 * areas that are unused or already full of prot_numa
2455 * PTEs, scan up to virtpages, to skip through those
2459 pages -= (end - start) >> PAGE_SHIFT;
2460 virtpages -= (end - start) >> PAGE_SHIFT;
2463 if (pages <= 0 || virtpages <= 0)
2467 } while (end != vma->vm_end);
2472 * It is possible to reach the end of the VMA list but the last few
2473 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2474 * would find the !migratable VMA on the next scan but not reset the
2475 * scanner to the start so check it now.
2478 mm->numa_scan_offset = start;
2480 reset_ptenuma_scan(p);
2481 up_read(&mm->mmap_sem);
2484 * Make sure tasks use at least 32x as much time to run other code
2485 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2486 * Usually update_task_scan_period slows down scanning enough; on an
2487 * overloaded system we need to limit overhead on a per task basis.
2489 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2490 u64 diff = p->se.sum_exec_runtime - runtime;
2491 p->node_stamp += 32 * diff;
2496 * Drive the periodic memory faults..
2498 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2500 struct callback_head *work = &curr->numa_work;
2504 * We don't care about NUMA placement if we don't have memory.
2506 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2510 * Using runtime rather than walltime has the dual advantage that
2511 * we (mostly) drive the selection from busy threads and that the
2512 * task needs to have done some actual work before we bother with
2515 now = curr->se.sum_exec_runtime;
2516 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2518 if (now > curr->node_stamp + period) {
2519 if (!curr->node_stamp)
2520 curr->numa_scan_period = task_scan_min(curr);
2521 curr->node_stamp += period;
2523 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2524 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2525 task_work_add(curr, work, true);
2530 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2534 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2538 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2541 #endif /* CONFIG_NUMA_BALANCING */
2544 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2546 update_load_add(&cfs_rq->load, se->load.weight);
2547 if (!parent_entity(se))
2548 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2550 if (entity_is_task(se)) {
2551 struct rq *rq = rq_of(cfs_rq);
2553 account_numa_enqueue(rq, task_of(se));
2554 list_add(&se->group_node, &rq->cfs_tasks);
2557 cfs_rq->nr_running++;
2561 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2563 update_load_sub(&cfs_rq->load, se->load.weight);
2564 if (!parent_entity(se))
2565 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2567 if (entity_is_task(se)) {
2568 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2569 list_del_init(&se->group_node);
2572 cfs_rq->nr_running--;
2575 #ifdef CONFIG_FAIR_GROUP_SCHED
2577 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2579 long tg_weight, load, shares;
2582 * This really should be: cfs_rq->avg.load_avg, but instead we use
2583 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2584 * the shares for small weight interactive tasks.
2586 load = scale_load_down(cfs_rq->load.weight);
2588 tg_weight = atomic_long_read(&tg->load_avg);
2590 /* Ensure tg_weight >= load */
2591 tg_weight -= cfs_rq->tg_load_avg_contrib;
2594 shares = (tg->shares * load);
2596 shares /= tg_weight;
2598 if (shares < MIN_SHARES)
2599 shares = MIN_SHARES;
2600 if (shares > tg->shares)
2601 shares = tg->shares;
2605 # else /* CONFIG_SMP */
2606 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2610 # endif /* CONFIG_SMP */
2612 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2613 unsigned long weight)
2616 /* commit outstanding execution time */
2617 if (cfs_rq->curr == se)
2618 update_curr(cfs_rq);
2619 account_entity_dequeue(cfs_rq, se);
2622 update_load_set(&se->load, weight);
2625 account_entity_enqueue(cfs_rq, se);
2628 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2630 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2632 struct task_group *tg;
2633 struct sched_entity *se;
2637 se = tg->se[cpu_of(rq_of(cfs_rq))];
2638 if (!se || throttled_hierarchy(cfs_rq))
2641 if (likely(se->load.weight == tg->shares))
2644 shares = calc_cfs_shares(cfs_rq, tg);
2646 reweight_entity(cfs_rq_of(se), se, shares);
2648 #else /* CONFIG_FAIR_GROUP_SCHED */
2649 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2652 #endif /* CONFIG_FAIR_GROUP_SCHED */
2655 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2656 static const u32 runnable_avg_yN_inv[] = {
2657 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2658 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2659 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2660 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2661 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2662 0x85aac367, 0x82cd8698,
2666 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2667 * over-estimates when re-combining.
2669 static const u32 runnable_avg_yN_sum[] = {
2670 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2671 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2672 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2676 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2677 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2680 static const u32 __accumulated_sum_N32[] = {
2681 0, 23371, 35056, 40899, 43820, 45281,
2682 46011, 46376, 46559, 46650, 46696, 46719,
2687 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2689 static __always_inline u64 decay_load(u64 val, u64 n)
2691 unsigned int local_n;
2695 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2698 /* after bounds checking we can collapse to 32-bit */
2702 * As y^PERIOD = 1/2, we can combine
2703 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2704 * With a look-up table which covers y^n (n<PERIOD)
2706 * To achieve constant time decay_load.
2708 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2709 val >>= local_n / LOAD_AVG_PERIOD;
2710 local_n %= LOAD_AVG_PERIOD;
2713 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2718 * For updates fully spanning n periods, the contribution to runnable
2719 * average will be: \Sum 1024*y^n
2721 * We can compute this reasonably efficiently by combining:
2722 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2724 static u32 __compute_runnable_contrib(u64 n)
2728 if (likely(n <= LOAD_AVG_PERIOD))
2729 return runnable_avg_yN_sum[n];
2730 else if (unlikely(n >= LOAD_AVG_MAX_N))
2731 return LOAD_AVG_MAX;
2733 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2734 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2735 n %= LOAD_AVG_PERIOD;
2736 contrib = decay_load(contrib, n);
2737 return contrib + runnable_avg_yN_sum[n];
2740 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2743 * We can represent the historical contribution to runnable average as the
2744 * coefficients of a geometric series. To do this we sub-divide our runnable
2745 * history into segments of approximately 1ms (1024us); label the segment that
2746 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2748 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2750 * (now) (~1ms ago) (~2ms ago)
2752 * Let u_i denote the fraction of p_i that the entity was runnable.
2754 * We then designate the fractions u_i as our co-efficients, yielding the
2755 * following representation of historical load:
2756 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2758 * We choose y based on the with of a reasonably scheduling period, fixing:
2761 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2762 * approximately half as much as the contribution to load within the last ms
2765 * When a period "rolls over" and we have new u_0`, multiplying the previous
2766 * sum again by y is sufficient to update:
2767 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2768 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2770 static __always_inline int
2771 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2772 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2774 u64 delta, scaled_delta, periods;
2776 unsigned int delta_w, scaled_delta_w, decayed = 0;
2777 unsigned long scale_freq, scale_cpu;
2779 delta = now - sa->last_update_time;
2781 * This should only happen when time goes backwards, which it
2782 * unfortunately does during sched clock init when we swap over to TSC.
2784 if ((s64)delta < 0) {
2785 sa->last_update_time = now;
2790 * Use 1024ns as the unit of measurement since it's a reasonable
2791 * approximation of 1us and fast to compute.
2796 sa->last_update_time = now;
2798 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2799 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2801 /* delta_w is the amount already accumulated against our next period */
2802 delta_w = sa->period_contrib;
2803 if (delta + delta_w >= 1024) {
2806 /* how much left for next period will start over, we don't know yet */
2807 sa->period_contrib = 0;
2810 * Now that we know we're crossing a period boundary, figure
2811 * out how much from delta we need to complete the current
2812 * period and accrue it.
2814 delta_w = 1024 - delta_w;
2815 scaled_delta_w = cap_scale(delta_w, scale_freq);
2817 sa->load_sum += weight * scaled_delta_w;
2819 cfs_rq->runnable_load_sum +=
2820 weight * scaled_delta_w;
2824 sa->util_sum += scaled_delta_w * scale_cpu;
2828 /* Figure out how many additional periods this update spans */
2829 periods = delta / 1024;
2832 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2834 cfs_rq->runnable_load_sum =
2835 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2837 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2839 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2840 contrib = __compute_runnable_contrib(periods);
2841 contrib = cap_scale(contrib, scale_freq);
2843 sa->load_sum += weight * contrib;
2845 cfs_rq->runnable_load_sum += weight * contrib;
2848 sa->util_sum += contrib * scale_cpu;
2851 /* Remainder of delta accrued against u_0` */
2852 scaled_delta = cap_scale(delta, scale_freq);
2854 sa->load_sum += weight * scaled_delta;
2856 cfs_rq->runnable_load_sum += weight * scaled_delta;
2859 sa->util_sum += scaled_delta * scale_cpu;
2861 sa->period_contrib += delta;
2864 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2866 cfs_rq->runnable_load_avg =
2867 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2869 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2875 #ifdef CONFIG_FAIR_GROUP_SCHED
2877 * update_tg_load_avg - update the tg's load avg
2878 * @cfs_rq: the cfs_rq whose avg changed
2879 * @force: update regardless of how small the difference
2881 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2882 * However, because tg->load_avg is a global value there are performance
2885 * In order to avoid having to look at the other cfs_rq's, we use a
2886 * differential update where we store the last value we propagated. This in
2887 * turn allows skipping updates if the differential is 'small'.
2889 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2890 * done) and effective_load() (which is not done because it is too costly).
2892 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2894 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2897 * No need to update load_avg for root_task_group as it is not used.
2899 if (cfs_rq->tg == &root_task_group)
2902 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2903 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2904 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2909 * Called within set_task_rq() right before setting a task's cpu. The
2910 * caller only guarantees p->pi_lock is held; no other assumptions,
2911 * including the state of rq->lock, should be made.
2913 void set_task_rq_fair(struct sched_entity *se,
2914 struct cfs_rq *prev, struct cfs_rq *next)
2916 if (!sched_feat(ATTACH_AGE_LOAD))
2920 * We are supposed to update the task to "current" time, then its up to
2921 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2922 * getting what current time is, so simply throw away the out-of-date
2923 * time. This will result in the wakee task is less decayed, but giving
2924 * the wakee more load sounds not bad.
2926 if (se->avg.last_update_time && prev) {
2927 u64 p_last_update_time;
2928 u64 n_last_update_time;
2930 #ifndef CONFIG_64BIT
2931 u64 p_last_update_time_copy;
2932 u64 n_last_update_time_copy;
2935 p_last_update_time_copy = prev->load_last_update_time_copy;
2936 n_last_update_time_copy = next->load_last_update_time_copy;
2940 p_last_update_time = prev->avg.last_update_time;
2941 n_last_update_time = next->avg.last_update_time;
2943 } while (p_last_update_time != p_last_update_time_copy ||
2944 n_last_update_time != n_last_update_time_copy);
2946 p_last_update_time = prev->avg.last_update_time;
2947 n_last_update_time = next->avg.last_update_time;
2949 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2950 &se->avg, 0, 0, NULL);
2951 se->avg.last_update_time = n_last_update_time;
2954 #else /* CONFIG_FAIR_GROUP_SCHED */
2955 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2956 #endif /* CONFIG_FAIR_GROUP_SCHED */
2958 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2960 struct rq *rq = rq_of(cfs_rq);
2961 int cpu = cpu_of(rq);
2963 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2964 unsigned long max = rq->cpu_capacity_orig;
2967 * There are a few boundary cases this might miss but it should
2968 * get called often enough that that should (hopefully) not be
2969 * a real problem -- added to that it only calls on the local
2970 * CPU, so if we enqueue remotely we'll miss an update, but
2971 * the next tick/schedule should update.
2973 * It will not get called when we go idle, because the idle
2974 * thread is a different class (!fair), nor will the utilization
2975 * number include things like RT tasks.
2977 * As is, the util number is not freq-invariant (we'd have to
2978 * implement arch_scale_freq_capacity() for that).
2982 cpufreq_update_util(rq_clock(rq),
2983 min(cfs_rq->avg.util_avg, max), max);
2988 * Unsigned subtract and clamp on underflow.
2990 * Explicitly do a load-store to ensure the intermediate value never hits
2991 * memory. This allows lockless observations without ever seeing the negative
2994 #define sub_positive(_ptr, _val) do { \
2995 typeof(_ptr) ptr = (_ptr); \
2996 typeof(*ptr) val = (_val); \
2997 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3001 WRITE_ONCE(*ptr, res); \
3005 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3006 * @now: current time, as per cfs_rq_clock_task()
3007 * @cfs_rq: cfs_rq to update
3008 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3010 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3011 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3012 * post_init_entity_util_avg().
3014 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3016 * Returns true if the load decayed or we removed load.
3018 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3019 * call update_tg_load_avg() when this function returns true.
3022 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3024 struct sched_avg *sa = &cfs_rq->avg;
3025 int decayed, removed_load = 0, removed_util = 0;
3027 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3028 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3029 sub_positive(&sa->load_avg, r);
3030 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3034 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3035 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3036 sub_positive(&sa->util_avg, r);
3037 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3041 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3042 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3044 #ifndef CONFIG_64BIT
3046 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3049 if (update_freq && (decayed || removed_util))
3050 cfs_rq_util_change(cfs_rq);
3052 return decayed || removed_load;
3055 /* Update task and its cfs_rq load average */
3056 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3058 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3059 u64 now = cfs_rq_clock_task(cfs_rq);
3060 struct rq *rq = rq_of(cfs_rq);
3061 int cpu = cpu_of(rq);
3064 * Track task load average for carrying it to new CPU after migrated, and
3065 * track group sched_entity load average for task_h_load calc in migration
3067 __update_load_avg(now, cpu, &se->avg,
3068 se->on_rq * scale_load_down(se->load.weight),
3069 cfs_rq->curr == se, NULL);
3071 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
3072 update_tg_load_avg(cfs_rq, 0);
3076 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3077 * @cfs_rq: cfs_rq to attach to
3078 * @se: sched_entity to attach
3080 * Must call update_cfs_rq_load_avg() before this, since we rely on
3081 * cfs_rq->avg.last_update_time being current.
3083 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3085 if (!sched_feat(ATTACH_AGE_LOAD))
3089 * If we got migrated (either between CPUs or between cgroups) we'll
3090 * have aged the average right before clearing @last_update_time.
3092 * Or we're fresh through post_init_entity_util_avg().
3094 if (se->avg.last_update_time) {
3095 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3096 &se->avg, 0, 0, NULL);
3099 * XXX: we could have just aged the entire load away if we've been
3100 * absent from the fair class for too long.
3105 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3106 cfs_rq->avg.load_avg += se->avg.load_avg;
3107 cfs_rq->avg.load_sum += se->avg.load_sum;
3108 cfs_rq->avg.util_avg += se->avg.util_avg;
3109 cfs_rq->avg.util_sum += se->avg.util_sum;
3111 cfs_rq_util_change(cfs_rq);
3115 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3116 * @cfs_rq: cfs_rq to detach from
3117 * @se: sched_entity to detach
3119 * Must call update_cfs_rq_load_avg() before this, since we rely on
3120 * cfs_rq->avg.last_update_time being current.
3122 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3124 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3125 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3126 cfs_rq->curr == se, NULL);
3128 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3129 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3130 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3131 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3133 cfs_rq_util_change(cfs_rq);
3136 /* Add the load generated by se into cfs_rq's load average */
3138 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3140 struct sched_avg *sa = &se->avg;
3141 u64 now = cfs_rq_clock_task(cfs_rq);
3142 int migrated, decayed;
3144 migrated = !sa->last_update_time;
3146 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3147 se->on_rq * scale_load_down(se->load.weight),
3148 cfs_rq->curr == se, NULL);
3151 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3153 cfs_rq->runnable_load_avg += sa->load_avg;
3154 cfs_rq->runnable_load_sum += sa->load_sum;
3157 attach_entity_load_avg(cfs_rq, se);
3159 if (decayed || migrated)
3160 update_tg_load_avg(cfs_rq, 0);
3163 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3165 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3167 update_load_avg(se, 1);
3169 cfs_rq->runnable_load_avg =
3170 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3171 cfs_rq->runnable_load_sum =
3172 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3175 #ifndef CONFIG_64BIT
3176 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3178 u64 last_update_time_copy;
3179 u64 last_update_time;
3182 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3184 last_update_time = cfs_rq->avg.last_update_time;
3185 } while (last_update_time != last_update_time_copy);
3187 return last_update_time;
3190 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3192 return cfs_rq->avg.last_update_time;
3197 * Task first catches up with cfs_rq, and then subtract
3198 * itself from the cfs_rq (task must be off the queue now).
3200 void remove_entity_load_avg(struct sched_entity *se)
3202 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3203 u64 last_update_time;
3206 * tasks cannot exit without having gone through wake_up_new_task() ->
3207 * post_init_entity_util_avg() which will have added things to the
3208 * cfs_rq, so we can remove unconditionally.
3210 * Similarly for groups, they will have passed through
3211 * post_init_entity_util_avg() before unregister_sched_fair_group()
3215 last_update_time = cfs_rq_last_update_time(cfs_rq);
3217 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3218 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3219 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3222 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3224 return cfs_rq->runnable_load_avg;
3227 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3229 return cfs_rq->avg.load_avg;
3232 static int idle_balance(struct rq *this_rq);
3234 #else /* CONFIG_SMP */
3237 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3242 static inline void update_load_avg(struct sched_entity *se, int not_used)
3244 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3245 struct rq *rq = rq_of(cfs_rq);
3247 cpufreq_trigger_update(rq_clock(rq));
3251 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3253 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3254 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3257 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3259 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3261 static inline int idle_balance(struct rq *rq)
3266 #endif /* CONFIG_SMP */
3268 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3270 #ifdef CONFIG_SCHED_DEBUG
3271 s64 d = se->vruntime - cfs_rq->min_vruntime;
3276 if (d > 3*sysctl_sched_latency)
3277 schedstat_inc(cfs_rq->nr_spread_over);
3282 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3284 u64 vruntime = cfs_rq->min_vruntime;
3287 * The 'current' period is already promised to the current tasks,
3288 * however the extra weight of the new task will slow them down a
3289 * little, place the new task so that it fits in the slot that
3290 * stays open at the end.
3292 if (initial && sched_feat(START_DEBIT))
3293 vruntime += sched_vslice(cfs_rq, se);
3295 /* sleeps up to a single latency don't count. */
3297 unsigned long thresh = sysctl_sched_latency;
3300 * Halve their sleep time's effect, to allow
3301 * for a gentler effect of sleepers:
3303 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3309 /* ensure we never gain time by being placed backwards. */
3310 se->vruntime = max_vruntime(se->vruntime, vruntime);
3313 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3315 static inline void check_schedstat_required(void)
3317 #ifdef CONFIG_SCHEDSTATS
3318 if (schedstat_enabled())
3321 /* Force schedstat enabled if a dependent tracepoint is active */
3322 if (trace_sched_stat_wait_enabled() ||
3323 trace_sched_stat_sleep_enabled() ||
3324 trace_sched_stat_iowait_enabled() ||
3325 trace_sched_stat_blocked_enabled() ||
3326 trace_sched_stat_runtime_enabled()) {
3327 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3328 "stat_blocked and stat_runtime require the "
3329 "kernel parameter schedstats=enabled or "
3330 "kernel.sched_schedstats=1\n");
3341 * update_min_vruntime()
3342 * vruntime -= min_vruntime
3346 * update_min_vruntime()
3347 * vruntime += min_vruntime
3349 * this way the vruntime transition between RQs is done when both
3350 * min_vruntime are up-to-date.
3354 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3355 * vruntime -= min_vruntime
3359 * update_min_vruntime()
3360 * vruntime += min_vruntime
3362 * this way we don't have the most up-to-date min_vruntime on the originating
3363 * CPU and an up-to-date min_vruntime on the destination CPU.
3367 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3369 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3370 bool curr = cfs_rq->curr == se;
3373 * If we're the current task, we must renormalise before calling
3377 se->vruntime += cfs_rq->min_vruntime;
3379 update_curr(cfs_rq);
3382 * Otherwise, renormalise after, such that we're placed at the current
3383 * moment in time, instead of some random moment in the past. Being
3384 * placed in the past could significantly boost this task to the
3385 * fairness detriment of existing tasks.
3387 if (renorm && !curr)
3388 se->vruntime += cfs_rq->min_vruntime;
3390 enqueue_entity_load_avg(cfs_rq, se);
3391 account_entity_enqueue(cfs_rq, se);
3392 update_cfs_shares(cfs_rq);
3394 if (flags & ENQUEUE_WAKEUP)
3395 place_entity(cfs_rq, se, 0);
3397 check_schedstat_required();
3398 update_stats_enqueue(cfs_rq, se, flags);
3399 check_spread(cfs_rq, se);
3401 __enqueue_entity(cfs_rq, se);
3404 if (cfs_rq->nr_running == 1) {
3405 list_add_leaf_cfs_rq(cfs_rq);
3406 check_enqueue_throttle(cfs_rq);
3410 static void __clear_buddies_last(struct sched_entity *se)
3412 for_each_sched_entity(se) {
3413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3414 if (cfs_rq->last != se)
3417 cfs_rq->last = NULL;
3421 static void __clear_buddies_next(struct sched_entity *se)
3423 for_each_sched_entity(se) {
3424 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3425 if (cfs_rq->next != se)
3428 cfs_rq->next = NULL;
3432 static void __clear_buddies_skip(struct sched_entity *se)
3434 for_each_sched_entity(se) {
3435 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3436 if (cfs_rq->skip != se)
3439 cfs_rq->skip = NULL;
3443 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3445 if (cfs_rq->last == se)
3446 __clear_buddies_last(se);
3448 if (cfs_rq->next == se)
3449 __clear_buddies_next(se);
3451 if (cfs_rq->skip == se)
3452 __clear_buddies_skip(se);
3455 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3458 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3461 * Update run-time statistics of the 'current'.
3463 update_curr(cfs_rq);
3464 dequeue_entity_load_avg(cfs_rq, se);
3466 update_stats_dequeue(cfs_rq, se, flags);
3468 clear_buddies(cfs_rq, se);
3470 if (se != cfs_rq->curr)
3471 __dequeue_entity(cfs_rq, se);
3473 account_entity_dequeue(cfs_rq, se);
3476 * Normalize the entity after updating the min_vruntime because the
3477 * update can refer to the ->curr item and we need to reflect this
3478 * movement in our normalized position.
3480 if (!(flags & DEQUEUE_SLEEP))
3481 se->vruntime -= cfs_rq->min_vruntime;
3483 /* return excess runtime on last dequeue */
3484 return_cfs_rq_runtime(cfs_rq);
3486 update_min_vruntime(cfs_rq);
3487 update_cfs_shares(cfs_rq);
3491 * Preempt the current task with a newly woken task if needed:
3494 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3496 unsigned long ideal_runtime, delta_exec;
3497 struct sched_entity *se;
3500 ideal_runtime = sched_slice(cfs_rq, curr);
3501 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3502 if (delta_exec > ideal_runtime) {
3503 resched_curr(rq_of(cfs_rq));
3505 * The current task ran long enough, ensure it doesn't get
3506 * re-elected due to buddy favours.
3508 clear_buddies(cfs_rq, curr);
3513 * Ensure that a task that missed wakeup preemption by a
3514 * narrow margin doesn't have to wait for a full slice.
3515 * This also mitigates buddy induced latencies under load.
3517 if (delta_exec < sysctl_sched_min_granularity)
3520 se = __pick_first_entity(cfs_rq);
3521 delta = curr->vruntime - se->vruntime;
3526 if (delta > ideal_runtime)
3527 resched_curr(rq_of(cfs_rq));
3531 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3533 /* 'current' is not kept within the tree. */
3536 * Any task has to be enqueued before it get to execute on
3537 * a CPU. So account for the time it spent waiting on the
3540 update_stats_wait_end(cfs_rq, se);
3541 __dequeue_entity(cfs_rq, se);
3542 update_load_avg(se, 1);
3545 update_stats_curr_start(cfs_rq, se);
3549 * Track our maximum slice length, if the CPU's load is at
3550 * least twice that of our own weight (i.e. dont track it
3551 * when there are only lesser-weight tasks around):
3553 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3554 schedstat_set(se->statistics.slice_max,
3555 max((u64)schedstat_val(se->statistics.slice_max),
3556 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3559 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3563 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3566 * Pick the next process, keeping these things in mind, in this order:
3567 * 1) keep things fair between processes/task groups
3568 * 2) pick the "next" process, since someone really wants that to run
3569 * 3) pick the "last" process, for cache locality
3570 * 4) do not run the "skip" process, if something else is available
3572 static struct sched_entity *
3573 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3575 struct sched_entity *left = __pick_first_entity(cfs_rq);
3576 struct sched_entity *se;
3579 * If curr is set we have to see if its left of the leftmost entity
3580 * still in the tree, provided there was anything in the tree at all.
3582 if (!left || (curr && entity_before(curr, left)))
3585 se = left; /* ideally we run the leftmost entity */
3588 * Avoid running the skip buddy, if running something else can
3589 * be done without getting too unfair.
3591 if (cfs_rq->skip == se) {
3592 struct sched_entity *second;
3595 second = __pick_first_entity(cfs_rq);
3597 second = __pick_next_entity(se);
3598 if (!second || (curr && entity_before(curr, second)))
3602 if (second && wakeup_preempt_entity(second, left) < 1)
3607 * Prefer last buddy, try to return the CPU to a preempted task.
3609 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3613 * Someone really wants this to run. If it's not unfair, run it.
3615 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3618 clear_buddies(cfs_rq, se);
3623 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3625 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3628 * If still on the runqueue then deactivate_task()
3629 * was not called and update_curr() has to be done:
3632 update_curr(cfs_rq);
3634 /* throttle cfs_rqs exceeding runtime */
3635 check_cfs_rq_runtime(cfs_rq);
3637 check_spread(cfs_rq, prev);
3640 update_stats_wait_start(cfs_rq, prev);
3641 /* Put 'current' back into the tree. */
3642 __enqueue_entity(cfs_rq, prev);
3643 /* in !on_rq case, update occurred at dequeue */
3644 update_load_avg(prev, 0);
3646 cfs_rq->curr = NULL;
3650 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3653 * Update run-time statistics of the 'current'.
3655 update_curr(cfs_rq);
3658 * Ensure that runnable average is periodically updated.
3660 update_load_avg(curr, 1);
3661 update_cfs_shares(cfs_rq);
3663 #ifdef CONFIG_SCHED_HRTICK
3665 * queued ticks are scheduled to match the slice, so don't bother
3666 * validating it and just reschedule.
3669 resched_curr(rq_of(cfs_rq));
3673 * don't let the period tick interfere with the hrtick preemption
3675 if (!sched_feat(DOUBLE_TICK) &&
3676 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3680 if (cfs_rq->nr_running > 1)
3681 check_preempt_tick(cfs_rq, curr);
3685 /**************************************************
3686 * CFS bandwidth control machinery
3689 #ifdef CONFIG_CFS_BANDWIDTH
3691 #ifdef HAVE_JUMP_LABEL
3692 static struct static_key __cfs_bandwidth_used;
3694 static inline bool cfs_bandwidth_used(void)
3696 return static_key_false(&__cfs_bandwidth_used);
3699 void cfs_bandwidth_usage_inc(void)
3701 static_key_slow_inc(&__cfs_bandwidth_used);
3704 void cfs_bandwidth_usage_dec(void)
3706 static_key_slow_dec(&__cfs_bandwidth_used);
3708 #else /* HAVE_JUMP_LABEL */
3709 static bool cfs_bandwidth_used(void)
3714 void cfs_bandwidth_usage_inc(void) {}
3715 void cfs_bandwidth_usage_dec(void) {}
3716 #endif /* HAVE_JUMP_LABEL */
3719 * default period for cfs group bandwidth.
3720 * default: 0.1s, units: nanoseconds
3722 static inline u64 default_cfs_period(void)
3724 return 100000000ULL;
3727 static inline u64 sched_cfs_bandwidth_slice(void)
3729 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3733 * Replenish runtime according to assigned quota and update expiration time.
3734 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3735 * additional synchronization around rq->lock.
3737 * requires cfs_b->lock
3739 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3743 if (cfs_b->quota == RUNTIME_INF)
3746 now = sched_clock_cpu(smp_processor_id());
3747 cfs_b->runtime = cfs_b->quota;
3748 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3751 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3753 return &tg->cfs_bandwidth;
3756 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3757 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3759 if (unlikely(cfs_rq->throttle_count))
3760 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3762 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3765 /* returns 0 on failure to allocate runtime */
3766 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3768 struct task_group *tg = cfs_rq->tg;
3769 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3770 u64 amount = 0, min_amount, expires;
3772 /* note: this is a positive sum as runtime_remaining <= 0 */
3773 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3775 raw_spin_lock(&cfs_b->lock);
3776 if (cfs_b->quota == RUNTIME_INF)
3777 amount = min_amount;
3779 start_cfs_bandwidth(cfs_b);
3781 if (cfs_b->runtime > 0) {
3782 amount = min(cfs_b->runtime, min_amount);
3783 cfs_b->runtime -= amount;
3787 expires = cfs_b->runtime_expires;
3788 raw_spin_unlock(&cfs_b->lock);
3790 cfs_rq->runtime_remaining += amount;
3792 * we may have advanced our local expiration to account for allowed
3793 * spread between our sched_clock and the one on which runtime was
3796 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3797 cfs_rq->runtime_expires = expires;
3799 return cfs_rq->runtime_remaining > 0;
3803 * Note: This depends on the synchronization provided by sched_clock and the
3804 * fact that rq->clock snapshots this value.
3806 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3808 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3810 /* if the deadline is ahead of our clock, nothing to do */
3811 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3814 if (cfs_rq->runtime_remaining < 0)
3818 * If the local deadline has passed we have to consider the
3819 * possibility that our sched_clock is 'fast' and the global deadline
3820 * has not truly expired.
3822 * Fortunately we can check determine whether this the case by checking
3823 * whether the global deadline has advanced. It is valid to compare
3824 * cfs_b->runtime_expires without any locks since we only care about
3825 * exact equality, so a partial write will still work.
3828 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3829 /* extend local deadline, drift is bounded above by 2 ticks */
3830 cfs_rq->runtime_expires += TICK_NSEC;
3832 /* global deadline is ahead, expiration has passed */
3833 cfs_rq->runtime_remaining = 0;
3837 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3839 /* dock delta_exec before expiring quota (as it could span periods) */
3840 cfs_rq->runtime_remaining -= delta_exec;
3841 expire_cfs_rq_runtime(cfs_rq);
3843 if (likely(cfs_rq->runtime_remaining > 0))
3847 * if we're unable to extend our runtime we resched so that the active
3848 * hierarchy can be throttled
3850 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3851 resched_curr(rq_of(cfs_rq));
3854 static __always_inline
3855 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3857 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3860 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3863 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3865 return cfs_bandwidth_used() && cfs_rq->throttled;
3868 /* check whether cfs_rq, or any parent, is throttled */
3869 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3871 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3875 * Ensure that neither of the group entities corresponding to src_cpu or
3876 * dest_cpu are members of a throttled hierarchy when performing group
3877 * load-balance operations.
3879 static inline int throttled_lb_pair(struct task_group *tg,
3880 int src_cpu, int dest_cpu)
3882 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3884 src_cfs_rq = tg->cfs_rq[src_cpu];
3885 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3887 return throttled_hierarchy(src_cfs_rq) ||
3888 throttled_hierarchy(dest_cfs_rq);
3891 /* updated child weight may affect parent so we have to do this bottom up */
3892 static int tg_unthrottle_up(struct task_group *tg, void *data)
3894 struct rq *rq = data;
3895 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3897 cfs_rq->throttle_count--;
3898 if (!cfs_rq->throttle_count) {
3899 /* adjust cfs_rq_clock_task() */
3900 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3901 cfs_rq->throttled_clock_task;
3907 static int tg_throttle_down(struct task_group *tg, void *data)
3909 struct rq *rq = data;
3910 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3912 /* group is entering throttled state, stop time */
3913 if (!cfs_rq->throttle_count)
3914 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3915 cfs_rq->throttle_count++;
3920 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3922 struct rq *rq = rq_of(cfs_rq);
3923 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3924 struct sched_entity *se;
3925 long task_delta, dequeue = 1;
3928 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3930 /* freeze hierarchy runnable averages while throttled */
3932 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3935 task_delta = cfs_rq->h_nr_running;
3936 for_each_sched_entity(se) {
3937 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3938 /* throttled entity or throttle-on-deactivate */
3943 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3944 qcfs_rq->h_nr_running -= task_delta;
3946 if (qcfs_rq->load.weight)
3951 sub_nr_running(rq, task_delta);
3953 cfs_rq->throttled = 1;
3954 cfs_rq->throttled_clock = rq_clock(rq);
3955 raw_spin_lock(&cfs_b->lock);
3956 empty = list_empty(&cfs_b->throttled_cfs_rq);
3959 * Add to the _head_ of the list, so that an already-started
3960 * distribute_cfs_runtime will not see us
3962 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3965 * If we're the first throttled task, make sure the bandwidth
3969 start_cfs_bandwidth(cfs_b);
3971 raw_spin_unlock(&cfs_b->lock);
3974 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3976 struct rq *rq = rq_of(cfs_rq);
3977 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3978 struct sched_entity *se;
3982 se = cfs_rq->tg->se[cpu_of(rq)];
3984 cfs_rq->throttled = 0;
3986 update_rq_clock(rq);
3988 raw_spin_lock(&cfs_b->lock);
3989 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3990 list_del_rcu(&cfs_rq->throttled_list);
3991 raw_spin_unlock(&cfs_b->lock);
3993 /* update hierarchical throttle state */
3994 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3996 if (!cfs_rq->load.weight)
3999 task_delta = cfs_rq->h_nr_running;
4000 for_each_sched_entity(se) {
4004 cfs_rq = cfs_rq_of(se);
4006 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4007 cfs_rq->h_nr_running += task_delta;
4009 if (cfs_rq_throttled(cfs_rq))
4014 add_nr_running(rq, task_delta);
4016 /* determine whether we need to wake up potentially idle cpu */
4017 if (rq->curr == rq->idle && rq->cfs.nr_running)
4021 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4022 u64 remaining, u64 expires)
4024 struct cfs_rq *cfs_rq;
4026 u64 starting_runtime = remaining;
4029 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4031 struct rq *rq = rq_of(cfs_rq);
4033 raw_spin_lock(&rq->lock);
4034 if (!cfs_rq_throttled(cfs_rq))
4037 runtime = -cfs_rq->runtime_remaining + 1;
4038 if (runtime > remaining)
4039 runtime = remaining;
4040 remaining -= runtime;
4042 cfs_rq->runtime_remaining += runtime;
4043 cfs_rq->runtime_expires = expires;
4045 /* we check whether we're throttled above */
4046 if (cfs_rq->runtime_remaining > 0)
4047 unthrottle_cfs_rq(cfs_rq);
4050 raw_spin_unlock(&rq->lock);
4057 return starting_runtime - remaining;
4061 * Responsible for refilling a task_group's bandwidth and unthrottling its
4062 * cfs_rqs as appropriate. If there has been no activity within the last
4063 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4064 * used to track this state.
4066 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4068 u64 runtime, runtime_expires;
4071 /* no need to continue the timer with no bandwidth constraint */
4072 if (cfs_b->quota == RUNTIME_INF)
4073 goto out_deactivate;
4075 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4076 cfs_b->nr_periods += overrun;
4079 * idle depends on !throttled (for the case of a large deficit), and if
4080 * we're going inactive then everything else can be deferred
4082 if (cfs_b->idle && !throttled)
4083 goto out_deactivate;
4085 __refill_cfs_bandwidth_runtime(cfs_b);
4088 /* mark as potentially idle for the upcoming period */
4093 /* account preceding periods in which throttling occurred */
4094 cfs_b->nr_throttled += overrun;
4096 runtime_expires = cfs_b->runtime_expires;
4099 * This check is repeated as we are holding onto the new bandwidth while
4100 * we unthrottle. This can potentially race with an unthrottled group
4101 * trying to acquire new bandwidth from the global pool. This can result
4102 * in us over-using our runtime if it is all used during this loop, but
4103 * only by limited amounts in that extreme case.
4105 while (throttled && cfs_b->runtime > 0) {
4106 runtime = cfs_b->runtime;
4107 raw_spin_unlock(&cfs_b->lock);
4108 /* we can't nest cfs_b->lock while distributing bandwidth */
4109 runtime = distribute_cfs_runtime(cfs_b, runtime,
4111 raw_spin_lock(&cfs_b->lock);
4113 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4115 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4119 * While we are ensured activity in the period following an
4120 * unthrottle, this also covers the case in which the new bandwidth is
4121 * insufficient to cover the existing bandwidth deficit. (Forcing the
4122 * timer to remain active while there are any throttled entities.)
4132 /* a cfs_rq won't donate quota below this amount */
4133 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4134 /* minimum remaining period time to redistribute slack quota */
4135 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4136 /* how long we wait to gather additional slack before distributing */
4137 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4140 * Are we near the end of the current quota period?
4142 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4143 * hrtimer base being cleared by hrtimer_start. In the case of
4144 * migrate_hrtimers, base is never cleared, so we are fine.
4146 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4148 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4151 /* if the call-back is running a quota refresh is already occurring */
4152 if (hrtimer_callback_running(refresh_timer))
4155 /* is a quota refresh about to occur? */
4156 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4157 if (remaining < min_expire)
4163 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4165 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4167 /* if there's a quota refresh soon don't bother with slack */
4168 if (runtime_refresh_within(cfs_b, min_left))
4171 hrtimer_start(&cfs_b->slack_timer,
4172 ns_to_ktime(cfs_bandwidth_slack_period),
4176 /* we know any runtime found here is valid as update_curr() precedes return */
4177 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4179 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4180 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4182 if (slack_runtime <= 0)
4185 raw_spin_lock(&cfs_b->lock);
4186 if (cfs_b->quota != RUNTIME_INF &&
4187 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4188 cfs_b->runtime += slack_runtime;
4190 /* we are under rq->lock, defer unthrottling using a timer */
4191 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4192 !list_empty(&cfs_b->throttled_cfs_rq))
4193 start_cfs_slack_bandwidth(cfs_b);
4195 raw_spin_unlock(&cfs_b->lock);
4197 /* even if it's not valid for return we don't want to try again */
4198 cfs_rq->runtime_remaining -= slack_runtime;
4201 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4203 if (!cfs_bandwidth_used())
4206 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4209 __return_cfs_rq_runtime(cfs_rq);
4213 * This is done with a timer (instead of inline with bandwidth return) since
4214 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4216 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4218 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4221 /* confirm we're still not at a refresh boundary */
4222 raw_spin_lock(&cfs_b->lock);
4223 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4224 raw_spin_unlock(&cfs_b->lock);
4228 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4229 runtime = cfs_b->runtime;
4231 expires = cfs_b->runtime_expires;
4232 raw_spin_unlock(&cfs_b->lock);
4237 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4239 raw_spin_lock(&cfs_b->lock);
4240 if (expires == cfs_b->runtime_expires)
4241 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4242 raw_spin_unlock(&cfs_b->lock);
4246 * When a group wakes up we want to make sure that its quota is not already
4247 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4248 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4250 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4252 if (!cfs_bandwidth_used())
4255 /* an active group must be handled by the update_curr()->put() path */
4256 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4259 /* ensure the group is not already throttled */
4260 if (cfs_rq_throttled(cfs_rq))
4263 /* update runtime allocation */
4264 account_cfs_rq_runtime(cfs_rq, 0);
4265 if (cfs_rq->runtime_remaining <= 0)
4266 throttle_cfs_rq(cfs_rq);
4269 static void sync_throttle(struct task_group *tg, int cpu)
4271 struct cfs_rq *pcfs_rq, *cfs_rq;
4273 if (!cfs_bandwidth_used())
4279 cfs_rq = tg->cfs_rq[cpu];
4280 pcfs_rq = tg->parent->cfs_rq[cpu];
4282 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4283 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4286 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4287 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4289 if (!cfs_bandwidth_used())
4292 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4296 * it's possible for a throttled entity to be forced into a running
4297 * state (e.g. set_curr_task), in this case we're finished.
4299 if (cfs_rq_throttled(cfs_rq))
4302 throttle_cfs_rq(cfs_rq);
4306 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4308 struct cfs_bandwidth *cfs_b =
4309 container_of(timer, struct cfs_bandwidth, slack_timer);
4311 do_sched_cfs_slack_timer(cfs_b);
4313 return HRTIMER_NORESTART;
4316 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4318 struct cfs_bandwidth *cfs_b =
4319 container_of(timer, struct cfs_bandwidth, period_timer);
4323 raw_spin_lock(&cfs_b->lock);
4325 overrun = hrtimer_forward_now(timer, cfs_b->period);
4329 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4332 cfs_b->period_active = 0;
4333 raw_spin_unlock(&cfs_b->lock);
4335 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4338 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4340 raw_spin_lock_init(&cfs_b->lock);
4342 cfs_b->quota = RUNTIME_INF;
4343 cfs_b->period = ns_to_ktime(default_cfs_period());
4345 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4346 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4347 cfs_b->period_timer.function = sched_cfs_period_timer;
4348 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4349 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4352 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4354 cfs_rq->runtime_enabled = 0;
4355 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4358 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4360 lockdep_assert_held(&cfs_b->lock);
4362 if (!cfs_b->period_active) {
4363 cfs_b->period_active = 1;
4364 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4365 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4369 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4371 /* init_cfs_bandwidth() was not called */
4372 if (!cfs_b->throttled_cfs_rq.next)
4375 hrtimer_cancel(&cfs_b->period_timer);
4376 hrtimer_cancel(&cfs_b->slack_timer);
4379 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4381 struct cfs_rq *cfs_rq;
4383 for_each_leaf_cfs_rq(rq, cfs_rq) {
4384 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4386 raw_spin_lock(&cfs_b->lock);
4387 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4388 raw_spin_unlock(&cfs_b->lock);
4392 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4394 struct cfs_rq *cfs_rq;
4396 for_each_leaf_cfs_rq(rq, cfs_rq) {
4397 if (!cfs_rq->runtime_enabled)
4401 * clock_task is not advancing so we just need to make sure
4402 * there's some valid quota amount
4404 cfs_rq->runtime_remaining = 1;
4406 * Offline rq is schedulable till cpu is completely disabled
4407 * in take_cpu_down(), so we prevent new cfs throttling here.
4409 cfs_rq->runtime_enabled = 0;
4411 if (cfs_rq_throttled(cfs_rq))
4412 unthrottle_cfs_rq(cfs_rq);
4416 #else /* CONFIG_CFS_BANDWIDTH */
4417 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4419 return rq_clock_task(rq_of(cfs_rq));
4422 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4423 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4424 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4425 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4426 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4428 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4433 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4438 static inline int throttled_lb_pair(struct task_group *tg,
4439 int src_cpu, int dest_cpu)
4444 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4446 #ifdef CONFIG_FAIR_GROUP_SCHED
4447 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4450 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4454 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4455 static inline void update_runtime_enabled(struct rq *rq) {}
4456 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4458 #endif /* CONFIG_CFS_BANDWIDTH */
4460 /**************************************************
4461 * CFS operations on tasks:
4464 #ifdef CONFIG_SCHED_HRTICK
4465 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4467 struct sched_entity *se = &p->se;
4468 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4470 WARN_ON(task_rq(p) != rq);
4472 if (rq->cfs.h_nr_running > 1) {
4473 u64 slice = sched_slice(cfs_rq, se);
4474 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4475 s64 delta = slice - ran;
4482 hrtick_start(rq, delta);
4487 * called from enqueue/dequeue and updates the hrtick when the
4488 * current task is from our class and nr_running is low enough
4491 static void hrtick_update(struct rq *rq)
4493 struct task_struct *curr = rq->curr;
4495 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4498 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4499 hrtick_start_fair(rq, curr);
4501 #else /* !CONFIG_SCHED_HRTICK */
4503 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4507 static inline void hrtick_update(struct rq *rq)
4513 * The enqueue_task method is called before nr_running is
4514 * increased. Here we update the fair scheduling stats and
4515 * then put the task into the rbtree:
4518 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4520 struct cfs_rq *cfs_rq;
4521 struct sched_entity *se = &p->se;
4523 for_each_sched_entity(se) {
4526 cfs_rq = cfs_rq_of(se);
4527 enqueue_entity(cfs_rq, se, flags);
4530 * end evaluation on encountering a throttled cfs_rq
4532 * note: in the case of encountering a throttled cfs_rq we will
4533 * post the final h_nr_running increment below.
4535 if (cfs_rq_throttled(cfs_rq))
4537 cfs_rq->h_nr_running++;
4539 flags = ENQUEUE_WAKEUP;
4542 for_each_sched_entity(se) {
4543 cfs_rq = cfs_rq_of(se);
4544 cfs_rq->h_nr_running++;
4546 if (cfs_rq_throttled(cfs_rq))
4549 update_load_avg(se, 1);
4550 update_cfs_shares(cfs_rq);
4554 add_nr_running(rq, 1);
4559 static void set_next_buddy(struct sched_entity *se);
4562 * The dequeue_task method is called before nr_running is
4563 * decreased. We remove the task from the rbtree and
4564 * update the fair scheduling stats:
4566 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4568 struct cfs_rq *cfs_rq;
4569 struct sched_entity *se = &p->se;
4570 int task_sleep = flags & DEQUEUE_SLEEP;
4572 for_each_sched_entity(se) {
4573 cfs_rq = cfs_rq_of(se);
4574 dequeue_entity(cfs_rq, se, flags);
4577 * end evaluation on encountering a throttled cfs_rq
4579 * note: in the case of encountering a throttled cfs_rq we will
4580 * post the final h_nr_running decrement below.
4582 if (cfs_rq_throttled(cfs_rq))
4584 cfs_rq->h_nr_running--;
4586 /* Don't dequeue parent if it has other entities besides us */
4587 if (cfs_rq->load.weight) {
4588 /* Avoid re-evaluating load for this entity: */
4589 se = parent_entity(se);
4591 * Bias pick_next to pick a task from this cfs_rq, as
4592 * p is sleeping when it is within its sched_slice.
4594 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4598 flags |= DEQUEUE_SLEEP;
4601 for_each_sched_entity(se) {
4602 cfs_rq = cfs_rq_of(se);
4603 cfs_rq->h_nr_running--;
4605 if (cfs_rq_throttled(cfs_rq))
4608 update_load_avg(se, 1);
4609 update_cfs_shares(cfs_rq);
4613 sub_nr_running(rq, 1);
4619 #ifdef CONFIG_NO_HZ_COMMON
4621 * per rq 'load' arrray crap; XXX kill this.
4625 * The exact cpuload calculated at every tick would be:
4627 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4629 * If a cpu misses updates for n ticks (as it was idle) and update gets
4630 * called on the n+1-th tick when cpu may be busy, then we have:
4632 * load_n = (1 - 1/2^i)^n * load_0
4633 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4635 * decay_load_missed() below does efficient calculation of
4637 * load' = (1 - 1/2^i)^n * load
4639 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4640 * This allows us to precompute the above in said factors, thereby allowing the
4641 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4642 * fixed_power_int())
4644 * The calculation is approximated on a 128 point scale.
4646 #define DEGRADE_SHIFT 7
4648 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4649 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4650 { 0, 0, 0, 0, 0, 0, 0, 0 },
4651 { 64, 32, 8, 0, 0, 0, 0, 0 },
4652 { 96, 72, 40, 12, 1, 0, 0, 0 },
4653 { 112, 98, 75, 43, 15, 1, 0, 0 },
4654 { 120, 112, 98, 76, 45, 16, 2, 0 }
4658 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4659 * would be when CPU is idle and so we just decay the old load without
4660 * adding any new load.
4662 static unsigned long
4663 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4667 if (!missed_updates)
4670 if (missed_updates >= degrade_zero_ticks[idx])
4674 return load >> missed_updates;
4676 while (missed_updates) {
4677 if (missed_updates % 2)
4678 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4680 missed_updates >>= 1;
4685 #endif /* CONFIG_NO_HZ_COMMON */
4688 * __cpu_load_update - update the rq->cpu_load[] statistics
4689 * @this_rq: The rq to update statistics for
4690 * @this_load: The current load
4691 * @pending_updates: The number of missed updates
4693 * Update rq->cpu_load[] statistics. This function is usually called every
4694 * scheduler tick (TICK_NSEC).
4696 * This function computes a decaying average:
4698 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4700 * Because of NOHZ it might not get called on every tick which gives need for
4701 * the @pending_updates argument.
4703 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4704 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4705 * = A * (A * load[i]_n-2 + B) + B
4706 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4707 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4708 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4709 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4710 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4712 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4713 * any change in load would have resulted in the tick being turned back on.
4715 * For regular NOHZ, this reduces to:
4717 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4719 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4722 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4723 unsigned long pending_updates)
4725 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4728 this_rq->nr_load_updates++;
4730 /* Update our load: */
4731 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4732 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4733 unsigned long old_load, new_load;
4735 /* scale is effectively 1 << i now, and >> i divides by scale */
4737 old_load = this_rq->cpu_load[i];
4738 #ifdef CONFIG_NO_HZ_COMMON
4739 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4740 if (tickless_load) {
4741 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4743 * old_load can never be a negative value because a
4744 * decayed tickless_load cannot be greater than the
4745 * original tickless_load.
4747 old_load += tickless_load;
4750 new_load = this_load;
4752 * Round up the averaging division if load is increasing. This
4753 * prevents us from getting stuck on 9 if the load is 10, for
4756 if (new_load > old_load)
4757 new_load += scale - 1;
4759 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4762 sched_avg_update(this_rq);
4765 /* Used instead of source_load when we know the type == 0 */
4766 static unsigned long weighted_cpuload(const int cpu)
4768 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4771 #ifdef CONFIG_NO_HZ_COMMON
4773 * There is no sane way to deal with nohz on smp when using jiffies because the
4774 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4775 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4777 * Therefore we need to avoid the delta approach from the regular tick when
4778 * possible since that would seriously skew the load calculation. This is why we
4779 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4780 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4781 * loop exit, nohz_idle_balance, nohz full exit...)
4783 * This means we might still be one tick off for nohz periods.
4786 static void cpu_load_update_nohz(struct rq *this_rq,
4787 unsigned long curr_jiffies,
4790 unsigned long pending_updates;
4792 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4793 if (pending_updates) {
4794 this_rq->last_load_update_tick = curr_jiffies;
4796 * In the regular NOHZ case, we were idle, this means load 0.
4797 * In the NOHZ_FULL case, we were non-idle, we should consider
4798 * its weighted load.
4800 cpu_load_update(this_rq, load, pending_updates);
4805 * Called from nohz_idle_balance() to update the load ratings before doing the
4808 static void cpu_load_update_idle(struct rq *this_rq)
4811 * bail if there's load or we're actually up-to-date.
4813 if (weighted_cpuload(cpu_of(this_rq)))
4816 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4820 * Record CPU load on nohz entry so we know the tickless load to account
4821 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4822 * than other cpu_load[idx] but it should be fine as cpu_load readers
4823 * shouldn't rely into synchronized cpu_load[*] updates.
4825 void cpu_load_update_nohz_start(void)
4827 struct rq *this_rq = this_rq();
4830 * This is all lockless but should be fine. If weighted_cpuload changes
4831 * concurrently we'll exit nohz. And cpu_load write can race with
4832 * cpu_load_update_idle() but both updater would be writing the same.
4834 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4838 * Account the tickless load in the end of a nohz frame.
4840 void cpu_load_update_nohz_stop(void)
4842 unsigned long curr_jiffies = READ_ONCE(jiffies);
4843 struct rq *this_rq = this_rq();
4846 if (curr_jiffies == this_rq->last_load_update_tick)
4849 load = weighted_cpuload(cpu_of(this_rq));
4850 raw_spin_lock(&this_rq->lock);
4851 update_rq_clock(this_rq);
4852 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4853 raw_spin_unlock(&this_rq->lock);
4855 #else /* !CONFIG_NO_HZ_COMMON */
4856 static inline void cpu_load_update_nohz(struct rq *this_rq,
4857 unsigned long curr_jiffies,
4858 unsigned long load) { }
4859 #endif /* CONFIG_NO_HZ_COMMON */
4861 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4863 #ifdef CONFIG_NO_HZ_COMMON
4864 /* See the mess around cpu_load_update_nohz(). */
4865 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4867 cpu_load_update(this_rq, load, 1);
4871 * Called from scheduler_tick()
4873 void cpu_load_update_active(struct rq *this_rq)
4875 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4877 if (tick_nohz_tick_stopped())
4878 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4880 cpu_load_update_periodic(this_rq, load);
4884 * Return a low guess at the load of a migration-source cpu weighted
4885 * according to the scheduling class and "nice" value.
4887 * We want to under-estimate the load of migration sources, to
4888 * balance conservatively.
4890 static unsigned long source_load(int cpu, int type)
4892 struct rq *rq = cpu_rq(cpu);
4893 unsigned long total = weighted_cpuload(cpu);
4895 if (type == 0 || !sched_feat(LB_BIAS))
4898 return min(rq->cpu_load[type-1], total);
4902 * Return a high guess at the load of a migration-target cpu weighted
4903 * according to the scheduling class and "nice" value.
4905 static unsigned long target_load(int cpu, int type)
4907 struct rq *rq = cpu_rq(cpu);
4908 unsigned long total = weighted_cpuload(cpu);
4910 if (type == 0 || !sched_feat(LB_BIAS))
4913 return max(rq->cpu_load[type-1], total);
4916 static unsigned long capacity_of(int cpu)
4918 return cpu_rq(cpu)->cpu_capacity;
4921 static unsigned long capacity_orig_of(int cpu)
4923 return cpu_rq(cpu)->cpu_capacity_orig;
4926 static unsigned long cpu_avg_load_per_task(int cpu)
4928 struct rq *rq = cpu_rq(cpu);
4929 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4930 unsigned long load_avg = weighted_cpuload(cpu);
4933 return load_avg / nr_running;
4938 #ifdef CONFIG_FAIR_GROUP_SCHED
4940 * effective_load() calculates the load change as seen from the root_task_group
4942 * Adding load to a group doesn't make a group heavier, but can cause movement
4943 * of group shares between cpus. Assuming the shares were perfectly aligned one
4944 * can calculate the shift in shares.
4946 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4947 * on this @cpu and results in a total addition (subtraction) of @wg to the
4948 * total group weight.
4950 * Given a runqueue weight distribution (rw_i) we can compute a shares
4951 * distribution (s_i) using:
4953 * s_i = rw_i / \Sum rw_j (1)
4955 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4956 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4957 * shares distribution (s_i):
4959 * rw_i = { 2, 4, 1, 0 }
4960 * s_i = { 2/7, 4/7, 1/7, 0 }
4962 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4963 * task used to run on and the CPU the waker is running on), we need to
4964 * compute the effect of waking a task on either CPU and, in case of a sync
4965 * wakeup, compute the effect of the current task going to sleep.
4967 * So for a change of @wl to the local @cpu with an overall group weight change
4968 * of @wl we can compute the new shares distribution (s'_i) using:
4970 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4972 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4973 * differences in waking a task to CPU 0. The additional task changes the
4974 * weight and shares distributions like:
4976 * rw'_i = { 3, 4, 1, 0 }
4977 * s'_i = { 3/8, 4/8, 1/8, 0 }
4979 * We can then compute the difference in effective weight by using:
4981 * dw_i = S * (s'_i - s_i) (3)
4983 * Where 'S' is the group weight as seen by its parent.
4985 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4986 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4987 * 4/7) times the weight of the group.
4989 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4991 struct sched_entity *se = tg->se[cpu];
4993 if (!tg->parent) /* the trivial, non-cgroup case */
4996 for_each_sched_entity(se) {
4997 struct cfs_rq *cfs_rq = se->my_q;
4998 long W, w = cfs_rq_load_avg(cfs_rq);
5003 * W = @wg + \Sum rw_j
5005 W = wg + atomic_long_read(&tg->load_avg);
5007 /* Ensure \Sum rw_j >= rw_i */
5008 W -= cfs_rq->tg_load_avg_contrib;
5017 * wl = S * s'_i; see (2)
5020 wl = (w * (long)scale_load_down(tg->shares)) / W;
5022 wl = scale_load_down(tg->shares);
5025 * Per the above, wl is the new se->load.weight value; since
5026 * those are clipped to [MIN_SHARES, ...) do so now. See
5027 * calc_cfs_shares().
5029 if (wl < MIN_SHARES)
5033 * wl = dw_i = S * (s'_i - s_i); see (3)
5035 wl -= se->avg.load_avg;
5038 * Recursively apply this logic to all parent groups to compute
5039 * the final effective load change on the root group. Since
5040 * only the @tg group gets extra weight, all parent groups can
5041 * only redistribute existing shares. @wl is the shift in shares
5042 * resulting from this level per the above.
5051 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5058 static void record_wakee(struct task_struct *p)
5061 * Only decay a single time; tasks that have less then 1 wakeup per
5062 * jiffy will not have built up many flips.
5064 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5065 current->wakee_flips >>= 1;
5066 current->wakee_flip_decay_ts = jiffies;
5069 if (current->last_wakee != p) {
5070 current->last_wakee = p;
5071 current->wakee_flips++;
5076 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5078 * A waker of many should wake a different task than the one last awakened
5079 * at a frequency roughly N times higher than one of its wakees.
5081 * In order to determine whether we should let the load spread vs consolidating
5082 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5083 * partner, and a factor of lls_size higher frequency in the other.
5085 * With both conditions met, we can be relatively sure that the relationship is
5086 * non-monogamous, with partner count exceeding socket size.
5088 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5089 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5092 static int wake_wide(struct task_struct *p)
5094 unsigned int master = current->wakee_flips;
5095 unsigned int slave = p->wakee_flips;
5096 int factor = this_cpu_read(sd_llc_size);
5099 swap(master, slave);
5100 if (slave < factor || master < slave * factor)
5105 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5106 int prev_cpu, int sync)
5108 s64 this_load, load;
5109 s64 this_eff_load, prev_eff_load;
5111 struct task_group *tg;
5112 unsigned long weight;
5116 this_cpu = smp_processor_id();
5117 load = source_load(prev_cpu, idx);
5118 this_load = target_load(this_cpu, idx);
5121 * If sync wakeup then subtract the (maximum possible)
5122 * effect of the currently running task from the load
5123 * of the current CPU:
5126 tg = task_group(current);
5127 weight = current->se.avg.load_avg;
5129 this_load += effective_load(tg, this_cpu, -weight, -weight);
5130 load += effective_load(tg, prev_cpu, 0, -weight);
5134 weight = p->se.avg.load_avg;
5137 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5138 * due to the sync cause above having dropped this_load to 0, we'll
5139 * always have an imbalance, but there's really nothing you can do
5140 * about that, so that's good too.
5142 * Otherwise check if either cpus are near enough in load to allow this
5143 * task to be woken on this_cpu.
5145 this_eff_load = 100;
5146 this_eff_load *= capacity_of(prev_cpu);
5148 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5149 prev_eff_load *= capacity_of(this_cpu);
5151 if (this_load > 0) {
5152 this_eff_load *= this_load +
5153 effective_load(tg, this_cpu, weight, weight);
5155 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5158 balanced = this_eff_load <= prev_eff_load;
5160 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5165 schedstat_inc(sd->ttwu_move_affine);
5166 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5172 * find_idlest_group finds and returns the least busy CPU group within the
5175 static struct sched_group *
5176 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5177 int this_cpu, int sd_flag)
5179 struct sched_group *idlest = NULL, *group = sd->groups;
5180 unsigned long min_load = ULONG_MAX, this_load = 0;
5181 int load_idx = sd->forkexec_idx;
5182 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5184 if (sd_flag & SD_BALANCE_WAKE)
5185 load_idx = sd->wake_idx;
5188 unsigned long load, avg_load;
5192 /* Skip over this group if it has no CPUs allowed */
5193 if (!cpumask_intersects(sched_group_cpus(group),
5194 tsk_cpus_allowed(p)))
5197 local_group = cpumask_test_cpu(this_cpu,
5198 sched_group_cpus(group));
5200 /* Tally up the load of all CPUs in the group */
5203 for_each_cpu(i, sched_group_cpus(group)) {
5204 /* Bias balancing toward cpus of our domain */
5206 load = source_load(i, load_idx);
5208 load = target_load(i, load_idx);
5213 /* Adjust by relative CPU capacity of the group */
5214 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5217 this_load = avg_load;
5218 } else if (avg_load < min_load) {
5219 min_load = avg_load;
5222 } while (group = group->next, group != sd->groups);
5224 if (!idlest || 100*this_load < imbalance*min_load)
5230 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5233 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5235 unsigned long load, min_load = ULONG_MAX;
5236 unsigned int min_exit_latency = UINT_MAX;
5237 u64 latest_idle_timestamp = 0;
5238 int least_loaded_cpu = this_cpu;
5239 int shallowest_idle_cpu = -1;
5242 /* Check if we have any choice: */
5243 if (group->group_weight == 1)
5244 return cpumask_first(sched_group_cpus(group));
5246 /* Traverse only the allowed CPUs */
5247 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5249 struct rq *rq = cpu_rq(i);
5250 struct cpuidle_state *idle = idle_get_state(rq);
5251 if (idle && idle->exit_latency < min_exit_latency) {
5253 * We give priority to a CPU whose idle state
5254 * has the smallest exit latency irrespective
5255 * of any idle timestamp.
5257 min_exit_latency = idle->exit_latency;
5258 latest_idle_timestamp = rq->idle_stamp;
5259 shallowest_idle_cpu = i;
5260 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5261 rq->idle_stamp > latest_idle_timestamp) {
5263 * If equal or no active idle state, then
5264 * the most recently idled CPU might have
5267 latest_idle_timestamp = rq->idle_stamp;
5268 shallowest_idle_cpu = i;
5270 } else if (shallowest_idle_cpu == -1) {
5271 load = weighted_cpuload(i);
5272 if (load < min_load || (load == min_load && i == this_cpu)) {
5274 least_loaded_cpu = i;
5279 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5283 * Try and locate an idle CPU in the sched_domain.
5285 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5287 struct sched_domain *sd;
5288 struct sched_group *sg;
5290 if (idle_cpu(target))
5294 * If the prevous cpu is cache affine and idle, don't be stupid.
5296 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5300 * Otherwise, iterate the domains and find an eligible idle cpu.
5302 * A completely idle sched group at higher domains is more
5303 * desirable than an idle group at a lower level, because lower
5304 * domains have smaller groups and usually share hardware
5305 * resources which causes tasks to contend on them, e.g. x86
5306 * hyperthread siblings in the lowest domain (SMT) can contend
5307 * on the shared cpu pipeline.
5309 * However, while we prefer idle groups at higher domains
5310 * finding an idle cpu at the lowest domain is still better than
5311 * returning 'target', which we've already established, isn't
5314 sd = rcu_dereference(per_cpu(sd_llc, target));
5315 for_each_lower_domain(sd) {
5320 if (!cpumask_intersects(sched_group_cpus(sg),
5321 tsk_cpus_allowed(p)))
5324 /* Ensure the entire group is idle */
5325 for_each_cpu(i, sched_group_cpus(sg)) {
5326 if (i == target || !idle_cpu(i))
5331 * It doesn't matter which cpu we pick, the
5332 * whole group is idle.
5334 target = cpumask_first_and(sched_group_cpus(sg),
5335 tsk_cpus_allowed(p));
5339 } while (sg != sd->groups);
5346 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5347 * tasks. The unit of the return value must be the one of capacity so we can
5348 * compare the utilization with the capacity of the CPU that is available for
5349 * CFS task (ie cpu_capacity).
5351 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5352 * recent utilization of currently non-runnable tasks on a CPU. It represents
5353 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5354 * capacity_orig is the cpu_capacity available at the highest frequency
5355 * (arch_scale_freq_capacity()).
5356 * The utilization of a CPU converges towards a sum equal to or less than the
5357 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5358 * the running time on this CPU scaled by capacity_curr.
5360 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5361 * higher than capacity_orig because of unfortunate rounding in
5362 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5363 * the average stabilizes with the new running time. We need to check that the
5364 * utilization stays within the range of [0..capacity_orig] and cap it if
5365 * necessary. Without utilization capping, a group could be seen as overloaded
5366 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5367 * available capacity. We allow utilization to overshoot capacity_curr (but not
5368 * capacity_orig) as it useful for predicting the capacity required after task
5369 * migrations (scheduler-driven DVFS).
5371 static int cpu_util(int cpu)
5373 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5374 unsigned long capacity = capacity_orig_of(cpu);
5376 return (util >= capacity) ? capacity : util;
5379 static inline int task_util(struct task_struct *p)
5381 return p->se.avg.util_avg;
5385 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5386 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5388 * In that case WAKE_AFFINE doesn't make sense and we'll let
5389 * BALANCE_WAKE sort things out.
5391 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5393 long min_cap, max_cap;
5395 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5396 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5398 /* Minimum capacity is close to max, no need to abort wake_affine */
5399 if (max_cap - min_cap < max_cap >> 3)
5402 return min_cap * 1024 < task_util(p) * capacity_margin;
5406 * select_task_rq_fair: Select target runqueue for the waking task in domains
5407 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5408 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5410 * Balances load by selecting the idlest cpu in the idlest group, or under
5411 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5413 * Returns the target cpu number.
5415 * preempt must be disabled.
5418 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5420 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5421 int cpu = smp_processor_id();
5422 int new_cpu = prev_cpu;
5423 int want_affine = 0;
5424 int sync = wake_flags & WF_SYNC;
5426 if (sd_flag & SD_BALANCE_WAKE) {
5428 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5429 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5433 for_each_domain(cpu, tmp) {
5434 if (!(tmp->flags & SD_LOAD_BALANCE))
5438 * If both cpu and prev_cpu are part of this domain,
5439 * cpu is a valid SD_WAKE_AFFINE target.
5441 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5442 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5447 if (tmp->flags & sd_flag)
5449 else if (!want_affine)
5454 sd = NULL; /* Prefer wake_affine over balance flags */
5455 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5460 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5461 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5464 struct sched_group *group;
5467 if (!(sd->flags & sd_flag)) {
5472 group = find_idlest_group(sd, p, cpu, sd_flag);
5478 new_cpu = find_idlest_cpu(group, p, cpu);
5479 if (new_cpu == -1 || new_cpu == cpu) {
5480 /* Now try balancing at a lower domain level of cpu */
5485 /* Now try balancing at a lower domain level of new_cpu */
5487 weight = sd->span_weight;
5489 for_each_domain(cpu, tmp) {
5490 if (weight <= tmp->span_weight)
5492 if (tmp->flags & sd_flag)
5495 /* while loop will break here if sd == NULL */
5503 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5504 * cfs_rq_of(p) references at time of call are still valid and identify the
5505 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5507 static void migrate_task_rq_fair(struct task_struct *p)
5510 * As blocked tasks retain absolute vruntime the migration needs to
5511 * deal with this by subtracting the old and adding the new
5512 * min_vruntime -- the latter is done by enqueue_entity() when placing
5513 * the task on the new runqueue.
5515 if (p->state == TASK_WAKING) {
5516 struct sched_entity *se = &p->se;
5517 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5520 #ifndef CONFIG_64BIT
5521 u64 min_vruntime_copy;
5524 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5526 min_vruntime = cfs_rq->min_vruntime;
5527 } while (min_vruntime != min_vruntime_copy);
5529 min_vruntime = cfs_rq->min_vruntime;
5532 se->vruntime -= min_vruntime;
5536 * We are supposed to update the task to "current" time, then its up to date
5537 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5538 * what current time is, so simply throw away the out-of-date time. This
5539 * will result in the wakee task is less decayed, but giving the wakee more
5540 * load sounds not bad.
5542 remove_entity_load_avg(&p->se);
5544 /* Tell new CPU we are migrated */
5545 p->se.avg.last_update_time = 0;
5547 /* We have migrated, no longer consider this task hot */
5548 p->se.exec_start = 0;
5551 static void task_dead_fair(struct task_struct *p)
5553 remove_entity_load_avg(&p->se);
5555 #endif /* CONFIG_SMP */
5557 static unsigned long
5558 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5560 unsigned long gran = sysctl_sched_wakeup_granularity;
5563 * Since its curr running now, convert the gran from real-time
5564 * to virtual-time in his units.
5566 * By using 'se' instead of 'curr' we penalize light tasks, so
5567 * they get preempted easier. That is, if 'se' < 'curr' then
5568 * the resulting gran will be larger, therefore penalizing the
5569 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5570 * be smaller, again penalizing the lighter task.
5572 * This is especially important for buddies when the leftmost
5573 * task is higher priority than the buddy.
5575 return calc_delta_fair(gran, se);
5579 * Should 'se' preempt 'curr'.
5593 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5595 s64 gran, vdiff = curr->vruntime - se->vruntime;
5600 gran = wakeup_gran(curr, se);
5607 static void set_last_buddy(struct sched_entity *se)
5609 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5612 for_each_sched_entity(se)
5613 cfs_rq_of(se)->last = se;
5616 static void set_next_buddy(struct sched_entity *se)
5618 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5621 for_each_sched_entity(se)
5622 cfs_rq_of(se)->next = se;
5625 static void set_skip_buddy(struct sched_entity *se)
5627 for_each_sched_entity(se)
5628 cfs_rq_of(se)->skip = se;
5632 * Preempt the current task with a newly woken task if needed:
5634 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5636 struct task_struct *curr = rq->curr;
5637 struct sched_entity *se = &curr->se, *pse = &p->se;
5638 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5639 int scale = cfs_rq->nr_running >= sched_nr_latency;
5640 int next_buddy_marked = 0;
5642 if (unlikely(se == pse))
5646 * This is possible from callers such as attach_tasks(), in which we
5647 * unconditionally check_prempt_curr() after an enqueue (which may have
5648 * lead to a throttle). This both saves work and prevents false
5649 * next-buddy nomination below.
5651 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5654 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5655 set_next_buddy(pse);
5656 next_buddy_marked = 1;
5660 * We can come here with TIF_NEED_RESCHED already set from new task
5663 * Note: this also catches the edge-case of curr being in a throttled
5664 * group (e.g. via set_curr_task), since update_curr() (in the
5665 * enqueue of curr) will have resulted in resched being set. This
5666 * prevents us from potentially nominating it as a false LAST_BUDDY
5669 if (test_tsk_need_resched(curr))
5672 /* Idle tasks are by definition preempted by non-idle tasks. */
5673 if (unlikely(curr->policy == SCHED_IDLE) &&
5674 likely(p->policy != SCHED_IDLE))
5678 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5679 * is driven by the tick):
5681 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5684 find_matching_se(&se, &pse);
5685 update_curr(cfs_rq_of(se));
5687 if (wakeup_preempt_entity(se, pse) == 1) {
5689 * Bias pick_next to pick the sched entity that is
5690 * triggering this preemption.
5692 if (!next_buddy_marked)
5693 set_next_buddy(pse);
5702 * Only set the backward buddy when the current task is still
5703 * on the rq. This can happen when a wakeup gets interleaved
5704 * with schedule on the ->pre_schedule() or idle_balance()
5705 * point, either of which can * drop the rq lock.
5707 * Also, during early boot the idle thread is in the fair class,
5708 * for obvious reasons its a bad idea to schedule back to it.
5710 if (unlikely(!se->on_rq || curr == rq->idle))
5713 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5717 static struct task_struct *
5718 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5720 struct cfs_rq *cfs_rq = &rq->cfs;
5721 struct sched_entity *se;
5722 struct task_struct *p;
5726 #ifdef CONFIG_FAIR_GROUP_SCHED
5727 if (!cfs_rq->nr_running)
5730 if (prev->sched_class != &fair_sched_class)
5734 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5735 * likely that a next task is from the same cgroup as the current.
5737 * Therefore attempt to avoid putting and setting the entire cgroup
5738 * hierarchy, only change the part that actually changes.
5742 struct sched_entity *curr = cfs_rq->curr;
5745 * Since we got here without doing put_prev_entity() we also
5746 * have to consider cfs_rq->curr. If it is still a runnable
5747 * entity, update_curr() will update its vruntime, otherwise
5748 * forget we've ever seen it.
5752 update_curr(cfs_rq);
5757 * This call to check_cfs_rq_runtime() will do the
5758 * throttle and dequeue its entity in the parent(s).
5759 * Therefore the 'simple' nr_running test will indeed
5762 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5766 se = pick_next_entity(cfs_rq, curr);
5767 cfs_rq = group_cfs_rq(se);
5773 * Since we haven't yet done put_prev_entity and if the selected task
5774 * is a different task than we started out with, try and touch the
5775 * least amount of cfs_rqs.
5778 struct sched_entity *pse = &prev->se;
5780 while (!(cfs_rq = is_same_group(se, pse))) {
5781 int se_depth = se->depth;
5782 int pse_depth = pse->depth;
5784 if (se_depth <= pse_depth) {
5785 put_prev_entity(cfs_rq_of(pse), pse);
5786 pse = parent_entity(pse);
5788 if (se_depth >= pse_depth) {
5789 set_next_entity(cfs_rq_of(se), se);
5790 se = parent_entity(se);
5794 put_prev_entity(cfs_rq, pse);
5795 set_next_entity(cfs_rq, se);
5798 if (hrtick_enabled(rq))
5799 hrtick_start_fair(rq, p);
5806 if (!cfs_rq->nr_running)
5809 put_prev_task(rq, prev);
5812 se = pick_next_entity(cfs_rq, NULL);
5813 set_next_entity(cfs_rq, se);
5814 cfs_rq = group_cfs_rq(se);
5819 if (hrtick_enabled(rq))
5820 hrtick_start_fair(rq, p);
5826 * This is OK, because current is on_cpu, which avoids it being picked
5827 * for load-balance and preemption/IRQs are still disabled avoiding
5828 * further scheduler activity on it and we're being very careful to
5829 * re-start the picking loop.
5831 lockdep_unpin_lock(&rq->lock, cookie);
5832 new_tasks = idle_balance(rq);
5833 lockdep_repin_lock(&rq->lock, cookie);
5835 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5836 * possible for any higher priority task to appear. In that case we
5837 * must re-start the pick_next_entity() loop.
5849 * Account for a descheduled task:
5851 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5853 struct sched_entity *se = &prev->se;
5854 struct cfs_rq *cfs_rq;
5856 for_each_sched_entity(se) {
5857 cfs_rq = cfs_rq_of(se);
5858 put_prev_entity(cfs_rq, se);
5863 * sched_yield() is very simple
5865 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5867 static void yield_task_fair(struct rq *rq)
5869 struct task_struct *curr = rq->curr;
5870 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5871 struct sched_entity *se = &curr->se;
5874 * Are we the only task in the tree?
5876 if (unlikely(rq->nr_running == 1))
5879 clear_buddies(cfs_rq, se);
5881 if (curr->policy != SCHED_BATCH) {
5882 update_rq_clock(rq);
5884 * Update run-time statistics of the 'current'.
5886 update_curr(cfs_rq);
5888 * Tell update_rq_clock() that we've just updated,
5889 * so we don't do microscopic update in schedule()
5890 * and double the fastpath cost.
5892 rq_clock_skip_update(rq, true);
5898 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5900 struct sched_entity *se = &p->se;
5902 /* throttled hierarchies are not runnable */
5903 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5906 /* Tell the scheduler that we'd really like pse to run next. */
5909 yield_task_fair(rq);
5915 /**************************************************
5916 * Fair scheduling class load-balancing methods.
5920 * The purpose of load-balancing is to achieve the same basic fairness the
5921 * per-cpu scheduler provides, namely provide a proportional amount of compute
5922 * time to each task. This is expressed in the following equation:
5924 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5926 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5927 * W_i,0 is defined as:
5929 * W_i,0 = \Sum_j w_i,j (2)
5931 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5932 * is derived from the nice value as per sched_prio_to_weight[].
5934 * The weight average is an exponential decay average of the instantaneous
5937 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5939 * C_i is the compute capacity of cpu i, typically it is the
5940 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5941 * can also include other factors [XXX].
5943 * To achieve this balance we define a measure of imbalance which follows
5944 * directly from (1):
5946 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5948 * We them move tasks around to minimize the imbalance. In the continuous
5949 * function space it is obvious this converges, in the discrete case we get
5950 * a few fun cases generally called infeasible weight scenarios.
5953 * - infeasible weights;
5954 * - local vs global optima in the discrete case. ]
5959 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5960 * for all i,j solution, we create a tree of cpus that follows the hardware
5961 * topology where each level pairs two lower groups (or better). This results
5962 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5963 * tree to only the first of the previous level and we decrease the frequency
5964 * of load-balance at each level inv. proportional to the number of cpus in
5970 * \Sum { --- * --- * 2^i } = O(n) (5)
5972 * `- size of each group
5973 * | | `- number of cpus doing load-balance
5975 * `- sum over all levels
5977 * Coupled with a limit on how many tasks we can migrate every balance pass,
5978 * this makes (5) the runtime complexity of the balancer.
5980 * An important property here is that each CPU is still (indirectly) connected
5981 * to every other cpu in at most O(log n) steps:
5983 * The adjacency matrix of the resulting graph is given by:
5986 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5989 * And you'll find that:
5991 * A^(log_2 n)_i,j != 0 for all i,j (7)
5993 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5994 * The task movement gives a factor of O(m), giving a convergence complexity
5997 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6002 * In order to avoid CPUs going idle while there's still work to do, new idle
6003 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6004 * tree itself instead of relying on other CPUs to bring it work.
6006 * This adds some complexity to both (5) and (8) but it reduces the total idle
6014 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6017 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6022 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6024 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6026 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6029 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6030 * rewrite all of this once again.]
6033 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6035 enum fbq_type { regular, remote, all };
6037 #define LBF_ALL_PINNED 0x01
6038 #define LBF_NEED_BREAK 0x02
6039 #define LBF_DST_PINNED 0x04
6040 #define LBF_SOME_PINNED 0x08
6043 struct sched_domain *sd;
6051 struct cpumask *dst_grpmask;
6053 enum cpu_idle_type idle;
6055 /* The set of CPUs under consideration for load-balancing */
6056 struct cpumask *cpus;
6061 unsigned int loop_break;
6062 unsigned int loop_max;
6064 enum fbq_type fbq_type;
6065 struct list_head tasks;
6069 * Is this task likely cache-hot:
6071 static int task_hot(struct task_struct *p, struct lb_env *env)
6075 lockdep_assert_held(&env->src_rq->lock);
6077 if (p->sched_class != &fair_sched_class)
6080 if (unlikely(p->policy == SCHED_IDLE))
6084 * Buddy candidates are cache hot:
6086 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6087 (&p->se == cfs_rq_of(&p->se)->next ||
6088 &p->se == cfs_rq_of(&p->se)->last))
6091 if (sysctl_sched_migration_cost == -1)
6093 if (sysctl_sched_migration_cost == 0)
6096 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6098 return delta < (s64)sysctl_sched_migration_cost;
6101 #ifdef CONFIG_NUMA_BALANCING
6103 * Returns 1, if task migration degrades locality
6104 * Returns 0, if task migration improves locality i.e migration preferred.
6105 * Returns -1, if task migration is not affected by locality.
6107 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6109 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6110 unsigned long src_faults, dst_faults;
6111 int src_nid, dst_nid;
6113 if (!static_branch_likely(&sched_numa_balancing))
6116 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6119 src_nid = cpu_to_node(env->src_cpu);
6120 dst_nid = cpu_to_node(env->dst_cpu);
6122 if (src_nid == dst_nid)
6125 /* Migrating away from the preferred node is always bad. */
6126 if (src_nid == p->numa_preferred_nid) {
6127 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6133 /* Encourage migration to the preferred node. */
6134 if (dst_nid == p->numa_preferred_nid)
6138 src_faults = group_faults(p, src_nid);
6139 dst_faults = group_faults(p, dst_nid);
6141 src_faults = task_faults(p, src_nid);
6142 dst_faults = task_faults(p, dst_nid);
6145 return dst_faults < src_faults;
6149 static inline int migrate_degrades_locality(struct task_struct *p,
6157 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6160 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6164 lockdep_assert_held(&env->src_rq->lock);
6167 * We do not migrate tasks that are:
6168 * 1) throttled_lb_pair, or
6169 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6170 * 3) running (obviously), or
6171 * 4) are cache-hot on their current CPU.
6173 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6176 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6179 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6181 env->flags |= LBF_SOME_PINNED;
6184 * Remember if this task can be migrated to any other cpu in
6185 * our sched_group. We may want to revisit it if we couldn't
6186 * meet load balance goals by pulling other tasks on src_cpu.
6188 * Also avoid computing new_dst_cpu if we have already computed
6189 * one in current iteration.
6191 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6194 /* Prevent to re-select dst_cpu via env's cpus */
6195 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6196 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6197 env->flags |= LBF_DST_PINNED;
6198 env->new_dst_cpu = cpu;
6206 /* Record that we found atleast one task that could run on dst_cpu */
6207 env->flags &= ~LBF_ALL_PINNED;
6209 if (task_running(env->src_rq, p)) {
6210 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6215 * Aggressive migration if:
6216 * 1) destination numa is preferred
6217 * 2) task is cache cold, or
6218 * 3) too many balance attempts have failed.
6220 tsk_cache_hot = migrate_degrades_locality(p, env);
6221 if (tsk_cache_hot == -1)
6222 tsk_cache_hot = task_hot(p, env);
6224 if (tsk_cache_hot <= 0 ||
6225 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6226 if (tsk_cache_hot == 1) {
6227 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6228 schedstat_inc(p->se.statistics.nr_forced_migrations);
6233 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6238 * detach_task() -- detach the task for the migration specified in env
6240 static void detach_task(struct task_struct *p, struct lb_env *env)
6242 lockdep_assert_held(&env->src_rq->lock);
6244 p->on_rq = TASK_ON_RQ_MIGRATING;
6245 deactivate_task(env->src_rq, p, 0);
6246 set_task_cpu(p, env->dst_cpu);
6250 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6251 * part of active balancing operations within "domain".
6253 * Returns a task if successful and NULL otherwise.
6255 static struct task_struct *detach_one_task(struct lb_env *env)
6257 struct task_struct *p, *n;
6259 lockdep_assert_held(&env->src_rq->lock);
6261 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6262 if (!can_migrate_task(p, env))
6265 detach_task(p, env);
6268 * Right now, this is only the second place where
6269 * lb_gained[env->idle] is updated (other is detach_tasks)
6270 * so we can safely collect stats here rather than
6271 * inside detach_tasks().
6273 schedstat_inc(env->sd->lb_gained[env->idle]);
6279 static const unsigned int sched_nr_migrate_break = 32;
6282 * detach_tasks() -- tries to detach up to imbalance weighted load from
6283 * busiest_rq, as part of a balancing operation within domain "sd".
6285 * Returns number of detached tasks if successful and 0 otherwise.
6287 static int detach_tasks(struct lb_env *env)
6289 struct list_head *tasks = &env->src_rq->cfs_tasks;
6290 struct task_struct *p;
6294 lockdep_assert_held(&env->src_rq->lock);
6296 if (env->imbalance <= 0)
6299 while (!list_empty(tasks)) {
6301 * We don't want to steal all, otherwise we may be treated likewise,
6302 * which could at worst lead to a livelock crash.
6304 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6307 p = list_first_entry(tasks, struct task_struct, se.group_node);
6310 /* We've more or less seen every task there is, call it quits */
6311 if (env->loop > env->loop_max)
6314 /* take a breather every nr_migrate tasks */
6315 if (env->loop > env->loop_break) {
6316 env->loop_break += sched_nr_migrate_break;
6317 env->flags |= LBF_NEED_BREAK;
6321 if (!can_migrate_task(p, env))
6324 load = task_h_load(p);
6326 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6329 if ((load / 2) > env->imbalance)
6332 detach_task(p, env);
6333 list_add(&p->se.group_node, &env->tasks);
6336 env->imbalance -= load;
6338 #ifdef CONFIG_PREEMPT
6340 * NEWIDLE balancing is a source of latency, so preemptible
6341 * kernels will stop after the first task is detached to minimize
6342 * the critical section.
6344 if (env->idle == CPU_NEWLY_IDLE)
6349 * We only want to steal up to the prescribed amount of
6352 if (env->imbalance <= 0)
6357 list_move_tail(&p->se.group_node, tasks);
6361 * Right now, this is one of only two places we collect this stat
6362 * so we can safely collect detach_one_task() stats here rather
6363 * than inside detach_one_task().
6365 schedstat_add(env->sd->lb_gained[env->idle], detached);
6371 * attach_task() -- attach the task detached by detach_task() to its new rq.
6373 static void attach_task(struct rq *rq, struct task_struct *p)
6375 lockdep_assert_held(&rq->lock);
6377 BUG_ON(task_rq(p) != rq);
6378 activate_task(rq, p, 0);
6379 p->on_rq = TASK_ON_RQ_QUEUED;
6380 check_preempt_curr(rq, p, 0);
6384 * attach_one_task() -- attaches the task returned from detach_one_task() to
6387 static void attach_one_task(struct rq *rq, struct task_struct *p)
6389 raw_spin_lock(&rq->lock);
6391 raw_spin_unlock(&rq->lock);
6395 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6398 static void attach_tasks(struct lb_env *env)
6400 struct list_head *tasks = &env->tasks;
6401 struct task_struct *p;
6403 raw_spin_lock(&env->dst_rq->lock);
6405 while (!list_empty(tasks)) {
6406 p = list_first_entry(tasks, struct task_struct, se.group_node);
6407 list_del_init(&p->se.group_node);
6409 attach_task(env->dst_rq, p);
6412 raw_spin_unlock(&env->dst_rq->lock);
6415 #ifdef CONFIG_FAIR_GROUP_SCHED
6416 static void update_blocked_averages(int cpu)
6418 struct rq *rq = cpu_rq(cpu);
6419 struct cfs_rq *cfs_rq;
6420 unsigned long flags;
6422 raw_spin_lock_irqsave(&rq->lock, flags);
6423 update_rq_clock(rq);
6426 * Iterates the task_group tree in a bottom up fashion, see
6427 * list_add_leaf_cfs_rq() for details.
6429 for_each_leaf_cfs_rq(rq, cfs_rq) {
6430 /* throttled entities do not contribute to load */
6431 if (throttled_hierarchy(cfs_rq))
6434 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6435 update_tg_load_avg(cfs_rq, 0);
6437 raw_spin_unlock_irqrestore(&rq->lock, flags);
6441 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6442 * This needs to be done in a top-down fashion because the load of a child
6443 * group is a fraction of its parents load.
6445 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6447 struct rq *rq = rq_of(cfs_rq);
6448 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6449 unsigned long now = jiffies;
6452 if (cfs_rq->last_h_load_update == now)
6455 cfs_rq->h_load_next = NULL;
6456 for_each_sched_entity(se) {
6457 cfs_rq = cfs_rq_of(se);
6458 cfs_rq->h_load_next = se;
6459 if (cfs_rq->last_h_load_update == now)
6464 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6465 cfs_rq->last_h_load_update = now;
6468 while ((se = cfs_rq->h_load_next) != NULL) {
6469 load = cfs_rq->h_load;
6470 load = div64_ul(load * se->avg.load_avg,
6471 cfs_rq_load_avg(cfs_rq) + 1);
6472 cfs_rq = group_cfs_rq(se);
6473 cfs_rq->h_load = load;
6474 cfs_rq->last_h_load_update = now;
6478 static unsigned long task_h_load(struct task_struct *p)
6480 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6482 update_cfs_rq_h_load(cfs_rq);
6483 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6484 cfs_rq_load_avg(cfs_rq) + 1);
6487 static inline void update_blocked_averages(int cpu)
6489 struct rq *rq = cpu_rq(cpu);
6490 struct cfs_rq *cfs_rq = &rq->cfs;
6491 unsigned long flags;
6493 raw_spin_lock_irqsave(&rq->lock, flags);
6494 update_rq_clock(rq);
6495 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6496 raw_spin_unlock_irqrestore(&rq->lock, flags);
6499 static unsigned long task_h_load(struct task_struct *p)
6501 return p->se.avg.load_avg;
6505 /********** Helpers for find_busiest_group ************************/
6514 * sg_lb_stats - stats of a sched_group required for load_balancing
6516 struct sg_lb_stats {
6517 unsigned long avg_load; /*Avg load across the CPUs of the group */
6518 unsigned long group_load; /* Total load over the CPUs of the group */
6519 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6520 unsigned long load_per_task;
6521 unsigned long group_capacity;
6522 unsigned long group_util; /* Total utilization of the group */
6523 unsigned int sum_nr_running; /* Nr tasks running in the group */
6524 unsigned int idle_cpus;
6525 unsigned int group_weight;
6526 enum group_type group_type;
6527 int group_no_capacity;
6528 #ifdef CONFIG_NUMA_BALANCING
6529 unsigned int nr_numa_running;
6530 unsigned int nr_preferred_running;
6535 * sd_lb_stats - Structure to store the statistics of a sched_domain
6536 * during load balancing.
6538 struct sd_lb_stats {
6539 struct sched_group *busiest; /* Busiest group in this sd */
6540 struct sched_group *local; /* Local group in this sd */
6541 unsigned long total_load; /* Total load of all groups in sd */
6542 unsigned long total_capacity; /* Total capacity of all groups in sd */
6543 unsigned long avg_load; /* Average load across all groups in sd */
6545 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6546 struct sg_lb_stats local_stat; /* Statistics of the local group */
6549 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6552 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6553 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6554 * We must however clear busiest_stat::avg_load because
6555 * update_sd_pick_busiest() reads this before assignment.
6557 *sds = (struct sd_lb_stats){
6561 .total_capacity = 0UL,
6564 .sum_nr_running = 0,
6565 .group_type = group_other,
6571 * get_sd_load_idx - Obtain the load index for a given sched domain.
6572 * @sd: The sched_domain whose load_idx is to be obtained.
6573 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6575 * Return: The load index.
6577 static inline int get_sd_load_idx(struct sched_domain *sd,
6578 enum cpu_idle_type idle)
6584 load_idx = sd->busy_idx;
6587 case CPU_NEWLY_IDLE:
6588 load_idx = sd->newidle_idx;
6591 load_idx = sd->idle_idx;
6598 static unsigned long scale_rt_capacity(int cpu)
6600 struct rq *rq = cpu_rq(cpu);
6601 u64 total, used, age_stamp, avg;
6605 * Since we're reading these variables without serialization make sure
6606 * we read them once before doing sanity checks on them.
6608 age_stamp = READ_ONCE(rq->age_stamp);
6609 avg = READ_ONCE(rq->rt_avg);
6610 delta = __rq_clock_broken(rq) - age_stamp;
6612 if (unlikely(delta < 0))
6615 total = sched_avg_period() + delta;
6617 used = div_u64(avg, total);
6619 if (likely(used < SCHED_CAPACITY_SCALE))
6620 return SCHED_CAPACITY_SCALE - used;
6625 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6627 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6628 struct sched_group *sdg = sd->groups;
6630 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6632 capacity *= scale_rt_capacity(cpu);
6633 capacity >>= SCHED_CAPACITY_SHIFT;
6638 cpu_rq(cpu)->cpu_capacity = capacity;
6639 sdg->sgc->capacity = capacity;
6642 void update_group_capacity(struct sched_domain *sd, int cpu)
6644 struct sched_domain *child = sd->child;
6645 struct sched_group *group, *sdg = sd->groups;
6646 unsigned long capacity;
6647 unsigned long interval;
6649 interval = msecs_to_jiffies(sd->balance_interval);
6650 interval = clamp(interval, 1UL, max_load_balance_interval);
6651 sdg->sgc->next_update = jiffies + interval;
6654 update_cpu_capacity(sd, cpu);
6660 if (child->flags & SD_OVERLAP) {
6662 * SD_OVERLAP domains cannot assume that child groups
6663 * span the current group.
6666 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6667 struct sched_group_capacity *sgc;
6668 struct rq *rq = cpu_rq(cpu);
6671 * build_sched_domains() -> init_sched_groups_capacity()
6672 * gets here before we've attached the domains to the
6675 * Use capacity_of(), which is set irrespective of domains
6676 * in update_cpu_capacity().
6678 * This avoids capacity from being 0 and
6679 * causing divide-by-zero issues on boot.
6681 if (unlikely(!rq->sd)) {
6682 capacity += capacity_of(cpu);
6686 sgc = rq->sd->groups->sgc;
6687 capacity += sgc->capacity;
6691 * !SD_OVERLAP domains can assume that child groups
6692 * span the current group.
6695 group = child->groups;
6697 capacity += group->sgc->capacity;
6698 group = group->next;
6699 } while (group != child->groups);
6702 sdg->sgc->capacity = capacity;
6706 * Check whether the capacity of the rq has been noticeably reduced by side
6707 * activity. The imbalance_pct is used for the threshold.
6708 * Return true is the capacity is reduced
6711 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6713 return ((rq->cpu_capacity * sd->imbalance_pct) <
6714 (rq->cpu_capacity_orig * 100));
6718 * Group imbalance indicates (and tries to solve) the problem where balancing
6719 * groups is inadequate due to tsk_cpus_allowed() constraints.
6721 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6722 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6725 * { 0 1 2 3 } { 4 5 6 7 }
6728 * If we were to balance group-wise we'd place two tasks in the first group and
6729 * two tasks in the second group. Clearly this is undesired as it will overload
6730 * cpu 3 and leave one of the cpus in the second group unused.
6732 * The current solution to this issue is detecting the skew in the first group
6733 * by noticing the lower domain failed to reach balance and had difficulty
6734 * moving tasks due to affinity constraints.
6736 * When this is so detected; this group becomes a candidate for busiest; see
6737 * update_sd_pick_busiest(). And calculate_imbalance() and
6738 * find_busiest_group() avoid some of the usual balance conditions to allow it
6739 * to create an effective group imbalance.
6741 * This is a somewhat tricky proposition since the next run might not find the
6742 * group imbalance and decide the groups need to be balanced again. A most
6743 * subtle and fragile situation.
6746 static inline int sg_imbalanced(struct sched_group *group)
6748 return group->sgc->imbalance;
6752 * group_has_capacity returns true if the group has spare capacity that could
6753 * be used by some tasks.
6754 * We consider that a group has spare capacity if the * number of task is
6755 * smaller than the number of CPUs or if the utilization is lower than the
6756 * available capacity for CFS tasks.
6757 * For the latter, we use a threshold to stabilize the state, to take into
6758 * account the variance of the tasks' load and to return true if the available
6759 * capacity in meaningful for the load balancer.
6760 * As an example, an available capacity of 1% can appear but it doesn't make
6761 * any benefit for the load balance.
6764 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6766 if (sgs->sum_nr_running < sgs->group_weight)
6769 if ((sgs->group_capacity * 100) >
6770 (sgs->group_util * env->sd->imbalance_pct))
6777 * group_is_overloaded returns true if the group has more tasks than it can
6779 * group_is_overloaded is not equals to !group_has_capacity because a group
6780 * with the exact right number of tasks, has no more spare capacity but is not
6781 * overloaded so both group_has_capacity and group_is_overloaded return
6785 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6787 if (sgs->sum_nr_running <= sgs->group_weight)
6790 if ((sgs->group_capacity * 100) <
6791 (sgs->group_util * env->sd->imbalance_pct))
6798 group_type group_classify(struct sched_group *group,
6799 struct sg_lb_stats *sgs)
6801 if (sgs->group_no_capacity)
6802 return group_overloaded;
6804 if (sg_imbalanced(group))
6805 return group_imbalanced;
6811 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6812 * @env: The load balancing environment.
6813 * @group: sched_group whose statistics are to be updated.
6814 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6815 * @local_group: Does group contain this_cpu.
6816 * @sgs: variable to hold the statistics for this group.
6817 * @overload: Indicate more than one runnable task for any CPU.
6819 static inline void update_sg_lb_stats(struct lb_env *env,
6820 struct sched_group *group, int load_idx,
6821 int local_group, struct sg_lb_stats *sgs,
6827 memset(sgs, 0, sizeof(*sgs));
6829 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6830 struct rq *rq = cpu_rq(i);
6832 /* Bias balancing toward cpus of our domain */
6834 load = target_load(i, load_idx);
6836 load = source_load(i, load_idx);
6838 sgs->group_load += load;
6839 sgs->group_util += cpu_util(i);
6840 sgs->sum_nr_running += rq->cfs.h_nr_running;
6842 nr_running = rq->nr_running;
6846 #ifdef CONFIG_NUMA_BALANCING
6847 sgs->nr_numa_running += rq->nr_numa_running;
6848 sgs->nr_preferred_running += rq->nr_preferred_running;
6850 sgs->sum_weighted_load += weighted_cpuload(i);
6852 * No need to call idle_cpu() if nr_running is not 0
6854 if (!nr_running && idle_cpu(i))
6858 /* Adjust by relative CPU capacity of the group */
6859 sgs->group_capacity = group->sgc->capacity;
6860 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6862 if (sgs->sum_nr_running)
6863 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6865 sgs->group_weight = group->group_weight;
6867 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6868 sgs->group_type = group_classify(group, sgs);
6872 * update_sd_pick_busiest - return 1 on busiest group
6873 * @env: The load balancing environment.
6874 * @sds: sched_domain statistics
6875 * @sg: sched_group candidate to be checked for being the busiest
6876 * @sgs: sched_group statistics
6878 * Determine if @sg is a busier group than the previously selected
6881 * Return: %true if @sg is a busier group than the previously selected
6882 * busiest group. %false otherwise.
6884 static bool update_sd_pick_busiest(struct lb_env *env,
6885 struct sd_lb_stats *sds,
6886 struct sched_group *sg,
6887 struct sg_lb_stats *sgs)
6889 struct sg_lb_stats *busiest = &sds->busiest_stat;
6891 if (sgs->group_type > busiest->group_type)
6894 if (sgs->group_type < busiest->group_type)
6897 if (sgs->avg_load <= busiest->avg_load)
6900 /* This is the busiest node in its class. */
6901 if (!(env->sd->flags & SD_ASYM_PACKING))
6904 /* No ASYM_PACKING if target cpu is already busy */
6905 if (env->idle == CPU_NOT_IDLE)
6908 * ASYM_PACKING needs to move all the work to the lowest
6909 * numbered CPUs in the group, therefore mark all groups
6910 * higher than ourself as busy.
6912 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6916 /* Prefer to move from highest possible cpu's work */
6917 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6924 #ifdef CONFIG_NUMA_BALANCING
6925 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6927 if (sgs->sum_nr_running > sgs->nr_numa_running)
6929 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6934 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6936 if (rq->nr_running > rq->nr_numa_running)
6938 if (rq->nr_running > rq->nr_preferred_running)
6943 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6948 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6952 #endif /* CONFIG_NUMA_BALANCING */
6955 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6956 * @env: The load balancing environment.
6957 * @sds: variable to hold the statistics for this sched_domain.
6959 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6961 struct sched_domain *child = env->sd->child;
6962 struct sched_group *sg = env->sd->groups;
6963 struct sg_lb_stats tmp_sgs;
6964 int load_idx, prefer_sibling = 0;
6965 bool overload = false;
6967 if (child && child->flags & SD_PREFER_SIBLING)
6970 load_idx = get_sd_load_idx(env->sd, env->idle);
6973 struct sg_lb_stats *sgs = &tmp_sgs;
6976 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6979 sgs = &sds->local_stat;
6981 if (env->idle != CPU_NEWLY_IDLE ||
6982 time_after_eq(jiffies, sg->sgc->next_update))
6983 update_group_capacity(env->sd, env->dst_cpu);
6986 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6993 * In case the child domain prefers tasks go to siblings
6994 * first, lower the sg capacity so that we'll try
6995 * and move all the excess tasks away. We lower the capacity
6996 * of a group only if the local group has the capacity to fit
6997 * these excess tasks. The extra check prevents the case where
6998 * you always pull from the heaviest group when it is already
6999 * under-utilized (possible with a large weight task outweighs
7000 * the tasks on the system).
7002 if (prefer_sibling && sds->local &&
7003 group_has_capacity(env, &sds->local_stat) &&
7004 (sgs->sum_nr_running > 1)) {
7005 sgs->group_no_capacity = 1;
7006 sgs->group_type = group_classify(sg, sgs);
7009 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7011 sds->busiest_stat = *sgs;
7015 /* Now, start updating sd_lb_stats */
7016 sds->total_load += sgs->group_load;
7017 sds->total_capacity += sgs->group_capacity;
7020 } while (sg != env->sd->groups);
7022 if (env->sd->flags & SD_NUMA)
7023 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7025 if (!env->sd->parent) {
7026 /* update overload indicator if we are at root domain */
7027 if (env->dst_rq->rd->overload != overload)
7028 env->dst_rq->rd->overload = overload;
7034 * check_asym_packing - Check to see if the group is packed into the
7037 * This is primarily intended to used at the sibling level. Some
7038 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7039 * case of POWER7, it can move to lower SMT modes only when higher
7040 * threads are idle. When in lower SMT modes, the threads will
7041 * perform better since they share less core resources. Hence when we
7042 * have idle threads, we want them to be the higher ones.
7044 * This packing function is run on idle threads. It checks to see if
7045 * the busiest CPU in this domain (core in the P7 case) has a higher
7046 * CPU number than the packing function is being run on. Here we are
7047 * assuming lower CPU number will be equivalent to lower a SMT thread
7050 * Return: 1 when packing is required and a task should be moved to
7051 * this CPU. The amount of the imbalance is returned in *imbalance.
7053 * @env: The load balancing environment.
7054 * @sds: Statistics of the sched_domain which is to be packed
7056 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7060 if (!(env->sd->flags & SD_ASYM_PACKING))
7063 if (env->idle == CPU_NOT_IDLE)
7069 busiest_cpu = group_first_cpu(sds->busiest);
7070 if (env->dst_cpu > busiest_cpu)
7073 env->imbalance = DIV_ROUND_CLOSEST(
7074 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7075 SCHED_CAPACITY_SCALE);
7081 * fix_small_imbalance - Calculate the minor imbalance that exists
7082 * amongst the groups of a sched_domain, during
7084 * @env: The load balancing environment.
7085 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7088 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7090 unsigned long tmp, capa_now = 0, capa_move = 0;
7091 unsigned int imbn = 2;
7092 unsigned long scaled_busy_load_per_task;
7093 struct sg_lb_stats *local, *busiest;
7095 local = &sds->local_stat;
7096 busiest = &sds->busiest_stat;
7098 if (!local->sum_nr_running)
7099 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7100 else if (busiest->load_per_task > local->load_per_task)
7103 scaled_busy_load_per_task =
7104 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7105 busiest->group_capacity;
7107 if (busiest->avg_load + scaled_busy_load_per_task >=
7108 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7109 env->imbalance = busiest->load_per_task;
7114 * OK, we don't have enough imbalance to justify moving tasks,
7115 * however we may be able to increase total CPU capacity used by
7119 capa_now += busiest->group_capacity *
7120 min(busiest->load_per_task, busiest->avg_load);
7121 capa_now += local->group_capacity *
7122 min(local->load_per_task, local->avg_load);
7123 capa_now /= SCHED_CAPACITY_SCALE;
7125 /* Amount of load we'd subtract */
7126 if (busiest->avg_load > scaled_busy_load_per_task) {
7127 capa_move += busiest->group_capacity *
7128 min(busiest->load_per_task,
7129 busiest->avg_load - scaled_busy_load_per_task);
7132 /* Amount of load we'd add */
7133 if (busiest->avg_load * busiest->group_capacity <
7134 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7135 tmp = (busiest->avg_load * busiest->group_capacity) /
7136 local->group_capacity;
7138 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7139 local->group_capacity;
7141 capa_move += local->group_capacity *
7142 min(local->load_per_task, local->avg_load + tmp);
7143 capa_move /= SCHED_CAPACITY_SCALE;
7145 /* Move if we gain throughput */
7146 if (capa_move > capa_now)
7147 env->imbalance = busiest->load_per_task;
7151 * calculate_imbalance - Calculate the amount of imbalance present within the
7152 * groups of a given sched_domain during load balance.
7153 * @env: load balance environment
7154 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7156 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7158 unsigned long max_pull, load_above_capacity = ~0UL;
7159 struct sg_lb_stats *local, *busiest;
7161 local = &sds->local_stat;
7162 busiest = &sds->busiest_stat;
7164 if (busiest->group_type == group_imbalanced) {
7166 * In the group_imb case we cannot rely on group-wide averages
7167 * to ensure cpu-load equilibrium, look at wider averages. XXX
7169 busiest->load_per_task =
7170 min(busiest->load_per_task, sds->avg_load);
7174 * Avg load of busiest sg can be less and avg load of local sg can
7175 * be greater than avg load across all sgs of sd because avg load
7176 * factors in sg capacity and sgs with smaller group_type are
7177 * skipped when updating the busiest sg:
7179 if (busiest->avg_load <= sds->avg_load ||
7180 local->avg_load >= sds->avg_load) {
7182 return fix_small_imbalance(env, sds);
7186 * If there aren't any idle cpus, avoid creating some.
7188 if (busiest->group_type == group_overloaded &&
7189 local->group_type == group_overloaded) {
7190 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7191 if (load_above_capacity > busiest->group_capacity) {
7192 load_above_capacity -= busiest->group_capacity;
7193 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7194 load_above_capacity /= busiest->group_capacity;
7196 load_above_capacity = ~0UL;
7200 * We're trying to get all the cpus to the average_load, so we don't
7201 * want to push ourselves above the average load, nor do we wish to
7202 * reduce the max loaded cpu below the average load. At the same time,
7203 * we also don't want to reduce the group load below the group
7204 * capacity. Thus we look for the minimum possible imbalance.
7206 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7208 /* How much load to actually move to equalise the imbalance */
7209 env->imbalance = min(
7210 max_pull * busiest->group_capacity,
7211 (sds->avg_load - local->avg_load) * local->group_capacity
7212 ) / SCHED_CAPACITY_SCALE;
7215 * if *imbalance is less than the average load per runnable task
7216 * there is no guarantee that any tasks will be moved so we'll have
7217 * a think about bumping its value to force at least one task to be
7220 if (env->imbalance < busiest->load_per_task)
7221 return fix_small_imbalance(env, sds);
7224 /******* find_busiest_group() helpers end here *********************/
7227 * find_busiest_group - Returns the busiest group within the sched_domain
7228 * if there is an imbalance.
7230 * Also calculates the amount of weighted load which should be moved
7231 * to restore balance.
7233 * @env: The load balancing environment.
7235 * Return: - The busiest group if imbalance exists.
7237 static struct sched_group *find_busiest_group(struct lb_env *env)
7239 struct sg_lb_stats *local, *busiest;
7240 struct sd_lb_stats sds;
7242 init_sd_lb_stats(&sds);
7245 * Compute the various statistics relavent for load balancing at
7248 update_sd_lb_stats(env, &sds);
7249 local = &sds.local_stat;
7250 busiest = &sds.busiest_stat;
7252 /* ASYM feature bypasses nice load balance check */
7253 if (check_asym_packing(env, &sds))
7256 /* There is no busy sibling group to pull tasks from */
7257 if (!sds.busiest || busiest->sum_nr_running == 0)
7260 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7261 / sds.total_capacity;
7264 * If the busiest group is imbalanced the below checks don't
7265 * work because they assume all things are equal, which typically
7266 * isn't true due to cpus_allowed constraints and the like.
7268 if (busiest->group_type == group_imbalanced)
7271 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7272 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7273 busiest->group_no_capacity)
7277 * If the local group is busier than the selected busiest group
7278 * don't try and pull any tasks.
7280 if (local->avg_load >= busiest->avg_load)
7284 * Don't pull any tasks if this group is already above the domain
7287 if (local->avg_load >= sds.avg_load)
7290 if (env->idle == CPU_IDLE) {
7292 * This cpu is idle. If the busiest group is not overloaded
7293 * and there is no imbalance between this and busiest group
7294 * wrt idle cpus, it is balanced. The imbalance becomes
7295 * significant if the diff is greater than 1 otherwise we
7296 * might end up to just move the imbalance on another group
7298 if ((busiest->group_type != group_overloaded) &&
7299 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7303 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7304 * imbalance_pct to be conservative.
7306 if (100 * busiest->avg_load <=
7307 env->sd->imbalance_pct * local->avg_load)
7312 /* Looks like there is an imbalance. Compute it */
7313 calculate_imbalance(env, &sds);
7322 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7324 static struct rq *find_busiest_queue(struct lb_env *env,
7325 struct sched_group *group)
7327 struct rq *busiest = NULL, *rq;
7328 unsigned long busiest_load = 0, busiest_capacity = 1;
7331 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7332 unsigned long capacity, wl;
7336 rt = fbq_classify_rq(rq);
7339 * We classify groups/runqueues into three groups:
7340 * - regular: there are !numa tasks
7341 * - remote: there are numa tasks that run on the 'wrong' node
7342 * - all: there is no distinction
7344 * In order to avoid migrating ideally placed numa tasks,
7345 * ignore those when there's better options.
7347 * If we ignore the actual busiest queue to migrate another
7348 * task, the next balance pass can still reduce the busiest
7349 * queue by moving tasks around inside the node.
7351 * If we cannot move enough load due to this classification
7352 * the next pass will adjust the group classification and
7353 * allow migration of more tasks.
7355 * Both cases only affect the total convergence complexity.
7357 if (rt > env->fbq_type)
7360 capacity = capacity_of(i);
7362 wl = weighted_cpuload(i);
7365 * When comparing with imbalance, use weighted_cpuload()
7366 * which is not scaled with the cpu capacity.
7369 if (rq->nr_running == 1 && wl > env->imbalance &&
7370 !check_cpu_capacity(rq, env->sd))
7374 * For the load comparisons with the other cpu's, consider
7375 * the weighted_cpuload() scaled with the cpu capacity, so
7376 * that the load can be moved away from the cpu that is
7377 * potentially running at a lower capacity.
7379 * Thus we're looking for max(wl_i / capacity_i), crosswise
7380 * multiplication to rid ourselves of the division works out
7381 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7382 * our previous maximum.
7384 if (wl * busiest_capacity > busiest_load * capacity) {
7386 busiest_capacity = capacity;
7395 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7396 * so long as it is large enough.
7398 #define MAX_PINNED_INTERVAL 512
7400 /* Working cpumask for load_balance and load_balance_newidle. */
7401 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7403 static int need_active_balance(struct lb_env *env)
7405 struct sched_domain *sd = env->sd;
7407 if (env->idle == CPU_NEWLY_IDLE) {
7410 * ASYM_PACKING needs to force migrate tasks from busy but
7411 * higher numbered CPUs in order to pack all tasks in the
7412 * lowest numbered CPUs.
7414 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7419 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7420 * It's worth migrating the task if the src_cpu's capacity is reduced
7421 * because of other sched_class or IRQs if more capacity stays
7422 * available on dst_cpu.
7424 if ((env->idle != CPU_NOT_IDLE) &&
7425 (env->src_rq->cfs.h_nr_running == 1)) {
7426 if ((check_cpu_capacity(env->src_rq, sd)) &&
7427 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7431 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7434 static int active_load_balance_cpu_stop(void *data);
7436 static int should_we_balance(struct lb_env *env)
7438 struct sched_group *sg = env->sd->groups;
7439 struct cpumask *sg_cpus, *sg_mask;
7440 int cpu, balance_cpu = -1;
7443 * In the newly idle case, we will allow all the cpu's
7444 * to do the newly idle load balance.
7446 if (env->idle == CPU_NEWLY_IDLE)
7449 sg_cpus = sched_group_cpus(sg);
7450 sg_mask = sched_group_mask(sg);
7451 /* Try to find first idle cpu */
7452 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7453 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7460 if (balance_cpu == -1)
7461 balance_cpu = group_balance_cpu(sg);
7464 * First idle cpu or the first cpu(busiest) in this sched group
7465 * is eligible for doing load balancing at this and above domains.
7467 return balance_cpu == env->dst_cpu;
7471 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7472 * tasks if there is an imbalance.
7474 static int load_balance(int this_cpu, struct rq *this_rq,
7475 struct sched_domain *sd, enum cpu_idle_type idle,
7476 int *continue_balancing)
7478 int ld_moved, cur_ld_moved, active_balance = 0;
7479 struct sched_domain *sd_parent = sd->parent;
7480 struct sched_group *group;
7482 unsigned long flags;
7483 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7485 struct lb_env env = {
7487 .dst_cpu = this_cpu,
7489 .dst_grpmask = sched_group_cpus(sd->groups),
7491 .loop_break = sched_nr_migrate_break,
7494 .tasks = LIST_HEAD_INIT(env.tasks),
7498 * For NEWLY_IDLE load_balancing, we don't need to consider
7499 * other cpus in our group
7501 if (idle == CPU_NEWLY_IDLE)
7502 env.dst_grpmask = NULL;
7504 cpumask_copy(cpus, cpu_active_mask);
7506 schedstat_inc(sd->lb_count[idle]);
7509 if (!should_we_balance(&env)) {
7510 *continue_balancing = 0;
7514 group = find_busiest_group(&env);
7516 schedstat_inc(sd->lb_nobusyg[idle]);
7520 busiest = find_busiest_queue(&env, group);
7522 schedstat_inc(sd->lb_nobusyq[idle]);
7526 BUG_ON(busiest == env.dst_rq);
7528 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
7530 env.src_cpu = busiest->cpu;
7531 env.src_rq = busiest;
7534 if (busiest->nr_running > 1) {
7536 * Attempt to move tasks. If find_busiest_group has found
7537 * an imbalance but busiest->nr_running <= 1, the group is
7538 * still unbalanced. ld_moved simply stays zero, so it is
7539 * correctly treated as an imbalance.
7541 env.flags |= LBF_ALL_PINNED;
7542 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7545 raw_spin_lock_irqsave(&busiest->lock, flags);
7548 * cur_ld_moved - load moved in current iteration
7549 * ld_moved - cumulative load moved across iterations
7551 cur_ld_moved = detach_tasks(&env);
7554 * We've detached some tasks from busiest_rq. Every
7555 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7556 * unlock busiest->lock, and we are able to be sure
7557 * that nobody can manipulate the tasks in parallel.
7558 * See task_rq_lock() family for the details.
7561 raw_spin_unlock(&busiest->lock);
7565 ld_moved += cur_ld_moved;
7568 local_irq_restore(flags);
7570 if (env.flags & LBF_NEED_BREAK) {
7571 env.flags &= ~LBF_NEED_BREAK;
7576 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7577 * us and move them to an alternate dst_cpu in our sched_group
7578 * where they can run. The upper limit on how many times we
7579 * iterate on same src_cpu is dependent on number of cpus in our
7582 * This changes load balance semantics a bit on who can move
7583 * load to a given_cpu. In addition to the given_cpu itself
7584 * (or a ilb_cpu acting on its behalf where given_cpu is
7585 * nohz-idle), we now have balance_cpu in a position to move
7586 * load to given_cpu. In rare situations, this may cause
7587 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7588 * _independently_ and at _same_ time to move some load to
7589 * given_cpu) causing exceess load to be moved to given_cpu.
7590 * This however should not happen so much in practice and
7591 * moreover subsequent load balance cycles should correct the
7592 * excess load moved.
7594 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7596 /* Prevent to re-select dst_cpu via env's cpus */
7597 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7599 env.dst_rq = cpu_rq(env.new_dst_cpu);
7600 env.dst_cpu = env.new_dst_cpu;
7601 env.flags &= ~LBF_DST_PINNED;
7603 env.loop_break = sched_nr_migrate_break;
7606 * Go back to "more_balance" rather than "redo" since we
7607 * need to continue with same src_cpu.
7613 * We failed to reach balance because of affinity.
7616 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7618 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7619 *group_imbalance = 1;
7622 /* All tasks on this runqueue were pinned by CPU affinity */
7623 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7624 cpumask_clear_cpu(cpu_of(busiest), cpus);
7625 if (!cpumask_empty(cpus)) {
7627 env.loop_break = sched_nr_migrate_break;
7630 goto out_all_pinned;
7635 schedstat_inc(sd->lb_failed[idle]);
7637 * Increment the failure counter only on periodic balance.
7638 * We do not want newidle balance, which can be very
7639 * frequent, pollute the failure counter causing
7640 * excessive cache_hot migrations and active balances.
7642 if (idle != CPU_NEWLY_IDLE)
7643 sd->nr_balance_failed++;
7645 if (need_active_balance(&env)) {
7646 raw_spin_lock_irqsave(&busiest->lock, flags);
7648 /* don't kick the active_load_balance_cpu_stop,
7649 * if the curr task on busiest cpu can't be
7652 if (!cpumask_test_cpu(this_cpu,
7653 tsk_cpus_allowed(busiest->curr))) {
7654 raw_spin_unlock_irqrestore(&busiest->lock,
7656 env.flags |= LBF_ALL_PINNED;
7657 goto out_one_pinned;
7661 * ->active_balance synchronizes accesses to
7662 * ->active_balance_work. Once set, it's cleared
7663 * only after active load balance is finished.
7665 if (!busiest->active_balance) {
7666 busiest->active_balance = 1;
7667 busiest->push_cpu = this_cpu;
7670 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7672 if (active_balance) {
7673 stop_one_cpu_nowait(cpu_of(busiest),
7674 active_load_balance_cpu_stop, busiest,
7675 &busiest->active_balance_work);
7678 /* We've kicked active balancing, force task migration. */
7679 sd->nr_balance_failed = sd->cache_nice_tries+1;
7682 sd->nr_balance_failed = 0;
7684 if (likely(!active_balance)) {
7685 /* We were unbalanced, so reset the balancing interval */
7686 sd->balance_interval = sd->min_interval;
7689 * If we've begun active balancing, start to back off. This
7690 * case may not be covered by the all_pinned logic if there
7691 * is only 1 task on the busy runqueue (because we don't call
7694 if (sd->balance_interval < sd->max_interval)
7695 sd->balance_interval *= 2;
7702 * We reach balance although we may have faced some affinity
7703 * constraints. Clear the imbalance flag if it was set.
7706 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7708 if (*group_imbalance)
7709 *group_imbalance = 0;
7714 * We reach balance because all tasks are pinned at this level so
7715 * we can't migrate them. Let the imbalance flag set so parent level
7716 * can try to migrate them.
7718 schedstat_inc(sd->lb_balanced[idle]);
7720 sd->nr_balance_failed = 0;
7723 /* tune up the balancing interval */
7724 if (((env.flags & LBF_ALL_PINNED) &&
7725 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7726 (sd->balance_interval < sd->max_interval))
7727 sd->balance_interval *= 2;
7734 static inline unsigned long
7735 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7737 unsigned long interval = sd->balance_interval;
7740 interval *= sd->busy_factor;
7742 /* scale ms to jiffies */
7743 interval = msecs_to_jiffies(interval);
7744 interval = clamp(interval, 1UL, max_load_balance_interval);
7750 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
7752 unsigned long interval, next;
7754 /* used by idle balance, so cpu_busy = 0 */
7755 interval = get_sd_balance_interval(sd, 0);
7756 next = sd->last_balance + interval;
7758 if (time_after(*next_balance, next))
7759 *next_balance = next;
7763 * idle_balance is called by schedule() if this_cpu is about to become
7764 * idle. Attempts to pull tasks from other CPUs.
7766 static int idle_balance(struct rq *this_rq)
7768 unsigned long next_balance = jiffies + HZ;
7769 int this_cpu = this_rq->cpu;
7770 struct sched_domain *sd;
7771 int pulled_task = 0;
7775 * We must set idle_stamp _before_ calling idle_balance(), such that we
7776 * measure the duration of idle_balance() as idle time.
7778 this_rq->idle_stamp = rq_clock(this_rq);
7780 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7781 !this_rq->rd->overload) {
7783 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7785 update_next_balance(sd, &next_balance);
7791 raw_spin_unlock(&this_rq->lock);
7793 update_blocked_averages(this_cpu);
7795 for_each_domain(this_cpu, sd) {
7796 int continue_balancing = 1;
7797 u64 t0, domain_cost;
7799 if (!(sd->flags & SD_LOAD_BALANCE))
7802 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7803 update_next_balance(sd, &next_balance);
7807 if (sd->flags & SD_BALANCE_NEWIDLE) {
7808 t0 = sched_clock_cpu(this_cpu);
7810 pulled_task = load_balance(this_cpu, this_rq,
7812 &continue_balancing);
7814 domain_cost = sched_clock_cpu(this_cpu) - t0;
7815 if (domain_cost > sd->max_newidle_lb_cost)
7816 sd->max_newidle_lb_cost = domain_cost;
7818 curr_cost += domain_cost;
7821 update_next_balance(sd, &next_balance);
7824 * Stop searching for tasks to pull if there are
7825 * now runnable tasks on this rq.
7827 if (pulled_task || this_rq->nr_running > 0)
7832 raw_spin_lock(&this_rq->lock);
7834 if (curr_cost > this_rq->max_idle_balance_cost)
7835 this_rq->max_idle_balance_cost = curr_cost;
7838 * While browsing the domains, we released the rq lock, a task could
7839 * have been enqueued in the meantime. Since we're not going idle,
7840 * pretend we pulled a task.
7842 if (this_rq->cfs.h_nr_running && !pulled_task)
7846 /* Move the next balance forward */
7847 if (time_after(this_rq->next_balance, next_balance))
7848 this_rq->next_balance = next_balance;
7850 /* Is there a task of a high priority class? */
7851 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7855 this_rq->idle_stamp = 0;
7861 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7862 * running tasks off the busiest CPU onto idle CPUs. It requires at
7863 * least 1 task to be running on each physical CPU where possible, and
7864 * avoids physical / logical imbalances.
7866 static int active_load_balance_cpu_stop(void *data)
7868 struct rq *busiest_rq = data;
7869 int busiest_cpu = cpu_of(busiest_rq);
7870 int target_cpu = busiest_rq->push_cpu;
7871 struct rq *target_rq = cpu_rq(target_cpu);
7872 struct sched_domain *sd;
7873 struct task_struct *p = NULL;
7875 raw_spin_lock_irq(&busiest_rq->lock);
7877 /* make sure the requested cpu hasn't gone down in the meantime */
7878 if (unlikely(busiest_cpu != smp_processor_id() ||
7879 !busiest_rq->active_balance))
7882 /* Is there any task to move? */
7883 if (busiest_rq->nr_running <= 1)
7887 * This condition is "impossible", if it occurs
7888 * we need to fix it. Originally reported by
7889 * Bjorn Helgaas on a 128-cpu setup.
7891 BUG_ON(busiest_rq == target_rq);
7893 /* Search for an sd spanning us and the target CPU. */
7895 for_each_domain(target_cpu, sd) {
7896 if ((sd->flags & SD_LOAD_BALANCE) &&
7897 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7902 struct lb_env env = {
7904 .dst_cpu = target_cpu,
7905 .dst_rq = target_rq,
7906 .src_cpu = busiest_rq->cpu,
7907 .src_rq = busiest_rq,
7911 schedstat_inc(sd->alb_count);
7913 p = detach_one_task(&env);
7915 schedstat_inc(sd->alb_pushed);
7916 /* Active balancing done, reset the failure counter. */
7917 sd->nr_balance_failed = 0;
7919 schedstat_inc(sd->alb_failed);
7924 busiest_rq->active_balance = 0;
7925 raw_spin_unlock(&busiest_rq->lock);
7928 attach_one_task(target_rq, p);
7935 static inline int on_null_domain(struct rq *rq)
7937 return unlikely(!rcu_dereference_sched(rq->sd));
7940 #ifdef CONFIG_NO_HZ_COMMON
7942 * idle load balancing details
7943 * - When one of the busy CPUs notice that there may be an idle rebalancing
7944 * needed, they will kick the idle load balancer, which then does idle
7945 * load balancing for all the idle CPUs.
7948 cpumask_var_t idle_cpus_mask;
7950 unsigned long next_balance; /* in jiffy units */
7951 } nohz ____cacheline_aligned;
7953 static inline int find_new_ilb(void)
7955 int ilb = cpumask_first(nohz.idle_cpus_mask);
7957 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7964 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7965 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7966 * CPU (if there is one).
7968 static void nohz_balancer_kick(void)
7972 nohz.next_balance++;
7974 ilb_cpu = find_new_ilb();
7976 if (ilb_cpu >= nr_cpu_ids)
7979 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7982 * Use smp_send_reschedule() instead of resched_cpu().
7983 * This way we generate a sched IPI on the target cpu which
7984 * is idle. And the softirq performing nohz idle load balance
7985 * will be run before returning from the IPI.
7987 smp_send_reschedule(ilb_cpu);
7991 void nohz_balance_exit_idle(unsigned int cpu)
7993 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7995 * Completely isolated CPUs don't ever set, so we must test.
7997 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7998 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7999 atomic_dec(&nohz.nr_cpus);
8001 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8005 static inline void set_cpu_sd_state_busy(void)
8007 struct sched_domain *sd;
8008 int cpu = smp_processor_id();
8011 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8013 if (!sd || !sd->nohz_idle)
8017 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8022 void set_cpu_sd_state_idle(void)
8024 struct sched_domain *sd;
8025 int cpu = smp_processor_id();
8028 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8030 if (!sd || sd->nohz_idle)
8034 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8040 * This routine will record that the cpu is going idle with tick stopped.
8041 * This info will be used in performing idle load balancing in the future.
8043 void nohz_balance_enter_idle(int cpu)
8046 * If this cpu is going down, then nothing needs to be done.
8048 if (!cpu_active(cpu))
8051 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8055 * If we're a completely isolated CPU, we don't play.
8057 if (on_null_domain(cpu_rq(cpu)))
8060 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8061 atomic_inc(&nohz.nr_cpus);
8062 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8066 static DEFINE_SPINLOCK(balancing);
8069 * Scale the max load_balance interval with the number of CPUs in the system.
8070 * This trades load-balance latency on larger machines for less cross talk.
8072 void update_max_interval(void)
8074 max_load_balance_interval = HZ*num_online_cpus()/10;
8078 * It checks each scheduling domain to see if it is due to be balanced,
8079 * and initiates a balancing operation if so.
8081 * Balancing parameters are set up in init_sched_domains.
8083 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8085 int continue_balancing = 1;
8087 unsigned long interval;
8088 struct sched_domain *sd;
8089 /* Earliest time when we have to do rebalance again */
8090 unsigned long next_balance = jiffies + 60*HZ;
8091 int update_next_balance = 0;
8092 int need_serialize, need_decay = 0;
8095 update_blocked_averages(cpu);
8098 for_each_domain(cpu, sd) {
8100 * Decay the newidle max times here because this is a regular
8101 * visit to all the domains. Decay ~1% per second.
8103 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8104 sd->max_newidle_lb_cost =
8105 (sd->max_newidle_lb_cost * 253) / 256;
8106 sd->next_decay_max_lb_cost = jiffies + HZ;
8109 max_cost += sd->max_newidle_lb_cost;
8111 if (!(sd->flags & SD_LOAD_BALANCE))
8115 * Stop the load balance at this level. There is another
8116 * CPU in our sched group which is doing load balancing more
8119 if (!continue_balancing) {
8125 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8127 need_serialize = sd->flags & SD_SERIALIZE;
8128 if (need_serialize) {
8129 if (!spin_trylock(&balancing))
8133 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8134 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8136 * The LBF_DST_PINNED logic could have changed
8137 * env->dst_cpu, so we can't know our idle
8138 * state even if we migrated tasks. Update it.
8140 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8142 sd->last_balance = jiffies;
8143 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8146 spin_unlock(&balancing);
8148 if (time_after(next_balance, sd->last_balance + interval)) {
8149 next_balance = sd->last_balance + interval;
8150 update_next_balance = 1;
8155 * Ensure the rq-wide value also decays but keep it at a
8156 * reasonable floor to avoid funnies with rq->avg_idle.
8158 rq->max_idle_balance_cost =
8159 max((u64)sysctl_sched_migration_cost, max_cost);
8164 * next_balance will be updated only when there is a need.
8165 * When the cpu is attached to null domain for ex, it will not be
8168 if (likely(update_next_balance)) {
8169 rq->next_balance = next_balance;
8171 #ifdef CONFIG_NO_HZ_COMMON
8173 * If this CPU has been elected to perform the nohz idle
8174 * balance. Other idle CPUs have already rebalanced with
8175 * nohz_idle_balance() and nohz.next_balance has been
8176 * updated accordingly. This CPU is now running the idle load
8177 * balance for itself and we need to update the
8178 * nohz.next_balance accordingly.
8180 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8181 nohz.next_balance = rq->next_balance;
8186 #ifdef CONFIG_NO_HZ_COMMON
8188 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8189 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8191 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8193 int this_cpu = this_rq->cpu;
8196 /* Earliest time when we have to do rebalance again */
8197 unsigned long next_balance = jiffies + 60*HZ;
8198 int update_next_balance = 0;
8200 if (idle != CPU_IDLE ||
8201 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8204 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8205 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8209 * If this cpu gets work to do, stop the load balancing
8210 * work being done for other cpus. Next load
8211 * balancing owner will pick it up.
8216 rq = cpu_rq(balance_cpu);
8219 * If time for next balance is due,
8222 if (time_after_eq(jiffies, rq->next_balance)) {
8223 raw_spin_lock_irq(&rq->lock);
8224 update_rq_clock(rq);
8225 cpu_load_update_idle(rq);
8226 raw_spin_unlock_irq(&rq->lock);
8227 rebalance_domains(rq, CPU_IDLE);
8230 if (time_after(next_balance, rq->next_balance)) {
8231 next_balance = rq->next_balance;
8232 update_next_balance = 1;
8237 * next_balance will be updated only when there is a need.
8238 * When the CPU is attached to null domain for ex, it will not be
8241 if (likely(update_next_balance))
8242 nohz.next_balance = next_balance;
8244 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8248 * Current heuristic for kicking the idle load balancer in the presence
8249 * of an idle cpu in the system.
8250 * - This rq has more than one task.
8251 * - This rq has at least one CFS task and the capacity of the CPU is
8252 * significantly reduced because of RT tasks or IRQs.
8253 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8254 * multiple busy cpu.
8255 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8256 * domain span are idle.
8258 static inline bool nohz_kick_needed(struct rq *rq)
8260 unsigned long now = jiffies;
8261 struct sched_domain *sd;
8262 struct sched_group_capacity *sgc;
8263 int nr_busy, cpu = rq->cpu;
8266 if (unlikely(rq->idle_balance))
8270 * We may be recently in ticked or tickless idle mode. At the first
8271 * busy tick after returning from idle, we will update the busy stats.
8273 set_cpu_sd_state_busy();
8274 nohz_balance_exit_idle(cpu);
8277 * None are in tickless mode and hence no need for NOHZ idle load
8280 if (likely(!atomic_read(&nohz.nr_cpus)))
8283 if (time_before(now, nohz.next_balance))
8286 if (rq->nr_running >= 2)
8290 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8292 sgc = sd->groups->sgc;
8293 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8302 sd = rcu_dereference(rq->sd);
8304 if ((rq->cfs.h_nr_running >= 1) &&
8305 check_cpu_capacity(rq, sd)) {
8311 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8312 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8313 sched_domain_span(sd)) < cpu)) {
8323 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8327 * run_rebalance_domains is triggered when needed from the scheduler tick.
8328 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8330 static void run_rebalance_domains(struct softirq_action *h)
8332 struct rq *this_rq = this_rq();
8333 enum cpu_idle_type idle = this_rq->idle_balance ?
8334 CPU_IDLE : CPU_NOT_IDLE;
8337 * If this cpu has a pending nohz_balance_kick, then do the
8338 * balancing on behalf of the other idle cpus whose ticks are
8339 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8340 * give the idle cpus a chance to load balance. Else we may
8341 * load balance only within the local sched_domain hierarchy
8342 * and abort nohz_idle_balance altogether if we pull some load.
8344 nohz_idle_balance(this_rq, idle);
8345 rebalance_domains(this_rq, idle);
8349 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8351 void trigger_load_balance(struct rq *rq)
8353 /* Don't need to rebalance while attached to NULL domain */
8354 if (unlikely(on_null_domain(rq)))
8357 if (time_after_eq(jiffies, rq->next_balance))
8358 raise_softirq(SCHED_SOFTIRQ);
8359 #ifdef CONFIG_NO_HZ_COMMON
8360 if (nohz_kick_needed(rq))
8361 nohz_balancer_kick();
8365 static void rq_online_fair(struct rq *rq)
8369 update_runtime_enabled(rq);
8372 static void rq_offline_fair(struct rq *rq)
8376 /* Ensure any throttled groups are reachable by pick_next_task */
8377 unthrottle_offline_cfs_rqs(rq);
8380 #endif /* CONFIG_SMP */
8383 * scheduler tick hitting a task of our scheduling class:
8385 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8387 struct cfs_rq *cfs_rq;
8388 struct sched_entity *se = &curr->se;
8390 for_each_sched_entity(se) {
8391 cfs_rq = cfs_rq_of(se);
8392 entity_tick(cfs_rq, se, queued);
8395 if (static_branch_unlikely(&sched_numa_balancing))
8396 task_tick_numa(rq, curr);
8400 * called on fork with the child task as argument from the parent's context
8401 * - child not yet on the tasklist
8402 * - preemption disabled
8404 static void task_fork_fair(struct task_struct *p)
8406 struct cfs_rq *cfs_rq;
8407 struct sched_entity *se = &p->se, *curr;
8408 struct rq *rq = this_rq();
8410 raw_spin_lock(&rq->lock);
8411 update_rq_clock(rq);
8413 cfs_rq = task_cfs_rq(current);
8414 curr = cfs_rq->curr;
8416 update_curr(cfs_rq);
8417 se->vruntime = curr->vruntime;
8419 place_entity(cfs_rq, se, 1);
8421 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8423 * Upon rescheduling, sched_class::put_prev_task() will place
8424 * 'current' within the tree based on its new key value.
8426 swap(curr->vruntime, se->vruntime);
8430 se->vruntime -= cfs_rq->min_vruntime;
8431 raw_spin_unlock(&rq->lock);
8435 * Priority of the task has changed. Check to see if we preempt
8439 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8441 if (!task_on_rq_queued(p))
8445 * Reschedule if we are currently running on this runqueue and
8446 * our priority decreased, or if we are not currently running on
8447 * this runqueue and our priority is higher than the current's
8449 if (rq->curr == p) {
8450 if (p->prio > oldprio)
8453 check_preempt_curr(rq, p, 0);
8456 static inline bool vruntime_normalized(struct task_struct *p)
8458 struct sched_entity *se = &p->se;
8461 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8462 * the dequeue_entity(.flags=0) will already have normalized the
8469 * When !on_rq, vruntime of the task has usually NOT been normalized.
8470 * But there are some cases where it has already been normalized:
8472 * - A forked child which is waiting for being woken up by
8473 * wake_up_new_task().
8474 * - A task which has been woken up by try_to_wake_up() and
8475 * waiting for actually being woken up by sched_ttwu_pending().
8477 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8483 static void detach_task_cfs_rq(struct task_struct *p)
8485 struct sched_entity *se = &p->se;
8486 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8487 u64 now = cfs_rq_clock_task(cfs_rq);
8489 if (!vruntime_normalized(p)) {
8491 * Fix up our vruntime so that the current sleep doesn't
8492 * cause 'unlimited' sleep bonus.
8494 place_entity(cfs_rq, se, 0);
8495 se->vruntime -= cfs_rq->min_vruntime;
8498 /* Catch up with the cfs_rq and remove our load when we leave */
8499 update_cfs_rq_load_avg(now, cfs_rq, false);
8500 detach_entity_load_avg(cfs_rq, se);
8501 update_tg_load_avg(cfs_rq, false);
8504 static void attach_task_cfs_rq(struct task_struct *p)
8506 struct sched_entity *se = &p->se;
8507 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8508 u64 now = cfs_rq_clock_task(cfs_rq);
8510 #ifdef CONFIG_FAIR_GROUP_SCHED
8512 * Since the real-depth could have been changed (only FAIR
8513 * class maintain depth value), reset depth properly.
8515 se->depth = se->parent ? se->parent->depth + 1 : 0;
8518 /* Synchronize task with its cfs_rq */
8519 update_cfs_rq_load_avg(now, cfs_rq, false);
8520 attach_entity_load_avg(cfs_rq, se);
8521 update_tg_load_avg(cfs_rq, false);
8523 if (!vruntime_normalized(p))
8524 se->vruntime += cfs_rq->min_vruntime;
8527 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8529 detach_task_cfs_rq(p);
8532 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8534 attach_task_cfs_rq(p);
8536 if (task_on_rq_queued(p)) {
8538 * We were most likely switched from sched_rt, so
8539 * kick off the schedule if running, otherwise just see
8540 * if we can still preempt the current task.
8545 check_preempt_curr(rq, p, 0);
8549 /* Account for a task changing its policy or group.
8551 * This routine is mostly called to set cfs_rq->curr field when a task
8552 * migrates between groups/classes.
8554 static void set_curr_task_fair(struct rq *rq)
8556 struct sched_entity *se = &rq->curr->se;
8558 for_each_sched_entity(se) {
8559 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8561 set_next_entity(cfs_rq, se);
8562 /* ensure bandwidth has been allocated on our new cfs_rq */
8563 account_cfs_rq_runtime(cfs_rq, 0);
8567 void init_cfs_rq(struct cfs_rq *cfs_rq)
8569 cfs_rq->tasks_timeline = RB_ROOT;
8570 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8571 #ifndef CONFIG_64BIT
8572 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8575 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8576 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8580 #ifdef CONFIG_FAIR_GROUP_SCHED
8581 static void task_set_group_fair(struct task_struct *p)
8583 struct sched_entity *se = &p->se;
8585 set_task_rq(p, task_cpu(p));
8586 se->depth = se->parent ? se->parent->depth + 1 : 0;
8589 static void task_move_group_fair(struct task_struct *p)
8591 detach_task_cfs_rq(p);
8592 set_task_rq(p, task_cpu(p));
8595 /* Tell se's cfs_rq has been changed -- migrated */
8596 p->se.avg.last_update_time = 0;
8598 attach_task_cfs_rq(p);
8601 static void task_change_group_fair(struct task_struct *p, int type)
8604 case TASK_SET_GROUP:
8605 task_set_group_fair(p);
8608 case TASK_MOVE_GROUP:
8609 task_move_group_fair(p);
8614 void free_fair_sched_group(struct task_group *tg)
8618 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8620 for_each_possible_cpu(i) {
8622 kfree(tg->cfs_rq[i]);
8631 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8633 struct sched_entity *se;
8634 struct cfs_rq *cfs_rq;
8638 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8641 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8645 tg->shares = NICE_0_LOAD;
8647 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8649 for_each_possible_cpu(i) {
8652 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8653 GFP_KERNEL, cpu_to_node(i));
8657 se = kzalloc_node(sizeof(struct sched_entity),
8658 GFP_KERNEL, cpu_to_node(i));
8662 init_cfs_rq(cfs_rq);
8663 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8664 init_entity_runnable_average(se);
8675 void online_fair_sched_group(struct task_group *tg)
8677 struct sched_entity *se;
8681 for_each_possible_cpu(i) {
8685 raw_spin_lock_irq(&rq->lock);
8686 post_init_entity_util_avg(se);
8687 sync_throttle(tg, i);
8688 raw_spin_unlock_irq(&rq->lock);
8692 void unregister_fair_sched_group(struct task_group *tg)
8694 unsigned long flags;
8698 for_each_possible_cpu(cpu) {
8700 remove_entity_load_avg(tg->se[cpu]);
8703 * Only empty task groups can be destroyed; so we can speculatively
8704 * check on_list without danger of it being re-added.
8706 if (!tg->cfs_rq[cpu]->on_list)
8711 raw_spin_lock_irqsave(&rq->lock, flags);
8712 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8713 raw_spin_unlock_irqrestore(&rq->lock, flags);
8717 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8718 struct sched_entity *se, int cpu,
8719 struct sched_entity *parent)
8721 struct rq *rq = cpu_rq(cpu);
8725 init_cfs_rq_runtime(cfs_rq);
8727 tg->cfs_rq[cpu] = cfs_rq;
8730 /* se could be NULL for root_task_group */
8735 se->cfs_rq = &rq->cfs;
8738 se->cfs_rq = parent->my_q;
8739 se->depth = parent->depth + 1;
8743 /* guarantee group entities always have weight */
8744 update_load_set(&se->load, NICE_0_LOAD);
8745 se->parent = parent;
8748 static DEFINE_MUTEX(shares_mutex);
8750 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8753 unsigned long flags;
8756 * We can't change the weight of the root cgroup.
8761 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8763 mutex_lock(&shares_mutex);
8764 if (tg->shares == shares)
8767 tg->shares = shares;
8768 for_each_possible_cpu(i) {
8769 struct rq *rq = cpu_rq(i);
8770 struct sched_entity *se;
8773 /* Propagate contribution to hierarchy */
8774 raw_spin_lock_irqsave(&rq->lock, flags);
8776 /* Possible calls to update_curr() need rq clock */
8777 update_rq_clock(rq);
8778 for_each_sched_entity(se)
8779 update_cfs_shares(group_cfs_rq(se));
8780 raw_spin_unlock_irqrestore(&rq->lock, flags);
8784 mutex_unlock(&shares_mutex);
8787 #else /* CONFIG_FAIR_GROUP_SCHED */
8789 void free_fair_sched_group(struct task_group *tg) { }
8791 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8796 void online_fair_sched_group(struct task_group *tg) { }
8798 void unregister_fair_sched_group(struct task_group *tg) { }
8800 #endif /* CONFIG_FAIR_GROUP_SCHED */
8803 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8805 struct sched_entity *se = &task->se;
8806 unsigned int rr_interval = 0;
8809 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8812 if (rq->cfs.load.weight)
8813 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8819 * All the scheduling class methods:
8821 const struct sched_class fair_sched_class = {
8822 .next = &idle_sched_class,
8823 .enqueue_task = enqueue_task_fair,
8824 .dequeue_task = dequeue_task_fair,
8825 .yield_task = yield_task_fair,
8826 .yield_to_task = yield_to_task_fair,
8828 .check_preempt_curr = check_preempt_wakeup,
8830 .pick_next_task = pick_next_task_fair,
8831 .put_prev_task = put_prev_task_fair,
8834 .select_task_rq = select_task_rq_fair,
8835 .migrate_task_rq = migrate_task_rq_fair,
8837 .rq_online = rq_online_fair,
8838 .rq_offline = rq_offline_fair,
8840 .task_dead = task_dead_fair,
8841 .set_cpus_allowed = set_cpus_allowed_common,
8844 .set_curr_task = set_curr_task_fair,
8845 .task_tick = task_tick_fair,
8846 .task_fork = task_fork_fair,
8848 .prio_changed = prio_changed_fair,
8849 .switched_from = switched_from_fair,
8850 .switched_to = switched_to_fair,
8852 .get_rr_interval = get_rr_interval_fair,
8854 .update_curr = update_curr_fair,
8856 #ifdef CONFIG_FAIR_GROUP_SCHED
8857 .task_change_group = task_change_group_fair,
8861 #ifdef CONFIG_SCHED_DEBUG
8862 void print_cfs_stats(struct seq_file *m, int cpu)
8864 struct cfs_rq *cfs_rq;
8867 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8868 print_cfs_rq(m, cpu, cfs_rq);
8872 #ifdef CONFIG_NUMA_BALANCING
8873 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8876 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8878 for_each_online_node(node) {
8879 if (p->numa_faults) {
8880 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8881 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8883 if (p->numa_group) {
8884 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8885 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8887 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8890 #endif /* CONFIG_NUMA_BALANCING */
8891 #endif /* CONFIG_SCHED_DEBUG */
8893 __init void init_sched_fair_class(void)
8896 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8898 #ifdef CONFIG_NO_HZ_COMMON
8899 nohz.next_balance = jiffies;
8900 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);