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.
1587 * select_idle_siblings() uses an per-cpu cpumask that
1588 * can be used from IRQ context.
1590 local_irq_disable();
1591 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1597 task_numa_assign(env, cur, imp);
1602 static void task_numa_find_cpu(struct task_numa_env *env,
1603 long taskimp, long groupimp)
1607 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1608 /* Skip this CPU if the source task cannot migrate */
1609 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1613 task_numa_compare(env, taskimp, groupimp);
1617 /* Only move tasks to a NUMA node less busy than the current node. */
1618 static bool numa_has_capacity(struct task_numa_env *env)
1620 struct numa_stats *src = &env->src_stats;
1621 struct numa_stats *dst = &env->dst_stats;
1623 if (src->has_free_capacity && !dst->has_free_capacity)
1627 * Only consider a task move if the source has a higher load
1628 * than the destination, corrected for CPU capacity on each node.
1630 * src->load dst->load
1631 * --------------------- vs ---------------------
1632 * src->compute_capacity dst->compute_capacity
1634 if (src->load * dst->compute_capacity * env->imbalance_pct >
1636 dst->load * src->compute_capacity * 100)
1642 static int task_numa_migrate(struct task_struct *p)
1644 struct task_numa_env env = {
1647 .src_cpu = task_cpu(p),
1648 .src_nid = task_node(p),
1650 .imbalance_pct = 112,
1656 struct sched_domain *sd;
1657 unsigned long taskweight, groupweight;
1659 long taskimp, groupimp;
1662 * Pick the lowest SD_NUMA domain, as that would have the smallest
1663 * imbalance and would be the first to start moving tasks about.
1665 * And we want to avoid any moving of tasks about, as that would create
1666 * random movement of tasks -- counter the numa conditions we're trying
1670 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1672 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1676 * Cpusets can break the scheduler domain tree into smaller
1677 * balance domains, some of which do not cross NUMA boundaries.
1678 * Tasks that are "trapped" in such domains cannot be migrated
1679 * elsewhere, so there is no point in (re)trying.
1681 if (unlikely(!sd)) {
1682 p->numa_preferred_nid = task_node(p);
1686 env.dst_nid = p->numa_preferred_nid;
1687 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1688 taskweight = task_weight(p, env.src_nid, dist);
1689 groupweight = group_weight(p, env.src_nid, dist);
1690 update_numa_stats(&env.src_stats, env.src_nid);
1691 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1692 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1693 update_numa_stats(&env.dst_stats, env.dst_nid);
1695 /* Try to find a spot on the preferred nid. */
1696 if (numa_has_capacity(&env))
1697 task_numa_find_cpu(&env, taskimp, groupimp);
1700 * Look at other nodes in these cases:
1701 * - there is no space available on the preferred_nid
1702 * - the task is part of a numa_group that is interleaved across
1703 * multiple NUMA nodes; in order to better consolidate the group,
1704 * we need to check other locations.
1706 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1707 for_each_online_node(nid) {
1708 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1711 dist = node_distance(env.src_nid, env.dst_nid);
1712 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1714 taskweight = task_weight(p, env.src_nid, dist);
1715 groupweight = group_weight(p, env.src_nid, dist);
1718 /* Only consider nodes where both task and groups benefit */
1719 taskimp = task_weight(p, nid, dist) - taskweight;
1720 groupimp = group_weight(p, nid, dist) - groupweight;
1721 if (taskimp < 0 && groupimp < 0)
1726 update_numa_stats(&env.dst_stats, env.dst_nid);
1727 if (numa_has_capacity(&env))
1728 task_numa_find_cpu(&env, taskimp, groupimp);
1733 * If the task is part of a workload that spans multiple NUMA nodes,
1734 * and is migrating into one of the workload's active nodes, remember
1735 * this node as the task's preferred numa node, so the workload can
1737 * A task that migrated to a second choice node will be better off
1738 * trying for a better one later. Do not set the preferred node here.
1740 if (p->numa_group) {
1741 struct numa_group *ng = p->numa_group;
1743 if (env.best_cpu == -1)
1748 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1749 sched_setnuma(p, env.dst_nid);
1752 /* No better CPU than the current one was found. */
1753 if (env.best_cpu == -1)
1757 * Reset the scan period if the task is being rescheduled on an
1758 * alternative node to recheck if the tasks is now properly placed.
1760 p->numa_scan_period = task_scan_min(p);
1762 if (env.best_task == NULL) {
1763 ret = migrate_task_to(p, env.best_cpu);
1765 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1769 ret = migrate_swap(p, env.best_task);
1771 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1772 put_task_struct(env.best_task);
1776 /* Attempt to migrate a task to a CPU on the preferred node. */
1777 static void numa_migrate_preferred(struct task_struct *p)
1779 unsigned long interval = HZ;
1781 /* This task has no NUMA fault statistics yet */
1782 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1785 /* Periodically retry migrating the task to the preferred node */
1786 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1787 p->numa_migrate_retry = jiffies + interval;
1789 /* Success if task is already running on preferred CPU */
1790 if (task_node(p) == p->numa_preferred_nid)
1793 /* Otherwise, try migrate to a CPU on the preferred node */
1794 task_numa_migrate(p);
1798 * Find out how many nodes on the workload is actively running on. Do this by
1799 * tracking the nodes from which NUMA hinting faults are triggered. This can
1800 * be different from the set of nodes where the workload's memory is currently
1803 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1805 unsigned long faults, max_faults = 0;
1806 int nid, active_nodes = 0;
1808 for_each_online_node(nid) {
1809 faults = group_faults_cpu(numa_group, nid);
1810 if (faults > max_faults)
1811 max_faults = faults;
1814 for_each_online_node(nid) {
1815 faults = group_faults_cpu(numa_group, nid);
1816 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1820 numa_group->max_faults_cpu = max_faults;
1821 numa_group->active_nodes = active_nodes;
1825 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1826 * increments. The more local the fault statistics are, the higher the scan
1827 * period will be for the next scan window. If local/(local+remote) ratio is
1828 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1829 * the scan period will decrease. Aim for 70% local accesses.
1831 #define NUMA_PERIOD_SLOTS 10
1832 #define NUMA_PERIOD_THRESHOLD 7
1835 * Increase the scan period (slow down scanning) if the majority of
1836 * our memory is already on our local node, or if the majority of
1837 * the page accesses are shared with other processes.
1838 * Otherwise, decrease the scan period.
1840 static void update_task_scan_period(struct task_struct *p,
1841 unsigned long shared, unsigned long private)
1843 unsigned int period_slot;
1847 unsigned long remote = p->numa_faults_locality[0];
1848 unsigned long local = p->numa_faults_locality[1];
1851 * If there were no record hinting faults then either the task is
1852 * completely idle or all activity is areas that are not of interest
1853 * to automatic numa balancing. Related to that, if there were failed
1854 * migration then it implies we are migrating too quickly or the local
1855 * node is overloaded. In either case, scan slower
1857 if (local + shared == 0 || p->numa_faults_locality[2]) {
1858 p->numa_scan_period = min(p->numa_scan_period_max,
1859 p->numa_scan_period << 1);
1861 p->mm->numa_next_scan = jiffies +
1862 msecs_to_jiffies(p->numa_scan_period);
1868 * Prepare to scale scan period relative to the current period.
1869 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1870 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1871 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1873 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1874 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1875 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1876 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1879 diff = slot * period_slot;
1881 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1884 * Scale scan rate increases based on sharing. There is an
1885 * inverse relationship between the degree of sharing and
1886 * the adjustment made to the scanning period. Broadly
1887 * speaking the intent is that there is little point
1888 * scanning faster if shared accesses dominate as it may
1889 * simply bounce migrations uselessly
1891 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1892 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1895 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1896 task_scan_min(p), task_scan_max(p));
1897 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1901 * Get the fraction of time the task has been running since the last
1902 * NUMA placement cycle. The scheduler keeps similar statistics, but
1903 * decays those on a 32ms period, which is orders of magnitude off
1904 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1905 * stats only if the task is so new there are no NUMA statistics yet.
1907 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1909 u64 runtime, delta, now;
1910 /* Use the start of this time slice to avoid calculations. */
1911 now = p->se.exec_start;
1912 runtime = p->se.sum_exec_runtime;
1914 if (p->last_task_numa_placement) {
1915 delta = runtime - p->last_sum_exec_runtime;
1916 *period = now - p->last_task_numa_placement;
1918 delta = p->se.avg.load_sum / p->se.load.weight;
1919 *period = LOAD_AVG_MAX;
1922 p->last_sum_exec_runtime = runtime;
1923 p->last_task_numa_placement = now;
1929 * Determine the preferred nid for a task in a numa_group. This needs to
1930 * be done in a way that produces consistent results with group_weight,
1931 * otherwise workloads might not converge.
1933 static int preferred_group_nid(struct task_struct *p, int nid)
1938 /* Direct connections between all NUMA nodes. */
1939 if (sched_numa_topology_type == NUMA_DIRECT)
1943 * On a system with glueless mesh NUMA topology, group_weight
1944 * scores nodes according to the number of NUMA hinting faults on
1945 * both the node itself, and on nearby nodes.
1947 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1948 unsigned long score, max_score = 0;
1949 int node, max_node = nid;
1951 dist = sched_max_numa_distance;
1953 for_each_online_node(node) {
1954 score = group_weight(p, node, dist);
1955 if (score > max_score) {
1964 * Finding the preferred nid in a system with NUMA backplane
1965 * interconnect topology is more involved. The goal is to locate
1966 * tasks from numa_groups near each other in the system, and
1967 * untangle workloads from different sides of the system. This requires
1968 * searching down the hierarchy of node groups, recursively searching
1969 * inside the highest scoring group of nodes. The nodemask tricks
1970 * keep the complexity of the search down.
1972 nodes = node_online_map;
1973 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1974 unsigned long max_faults = 0;
1975 nodemask_t max_group = NODE_MASK_NONE;
1978 /* Are there nodes at this distance from each other? */
1979 if (!find_numa_distance(dist))
1982 for_each_node_mask(a, nodes) {
1983 unsigned long faults = 0;
1984 nodemask_t this_group;
1985 nodes_clear(this_group);
1987 /* Sum group's NUMA faults; includes a==b case. */
1988 for_each_node_mask(b, nodes) {
1989 if (node_distance(a, b) < dist) {
1990 faults += group_faults(p, b);
1991 node_set(b, this_group);
1992 node_clear(b, nodes);
1996 /* Remember the top group. */
1997 if (faults > max_faults) {
1998 max_faults = faults;
1999 max_group = this_group;
2001 * subtle: at the smallest distance there is
2002 * just one node left in each "group", the
2003 * winner is the preferred nid.
2008 /* Next round, evaluate the nodes within max_group. */
2016 static void task_numa_placement(struct task_struct *p)
2018 int seq, nid, max_nid = -1, max_group_nid = -1;
2019 unsigned long max_faults = 0, max_group_faults = 0;
2020 unsigned long fault_types[2] = { 0, 0 };
2021 unsigned long total_faults;
2022 u64 runtime, period;
2023 spinlock_t *group_lock = NULL;
2026 * The p->mm->numa_scan_seq field gets updated without
2027 * exclusive access. Use READ_ONCE() here to ensure
2028 * that the field is read in a single access:
2030 seq = READ_ONCE(p->mm->numa_scan_seq);
2031 if (p->numa_scan_seq == seq)
2033 p->numa_scan_seq = seq;
2034 p->numa_scan_period_max = task_scan_max(p);
2036 total_faults = p->numa_faults_locality[0] +
2037 p->numa_faults_locality[1];
2038 runtime = numa_get_avg_runtime(p, &period);
2040 /* If the task is part of a group prevent parallel updates to group stats */
2041 if (p->numa_group) {
2042 group_lock = &p->numa_group->lock;
2043 spin_lock_irq(group_lock);
2046 /* Find the node with the highest number of faults */
2047 for_each_online_node(nid) {
2048 /* Keep track of the offsets in numa_faults array */
2049 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2050 unsigned long faults = 0, group_faults = 0;
2053 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2054 long diff, f_diff, f_weight;
2056 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2057 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2058 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2059 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2061 /* Decay existing window, copy faults since last scan */
2062 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2063 fault_types[priv] += p->numa_faults[membuf_idx];
2064 p->numa_faults[membuf_idx] = 0;
2067 * Normalize the faults_from, so all tasks in a group
2068 * count according to CPU use, instead of by the raw
2069 * number of faults. Tasks with little runtime have
2070 * little over-all impact on throughput, and thus their
2071 * faults are less important.
2073 f_weight = div64_u64(runtime << 16, period + 1);
2074 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2076 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2077 p->numa_faults[cpubuf_idx] = 0;
2079 p->numa_faults[mem_idx] += diff;
2080 p->numa_faults[cpu_idx] += f_diff;
2081 faults += p->numa_faults[mem_idx];
2082 p->total_numa_faults += diff;
2083 if (p->numa_group) {
2085 * safe because we can only change our own group
2087 * mem_idx represents the offset for a given
2088 * nid and priv in a specific region because it
2089 * is at the beginning of the numa_faults array.
2091 p->numa_group->faults[mem_idx] += diff;
2092 p->numa_group->faults_cpu[mem_idx] += f_diff;
2093 p->numa_group->total_faults += diff;
2094 group_faults += p->numa_group->faults[mem_idx];
2098 if (faults > max_faults) {
2099 max_faults = faults;
2103 if (group_faults > max_group_faults) {
2104 max_group_faults = group_faults;
2105 max_group_nid = nid;
2109 update_task_scan_period(p, fault_types[0], fault_types[1]);
2111 if (p->numa_group) {
2112 numa_group_count_active_nodes(p->numa_group);
2113 spin_unlock_irq(group_lock);
2114 max_nid = preferred_group_nid(p, max_group_nid);
2118 /* Set the new preferred node */
2119 if (max_nid != p->numa_preferred_nid)
2120 sched_setnuma(p, max_nid);
2122 if (task_node(p) != p->numa_preferred_nid)
2123 numa_migrate_preferred(p);
2127 static inline int get_numa_group(struct numa_group *grp)
2129 return atomic_inc_not_zero(&grp->refcount);
2132 static inline void put_numa_group(struct numa_group *grp)
2134 if (atomic_dec_and_test(&grp->refcount))
2135 kfree_rcu(grp, rcu);
2138 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2141 struct numa_group *grp, *my_grp;
2142 struct task_struct *tsk;
2144 int cpu = cpupid_to_cpu(cpupid);
2147 if (unlikely(!p->numa_group)) {
2148 unsigned int size = sizeof(struct numa_group) +
2149 4*nr_node_ids*sizeof(unsigned long);
2151 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2155 atomic_set(&grp->refcount, 1);
2156 grp->active_nodes = 1;
2157 grp->max_faults_cpu = 0;
2158 spin_lock_init(&grp->lock);
2160 /* Second half of the array tracks nids where faults happen */
2161 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2164 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2165 grp->faults[i] = p->numa_faults[i];
2167 grp->total_faults = p->total_numa_faults;
2170 rcu_assign_pointer(p->numa_group, grp);
2174 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2176 if (!cpupid_match_pid(tsk, cpupid))
2179 grp = rcu_dereference(tsk->numa_group);
2183 my_grp = p->numa_group;
2188 * Only join the other group if its bigger; if we're the bigger group,
2189 * the other task will join us.
2191 if (my_grp->nr_tasks > grp->nr_tasks)
2195 * Tie-break on the grp address.
2197 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2200 /* Always join threads in the same process. */
2201 if (tsk->mm == current->mm)
2204 /* Simple filter to avoid false positives due to PID collisions */
2205 if (flags & TNF_SHARED)
2208 /* Update priv based on whether false sharing was detected */
2211 if (join && !get_numa_group(grp))
2219 BUG_ON(irqs_disabled());
2220 double_lock_irq(&my_grp->lock, &grp->lock);
2222 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2223 my_grp->faults[i] -= p->numa_faults[i];
2224 grp->faults[i] += p->numa_faults[i];
2226 my_grp->total_faults -= p->total_numa_faults;
2227 grp->total_faults += p->total_numa_faults;
2232 spin_unlock(&my_grp->lock);
2233 spin_unlock_irq(&grp->lock);
2235 rcu_assign_pointer(p->numa_group, grp);
2237 put_numa_group(my_grp);
2245 void task_numa_free(struct task_struct *p)
2247 struct numa_group *grp = p->numa_group;
2248 void *numa_faults = p->numa_faults;
2249 unsigned long flags;
2253 spin_lock_irqsave(&grp->lock, flags);
2254 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2255 grp->faults[i] -= p->numa_faults[i];
2256 grp->total_faults -= p->total_numa_faults;
2259 spin_unlock_irqrestore(&grp->lock, flags);
2260 RCU_INIT_POINTER(p->numa_group, NULL);
2261 put_numa_group(grp);
2264 p->numa_faults = NULL;
2269 * Got a PROT_NONE fault for a page on @node.
2271 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2273 struct task_struct *p = current;
2274 bool migrated = flags & TNF_MIGRATED;
2275 int cpu_node = task_node(current);
2276 int local = !!(flags & TNF_FAULT_LOCAL);
2277 struct numa_group *ng;
2280 if (!static_branch_likely(&sched_numa_balancing))
2283 /* for example, ksmd faulting in a user's mm */
2287 /* Allocate buffer to track faults on a per-node basis */
2288 if (unlikely(!p->numa_faults)) {
2289 int size = sizeof(*p->numa_faults) *
2290 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2292 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2293 if (!p->numa_faults)
2296 p->total_numa_faults = 0;
2297 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2301 * First accesses are treated as private, otherwise consider accesses
2302 * to be private if the accessing pid has not changed
2304 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2307 priv = cpupid_match_pid(p, last_cpupid);
2308 if (!priv && !(flags & TNF_NO_GROUP))
2309 task_numa_group(p, last_cpupid, flags, &priv);
2313 * If a workload spans multiple NUMA nodes, a shared fault that
2314 * occurs wholly within the set of nodes that the workload is
2315 * actively using should be counted as local. This allows the
2316 * scan rate to slow down when a workload has settled down.
2319 if (!priv && !local && ng && ng->active_nodes > 1 &&
2320 numa_is_active_node(cpu_node, ng) &&
2321 numa_is_active_node(mem_node, ng))
2324 task_numa_placement(p);
2327 * Retry task to preferred node migration periodically, in case it
2328 * case it previously failed, or the scheduler moved us.
2330 if (time_after(jiffies, p->numa_migrate_retry))
2331 numa_migrate_preferred(p);
2334 p->numa_pages_migrated += pages;
2335 if (flags & TNF_MIGRATE_FAIL)
2336 p->numa_faults_locality[2] += pages;
2338 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2339 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2340 p->numa_faults_locality[local] += pages;
2343 static void reset_ptenuma_scan(struct task_struct *p)
2346 * We only did a read acquisition of the mmap sem, so
2347 * p->mm->numa_scan_seq is written to without exclusive access
2348 * and the update is not guaranteed to be atomic. That's not
2349 * much of an issue though, since this is just used for
2350 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2351 * expensive, to avoid any form of compiler optimizations:
2353 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2354 p->mm->numa_scan_offset = 0;
2358 * The expensive part of numa migration is done from task_work context.
2359 * Triggered from task_tick_numa().
2361 void task_numa_work(struct callback_head *work)
2363 unsigned long migrate, next_scan, now = jiffies;
2364 struct task_struct *p = current;
2365 struct mm_struct *mm = p->mm;
2366 u64 runtime = p->se.sum_exec_runtime;
2367 struct vm_area_struct *vma;
2368 unsigned long start, end;
2369 unsigned long nr_pte_updates = 0;
2370 long pages, virtpages;
2372 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2374 work->next = work; /* protect against double add */
2376 * Who cares about NUMA placement when they're dying.
2378 * NOTE: make sure not to dereference p->mm before this check,
2379 * exit_task_work() happens _after_ exit_mm() so we could be called
2380 * without p->mm even though we still had it when we enqueued this
2383 if (p->flags & PF_EXITING)
2386 if (!mm->numa_next_scan) {
2387 mm->numa_next_scan = now +
2388 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2392 * Enforce maximal scan/migration frequency..
2394 migrate = mm->numa_next_scan;
2395 if (time_before(now, migrate))
2398 if (p->numa_scan_period == 0) {
2399 p->numa_scan_period_max = task_scan_max(p);
2400 p->numa_scan_period = task_scan_min(p);
2403 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2404 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2408 * Delay this task enough that another task of this mm will likely win
2409 * the next time around.
2411 p->node_stamp += 2 * TICK_NSEC;
2413 start = mm->numa_scan_offset;
2414 pages = sysctl_numa_balancing_scan_size;
2415 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2416 virtpages = pages * 8; /* Scan up to this much virtual space */
2421 down_read(&mm->mmap_sem);
2422 vma = find_vma(mm, start);
2424 reset_ptenuma_scan(p);
2428 for (; vma; vma = vma->vm_next) {
2429 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2430 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2435 * Shared library pages mapped by multiple processes are not
2436 * migrated as it is expected they are cache replicated. Avoid
2437 * hinting faults in read-only file-backed mappings or the vdso
2438 * as migrating the pages will be of marginal benefit.
2441 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2445 * Skip inaccessible VMAs to avoid any confusion between
2446 * PROT_NONE and NUMA hinting ptes
2448 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2452 start = max(start, vma->vm_start);
2453 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2454 end = min(end, vma->vm_end);
2455 nr_pte_updates = change_prot_numa(vma, start, end);
2458 * Try to scan sysctl_numa_balancing_size worth of
2459 * hpages that have at least one present PTE that
2460 * is not already pte-numa. If the VMA contains
2461 * areas that are unused or already full of prot_numa
2462 * PTEs, scan up to virtpages, to skip through those
2466 pages -= (end - start) >> PAGE_SHIFT;
2467 virtpages -= (end - start) >> PAGE_SHIFT;
2470 if (pages <= 0 || virtpages <= 0)
2474 } while (end != vma->vm_end);
2479 * It is possible to reach the end of the VMA list but the last few
2480 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2481 * would find the !migratable VMA on the next scan but not reset the
2482 * scanner to the start so check it now.
2485 mm->numa_scan_offset = start;
2487 reset_ptenuma_scan(p);
2488 up_read(&mm->mmap_sem);
2491 * Make sure tasks use at least 32x as much time to run other code
2492 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2493 * Usually update_task_scan_period slows down scanning enough; on an
2494 * overloaded system we need to limit overhead on a per task basis.
2496 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2497 u64 diff = p->se.sum_exec_runtime - runtime;
2498 p->node_stamp += 32 * diff;
2503 * Drive the periodic memory faults..
2505 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2507 struct callback_head *work = &curr->numa_work;
2511 * We don't care about NUMA placement if we don't have memory.
2513 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2517 * Using runtime rather than walltime has the dual advantage that
2518 * we (mostly) drive the selection from busy threads and that the
2519 * task needs to have done some actual work before we bother with
2522 now = curr->se.sum_exec_runtime;
2523 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2525 if (now > curr->node_stamp + period) {
2526 if (!curr->node_stamp)
2527 curr->numa_scan_period = task_scan_min(curr);
2528 curr->node_stamp += period;
2530 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2531 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2532 task_work_add(curr, work, true);
2537 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2541 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2545 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2548 #endif /* CONFIG_NUMA_BALANCING */
2551 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2553 update_load_add(&cfs_rq->load, se->load.weight);
2554 if (!parent_entity(se))
2555 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2557 if (entity_is_task(se)) {
2558 struct rq *rq = rq_of(cfs_rq);
2560 account_numa_enqueue(rq, task_of(se));
2561 list_add(&se->group_node, &rq->cfs_tasks);
2564 cfs_rq->nr_running++;
2568 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2570 update_load_sub(&cfs_rq->load, se->load.weight);
2571 if (!parent_entity(se))
2572 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2574 if (entity_is_task(se)) {
2575 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2576 list_del_init(&se->group_node);
2579 cfs_rq->nr_running--;
2582 #ifdef CONFIG_FAIR_GROUP_SCHED
2584 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2586 long tg_weight, load, shares;
2589 * This really should be: cfs_rq->avg.load_avg, but instead we use
2590 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2591 * the shares for small weight interactive tasks.
2593 load = scale_load_down(cfs_rq->load.weight);
2595 tg_weight = atomic_long_read(&tg->load_avg);
2597 /* Ensure tg_weight >= load */
2598 tg_weight -= cfs_rq->tg_load_avg_contrib;
2601 shares = (tg->shares * load);
2603 shares /= tg_weight;
2605 if (shares < MIN_SHARES)
2606 shares = MIN_SHARES;
2607 if (shares > tg->shares)
2608 shares = tg->shares;
2612 # else /* CONFIG_SMP */
2613 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2617 # endif /* CONFIG_SMP */
2619 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2620 unsigned long weight)
2623 /* commit outstanding execution time */
2624 if (cfs_rq->curr == se)
2625 update_curr(cfs_rq);
2626 account_entity_dequeue(cfs_rq, se);
2629 update_load_set(&se->load, weight);
2632 account_entity_enqueue(cfs_rq, se);
2635 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2637 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2639 struct task_group *tg;
2640 struct sched_entity *se;
2644 se = tg->se[cpu_of(rq_of(cfs_rq))];
2645 if (!se || throttled_hierarchy(cfs_rq))
2648 if (likely(se->load.weight == tg->shares))
2651 shares = calc_cfs_shares(cfs_rq, tg);
2653 reweight_entity(cfs_rq_of(se), se, shares);
2655 #else /* CONFIG_FAIR_GROUP_SCHED */
2656 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2659 #endif /* CONFIG_FAIR_GROUP_SCHED */
2662 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2663 static const u32 runnable_avg_yN_inv[] = {
2664 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2665 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2666 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2667 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2668 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2669 0x85aac367, 0x82cd8698,
2673 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2674 * over-estimates when re-combining.
2676 static const u32 runnable_avg_yN_sum[] = {
2677 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2678 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2679 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2683 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2684 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2687 static const u32 __accumulated_sum_N32[] = {
2688 0, 23371, 35056, 40899, 43820, 45281,
2689 46011, 46376, 46559, 46650, 46696, 46719,
2694 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2696 static __always_inline u64 decay_load(u64 val, u64 n)
2698 unsigned int local_n;
2702 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2705 /* after bounds checking we can collapse to 32-bit */
2709 * As y^PERIOD = 1/2, we can combine
2710 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2711 * With a look-up table which covers y^n (n<PERIOD)
2713 * To achieve constant time decay_load.
2715 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2716 val >>= local_n / LOAD_AVG_PERIOD;
2717 local_n %= LOAD_AVG_PERIOD;
2720 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2725 * For updates fully spanning n periods, the contribution to runnable
2726 * average will be: \Sum 1024*y^n
2728 * We can compute this reasonably efficiently by combining:
2729 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2731 static u32 __compute_runnable_contrib(u64 n)
2735 if (likely(n <= LOAD_AVG_PERIOD))
2736 return runnable_avg_yN_sum[n];
2737 else if (unlikely(n >= LOAD_AVG_MAX_N))
2738 return LOAD_AVG_MAX;
2740 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2741 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2742 n %= LOAD_AVG_PERIOD;
2743 contrib = decay_load(contrib, n);
2744 return contrib + runnable_avg_yN_sum[n];
2747 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2750 * We can represent the historical contribution to runnable average as the
2751 * coefficients of a geometric series. To do this we sub-divide our runnable
2752 * history into segments of approximately 1ms (1024us); label the segment that
2753 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2755 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2757 * (now) (~1ms ago) (~2ms ago)
2759 * Let u_i denote the fraction of p_i that the entity was runnable.
2761 * We then designate the fractions u_i as our co-efficients, yielding the
2762 * following representation of historical load:
2763 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2765 * We choose y based on the with of a reasonably scheduling period, fixing:
2768 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2769 * approximately half as much as the contribution to load within the last ms
2772 * When a period "rolls over" and we have new u_0`, multiplying the previous
2773 * sum again by y is sufficient to update:
2774 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2775 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2777 static __always_inline int
2778 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2779 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2781 u64 delta, scaled_delta, periods;
2783 unsigned int delta_w, scaled_delta_w, decayed = 0;
2784 unsigned long scale_freq, scale_cpu;
2786 delta = now - sa->last_update_time;
2788 * This should only happen when time goes backwards, which it
2789 * unfortunately does during sched clock init when we swap over to TSC.
2791 if ((s64)delta < 0) {
2792 sa->last_update_time = now;
2797 * Use 1024ns as the unit of measurement since it's a reasonable
2798 * approximation of 1us and fast to compute.
2803 sa->last_update_time = now;
2805 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2806 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2808 /* delta_w is the amount already accumulated against our next period */
2809 delta_w = sa->period_contrib;
2810 if (delta + delta_w >= 1024) {
2813 /* how much left for next period will start over, we don't know yet */
2814 sa->period_contrib = 0;
2817 * Now that we know we're crossing a period boundary, figure
2818 * out how much from delta we need to complete the current
2819 * period and accrue it.
2821 delta_w = 1024 - delta_w;
2822 scaled_delta_w = cap_scale(delta_w, scale_freq);
2824 sa->load_sum += weight * scaled_delta_w;
2826 cfs_rq->runnable_load_sum +=
2827 weight * scaled_delta_w;
2831 sa->util_sum += scaled_delta_w * scale_cpu;
2835 /* Figure out how many additional periods this update spans */
2836 periods = delta / 1024;
2839 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2841 cfs_rq->runnable_load_sum =
2842 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2844 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2846 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2847 contrib = __compute_runnable_contrib(periods);
2848 contrib = cap_scale(contrib, scale_freq);
2850 sa->load_sum += weight * contrib;
2852 cfs_rq->runnable_load_sum += weight * contrib;
2855 sa->util_sum += contrib * scale_cpu;
2858 /* Remainder of delta accrued against u_0` */
2859 scaled_delta = cap_scale(delta, scale_freq);
2861 sa->load_sum += weight * scaled_delta;
2863 cfs_rq->runnable_load_sum += weight * scaled_delta;
2866 sa->util_sum += scaled_delta * scale_cpu;
2868 sa->period_contrib += delta;
2871 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2873 cfs_rq->runnable_load_avg =
2874 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2876 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2882 #ifdef CONFIG_FAIR_GROUP_SCHED
2884 * update_tg_load_avg - update the tg's load avg
2885 * @cfs_rq: the cfs_rq whose avg changed
2886 * @force: update regardless of how small the difference
2888 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2889 * However, because tg->load_avg is a global value there are performance
2892 * In order to avoid having to look at the other cfs_rq's, we use a
2893 * differential update where we store the last value we propagated. This in
2894 * turn allows skipping updates if the differential is 'small'.
2896 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2897 * done) and effective_load() (which is not done because it is too costly).
2899 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2901 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2904 * No need to update load_avg for root_task_group as it is not used.
2906 if (cfs_rq->tg == &root_task_group)
2909 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2910 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2911 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2916 * Called within set_task_rq() right before setting a task's cpu. The
2917 * caller only guarantees p->pi_lock is held; no other assumptions,
2918 * including the state of rq->lock, should be made.
2920 void set_task_rq_fair(struct sched_entity *se,
2921 struct cfs_rq *prev, struct cfs_rq *next)
2923 if (!sched_feat(ATTACH_AGE_LOAD))
2927 * We are supposed to update the task to "current" time, then its up to
2928 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2929 * getting what current time is, so simply throw away the out-of-date
2930 * time. This will result in the wakee task is less decayed, but giving
2931 * the wakee more load sounds not bad.
2933 if (se->avg.last_update_time && prev) {
2934 u64 p_last_update_time;
2935 u64 n_last_update_time;
2937 #ifndef CONFIG_64BIT
2938 u64 p_last_update_time_copy;
2939 u64 n_last_update_time_copy;
2942 p_last_update_time_copy = prev->load_last_update_time_copy;
2943 n_last_update_time_copy = next->load_last_update_time_copy;
2947 p_last_update_time = prev->avg.last_update_time;
2948 n_last_update_time = next->avg.last_update_time;
2950 } while (p_last_update_time != p_last_update_time_copy ||
2951 n_last_update_time != n_last_update_time_copy);
2953 p_last_update_time = prev->avg.last_update_time;
2954 n_last_update_time = next->avg.last_update_time;
2956 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2957 &se->avg, 0, 0, NULL);
2958 se->avg.last_update_time = n_last_update_time;
2961 #else /* CONFIG_FAIR_GROUP_SCHED */
2962 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2963 #endif /* CONFIG_FAIR_GROUP_SCHED */
2965 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2967 struct rq *rq = rq_of(cfs_rq);
2968 int cpu = cpu_of(rq);
2970 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2971 unsigned long max = rq->cpu_capacity_orig;
2974 * There are a few boundary cases this might miss but it should
2975 * get called often enough that that should (hopefully) not be
2976 * a real problem -- added to that it only calls on the local
2977 * CPU, so if we enqueue remotely we'll miss an update, but
2978 * the next tick/schedule should update.
2980 * It will not get called when we go idle, because the idle
2981 * thread is a different class (!fair), nor will the utilization
2982 * number include things like RT tasks.
2984 * As is, the util number is not freq-invariant (we'd have to
2985 * implement arch_scale_freq_capacity() for that).
2989 cpufreq_update_util(rq_clock(rq),
2990 min(cfs_rq->avg.util_avg, max), max);
2995 * Unsigned subtract and clamp on underflow.
2997 * Explicitly do a load-store to ensure the intermediate value never hits
2998 * memory. This allows lockless observations without ever seeing the negative
3001 #define sub_positive(_ptr, _val) do { \
3002 typeof(_ptr) ptr = (_ptr); \
3003 typeof(*ptr) val = (_val); \
3004 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3008 WRITE_ONCE(*ptr, res); \
3012 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3013 * @now: current time, as per cfs_rq_clock_task()
3014 * @cfs_rq: cfs_rq to update
3015 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3017 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3018 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3019 * post_init_entity_util_avg().
3021 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3023 * Returns true if the load decayed or we removed load.
3025 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3026 * call update_tg_load_avg() when this function returns true.
3029 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3031 struct sched_avg *sa = &cfs_rq->avg;
3032 int decayed, removed_load = 0, removed_util = 0;
3034 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3035 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3036 sub_positive(&sa->load_avg, r);
3037 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3041 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3042 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3043 sub_positive(&sa->util_avg, r);
3044 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3048 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3049 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3051 #ifndef CONFIG_64BIT
3053 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3056 if (update_freq && (decayed || removed_util))
3057 cfs_rq_util_change(cfs_rq);
3059 return decayed || removed_load;
3062 /* Update task and its cfs_rq load average */
3063 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3065 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3066 u64 now = cfs_rq_clock_task(cfs_rq);
3067 struct rq *rq = rq_of(cfs_rq);
3068 int cpu = cpu_of(rq);
3071 * Track task load average for carrying it to new CPU after migrated, and
3072 * track group sched_entity load average for task_h_load calc in migration
3074 __update_load_avg(now, cpu, &se->avg,
3075 se->on_rq * scale_load_down(se->load.weight),
3076 cfs_rq->curr == se, NULL);
3078 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
3079 update_tg_load_avg(cfs_rq, 0);
3083 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3084 * @cfs_rq: cfs_rq to attach to
3085 * @se: sched_entity to attach
3087 * Must call update_cfs_rq_load_avg() before this, since we rely on
3088 * cfs_rq->avg.last_update_time being current.
3090 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3092 if (!sched_feat(ATTACH_AGE_LOAD))
3096 * If we got migrated (either between CPUs or between cgroups) we'll
3097 * have aged the average right before clearing @last_update_time.
3099 * Or we're fresh through post_init_entity_util_avg().
3101 if (se->avg.last_update_time) {
3102 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3103 &se->avg, 0, 0, NULL);
3106 * XXX: we could have just aged the entire load away if we've been
3107 * absent from the fair class for too long.
3112 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3113 cfs_rq->avg.load_avg += se->avg.load_avg;
3114 cfs_rq->avg.load_sum += se->avg.load_sum;
3115 cfs_rq->avg.util_avg += se->avg.util_avg;
3116 cfs_rq->avg.util_sum += se->avg.util_sum;
3118 cfs_rq_util_change(cfs_rq);
3122 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3123 * @cfs_rq: cfs_rq to detach from
3124 * @se: sched_entity to detach
3126 * Must call update_cfs_rq_load_avg() before this, since we rely on
3127 * cfs_rq->avg.last_update_time being current.
3129 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3131 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3132 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3133 cfs_rq->curr == se, NULL);
3135 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3136 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3137 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3138 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3140 cfs_rq_util_change(cfs_rq);
3143 /* Add the load generated by se into cfs_rq's load average */
3145 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3147 struct sched_avg *sa = &se->avg;
3148 u64 now = cfs_rq_clock_task(cfs_rq);
3149 int migrated, decayed;
3151 migrated = !sa->last_update_time;
3153 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3154 se->on_rq * scale_load_down(se->load.weight),
3155 cfs_rq->curr == se, NULL);
3158 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3160 cfs_rq->runnable_load_avg += sa->load_avg;
3161 cfs_rq->runnable_load_sum += sa->load_sum;
3164 attach_entity_load_avg(cfs_rq, se);
3166 if (decayed || migrated)
3167 update_tg_load_avg(cfs_rq, 0);
3170 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3172 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3174 update_load_avg(se, 1);
3176 cfs_rq->runnable_load_avg =
3177 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3178 cfs_rq->runnable_load_sum =
3179 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3182 #ifndef CONFIG_64BIT
3183 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3185 u64 last_update_time_copy;
3186 u64 last_update_time;
3189 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3191 last_update_time = cfs_rq->avg.last_update_time;
3192 } while (last_update_time != last_update_time_copy);
3194 return last_update_time;
3197 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3199 return cfs_rq->avg.last_update_time;
3204 * Task first catches up with cfs_rq, and then subtract
3205 * itself from the cfs_rq (task must be off the queue now).
3207 void remove_entity_load_avg(struct sched_entity *se)
3209 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3210 u64 last_update_time;
3213 * tasks cannot exit without having gone through wake_up_new_task() ->
3214 * post_init_entity_util_avg() which will have added things to the
3215 * cfs_rq, so we can remove unconditionally.
3217 * Similarly for groups, they will have passed through
3218 * post_init_entity_util_avg() before unregister_sched_fair_group()
3222 last_update_time = cfs_rq_last_update_time(cfs_rq);
3224 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3225 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3226 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3229 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3231 return cfs_rq->runnable_load_avg;
3234 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3236 return cfs_rq->avg.load_avg;
3239 static int idle_balance(struct rq *this_rq);
3241 #else /* CONFIG_SMP */
3244 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3249 static inline void update_load_avg(struct sched_entity *se, int not_used)
3251 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3252 struct rq *rq = rq_of(cfs_rq);
3254 cpufreq_trigger_update(rq_clock(rq));
3258 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3260 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3261 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3264 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3266 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3268 static inline int idle_balance(struct rq *rq)
3273 #endif /* CONFIG_SMP */
3275 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3277 #ifdef CONFIG_SCHED_DEBUG
3278 s64 d = se->vruntime - cfs_rq->min_vruntime;
3283 if (d > 3*sysctl_sched_latency)
3284 schedstat_inc(cfs_rq->nr_spread_over);
3289 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3291 u64 vruntime = cfs_rq->min_vruntime;
3294 * The 'current' period is already promised to the current tasks,
3295 * however the extra weight of the new task will slow them down a
3296 * little, place the new task so that it fits in the slot that
3297 * stays open at the end.
3299 if (initial && sched_feat(START_DEBIT))
3300 vruntime += sched_vslice(cfs_rq, se);
3302 /* sleeps up to a single latency don't count. */
3304 unsigned long thresh = sysctl_sched_latency;
3307 * Halve their sleep time's effect, to allow
3308 * for a gentler effect of sleepers:
3310 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3316 /* ensure we never gain time by being placed backwards. */
3317 se->vruntime = max_vruntime(se->vruntime, vruntime);
3320 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3322 static inline void check_schedstat_required(void)
3324 #ifdef CONFIG_SCHEDSTATS
3325 if (schedstat_enabled())
3328 /* Force schedstat enabled if a dependent tracepoint is active */
3329 if (trace_sched_stat_wait_enabled() ||
3330 trace_sched_stat_sleep_enabled() ||
3331 trace_sched_stat_iowait_enabled() ||
3332 trace_sched_stat_blocked_enabled() ||
3333 trace_sched_stat_runtime_enabled()) {
3334 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3335 "stat_blocked and stat_runtime require the "
3336 "kernel parameter schedstats=enabled or "
3337 "kernel.sched_schedstats=1\n");
3348 * update_min_vruntime()
3349 * vruntime -= min_vruntime
3353 * update_min_vruntime()
3354 * vruntime += min_vruntime
3356 * this way the vruntime transition between RQs is done when both
3357 * min_vruntime are up-to-date.
3361 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3362 * vruntime -= min_vruntime
3366 * update_min_vruntime()
3367 * vruntime += min_vruntime
3369 * this way we don't have the most up-to-date min_vruntime on the originating
3370 * CPU and an up-to-date min_vruntime on the destination CPU.
3374 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3376 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3377 bool curr = cfs_rq->curr == se;
3380 * If we're the current task, we must renormalise before calling
3384 se->vruntime += cfs_rq->min_vruntime;
3386 update_curr(cfs_rq);
3389 * Otherwise, renormalise after, such that we're placed at the current
3390 * moment in time, instead of some random moment in the past. Being
3391 * placed in the past could significantly boost this task to the
3392 * fairness detriment of existing tasks.
3394 if (renorm && !curr)
3395 se->vruntime += cfs_rq->min_vruntime;
3397 enqueue_entity_load_avg(cfs_rq, se);
3398 account_entity_enqueue(cfs_rq, se);
3399 update_cfs_shares(cfs_rq);
3401 if (flags & ENQUEUE_WAKEUP)
3402 place_entity(cfs_rq, se, 0);
3404 check_schedstat_required();
3405 update_stats_enqueue(cfs_rq, se, flags);
3406 check_spread(cfs_rq, se);
3408 __enqueue_entity(cfs_rq, se);
3411 if (cfs_rq->nr_running == 1) {
3412 list_add_leaf_cfs_rq(cfs_rq);
3413 check_enqueue_throttle(cfs_rq);
3417 static void __clear_buddies_last(struct sched_entity *se)
3419 for_each_sched_entity(se) {
3420 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3421 if (cfs_rq->last != se)
3424 cfs_rq->last = NULL;
3428 static void __clear_buddies_next(struct sched_entity *se)
3430 for_each_sched_entity(se) {
3431 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3432 if (cfs_rq->next != se)
3435 cfs_rq->next = NULL;
3439 static void __clear_buddies_skip(struct sched_entity *se)
3441 for_each_sched_entity(se) {
3442 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3443 if (cfs_rq->skip != se)
3446 cfs_rq->skip = NULL;
3450 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3452 if (cfs_rq->last == se)
3453 __clear_buddies_last(se);
3455 if (cfs_rq->next == se)
3456 __clear_buddies_next(se);
3458 if (cfs_rq->skip == se)
3459 __clear_buddies_skip(se);
3462 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3465 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3468 * Update run-time statistics of the 'current'.
3470 update_curr(cfs_rq);
3471 dequeue_entity_load_avg(cfs_rq, se);
3473 update_stats_dequeue(cfs_rq, se, flags);
3475 clear_buddies(cfs_rq, se);
3477 if (se != cfs_rq->curr)
3478 __dequeue_entity(cfs_rq, se);
3480 account_entity_dequeue(cfs_rq, se);
3483 * Normalize the entity after updating the min_vruntime because the
3484 * update can refer to the ->curr item and we need to reflect this
3485 * movement in our normalized position.
3487 if (!(flags & DEQUEUE_SLEEP))
3488 se->vruntime -= cfs_rq->min_vruntime;
3490 /* return excess runtime on last dequeue */
3491 return_cfs_rq_runtime(cfs_rq);
3493 update_min_vruntime(cfs_rq);
3494 update_cfs_shares(cfs_rq);
3498 * Preempt the current task with a newly woken task if needed:
3501 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3503 unsigned long ideal_runtime, delta_exec;
3504 struct sched_entity *se;
3507 ideal_runtime = sched_slice(cfs_rq, curr);
3508 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3509 if (delta_exec > ideal_runtime) {
3510 resched_curr(rq_of(cfs_rq));
3512 * The current task ran long enough, ensure it doesn't get
3513 * re-elected due to buddy favours.
3515 clear_buddies(cfs_rq, curr);
3520 * Ensure that a task that missed wakeup preemption by a
3521 * narrow margin doesn't have to wait for a full slice.
3522 * This also mitigates buddy induced latencies under load.
3524 if (delta_exec < sysctl_sched_min_granularity)
3527 se = __pick_first_entity(cfs_rq);
3528 delta = curr->vruntime - se->vruntime;
3533 if (delta > ideal_runtime)
3534 resched_curr(rq_of(cfs_rq));
3538 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3540 /* 'current' is not kept within the tree. */
3543 * Any task has to be enqueued before it get to execute on
3544 * a CPU. So account for the time it spent waiting on the
3547 update_stats_wait_end(cfs_rq, se);
3548 __dequeue_entity(cfs_rq, se);
3549 update_load_avg(se, 1);
3552 update_stats_curr_start(cfs_rq, se);
3556 * Track our maximum slice length, if the CPU's load is at
3557 * least twice that of our own weight (i.e. dont track it
3558 * when there are only lesser-weight tasks around):
3560 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3561 schedstat_set(se->statistics.slice_max,
3562 max((u64)schedstat_val(se->statistics.slice_max),
3563 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3566 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3570 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3573 * Pick the next process, keeping these things in mind, in this order:
3574 * 1) keep things fair between processes/task groups
3575 * 2) pick the "next" process, since someone really wants that to run
3576 * 3) pick the "last" process, for cache locality
3577 * 4) do not run the "skip" process, if something else is available
3579 static struct sched_entity *
3580 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3582 struct sched_entity *left = __pick_first_entity(cfs_rq);
3583 struct sched_entity *se;
3586 * If curr is set we have to see if its left of the leftmost entity
3587 * still in the tree, provided there was anything in the tree at all.
3589 if (!left || (curr && entity_before(curr, left)))
3592 se = left; /* ideally we run the leftmost entity */
3595 * Avoid running the skip buddy, if running something else can
3596 * be done without getting too unfair.
3598 if (cfs_rq->skip == se) {
3599 struct sched_entity *second;
3602 second = __pick_first_entity(cfs_rq);
3604 second = __pick_next_entity(se);
3605 if (!second || (curr && entity_before(curr, second)))
3609 if (second && wakeup_preempt_entity(second, left) < 1)
3614 * Prefer last buddy, try to return the CPU to a preempted task.
3616 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3620 * Someone really wants this to run. If it's not unfair, run it.
3622 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3625 clear_buddies(cfs_rq, se);
3630 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3632 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3635 * If still on the runqueue then deactivate_task()
3636 * was not called and update_curr() has to be done:
3639 update_curr(cfs_rq);
3641 /* throttle cfs_rqs exceeding runtime */
3642 check_cfs_rq_runtime(cfs_rq);
3644 check_spread(cfs_rq, prev);
3647 update_stats_wait_start(cfs_rq, prev);
3648 /* Put 'current' back into the tree. */
3649 __enqueue_entity(cfs_rq, prev);
3650 /* in !on_rq case, update occurred at dequeue */
3651 update_load_avg(prev, 0);
3653 cfs_rq->curr = NULL;
3657 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3660 * Update run-time statistics of the 'current'.
3662 update_curr(cfs_rq);
3665 * Ensure that runnable average is periodically updated.
3667 update_load_avg(curr, 1);
3668 update_cfs_shares(cfs_rq);
3670 #ifdef CONFIG_SCHED_HRTICK
3672 * queued ticks are scheduled to match the slice, so don't bother
3673 * validating it and just reschedule.
3676 resched_curr(rq_of(cfs_rq));
3680 * don't let the period tick interfere with the hrtick preemption
3682 if (!sched_feat(DOUBLE_TICK) &&
3683 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3687 if (cfs_rq->nr_running > 1)
3688 check_preempt_tick(cfs_rq, curr);
3692 /**************************************************
3693 * CFS bandwidth control machinery
3696 #ifdef CONFIG_CFS_BANDWIDTH
3698 #ifdef HAVE_JUMP_LABEL
3699 static struct static_key __cfs_bandwidth_used;
3701 static inline bool cfs_bandwidth_used(void)
3703 return static_key_false(&__cfs_bandwidth_used);
3706 void cfs_bandwidth_usage_inc(void)
3708 static_key_slow_inc(&__cfs_bandwidth_used);
3711 void cfs_bandwidth_usage_dec(void)
3713 static_key_slow_dec(&__cfs_bandwidth_used);
3715 #else /* HAVE_JUMP_LABEL */
3716 static bool cfs_bandwidth_used(void)
3721 void cfs_bandwidth_usage_inc(void) {}
3722 void cfs_bandwidth_usage_dec(void) {}
3723 #endif /* HAVE_JUMP_LABEL */
3726 * default period for cfs group bandwidth.
3727 * default: 0.1s, units: nanoseconds
3729 static inline u64 default_cfs_period(void)
3731 return 100000000ULL;
3734 static inline u64 sched_cfs_bandwidth_slice(void)
3736 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3740 * Replenish runtime according to assigned quota and update expiration time.
3741 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3742 * additional synchronization around rq->lock.
3744 * requires cfs_b->lock
3746 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3750 if (cfs_b->quota == RUNTIME_INF)
3753 now = sched_clock_cpu(smp_processor_id());
3754 cfs_b->runtime = cfs_b->quota;
3755 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3758 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3760 return &tg->cfs_bandwidth;
3763 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3764 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3766 if (unlikely(cfs_rq->throttle_count))
3767 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3769 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3772 /* returns 0 on failure to allocate runtime */
3773 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3775 struct task_group *tg = cfs_rq->tg;
3776 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3777 u64 amount = 0, min_amount, expires;
3779 /* note: this is a positive sum as runtime_remaining <= 0 */
3780 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3782 raw_spin_lock(&cfs_b->lock);
3783 if (cfs_b->quota == RUNTIME_INF)
3784 amount = min_amount;
3786 start_cfs_bandwidth(cfs_b);
3788 if (cfs_b->runtime > 0) {
3789 amount = min(cfs_b->runtime, min_amount);
3790 cfs_b->runtime -= amount;
3794 expires = cfs_b->runtime_expires;
3795 raw_spin_unlock(&cfs_b->lock);
3797 cfs_rq->runtime_remaining += amount;
3799 * we may have advanced our local expiration to account for allowed
3800 * spread between our sched_clock and the one on which runtime was
3803 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3804 cfs_rq->runtime_expires = expires;
3806 return cfs_rq->runtime_remaining > 0;
3810 * Note: This depends on the synchronization provided by sched_clock and the
3811 * fact that rq->clock snapshots this value.
3813 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3815 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3817 /* if the deadline is ahead of our clock, nothing to do */
3818 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3821 if (cfs_rq->runtime_remaining < 0)
3825 * If the local deadline has passed we have to consider the
3826 * possibility that our sched_clock is 'fast' and the global deadline
3827 * has not truly expired.
3829 * Fortunately we can check determine whether this the case by checking
3830 * whether the global deadline has advanced. It is valid to compare
3831 * cfs_b->runtime_expires without any locks since we only care about
3832 * exact equality, so a partial write will still work.
3835 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3836 /* extend local deadline, drift is bounded above by 2 ticks */
3837 cfs_rq->runtime_expires += TICK_NSEC;
3839 /* global deadline is ahead, expiration has passed */
3840 cfs_rq->runtime_remaining = 0;
3844 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3846 /* dock delta_exec before expiring quota (as it could span periods) */
3847 cfs_rq->runtime_remaining -= delta_exec;
3848 expire_cfs_rq_runtime(cfs_rq);
3850 if (likely(cfs_rq->runtime_remaining > 0))
3854 * if we're unable to extend our runtime we resched so that the active
3855 * hierarchy can be throttled
3857 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3858 resched_curr(rq_of(cfs_rq));
3861 static __always_inline
3862 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3864 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3867 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3870 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3872 return cfs_bandwidth_used() && cfs_rq->throttled;
3875 /* check whether cfs_rq, or any parent, is throttled */
3876 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3878 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3882 * Ensure that neither of the group entities corresponding to src_cpu or
3883 * dest_cpu are members of a throttled hierarchy when performing group
3884 * load-balance operations.
3886 static inline int throttled_lb_pair(struct task_group *tg,
3887 int src_cpu, int dest_cpu)
3889 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3891 src_cfs_rq = tg->cfs_rq[src_cpu];
3892 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3894 return throttled_hierarchy(src_cfs_rq) ||
3895 throttled_hierarchy(dest_cfs_rq);
3898 /* updated child weight may affect parent so we have to do this bottom up */
3899 static int tg_unthrottle_up(struct task_group *tg, void *data)
3901 struct rq *rq = data;
3902 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3904 cfs_rq->throttle_count--;
3905 if (!cfs_rq->throttle_count) {
3906 /* adjust cfs_rq_clock_task() */
3907 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3908 cfs_rq->throttled_clock_task;
3914 static int tg_throttle_down(struct task_group *tg, void *data)
3916 struct rq *rq = data;
3917 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3919 /* group is entering throttled state, stop time */
3920 if (!cfs_rq->throttle_count)
3921 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3922 cfs_rq->throttle_count++;
3927 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3929 struct rq *rq = rq_of(cfs_rq);
3930 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3931 struct sched_entity *se;
3932 long task_delta, dequeue = 1;
3935 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3937 /* freeze hierarchy runnable averages while throttled */
3939 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3942 task_delta = cfs_rq->h_nr_running;
3943 for_each_sched_entity(se) {
3944 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3945 /* throttled entity or throttle-on-deactivate */
3950 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3951 qcfs_rq->h_nr_running -= task_delta;
3953 if (qcfs_rq->load.weight)
3958 sub_nr_running(rq, task_delta);
3960 cfs_rq->throttled = 1;
3961 cfs_rq->throttled_clock = rq_clock(rq);
3962 raw_spin_lock(&cfs_b->lock);
3963 empty = list_empty(&cfs_b->throttled_cfs_rq);
3966 * Add to the _head_ of the list, so that an already-started
3967 * distribute_cfs_runtime will not see us
3969 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3972 * If we're the first throttled task, make sure the bandwidth
3976 start_cfs_bandwidth(cfs_b);
3978 raw_spin_unlock(&cfs_b->lock);
3981 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3983 struct rq *rq = rq_of(cfs_rq);
3984 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3985 struct sched_entity *se;
3989 se = cfs_rq->tg->se[cpu_of(rq)];
3991 cfs_rq->throttled = 0;
3993 update_rq_clock(rq);
3995 raw_spin_lock(&cfs_b->lock);
3996 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3997 list_del_rcu(&cfs_rq->throttled_list);
3998 raw_spin_unlock(&cfs_b->lock);
4000 /* update hierarchical throttle state */
4001 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4003 if (!cfs_rq->load.weight)
4006 task_delta = cfs_rq->h_nr_running;
4007 for_each_sched_entity(se) {
4011 cfs_rq = cfs_rq_of(se);
4013 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4014 cfs_rq->h_nr_running += task_delta;
4016 if (cfs_rq_throttled(cfs_rq))
4021 add_nr_running(rq, task_delta);
4023 /* determine whether we need to wake up potentially idle cpu */
4024 if (rq->curr == rq->idle && rq->cfs.nr_running)
4028 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4029 u64 remaining, u64 expires)
4031 struct cfs_rq *cfs_rq;
4033 u64 starting_runtime = remaining;
4036 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4038 struct rq *rq = rq_of(cfs_rq);
4040 raw_spin_lock(&rq->lock);
4041 if (!cfs_rq_throttled(cfs_rq))
4044 runtime = -cfs_rq->runtime_remaining + 1;
4045 if (runtime > remaining)
4046 runtime = remaining;
4047 remaining -= runtime;
4049 cfs_rq->runtime_remaining += runtime;
4050 cfs_rq->runtime_expires = expires;
4052 /* we check whether we're throttled above */
4053 if (cfs_rq->runtime_remaining > 0)
4054 unthrottle_cfs_rq(cfs_rq);
4057 raw_spin_unlock(&rq->lock);
4064 return starting_runtime - remaining;
4068 * Responsible for refilling a task_group's bandwidth and unthrottling its
4069 * cfs_rqs as appropriate. If there has been no activity within the last
4070 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4071 * used to track this state.
4073 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4075 u64 runtime, runtime_expires;
4078 /* no need to continue the timer with no bandwidth constraint */
4079 if (cfs_b->quota == RUNTIME_INF)
4080 goto out_deactivate;
4082 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4083 cfs_b->nr_periods += overrun;
4086 * idle depends on !throttled (for the case of a large deficit), and if
4087 * we're going inactive then everything else can be deferred
4089 if (cfs_b->idle && !throttled)
4090 goto out_deactivate;
4092 __refill_cfs_bandwidth_runtime(cfs_b);
4095 /* mark as potentially idle for the upcoming period */
4100 /* account preceding periods in which throttling occurred */
4101 cfs_b->nr_throttled += overrun;
4103 runtime_expires = cfs_b->runtime_expires;
4106 * This check is repeated as we are holding onto the new bandwidth while
4107 * we unthrottle. This can potentially race with an unthrottled group
4108 * trying to acquire new bandwidth from the global pool. This can result
4109 * in us over-using our runtime if it is all used during this loop, but
4110 * only by limited amounts in that extreme case.
4112 while (throttled && cfs_b->runtime > 0) {
4113 runtime = cfs_b->runtime;
4114 raw_spin_unlock(&cfs_b->lock);
4115 /* we can't nest cfs_b->lock while distributing bandwidth */
4116 runtime = distribute_cfs_runtime(cfs_b, runtime,
4118 raw_spin_lock(&cfs_b->lock);
4120 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4122 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4126 * While we are ensured activity in the period following an
4127 * unthrottle, this also covers the case in which the new bandwidth is
4128 * insufficient to cover the existing bandwidth deficit. (Forcing the
4129 * timer to remain active while there are any throttled entities.)
4139 /* a cfs_rq won't donate quota below this amount */
4140 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4141 /* minimum remaining period time to redistribute slack quota */
4142 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4143 /* how long we wait to gather additional slack before distributing */
4144 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4147 * Are we near the end of the current quota period?
4149 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4150 * hrtimer base being cleared by hrtimer_start. In the case of
4151 * migrate_hrtimers, base is never cleared, so we are fine.
4153 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4155 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4158 /* if the call-back is running a quota refresh is already occurring */
4159 if (hrtimer_callback_running(refresh_timer))
4162 /* is a quota refresh about to occur? */
4163 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4164 if (remaining < min_expire)
4170 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4172 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4174 /* if there's a quota refresh soon don't bother with slack */
4175 if (runtime_refresh_within(cfs_b, min_left))
4178 hrtimer_start(&cfs_b->slack_timer,
4179 ns_to_ktime(cfs_bandwidth_slack_period),
4183 /* we know any runtime found here is valid as update_curr() precedes return */
4184 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4186 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4187 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4189 if (slack_runtime <= 0)
4192 raw_spin_lock(&cfs_b->lock);
4193 if (cfs_b->quota != RUNTIME_INF &&
4194 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4195 cfs_b->runtime += slack_runtime;
4197 /* we are under rq->lock, defer unthrottling using a timer */
4198 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4199 !list_empty(&cfs_b->throttled_cfs_rq))
4200 start_cfs_slack_bandwidth(cfs_b);
4202 raw_spin_unlock(&cfs_b->lock);
4204 /* even if it's not valid for return we don't want to try again */
4205 cfs_rq->runtime_remaining -= slack_runtime;
4208 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4210 if (!cfs_bandwidth_used())
4213 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4216 __return_cfs_rq_runtime(cfs_rq);
4220 * This is done with a timer (instead of inline with bandwidth return) since
4221 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4223 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4225 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4228 /* confirm we're still not at a refresh boundary */
4229 raw_spin_lock(&cfs_b->lock);
4230 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4231 raw_spin_unlock(&cfs_b->lock);
4235 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4236 runtime = cfs_b->runtime;
4238 expires = cfs_b->runtime_expires;
4239 raw_spin_unlock(&cfs_b->lock);
4244 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4246 raw_spin_lock(&cfs_b->lock);
4247 if (expires == cfs_b->runtime_expires)
4248 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4249 raw_spin_unlock(&cfs_b->lock);
4253 * When a group wakes up we want to make sure that its quota is not already
4254 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4255 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4257 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4259 if (!cfs_bandwidth_used())
4262 /* an active group must be handled by the update_curr()->put() path */
4263 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4266 /* ensure the group is not already throttled */
4267 if (cfs_rq_throttled(cfs_rq))
4270 /* update runtime allocation */
4271 account_cfs_rq_runtime(cfs_rq, 0);
4272 if (cfs_rq->runtime_remaining <= 0)
4273 throttle_cfs_rq(cfs_rq);
4276 static void sync_throttle(struct task_group *tg, int cpu)
4278 struct cfs_rq *pcfs_rq, *cfs_rq;
4280 if (!cfs_bandwidth_used())
4286 cfs_rq = tg->cfs_rq[cpu];
4287 pcfs_rq = tg->parent->cfs_rq[cpu];
4289 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4290 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4293 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4294 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4296 if (!cfs_bandwidth_used())
4299 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4303 * it's possible for a throttled entity to be forced into a running
4304 * state (e.g. set_curr_task), in this case we're finished.
4306 if (cfs_rq_throttled(cfs_rq))
4309 throttle_cfs_rq(cfs_rq);
4313 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4315 struct cfs_bandwidth *cfs_b =
4316 container_of(timer, struct cfs_bandwidth, slack_timer);
4318 do_sched_cfs_slack_timer(cfs_b);
4320 return HRTIMER_NORESTART;
4323 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4325 struct cfs_bandwidth *cfs_b =
4326 container_of(timer, struct cfs_bandwidth, period_timer);
4330 raw_spin_lock(&cfs_b->lock);
4332 overrun = hrtimer_forward_now(timer, cfs_b->period);
4336 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4339 cfs_b->period_active = 0;
4340 raw_spin_unlock(&cfs_b->lock);
4342 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4345 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4347 raw_spin_lock_init(&cfs_b->lock);
4349 cfs_b->quota = RUNTIME_INF;
4350 cfs_b->period = ns_to_ktime(default_cfs_period());
4352 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4353 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4354 cfs_b->period_timer.function = sched_cfs_period_timer;
4355 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4356 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4359 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4361 cfs_rq->runtime_enabled = 0;
4362 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4365 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4367 lockdep_assert_held(&cfs_b->lock);
4369 if (!cfs_b->period_active) {
4370 cfs_b->period_active = 1;
4371 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4372 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4376 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4378 /* init_cfs_bandwidth() was not called */
4379 if (!cfs_b->throttled_cfs_rq.next)
4382 hrtimer_cancel(&cfs_b->period_timer);
4383 hrtimer_cancel(&cfs_b->slack_timer);
4386 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4388 struct cfs_rq *cfs_rq;
4390 for_each_leaf_cfs_rq(rq, cfs_rq) {
4391 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4393 raw_spin_lock(&cfs_b->lock);
4394 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4395 raw_spin_unlock(&cfs_b->lock);
4399 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4401 struct cfs_rq *cfs_rq;
4403 for_each_leaf_cfs_rq(rq, cfs_rq) {
4404 if (!cfs_rq->runtime_enabled)
4408 * clock_task is not advancing so we just need to make sure
4409 * there's some valid quota amount
4411 cfs_rq->runtime_remaining = 1;
4413 * Offline rq is schedulable till cpu is completely disabled
4414 * in take_cpu_down(), so we prevent new cfs throttling here.
4416 cfs_rq->runtime_enabled = 0;
4418 if (cfs_rq_throttled(cfs_rq))
4419 unthrottle_cfs_rq(cfs_rq);
4423 #else /* CONFIG_CFS_BANDWIDTH */
4424 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4426 return rq_clock_task(rq_of(cfs_rq));
4429 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4430 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4431 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4432 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4433 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4435 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4440 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4445 static inline int throttled_lb_pair(struct task_group *tg,
4446 int src_cpu, int dest_cpu)
4451 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4453 #ifdef CONFIG_FAIR_GROUP_SCHED
4454 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4457 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4461 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4462 static inline void update_runtime_enabled(struct rq *rq) {}
4463 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4465 #endif /* CONFIG_CFS_BANDWIDTH */
4467 /**************************************************
4468 * CFS operations on tasks:
4471 #ifdef CONFIG_SCHED_HRTICK
4472 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4474 struct sched_entity *se = &p->se;
4475 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4477 WARN_ON(task_rq(p) != rq);
4479 if (rq->cfs.h_nr_running > 1) {
4480 u64 slice = sched_slice(cfs_rq, se);
4481 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4482 s64 delta = slice - ran;
4489 hrtick_start(rq, delta);
4494 * called from enqueue/dequeue and updates the hrtick when the
4495 * current task is from our class and nr_running is low enough
4498 static void hrtick_update(struct rq *rq)
4500 struct task_struct *curr = rq->curr;
4502 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4505 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4506 hrtick_start_fair(rq, curr);
4508 #else /* !CONFIG_SCHED_HRTICK */
4510 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4514 static inline void hrtick_update(struct rq *rq)
4520 * The enqueue_task method is called before nr_running is
4521 * increased. Here we update the fair scheduling stats and
4522 * then put the task into the rbtree:
4525 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4527 struct cfs_rq *cfs_rq;
4528 struct sched_entity *se = &p->se;
4530 for_each_sched_entity(se) {
4533 cfs_rq = cfs_rq_of(se);
4534 enqueue_entity(cfs_rq, se, flags);
4537 * end evaluation on encountering a throttled cfs_rq
4539 * note: in the case of encountering a throttled cfs_rq we will
4540 * post the final h_nr_running increment below.
4542 if (cfs_rq_throttled(cfs_rq))
4544 cfs_rq->h_nr_running++;
4546 flags = ENQUEUE_WAKEUP;
4549 for_each_sched_entity(se) {
4550 cfs_rq = cfs_rq_of(se);
4551 cfs_rq->h_nr_running++;
4553 if (cfs_rq_throttled(cfs_rq))
4556 update_load_avg(se, 1);
4557 update_cfs_shares(cfs_rq);
4561 add_nr_running(rq, 1);
4566 static void set_next_buddy(struct sched_entity *se);
4569 * The dequeue_task method is called before nr_running is
4570 * decreased. We remove the task from the rbtree and
4571 * update the fair scheduling stats:
4573 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4575 struct cfs_rq *cfs_rq;
4576 struct sched_entity *se = &p->se;
4577 int task_sleep = flags & DEQUEUE_SLEEP;
4579 for_each_sched_entity(se) {
4580 cfs_rq = cfs_rq_of(se);
4581 dequeue_entity(cfs_rq, se, flags);
4584 * end evaluation on encountering a throttled cfs_rq
4586 * note: in the case of encountering a throttled cfs_rq we will
4587 * post the final h_nr_running decrement below.
4589 if (cfs_rq_throttled(cfs_rq))
4591 cfs_rq->h_nr_running--;
4593 /* Don't dequeue parent if it has other entities besides us */
4594 if (cfs_rq->load.weight) {
4595 /* Avoid re-evaluating load for this entity: */
4596 se = parent_entity(se);
4598 * Bias pick_next to pick a task from this cfs_rq, as
4599 * p is sleeping when it is within its sched_slice.
4601 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4605 flags |= DEQUEUE_SLEEP;
4608 for_each_sched_entity(se) {
4609 cfs_rq = cfs_rq_of(se);
4610 cfs_rq->h_nr_running--;
4612 if (cfs_rq_throttled(cfs_rq))
4615 update_load_avg(se, 1);
4616 update_cfs_shares(cfs_rq);
4620 sub_nr_running(rq, 1);
4627 /* Working cpumask for: load_balance, load_balance_newidle. */
4628 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4629 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4631 #ifdef CONFIG_NO_HZ_COMMON
4633 * per rq 'load' arrray crap; XXX kill this.
4637 * The exact cpuload calculated at every tick would be:
4639 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4641 * If a cpu misses updates for n ticks (as it was idle) and update gets
4642 * called on the n+1-th tick when cpu may be busy, then we have:
4644 * load_n = (1 - 1/2^i)^n * load_0
4645 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4647 * decay_load_missed() below does efficient calculation of
4649 * load' = (1 - 1/2^i)^n * load
4651 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4652 * This allows us to precompute the above in said factors, thereby allowing the
4653 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4654 * fixed_power_int())
4656 * The calculation is approximated on a 128 point scale.
4658 #define DEGRADE_SHIFT 7
4660 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4661 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4662 { 0, 0, 0, 0, 0, 0, 0, 0 },
4663 { 64, 32, 8, 0, 0, 0, 0, 0 },
4664 { 96, 72, 40, 12, 1, 0, 0, 0 },
4665 { 112, 98, 75, 43, 15, 1, 0, 0 },
4666 { 120, 112, 98, 76, 45, 16, 2, 0 }
4670 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4671 * would be when CPU is idle and so we just decay the old load without
4672 * adding any new load.
4674 static unsigned long
4675 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4679 if (!missed_updates)
4682 if (missed_updates >= degrade_zero_ticks[idx])
4686 return load >> missed_updates;
4688 while (missed_updates) {
4689 if (missed_updates % 2)
4690 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4692 missed_updates >>= 1;
4697 #endif /* CONFIG_NO_HZ_COMMON */
4700 * __cpu_load_update - update the rq->cpu_load[] statistics
4701 * @this_rq: The rq to update statistics for
4702 * @this_load: The current load
4703 * @pending_updates: The number of missed updates
4705 * Update rq->cpu_load[] statistics. This function is usually called every
4706 * scheduler tick (TICK_NSEC).
4708 * This function computes a decaying average:
4710 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4712 * Because of NOHZ it might not get called on every tick which gives need for
4713 * the @pending_updates argument.
4715 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4716 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4717 * = A * (A * load[i]_n-2 + B) + B
4718 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4719 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4720 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4721 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4722 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4724 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4725 * any change in load would have resulted in the tick being turned back on.
4727 * For regular NOHZ, this reduces to:
4729 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4731 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4734 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4735 unsigned long pending_updates)
4737 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4740 this_rq->nr_load_updates++;
4742 /* Update our load: */
4743 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4744 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4745 unsigned long old_load, new_load;
4747 /* scale is effectively 1 << i now, and >> i divides by scale */
4749 old_load = this_rq->cpu_load[i];
4750 #ifdef CONFIG_NO_HZ_COMMON
4751 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4752 if (tickless_load) {
4753 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4755 * old_load can never be a negative value because a
4756 * decayed tickless_load cannot be greater than the
4757 * original tickless_load.
4759 old_load += tickless_load;
4762 new_load = this_load;
4764 * Round up the averaging division if load is increasing. This
4765 * prevents us from getting stuck on 9 if the load is 10, for
4768 if (new_load > old_load)
4769 new_load += scale - 1;
4771 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4774 sched_avg_update(this_rq);
4777 /* Used instead of source_load when we know the type == 0 */
4778 static unsigned long weighted_cpuload(const int cpu)
4780 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4783 #ifdef CONFIG_NO_HZ_COMMON
4785 * There is no sane way to deal with nohz on smp when using jiffies because the
4786 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4787 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4789 * Therefore we need to avoid the delta approach from the regular tick when
4790 * possible since that would seriously skew the load calculation. This is why we
4791 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4792 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4793 * loop exit, nohz_idle_balance, nohz full exit...)
4795 * This means we might still be one tick off for nohz periods.
4798 static void cpu_load_update_nohz(struct rq *this_rq,
4799 unsigned long curr_jiffies,
4802 unsigned long pending_updates;
4804 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4805 if (pending_updates) {
4806 this_rq->last_load_update_tick = curr_jiffies;
4808 * In the regular NOHZ case, we were idle, this means load 0.
4809 * In the NOHZ_FULL case, we were non-idle, we should consider
4810 * its weighted load.
4812 cpu_load_update(this_rq, load, pending_updates);
4817 * Called from nohz_idle_balance() to update the load ratings before doing the
4820 static void cpu_load_update_idle(struct rq *this_rq)
4823 * bail if there's load or we're actually up-to-date.
4825 if (weighted_cpuload(cpu_of(this_rq)))
4828 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4832 * Record CPU load on nohz entry so we know the tickless load to account
4833 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4834 * than other cpu_load[idx] but it should be fine as cpu_load readers
4835 * shouldn't rely into synchronized cpu_load[*] updates.
4837 void cpu_load_update_nohz_start(void)
4839 struct rq *this_rq = this_rq();
4842 * This is all lockless but should be fine. If weighted_cpuload changes
4843 * concurrently we'll exit nohz. And cpu_load write can race with
4844 * cpu_load_update_idle() but both updater would be writing the same.
4846 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4850 * Account the tickless load in the end of a nohz frame.
4852 void cpu_load_update_nohz_stop(void)
4854 unsigned long curr_jiffies = READ_ONCE(jiffies);
4855 struct rq *this_rq = this_rq();
4858 if (curr_jiffies == this_rq->last_load_update_tick)
4861 load = weighted_cpuload(cpu_of(this_rq));
4862 raw_spin_lock(&this_rq->lock);
4863 update_rq_clock(this_rq);
4864 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4865 raw_spin_unlock(&this_rq->lock);
4867 #else /* !CONFIG_NO_HZ_COMMON */
4868 static inline void cpu_load_update_nohz(struct rq *this_rq,
4869 unsigned long curr_jiffies,
4870 unsigned long load) { }
4871 #endif /* CONFIG_NO_HZ_COMMON */
4873 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4875 #ifdef CONFIG_NO_HZ_COMMON
4876 /* See the mess around cpu_load_update_nohz(). */
4877 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4879 cpu_load_update(this_rq, load, 1);
4883 * Called from scheduler_tick()
4885 void cpu_load_update_active(struct rq *this_rq)
4887 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4889 if (tick_nohz_tick_stopped())
4890 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4892 cpu_load_update_periodic(this_rq, load);
4896 * Return a low guess at the load of a migration-source cpu weighted
4897 * according to the scheduling class and "nice" value.
4899 * We want to under-estimate the load of migration sources, to
4900 * balance conservatively.
4902 static unsigned long source_load(int cpu, int type)
4904 struct rq *rq = cpu_rq(cpu);
4905 unsigned long total = weighted_cpuload(cpu);
4907 if (type == 0 || !sched_feat(LB_BIAS))
4910 return min(rq->cpu_load[type-1], total);
4914 * Return a high guess at the load of a migration-target cpu weighted
4915 * according to the scheduling class and "nice" value.
4917 static unsigned long target_load(int cpu, int type)
4919 struct rq *rq = cpu_rq(cpu);
4920 unsigned long total = weighted_cpuload(cpu);
4922 if (type == 0 || !sched_feat(LB_BIAS))
4925 return max(rq->cpu_load[type-1], total);
4928 static unsigned long capacity_of(int cpu)
4930 return cpu_rq(cpu)->cpu_capacity;
4933 static unsigned long capacity_orig_of(int cpu)
4935 return cpu_rq(cpu)->cpu_capacity_orig;
4938 static unsigned long cpu_avg_load_per_task(int cpu)
4940 struct rq *rq = cpu_rq(cpu);
4941 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4942 unsigned long load_avg = weighted_cpuload(cpu);
4945 return load_avg / nr_running;
4950 #ifdef CONFIG_FAIR_GROUP_SCHED
4952 * effective_load() calculates the load change as seen from the root_task_group
4954 * Adding load to a group doesn't make a group heavier, but can cause movement
4955 * of group shares between cpus. Assuming the shares were perfectly aligned one
4956 * can calculate the shift in shares.
4958 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4959 * on this @cpu and results in a total addition (subtraction) of @wg to the
4960 * total group weight.
4962 * Given a runqueue weight distribution (rw_i) we can compute a shares
4963 * distribution (s_i) using:
4965 * s_i = rw_i / \Sum rw_j (1)
4967 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4968 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4969 * shares distribution (s_i):
4971 * rw_i = { 2, 4, 1, 0 }
4972 * s_i = { 2/7, 4/7, 1/7, 0 }
4974 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4975 * task used to run on and the CPU the waker is running on), we need to
4976 * compute the effect of waking a task on either CPU and, in case of a sync
4977 * wakeup, compute the effect of the current task going to sleep.
4979 * So for a change of @wl to the local @cpu with an overall group weight change
4980 * of @wl we can compute the new shares distribution (s'_i) using:
4982 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4984 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4985 * differences in waking a task to CPU 0. The additional task changes the
4986 * weight and shares distributions like:
4988 * rw'_i = { 3, 4, 1, 0 }
4989 * s'_i = { 3/8, 4/8, 1/8, 0 }
4991 * We can then compute the difference in effective weight by using:
4993 * dw_i = S * (s'_i - s_i) (3)
4995 * Where 'S' is the group weight as seen by its parent.
4997 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4998 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4999 * 4/7) times the weight of the group.
5001 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5003 struct sched_entity *se = tg->se[cpu];
5005 if (!tg->parent) /* the trivial, non-cgroup case */
5008 for_each_sched_entity(se) {
5009 struct cfs_rq *cfs_rq = se->my_q;
5010 long W, w = cfs_rq_load_avg(cfs_rq);
5015 * W = @wg + \Sum rw_j
5017 W = wg + atomic_long_read(&tg->load_avg);
5019 /* Ensure \Sum rw_j >= rw_i */
5020 W -= cfs_rq->tg_load_avg_contrib;
5029 * wl = S * s'_i; see (2)
5032 wl = (w * (long)scale_load_down(tg->shares)) / W;
5034 wl = scale_load_down(tg->shares);
5037 * Per the above, wl is the new se->load.weight value; since
5038 * those are clipped to [MIN_SHARES, ...) do so now. See
5039 * calc_cfs_shares().
5041 if (wl < MIN_SHARES)
5045 * wl = dw_i = S * (s'_i - s_i); see (3)
5047 wl -= se->avg.load_avg;
5050 * Recursively apply this logic to all parent groups to compute
5051 * the final effective load change on the root group. Since
5052 * only the @tg group gets extra weight, all parent groups can
5053 * only redistribute existing shares. @wl is the shift in shares
5054 * resulting from this level per the above.
5063 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5070 static void record_wakee(struct task_struct *p)
5073 * Only decay a single time; tasks that have less then 1 wakeup per
5074 * jiffy will not have built up many flips.
5076 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5077 current->wakee_flips >>= 1;
5078 current->wakee_flip_decay_ts = jiffies;
5081 if (current->last_wakee != p) {
5082 current->last_wakee = p;
5083 current->wakee_flips++;
5088 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5090 * A waker of many should wake a different task than the one last awakened
5091 * at a frequency roughly N times higher than one of its wakees.
5093 * In order to determine whether we should let the load spread vs consolidating
5094 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5095 * partner, and a factor of lls_size higher frequency in the other.
5097 * With both conditions met, we can be relatively sure that the relationship is
5098 * non-monogamous, with partner count exceeding socket size.
5100 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5101 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5104 static int wake_wide(struct task_struct *p)
5106 unsigned int master = current->wakee_flips;
5107 unsigned int slave = p->wakee_flips;
5108 int factor = this_cpu_read(sd_llc_size);
5111 swap(master, slave);
5112 if (slave < factor || master < slave * factor)
5117 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5118 int prev_cpu, int sync)
5120 s64 this_load, load;
5121 s64 this_eff_load, prev_eff_load;
5123 struct task_group *tg;
5124 unsigned long weight;
5128 this_cpu = smp_processor_id();
5129 load = source_load(prev_cpu, idx);
5130 this_load = target_load(this_cpu, idx);
5133 * If sync wakeup then subtract the (maximum possible)
5134 * effect of the currently running task from the load
5135 * of the current CPU:
5138 tg = task_group(current);
5139 weight = current->se.avg.load_avg;
5141 this_load += effective_load(tg, this_cpu, -weight, -weight);
5142 load += effective_load(tg, prev_cpu, 0, -weight);
5146 weight = p->se.avg.load_avg;
5149 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5150 * due to the sync cause above having dropped this_load to 0, we'll
5151 * always have an imbalance, but there's really nothing you can do
5152 * about that, so that's good too.
5154 * Otherwise check if either cpus are near enough in load to allow this
5155 * task to be woken on this_cpu.
5157 this_eff_load = 100;
5158 this_eff_load *= capacity_of(prev_cpu);
5160 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5161 prev_eff_load *= capacity_of(this_cpu);
5163 if (this_load > 0) {
5164 this_eff_load *= this_load +
5165 effective_load(tg, this_cpu, weight, weight);
5167 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5170 balanced = this_eff_load <= prev_eff_load;
5172 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5177 schedstat_inc(sd->ttwu_move_affine);
5178 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5184 * find_idlest_group finds and returns the least busy CPU group within the
5187 static struct sched_group *
5188 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5189 int this_cpu, int sd_flag)
5191 struct sched_group *idlest = NULL, *group = sd->groups;
5192 unsigned long min_load = ULONG_MAX, this_load = 0;
5193 int load_idx = sd->forkexec_idx;
5194 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5196 if (sd_flag & SD_BALANCE_WAKE)
5197 load_idx = sd->wake_idx;
5200 unsigned long load, avg_load;
5204 /* Skip over this group if it has no CPUs allowed */
5205 if (!cpumask_intersects(sched_group_cpus(group),
5206 tsk_cpus_allowed(p)))
5209 local_group = cpumask_test_cpu(this_cpu,
5210 sched_group_cpus(group));
5212 /* Tally up the load of all CPUs in the group */
5215 for_each_cpu(i, sched_group_cpus(group)) {
5216 /* Bias balancing toward cpus of our domain */
5218 load = source_load(i, load_idx);
5220 load = target_load(i, load_idx);
5225 /* Adjust by relative CPU capacity of the group */
5226 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5229 this_load = avg_load;
5230 } else if (avg_load < min_load) {
5231 min_load = avg_load;
5234 } while (group = group->next, group != sd->groups);
5236 if (!idlest || 100*this_load < imbalance*min_load)
5242 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5245 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5247 unsigned long load, min_load = ULONG_MAX;
5248 unsigned int min_exit_latency = UINT_MAX;
5249 u64 latest_idle_timestamp = 0;
5250 int least_loaded_cpu = this_cpu;
5251 int shallowest_idle_cpu = -1;
5254 /* Check if we have any choice: */
5255 if (group->group_weight == 1)
5256 return cpumask_first(sched_group_cpus(group));
5258 /* Traverse only the allowed CPUs */
5259 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5261 struct rq *rq = cpu_rq(i);
5262 struct cpuidle_state *idle = idle_get_state(rq);
5263 if (idle && idle->exit_latency < min_exit_latency) {
5265 * We give priority to a CPU whose idle state
5266 * has the smallest exit latency irrespective
5267 * of any idle timestamp.
5269 min_exit_latency = idle->exit_latency;
5270 latest_idle_timestamp = rq->idle_stamp;
5271 shallowest_idle_cpu = i;
5272 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5273 rq->idle_stamp > latest_idle_timestamp) {
5275 * If equal or no active idle state, then
5276 * the most recently idled CPU might have
5279 latest_idle_timestamp = rq->idle_stamp;
5280 shallowest_idle_cpu = i;
5282 } else if (shallowest_idle_cpu == -1) {
5283 load = weighted_cpuload(i);
5284 if (load < min_load || (load == min_load && i == this_cpu)) {
5286 least_loaded_cpu = i;
5291 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5295 * Implement a for_each_cpu() variant that starts the scan at a given cpu
5296 * (@start), and wraps around.
5298 * This is used to scan for idle CPUs; such that not all CPUs looking for an
5299 * idle CPU find the same CPU. The down-side is that tasks tend to cycle
5300 * through the LLC domain.
5302 * Especially tbench is found sensitive to this.
5305 static int cpumask_next_wrap(int n, const struct cpumask *mask, int start, int *wrapped)
5310 next = find_next_bit(cpumask_bits(mask), nr_cpumask_bits, n+1);
5314 return nr_cpumask_bits;
5316 if (next >= nr_cpumask_bits) {
5326 #define for_each_cpu_wrap(cpu, mask, start, wrap) \
5327 for ((wrap) = 0, (cpu) = (start)-1; \
5328 (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)), \
5329 (cpu) < nr_cpumask_bits; )
5331 #ifdef CONFIG_SCHED_SMT
5333 static inline void set_idle_cores(int cpu, int val)
5335 struct sched_domain_shared *sds;
5337 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5339 WRITE_ONCE(sds->has_idle_cores, val);
5342 static inline bool test_idle_cores(int cpu, bool def)
5344 struct sched_domain_shared *sds;
5346 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5348 return READ_ONCE(sds->has_idle_cores);
5354 * Scans the local SMT mask to see if the entire core is idle, and records this
5355 * information in sd_llc_shared->has_idle_cores.
5357 * Since SMT siblings share all cache levels, inspecting this limited remote
5358 * state should be fairly cheap.
5360 void __update_idle_core(struct rq *rq)
5362 int core = cpu_of(rq);
5366 if (test_idle_cores(core, true))
5369 for_each_cpu(cpu, cpu_smt_mask(core)) {
5377 set_idle_cores(core, 1);
5383 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5384 * there are no idle cores left in the system; tracked through
5385 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5387 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5389 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5390 int core, cpu, wrap;
5392 if (!static_branch_likely(&sched_smt_present))
5395 if (!test_idle_cores(target, false))
5398 cpumask_and(cpus, sched_domain_span(sd), tsk_cpus_allowed(p));
5400 for_each_cpu_wrap(core, cpus, target, wrap) {
5403 for_each_cpu(cpu, cpu_smt_mask(core)) {
5404 cpumask_clear_cpu(cpu, cpus);
5414 * Failed to find an idle core; stop looking for one.
5416 set_idle_cores(target, 0);
5422 * Scan the local SMT mask for idle CPUs.
5424 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5428 if (!static_branch_likely(&sched_smt_present))
5431 for_each_cpu(cpu, cpu_smt_mask(target)) {
5432 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5441 #else /* CONFIG_SCHED_SMT */
5443 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5448 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5453 #endif /* CONFIG_SCHED_SMT */
5456 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5457 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5458 * average idle time for this rq (as found in rq->avg_idle).
5460 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5462 struct sched_domain *this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5463 u64 avg_idle = this_rq()->avg_idle;
5464 u64 avg_cost = this_sd->avg_scan_cost;
5470 * Due to large variance we need a large fuzz factor; hackbench in
5471 * particularly is sensitive here.
5473 if ((avg_idle / 512) < avg_cost)
5476 time = local_clock();
5478 for_each_cpu_wrap(cpu, sched_domain_span(sd), target, wrap) {
5479 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5485 time = local_clock() - time;
5486 cost = this_sd->avg_scan_cost;
5487 delta = (s64)(time - cost) / 8;
5488 this_sd->avg_scan_cost += delta;
5494 * Try and locate an idle core/thread in the LLC cache domain.
5496 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5498 struct sched_domain *sd;
5501 if (idle_cpu(target))
5505 * If the previous cpu is cache affine and idle, don't be stupid.
5507 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5510 sd = rcu_dereference(per_cpu(sd_llc, target));
5514 i = select_idle_core(p, sd, target);
5515 if ((unsigned)i < nr_cpumask_bits)
5518 i = select_idle_cpu(p, sd, target);
5519 if ((unsigned)i < nr_cpumask_bits)
5522 i = select_idle_smt(p, sd, target);
5523 if ((unsigned)i < nr_cpumask_bits)
5530 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5531 * tasks. The unit of the return value must be the one of capacity so we can
5532 * compare the utilization with the capacity of the CPU that is available for
5533 * CFS task (ie cpu_capacity).
5535 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5536 * recent utilization of currently non-runnable tasks on a CPU. It represents
5537 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5538 * capacity_orig is the cpu_capacity available at the highest frequency
5539 * (arch_scale_freq_capacity()).
5540 * The utilization of a CPU converges towards a sum equal to or less than the
5541 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5542 * the running time on this CPU scaled by capacity_curr.
5544 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5545 * higher than capacity_orig because of unfortunate rounding in
5546 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5547 * the average stabilizes with the new running time. We need to check that the
5548 * utilization stays within the range of [0..capacity_orig] and cap it if
5549 * necessary. Without utilization capping, a group could be seen as overloaded
5550 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5551 * available capacity. We allow utilization to overshoot capacity_curr (but not
5552 * capacity_orig) as it useful for predicting the capacity required after task
5553 * migrations (scheduler-driven DVFS).
5555 static int cpu_util(int cpu)
5557 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5558 unsigned long capacity = capacity_orig_of(cpu);
5560 return (util >= capacity) ? capacity : util;
5563 static inline int task_util(struct task_struct *p)
5565 return p->se.avg.util_avg;
5569 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5570 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5572 * In that case WAKE_AFFINE doesn't make sense and we'll let
5573 * BALANCE_WAKE sort things out.
5575 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5577 long min_cap, max_cap;
5579 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5580 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5582 /* Minimum capacity is close to max, no need to abort wake_affine */
5583 if (max_cap - min_cap < max_cap >> 3)
5586 return min_cap * 1024 < task_util(p) * capacity_margin;
5590 * select_task_rq_fair: Select target runqueue for the waking task in domains
5591 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5592 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5594 * Balances load by selecting the idlest cpu in the idlest group, or under
5595 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5597 * Returns the target cpu number.
5599 * preempt must be disabled.
5602 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5604 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5605 int cpu = smp_processor_id();
5606 int new_cpu = prev_cpu;
5607 int want_affine = 0;
5608 int sync = wake_flags & WF_SYNC;
5610 if (sd_flag & SD_BALANCE_WAKE) {
5612 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5613 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5617 for_each_domain(cpu, tmp) {
5618 if (!(tmp->flags & SD_LOAD_BALANCE))
5622 * If both cpu and prev_cpu are part of this domain,
5623 * cpu is a valid SD_WAKE_AFFINE target.
5625 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5626 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5631 if (tmp->flags & sd_flag)
5633 else if (!want_affine)
5638 sd = NULL; /* Prefer wake_affine over balance flags */
5639 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5644 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5645 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5648 struct sched_group *group;
5651 if (!(sd->flags & sd_flag)) {
5656 group = find_idlest_group(sd, p, cpu, sd_flag);
5662 new_cpu = find_idlest_cpu(group, p, cpu);
5663 if (new_cpu == -1 || new_cpu == cpu) {
5664 /* Now try balancing at a lower domain level of cpu */
5669 /* Now try balancing at a lower domain level of new_cpu */
5671 weight = sd->span_weight;
5673 for_each_domain(cpu, tmp) {
5674 if (weight <= tmp->span_weight)
5676 if (tmp->flags & sd_flag)
5679 /* while loop will break here if sd == NULL */
5687 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5688 * cfs_rq_of(p) references at time of call are still valid and identify the
5689 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5691 static void migrate_task_rq_fair(struct task_struct *p)
5694 * As blocked tasks retain absolute vruntime the migration needs to
5695 * deal with this by subtracting the old and adding the new
5696 * min_vruntime -- the latter is done by enqueue_entity() when placing
5697 * the task on the new runqueue.
5699 if (p->state == TASK_WAKING) {
5700 struct sched_entity *se = &p->se;
5701 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5704 #ifndef CONFIG_64BIT
5705 u64 min_vruntime_copy;
5708 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5710 min_vruntime = cfs_rq->min_vruntime;
5711 } while (min_vruntime != min_vruntime_copy);
5713 min_vruntime = cfs_rq->min_vruntime;
5716 se->vruntime -= min_vruntime;
5720 * We are supposed to update the task to "current" time, then its up to date
5721 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5722 * what current time is, so simply throw away the out-of-date time. This
5723 * will result in the wakee task is less decayed, but giving the wakee more
5724 * load sounds not bad.
5726 remove_entity_load_avg(&p->se);
5728 /* Tell new CPU we are migrated */
5729 p->se.avg.last_update_time = 0;
5731 /* We have migrated, no longer consider this task hot */
5732 p->se.exec_start = 0;
5735 static void task_dead_fair(struct task_struct *p)
5737 remove_entity_load_avg(&p->se);
5739 #endif /* CONFIG_SMP */
5741 static unsigned long
5742 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5744 unsigned long gran = sysctl_sched_wakeup_granularity;
5747 * Since its curr running now, convert the gran from real-time
5748 * to virtual-time in his units.
5750 * By using 'se' instead of 'curr' we penalize light tasks, so
5751 * they get preempted easier. That is, if 'se' < 'curr' then
5752 * the resulting gran will be larger, therefore penalizing the
5753 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5754 * be smaller, again penalizing the lighter task.
5756 * This is especially important for buddies when the leftmost
5757 * task is higher priority than the buddy.
5759 return calc_delta_fair(gran, se);
5763 * Should 'se' preempt 'curr'.
5777 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5779 s64 gran, vdiff = curr->vruntime - se->vruntime;
5784 gran = wakeup_gran(curr, se);
5791 static void set_last_buddy(struct sched_entity *se)
5793 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5796 for_each_sched_entity(se)
5797 cfs_rq_of(se)->last = se;
5800 static void set_next_buddy(struct sched_entity *se)
5802 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5805 for_each_sched_entity(se)
5806 cfs_rq_of(se)->next = se;
5809 static void set_skip_buddy(struct sched_entity *se)
5811 for_each_sched_entity(se)
5812 cfs_rq_of(se)->skip = se;
5816 * Preempt the current task with a newly woken task if needed:
5818 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5820 struct task_struct *curr = rq->curr;
5821 struct sched_entity *se = &curr->se, *pse = &p->se;
5822 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5823 int scale = cfs_rq->nr_running >= sched_nr_latency;
5824 int next_buddy_marked = 0;
5826 if (unlikely(se == pse))
5830 * This is possible from callers such as attach_tasks(), in which we
5831 * unconditionally check_prempt_curr() after an enqueue (which may have
5832 * lead to a throttle). This both saves work and prevents false
5833 * next-buddy nomination below.
5835 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5838 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5839 set_next_buddy(pse);
5840 next_buddy_marked = 1;
5844 * We can come here with TIF_NEED_RESCHED already set from new task
5847 * Note: this also catches the edge-case of curr being in a throttled
5848 * group (e.g. via set_curr_task), since update_curr() (in the
5849 * enqueue of curr) will have resulted in resched being set. This
5850 * prevents us from potentially nominating it as a false LAST_BUDDY
5853 if (test_tsk_need_resched(curr))
5856 /* Idle tasks are by definition preempted by non-idle tasks. */
5857 if (unlikely(curr->policy == SCHED_IDLE) &&
5858 likely(p->policy != SCHED_IDLE))
5862 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5863 * is driven by the tick):
5865 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5868 find_matching_se(&se, &pse);
5869 update_curr(cfs_rq_of(se));
5871 if (wakeup_preempt_entity(se, pse) == 1) {
5873 * Bias pick_next to pick the sched entity that is
5874 * triggering this preemption.
5876 if (!next_buddy_marked)
5877 set_next_buddy(pse);
5886 * Only set the backward buddy when the current task is still
5887 * on the rq. This can happen when a wakeup gets interleaved
5888 * with schedule on the ->pre_schedule() or idle_balance()
5889 * point, either of which can * drop the rq lock.
5891 * Also, during early boot the idle thread is in the fair class,
5892 * for obvious reasons its a bad idea to schedule back to it.
5894 if (unlikely(!se->on_rq || curr == rq->idle))
5897 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5901 static struct task_struct *
5902 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5904 struct cfs_rq *cfs_rq = &rq->cfs;
5905 struct sched_entity *se;
5906 struct task_struct *p;
5910 #ifdef CONFIG_FAIR_GROUP_SCHED
5911 if (!cfs_rq->nr_running)
5914 if (prev->sched_class != &fair_sched_class)
5918 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5919 * likely that a next task is from the same cgroup as the current.
5921 * Therefore attempt to avoid putting and setting the entire cgroup
5922 * hierarchy, only change the part that actually changes.
5926 struct sched_entity *curr = cfs_rq->curr;
5929 * Since we got here without doing put_prev_entity() we also
5930 * have to consider cfs_rq->curr. If it is still a runnable
5931 * entity, update_curr() will update its vruntime, otherwise
5932 * forget we've ever seen it.
5936 update_curr(cfs_rq);
5941 * This call to check_cfs_rq_runtime() will do the
5942 * throttle and dequeue its entity in the parent(s).
5943 * Therefore the 'simple' nr_running test will indeed
5946 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5950 se = pick_next_entity(cfs_rq, curr);
5951 cfs_rq = group_cfs_rq(se);
5957 * Since we haven't yet done put_prev_entity and if the selected task
5958 * is a different task than we started out with, try and touch the
5959 * least amount of cfs_rqs.
5962 struct sched_entity *pse = &prev->se;
5964 while (!(cfs_rq = is_same_group(se, pse))) {
5965 int se_depth = se->depth;
5966 int pse_depth = pse->depth;
5968 if (se_depth <= pse_depth) {
5969 put_prev_entity(cfs_rq_of(pse), pse);
5970 pse = parent_entity(pse);
5972 if (se_depth >= pse_depth) {
5973 set_next_entity(cfs_rq_of(se), se);
5974 se = parent_entity(se);
5978 put_prev_entity(cfs_rq, pse);
5979 set_next_entity(cfs_rq, se);
5982 if (hrtick_enabled(rq))
5983 hrtick_start_fair(rq, p);
5990 if (!cfs_rq->nr_running)
5993 put_prev_task(rq, prev);
5996 se = pick_next_entity(cfs_rq, NULL);
5997 set_next_entity(cfs_rq, se);
5998 cfs_rq = group_cfs_rq(se);
6003 if (hrtick_enabled(rq))
6004 hrtick_start_fair(rq, p);
6010 * This is OK, because current is on_cpu, which avoids it being picked
6011 * for load-balance and preemption/IRQs are still disabled avoiding
6012 * further scheduler activity on it and we're being very careful to
6013 * re-start the picking loop.
6015 lockdep_unpin_lock(&rq->lock, cookie);
6016 new_tasks = idle_balance(rq);
6017 lockdep_repin_lock(&rq->lock, cookie);
6019 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6020 * possible for any higher priority task to appear. In that case we
6021 * must re-start the pick_next_entity() loop.
6033 * Account for a descheduled task:
6035 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6037 struct sched_entity *se = &prev->se;
6038 struct cfs_rq *cfs_rq;
6040 for_each_sched_entity(se) {
6041 cfs_rq = cfs_rq_of(se);
6042 put_prev_entity(cfs_rq, se);
6047 * sched_yield() is very simple
6049 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6051 static void yield_task_fair(struct rq *rq)
6053 struct task_struct *curr = rq->curr;
6054 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6055 struct sched_entity *se = &curr->se;
6058 * Are we the only task in the tree?
6060 if (unlikely(rq->nr_running == 1))
6063 clear_buddies(cfs_rq, se);
6065 if (curr->policy != SCHED_BATCH) {
6066 update_rq_clock(rq);
6068 * Update run-time statistics of the 'current'.
6070 update_curr(cfs_rq);
6072 * Tell update_rq_clock() that we've just updated,
6073 * so we don't do microscopic update in schedule()
6074 * and double the fastpath cost.
6076 rq_clock_skip_update(rq, true);
6082 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6084 struct sched_entity *se = &p->se;
6086 /* throttled hierarchies are not runnable */
6087 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6090 /* Tell the scheduler that we'd really like pse to run next. */
6093 yield_task_fair(rq);
6099 /**************************************************
6100 * Fair scheduling class load-balancing methods.
6104 * The purpose of load-balancing is to achieve the same basic fairness the
6105 * per-cpu scheduler provides, namely provide a proportional amount of compute
6106 * time to each task. This is expressed in the following equation:
6108 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6110 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6111 * W_i,0 is defined as:
6113 * W_i,0 = \Sum_j w_i,j (2)
6115 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6116 * is derived from the nice value as per sched_prio_to_weight[].
6118 * The weight average is an exponential decay average of the instantaneous
6121 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6123 * C_i is the compute capacity of cpu i, typically it is the
6124 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6125 * can also include other factors [XXX].
6127 * To achieve this balance we define a measure of imbalance which follows
6128 * directly from (1):
6130 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6132 * We them move tasks around to minimize the imbalance. In the continuous
6133 * function space it is obvious this converges, in the discrete case we get
6134 * a few fun cases generally called infeasible weight scenarios.
6137 * - infeasible weights;
6138 * - local vs global optima in the discrete case. ]
6143 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6144 * for all i,j solution, we create a tree of cpus that follows the hardware
6145 * topology where each level pairs two lower groups (or better). This results
6146 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6147 * tree to only the first of the previous level and we decrease the frequency
6148 * of load-balance at each level inv. proportional to the number of cpus in
6154 * \Sum { --- * --- * 2^i } = O(n) (5)
6156 * `- size of each group
6157 * | | `- number of cpus doing load-balance
6159 * `- sum over all levels
6161 * Coupled with a limit on how many tasks we can migrate every balance pass,
6162 * this makes (5) the runtime complexity of the balancer.
6164 * An important property here is that each CPU is still (indirectly) connected
6165 * to every other cpu in at most O(log n) steps:
6167 * The adjacency matrix of the resulting graph is given by:
6170 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6173 * And you'll find that:
6175 * A^(log_2 n)_i,j != 0 for all i,j (7)
6177 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6178 * The task movement gives a factor of O(m), giving a convergence complexity
6181 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6186 * In order to avoid CPUs going idle while there's still work to do, new idle
6187 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6188 * tree itself instead of relying on other CPUs to bring it work.
6190 * This adds some complexity to both (5) and (8) but it reduces the total idle
6198 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6201 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6206 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6208 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6210 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6213 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6214 * rewrite all of this once again.]
6217 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6219 enum fbq_type { regular, remote, all };
6221 #define LBF_ALL_PINNED 0x01
6222 #define LBF_NEED_BREAK 0x02
6223 #define LBF_DST_PINNED 0x04
6224 #define LBF_SOME_PINNED 0x08
6227 struct sched_domain *sd;
6235 struct cpumask *dst_grpmask;
6237 enum cpu_idle_type idle;
6239 /* The set of CPUs under consideration for load-balancing */
6240 struct cpumask *cpus;
6245 unsigned int loop_break;
6246 unsigned int loop_max;
6248 enum fbq_type fbq_type;
6249 struct list_head tasks;
6253 * Is this task likely cache-hot:
6255 static int task_hot(struct task_struct *p, struct lb_env *env)
6259 lockdep_assert_held(&env->src_rq->lock);
6261 if (p->sched_class != &fair_sched_class)
6264 if (unlikely(p->policy == SCHED_IDLE))
6268 * Buddy candidates are cache hot:
6270 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6271 (&p->se == cfs_rq_of(&p->se)->next ||
6272 &p->se == cfs_rq_of(&p->se)->last))
6275 if (sysctl_sched_migration_cost == -1)
6277 if (sysctl_sched_migration_cost == 0)
6280 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6282 return delta < (s64)sysctl_sched_migration_cost;
6285 #ifdef CONFIG_NUMA_BALANCING
6287 * Returns 1, if task migration degrades locality
6288 * Returns 0, if task migration improves locality i.e migration preferred.
6289 * Returns -1, if task migration is not affected by locality.
6291 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6293 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6294 unsigned long src_faults, dst_faults;
6295 int src_nid, dst_nid;
6297 if (!static_branch_likely(&sched_numa_balancing))
6300 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6303 src_nid = cpu_to_node(env->src_cpu);
6304 dst_nid = cpu_to_node(env->dst_cpu);
6306 if (src_nid == dst_nid)
6309 /* Migrating away from the preferred node is always bad. */
6310 if (src_nid == p->numa_preferred_nid) {
6311 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6317 /* Encourage migration to the preferred node. */
6318 if (dst_nid == p->numa_preferred_nid)
6322 src_faults = group_faults(p, src_nid);
6323 dst_faults = group_faults(p, dst_nid);
6325 src_faults = task_faults(p, src_nid);
6326 dst_faults = task_faults(p, dst_nid);
6329 return dst_faults < src_faults;
6333 static inline int migrate_degrades_locality(struct task_struct *p,
6341 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6344 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6348 lockdep_assert_held(&env->src_rq->lock);
6351 * We do not migrate tasks that are:
6352 * 1) throttled_lb_pair, or
6353 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6354 * 3) running (obviously), or
6355 * 4) are cache-hot on their current CPU.
6357 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6360 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6363 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6365 env->flags |= LBF_SOME_PINNED;
6368 * Remember if this task can be migrated to any other cpu in
6369 * our sched_group. We may want to revisit it if we couldn't
6370 * meet load balance goals by pulling other tasks on src_cpu.
6372 * Also avoid computing new_dst_cpu if we have already computed
6373 * one in current iteration.
6375 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6378 /* Prevent to re-select dst_cpu via env's cpus */
6379 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6380 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6381 env->flags |= LBF_DST_PINNED;
6382 env->new_dst_cpu = cpu;
6390 /* Record that we found atleast one task that could run on dst_cpu */
6391 env->flags &= ~LBF_ALL_PINNED;
6393 if (task_running(env->src_rq, p)) {
6394 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6399 * Aggressive migration if:
6400 * 1) destination numa is preferred
6401 * 2) task is cache cold, or
6402 * 3) too many balance attempts have failed.
6404 tsk_cache_hot = migrate_degrades_locality(p, env);
6405 if (tsk_cache_hot == -1)
6406 tsk_cache_hot = task_hot(p, env);
6408 if (tsk_cache_hot <= 0 ||
6409 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6410 if (tsk_cache_hot == 1) {
6411 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6412 schedstat_inc(p->se.statistics.nr_forced_migrations);
6417 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6422 * detach_task() -- detach the task for the migration specified in env
6424 static void detach_task(struct task_struct *p, struct lb_env *env)
6426 lockdep_assert_held(&env->src_rq->lock);
6428 p->on_rq = TASK_ON_RQ_MIGRATING;
6429 deactivate_task(env->src_rq, p, 0);
6430 set_task_cpu(p, env->dst_cpu);
6434 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6435 * part of active balancing operations within "domain".
6437 * Returns a task if successful and NULL otherwise.
6439 static struct task_struct *detach_one_task(struct lb_env *env)
6441 struct task_struct *p, *n;
6443 lockdep_assert_held(&env->src_rq->lock);
6445 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6446 if (!can_migrate_task(p, env))
6449 detach_task(p, env);
6452 * Right now, this is only the second place where
6453 * lb_gained[env->idle] is updated (other is detach_tasks)
6454 * so we can safely collect stats here rather than
6455 * inside detach_tasks().
6457 schedstat_inc(env->sd->lb_gained[env->idle]);
6463 static const unsigned int sched_nr_migrate_break = 32;
6466 * detach_tasks() -- tries to detach up to imbalance weighted load from
6467 * busiest_rq, as part of a balancing operation within domain "sd".
6469 * Returns number of detached tasks if successful and 0 otherwise.
6471 static int detach_tasks(struct lb_env *env)
6473 struct list_head *tasks = &env->src_rq->cfs_tasks;
6474 struct task_struct *p;
6478 lockdep_assert_held(&env->src_rq->lock);
6480 if (env->imbalance <= 0)
6483 while (!list_empty(tasks)) {
6485 * We don't want to steal all, otherwise we may be treated likewise,
6486 * which could at worst lead to a livelock crash.
6488 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6491 p = list_first_entry(tasks, struct task_struct, se.group_node);
6494 /* We've more or less seen every task there is, call it quits */
6495 if (env->loop > env->loop_max)
6498 /* take a breather every nr_migrate tasks */
6499 if (env->loop > env->loop_break) {
6500 env->loop_break += sched_nr_migrate_break;
6501 env->flags |= LBF_NEED_BREAK;
6505 if (!can_migrate_task(p, env))
6508 load = task_h_load(p);
6510 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6513 if ((load / 2) > env->imbalance)
6516 detach_task(p, env);
6517 list_add(&p->se.group_node, &env->tasks);
6520 env->imbalance -= load;
6522 #ifdef CONFIG_PREEMPT
6524 * NEWIDLE balancing is a source of latency, so preemptible
6525 * kernels will stop after the first task is detached to minimize
6526 * the critical section.
6528 if (env->idle == CPU_NEWLY_IDLE)
6533 * We only want to steal up to the prescribed amount of
6536 if (env->imbalance <= 0)
6541 list_move_tail(&p->se.group_node, tasks);
6545 * Right now, this is one of only two places we collect this stat
6546 * so we can safely collect detach_one_task() stats here rather
6547 * than inside detach_one_task().
6549 schedstat_add(env->sd->lb_gained[env->idle], detached);
6555 * attach_task() -- attach the task detached by detach_task() to its new rq.
6557 static void attach_task(struct rq *rq, struct task_struct *p)
6559 lockdep_assert_held(&rq->lock);
6561 BUG_ON(task_rq(p) != rq);
6562 activate_task(rq, p, 0);
6563 p->on_rq = TASK_ON_RQ_QUEUED;
6564 check_preempt_curr(rq, p, 0);
6568 * attach_one_task() -- attaches the task returned from detach_one_task() to
6571 static void attach_one_task(struct rq *rq, struct task_struct *p)
6573 raw_spin_lock(&rq->lock);
6575 raw_spin_unlock(&rq->lock);
6579 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6582 static void attach_tasks(struct lb_env *env)
6584 struct list_head *tasks = &env->tasks;
6585 struct task_struct *p;
6587 raw_spin_lock(&env->dst_rq->lock);
6589 while (!list_empty(tasks)) {
6590 p = list_first_entry(tasks, struct task_struct, se.group_node);
6591 list_del_init(&p->se.group_node);
6593 attach_task(env->dst_rq, p);
6596 raw_spin_unlock(&env->dst_rq->lock);
6599 #ifdef CONFIG_FAIR_GROUP_SCHED
6600 static void update_blocked_averages(int cpu)
6602 struct rq *rq = cpu_rq(cpu);
6603 struct cfs_rq *cfs_rq;
6604 unsigned long flags;
6606 raw_spin_lock_irqsave(&rq->lock, flags);
6607 update_rq_clock(rq);
6610 * Iterates the task_group tree in a bottom up fashion, see
6611 * list_add_leaf_cfs_rq() for details.
6613 for_each_leaf_cfs_rq(rq, cfs_rq) {
6614 /* throttled entities do not contribute to load */
6615 if (throttled_hierarchy(cfs_rq))
6618 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6619 update_tg_load_avg(cfs_rq, 0);
6621 raw_spin_unlock_irqrestore(&rq->lock, flags);
6625 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6626 * This needs to be done in a top-down fashion because the load of a child
6627 * group is a fraction of its parents load.
6629 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6631 struct rq *rq = rq_of(cfs_rq);
6632 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6633 unsigned long now = jiffies;
6636 if (cfs_rq->last_h_load_update == now)
6639 cfs_rq->h_load_next = NULL;
6640 for_each_sched_entity(se) {
6641 cfs_rq = cfs_rq_of(se);
6642 cfs_rq->h_load_next = se;
6643 if (cfs_rq->last_h_load_update == now)
6648 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6649 cfs_rq->last_h_load_update = now;
6652 while ((se = cfs_rq->h_load_next) != NULL) {
6653 load = cfs_rq->h_load;
6654 load = div64_ul(load * se->avg.load_avg,
6655 cfs_rq_load_avg(cfs_rq) + 1);
6656 cfs_rq = group_cfs_rq(se);
6657 cfs_rq->h_load = load;
6658 cfs_rq->last_h_load_update = now;
6662 static unsigned long task_h_load(struct task_struct *p)
6664 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6666 update_cfs_rq_h_load(cfs_rq);
6667 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6668 cfs_rq_load_avg(cfs_rq) + 1);
6671 static inline void update_blocked_averages(int cpu)
6673 struct rq *rq = cpu_rq(cpu);
6674 struct cfs_rq *cfs_rq = &rq->cfs;
6675 unsigned long flags;
6677 raw_spin_lock_irqsave(&rq->lock, flags);
6678 update_rq_clock(rq);
6679 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6680 raw_spin_unlock_irqrestore(&rq->lock, flags);
6683 static unsigned long task_h_load(struct task_struct *p)
6685 return p->se.avg.load_avg;
6689 /********** Helpers for find_busiest_group ************************/
6698 * sg_lb_stats - stats of a sched_group required for load_balancing
6700 struct sg_lb_stats {
6701 unsigned long avg_load; /*Avg load across the CPUs of the group */
6702 unsigned long group_load; /* Total load over the CPUs of the group */
6703 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6704 unsigned long load_per_task;
6705 unsigned long group_capacity;
6706 unsigned long group_util; /* Total utilization of the group */
6707 unsigned int sum_nr_running; /* Nr tasks running in the group */
6708 unsigned int idle_cpus;
6709 unsigned int group_weight;
6710 enum group_type group_type;
6711 int group_no_capacity;
6712 #ifdef CONFIG_NUMA_BALANCING
6713 unsigned int nr_numa_running;
6714 unsigned int nr_preferred_running;
6719 * sd_lb_stats - Structure to store the statistics of a sched_domain
6720 * during load balancing.
6722 struct sd_lb_stats {
6723 struct sched_group *busiest; /* Busiest group in this sd */
6724 struct sched_group *local; /* Local group in this sd */
6725 unsigned long total_load; /* Total load of all groups in sd */
6726 unsigned long total_capacity; /* Total capacity of all groups in sd */
6727 unsigned long avg_load; /* Average load across all groups in sd */
6729 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6730 struct sg_lb_stats local_stat; /* Statistics of the local group */
6733 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6736 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6737 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6738 * We must however clear busiest_stat::avg_load because
6739 * update_sd_pick_busiest() reads this before assignment.
6741 *sds = (struct sd_lb_stats){
6745 .total_capacity = 0UL,
6748 .sum_nr_running = 0,
6749 .group_type = group_other,
6755 * get_sd_load_idx - Obtain the load index for a given sched domain.
6756 * @sd: The sched_domain whose load_idx is to be obtained.
6757 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6759 * Return: The load index.
6761 static inline int get_sd_load_idx(struct sched_domain *sd,
6762 enum cpu_idle_type idle)
6768 load_idx = sd->busy_idx;
6771 case CPU_NEWLY_IDLE:
6772 load_idx = sd->newidle_idx;
6775 load_idx = sd->idle_idx;
6782 static unsigned long scale_rt_capacity(int cpu)
6784 struct rq *rq = cpu_rq(cpu);
6785 u64 total, used, age_stamp, avg;
6789 * Since we're reading these variables without serialization make sure
6790 * we read them once before doing sanity checks on them.
6792 age_stamp = READ_ONCE(rq->age_stamp);
6793 avg = READ_ONCE(rq->rt_avg);
6794 delta = __rq_clock_broken(rq) - age_stamp;
6796 if (unlikely(delta < 0))
6799 total = sched_avg_period() + delta;
6801 used = div_u64(avg, total);
6803 if (likely(used < SCHED_CAPACITY_SCALE))
6804 return SCHED_CAPACITY_SCALE - used;
6809 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6811 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6812 struct sched_group *sdg = sd->groups;
6814 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6816 capacity *= scale_rt_capacity(cpu);
6817 capacity >>= SCHED_CAPACITY_SHIFT;
6822 cpu_rq(cpu)->cpu_capacity = capacity;
6823 sdg->sgc->capacity = capacity;
6826 void update_group_capacity(struct sched_domain *sd, int cpu)
6828 struct sched_domain *child = sd->child;
6829 struct sched_group *group, *sdg = sd->groups;
6830 unsigned long capacity;
6831 unsigned long interval;
6833 interval = msecs_to_jiffies(sd->balance_interval);
6834 interval = clamp(interval, 1UL, max_load_balance_interval);
6835 sdg->sgc->next_update = jiffies + interval;
6838 update_cpu_capacity(sd, cpu);
6844 if (child->flags & SD_OVERLAP) {
6846 * SD_OVERLAP domains cannot assume that child groups
6847 * span the current group.
6850 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6851 struct sched_group_capacity *sgc;
6852 struct rq *rq = cpu_rq(cpu);
6855 * build_sched_domains() -> init_sched_groups_capacity()
6856 * gets here before we've attached the domains to the
6859 * Use capacity_of(), which is set irrespective of domains
6860 * in update_cpu_capacity().
6862 * This avoids capacity from being 0 and
6863 * causing divide-by-zero issues on boot.
6865 if (unlikely(!rq->sd)) {
6866 capacity += capacity_of(cpu);
6870 sgc = rq->sd->groups->sgc;
6871 capacity += sgc->capacity;
6875 * !SD_OVERLAP domains can assume that child groups
6876 * span the current group.
6879 group = child->groups;
6881 capacity += group->sgc->capacity;
6882 group = group->next;
6883 } while (group != child->groups);
6886 sdg->sgc->capacity = capacity;
6890 * Check whether the capacity of the rq has been noticeably reduced by side
6891 * activity. The imbalance_pct is used for the threshold.
6892 * Return true is the capacity is reduced
6895 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6897 return ((rq->cpu_capacity * sd->imbalance_pct) <
6898 (rq->cpu_capacity_orig * 100));
6902 * Group imbalance indicates (and tries to solve) the problem where balancing
6903 * groups is inadequate due to tsk_cpus_allowed() constraints.
6905 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6906 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6909 * { 0 1 2 3 } { 4 5 6 7 }
6912 * If we were to balance group-wise we'd place two tasks in the first group and
6913 * two tasks in the second group. Clearly this is undesired as it will overload
6914 * cpu 3 and leave one of the cpus in the second group unused.
6916 * The current solution to this issue is detecting the skew in the first group
6917 * by noticing the lower domain failed to reach balance and had difficulty
6918 * moving tasks due to affinity constraints.
6920 * When this is so detected; this group becomes a candidate for busiest; see
6921 * update_sd_pick_busiest(). And calculate_imbalance() and
6922 * find_busiest_group() avoid some of the usual balance conditions to allow it
6923 * to create an effective group imbalance.
6925 * This is a somewhat tricky proposition since the next run might not find the
6926 * group imbalance and decide the groups need to be balanced again. A most
6927 * subtle and fragile situation.
6930 static inline int sg_imbalanced(struct sched_group *group)
6932 return group->sgc->imbalance;
6936 * group_has_capacity returns true if the group has spare capacity that could
6937 * be used by some tasks.
6938 * We consider that a group has spare capacity if the * number of task is
6939 * smaller than the number of CPUs or if the utilization is lower than the
6940 * available capacity for CFS tasks.
6941 * For the latter, we use a threshold to stabilize the state, to take into
6942 * account the variance of the tasks' load and to return true if the available
6943 * capacity in meaningful for the load balancer.
6944 * As an example, an available capacity of 1% can appear but it doesn't make
6945 * any benefit for the load balance.
6948 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6950 if (sgs->sum_nr_running < sgs->group_weight)
6953 if ((sgs->group_capacity * 100) >
6954 (sgs->group_util * env->sd->imbalance_pct))
6961 * group_is_overloaded returns true if the group has more tasks than it can
6963 * group_is_overloaded is not equals to !group_has_capacity because a group
6964 * with the exact right number of tasks, has no more spare capacity but is not
6965 * overloaded so both group_has_capacity and group_is_overloaded return
6969 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6971 if (sgs->sum_nr_running <= sgs->group_weight)
6974 if ((sgs->group_capacity * 100) <
6975 (sgs->group_util * env->sd->imbalance_pct))
6982 group_type group_classify(struct sched_group *group,
6983 struct sg_lb_stats *sgs)
6985 if (sgs->group_no_capacity)
6986 return group_overloaded;
6988 if (sg_imbalanced(group))
6989 return group_imbalanced;
6995 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6996 * @env: The load balancing environment.
6997 * @group: sched_group whose statistics are to be updated.
6998 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6999 * @local_group: Does group contain this_cpu.
7000 * @sgs: variable to hold the statistics for this group.
7001 * @overload: Indicate more than one runnable task for any CPU.
7003 static inline void update_sg_lb_stats(struct lb_env *env,
7004 struct sched_group *group, int load_idx,
7005 int local_group, struct sg_lb_stats *sgs,
7011 memset(sgs, 0, sizeof(*sgs));
7013 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7014 struct rq *rq = cpu_rq(i);
7016 /* Bias balancing toward cpus of our domain */
7018 load = target_load(i, load_idx);
7020 load = source_load(i, load_idx);
7022 sgs->group_load += load;
7023 sgs->group_util += cpu_util(i);
7024 sgs->sum_nr_running += rq->cfs.h_nr_running;
7026 nr_running = rq->nr_running;
7030 #ifdef CONFIG_NUMA_BALANCING
7031 sgs->nr_numa_running += rq->nr_numa_running;
7032 sgs->nr_preferred_running += rq->nr_preferred_running;
7034 sgs->sum_weighted_load += weighted_cpuload(i);
7036 * No need to call idle_cpu() if nr_running is not 0
7038 if (!nr_running && idle_cpu(i))
7042 /* Adjust by relative CPU capacity of the group */
7043 sgs->group_capacity = group->sgc->capacity;
7044 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7046 if (sgs->sum_nr_running)
7047 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7049 sgs->group_weight = group->group_weight;
7051 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7052 sgs->group_type = group_classify(group, sgs);
7056 * update_sd_pick_busiest - return 1 on busiest group
7057 * @env: The load balancing environment.
7058 * @sds: sched_domain statistics
7059 * @sg: sched_group candidate to be checked for being the busiest
7060 * @sgs: sched_group statistics
7062 * Determine if @sg is a busier group than the previously selected
7065 * Return: %true if @sg is a busier group than the previously selected
7066 * busiest group. %false otherwise.
7068 static bool update_sd_pick_busiest(struct lb_env *env,
7069 struct sd_lb_stats *sds,
7070 struct sched_group *sg,
7071 struct sg_lb_stats *sgs)
7073 struct sg_lb_stats *busiest = &sds->busiest_stat;
7075 if (sgs->group_type > busiest->group_type)
7078 if (sgs->group_type < busiest->group_type)
7081 if (sgs->avg_load <= busiest->avg_load)
7084 /* This is the busiest node in its class. */
7085 if (!(env->sd->flags & SD_ASYM_PACKING))
7088 /* No ASYM_PACKING if target cpu is already busy */
7089 if (env->idle == CPU_NOT_IDLE)
7092 * ASYM_PACKING needs to move all the work to the lowest
7093 * numbered CPUs in the group, therefore mark all groups
7094 * higher than ourself as busy.
7096 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7100 /* Prefer to move from highest possible cpu's work */
7101 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
7108 #ifdef CONFIG_NUMA_BALANCING
7109 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7111 if (sgs->sum_nr_running > sgs->nr_numa_running)
7113 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7118 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7120 if (rq->nr_running > rq->nr_numa_running)
7122 if (rq->nr_running > rq->nr_preferred_running)
7127 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7132 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7136 #endif /* CONFIG_NUMA_BALANCING */
7139 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7140 * @env: The load balancing environment.
7141 * @sds: variable to hold the statistics for this sched_domain.
7143 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7145 struct sched_domain *child = env->sd->child;
7146 struct sched_group *sg = env->sd->groups;
7147 struct sg_lb_stats tmp_sgs;
7148 int load_idx, prefer_sibling = 0;
7149 bool overload = false;
7151 if (child && child->flags & SD_PREFER_SIBLING)
7154 load_idx = get_sd_load_idx(env->sd, env->idle);
7157 struct sg_lb_stats *sgs = &tmp_sgs;
7160 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7163 sgs = &sds->local_stat;
7165 if (env->idle != CPU_NEWLY_IDLE ||
7166 time_after_eq(jiffies, sg->sgc->next_update))
7167 update_group_capacity(env->sd, env->dst_cpu);
7170 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7177 * In case the child domain prefers tasks go to siblings
7178 * first, lower the sg capacity so that we'll try
7179 * and move all the excess tasks away. We lower the capacity
7180 * of a group only if the local group has the capacity to fit
7181 * these excess tasks. The extra check prevents the case where
7182 * you always pull from the heaviest group when it is already
7183 * under-utilized (possible with a large weight task outweighs
7184 * the tasks on the system).
7186 if (prefer_sibling && sds->local &&
7187 group_has_capacity(env, &sds->local_stat) &&
7188 (sgs->sum_nr_running > 1)) {
7189 sgs->group_no_capacity = 1;
7190 sgs->group_type = group_classify(sg, sgs);
7193 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7195 sds->busiest_stat = *sgs;
7199 /* Now, start updating sd_lb_stats */
7200 sds->total_load += sgs->group_load;
7201 sds->total_capacity += sgs->group_capacity;
7204 } while (sg != env->sd->groups);
7206 if (env->sd->flags & SD_NUMA)
7207 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7209 if (!env->sd->parent) {
7210 /* update overload indicator if we are at root domain */
7211 if (env->dst_rq->rd->overload != overload)
7212 env->dst_rq->rd->overload = overload;
7218 * check_asym_packing - Check to see if the group is packed into the
7221 * This is primarily intended to used at the sibling level. Some
7222 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7223 * case of POWER7, it can move to lower SMT modes only when higher
7224 * threads are idle. When in lower SMT modes, the threads will
7225 * perform better since they share less core resources. Hence when we
7226 * have idle threads, we want them to be the higher ones.
7228 * This packing function is run on idle threads. It checks to see if
7229 * the busiest CPU in this domain (core in the P7 case) has a higher
7230 * CPU number than the packing function is being run on. Here we are
7231 * assuming lower CPU number will be equivalent to lower a SMT thread
7234 * Return: 1 when packing is required and a task should be moved to
7235 * this CPU. The amount of the imbalance is returned in *imbalance.
7237 * @env: The load balancing environment.
7238 * @sds: Statistics of the sched_domain which is to be packed
7240 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7244 if (!(env->sd->flags & SD_ASYM_PACKING))
7247 if (env->idle == CPU_NOT_IDLE)
7253 busiest_cpu = group_first_cpu(sds->busiest);
7254 if (env->dst_cpu > busiest_cpu)
7257 env->imbalance = DIV_ROUND_CLOSEST(
7258 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7259 SCHED_CAPACITY_SCALE);
7265 * fix_small_imbalance - Calculate the minor imbalance that exists
7266 * amongst the groups of a sched_domain, during
7268 * @env: The load balancing environment.
7269 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7272 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7274 unsigned long tmp, capa_now = 0, capa_move = 0;
7275 unsigned int imbn = 2;
7276 unsigned long scaled_busy_load_per_task;
7277 struct sg_lb_stats *local, *busiest;
7279 local = &sds->local_stat;
7280 busiest = &sds->busiest_stat;
7282 if (!local->sum_nr_running)
7283 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7284 else if (busiest->load_per_task > local->load_per_task)
7287 scaled_busy_load_per_task =
7288 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7289 busiest->group_capacity;
7291 if (busiest->avg_load + scaled_busy_load_per_task >=
7292 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7293 env->imbalance = busiest->load_per_task;
7298 * OK, we don't have enough imbalance to justify moving tasks,
7299 * however we may be able to increase total CPU capacity used by
7303 capa_now += busiest->group_capacity *
7304 min(busiest->load_per_task, busiest->avg_load);
7305 capa_now += local->group_capacity *
7306 min(local->load_per_task, local->avg_load);
7307 capa_now /= SCHED_CAPACITY_SCALE;
7309 /* Amount of load we'd subtract */
7310 if (busiest->avg_load > scaled_busy_load_per_task) {
7311 capa_move += busiest->group_capacity *
7312 min(busiest->load_per_task,
7313 busiest->avg_load - scaled_busy_load_per_task);
7316 /* Amount of load we'd add */
7317 if (busiest->avg_load * busiest->group_capacity <
7318 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7319 tmp = (busiest->avg_load * busiest->group_capacity) /
7320 local->group_capacity;
7322 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7323 local->group_capacity;
7325 capa_move += local->group_capacity *
7326 min(local->load_per_task, local->avg_load + tmp);
7327 capa_move /= SCHED_CAPACITY_SCALE;
7329 /* Move if we gain throughput */
7330 if (capa_move > capa_now)
7331 env->imbalance = busiest->load_per_task;
7335 * calculate_imbalance - Calculate the amount of imbalance present within the
7336 * groups of a given sched_domain during load balance.
7337 * @env: load balance environment
7338 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7340 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7342 unsigned long max_pull, load_above_capacity = ~0UL;
7343 struct sg_lb_stats *local, *busiest;
7345 local = &sds->local_stat;
7346 busiest = &sds->busiest_stat;
7348 if (busiest->group_type == group_imbalanced) {
7350 * In the group_imb case we cannot rely on group-wide averages
7351 * to ensure cpu-load equilibrium, look at wider averages. XXX
7353 busiest->load_per_task =
7354 min(busiest->load_per_task, sds->avg_load);
7358 * Avg load of busiest sg can be less and avg load of local sg can
7359 * be greater than avg load across all sgs of sd because avg load
7360 * factors in sg capacity and sgs with smaller group_type are
7361 * skipped when updating the busiest sg:
7363 if (busiest->avg_load <= sds->avg_load ||
7364 local->avg_load >= sds->avg_load) {
7366 return fix_small_imbalance(env, sds);
7370 * If there aren't any idle cpus, avoid creating some.
7372 if (busiest->group_type == group_overloaded &&
7373 local->group_type == group_overloaded) {
7374 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7375 if (load_above_capacity > busiest->group_capacity) {
7376 load_above_capacity -= busiest->group_capacity;
7377 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7378 load_above_capacity /= busiest->group_capacity;
7380 load_above_capacity = ~0UL;
7384 * We're trying to get all the cpus to the average_load, so we don't
7385 * want to push ourselves above the average load, nor do we wish to
7386 * reduce the max loaded cpu below the average load. At the same time,
7387 * we also don't want to reduce the group load below the group
7388 * capacity. Thus we look for the minimum possible imbalance.
7390 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7392 /* How much load to actually move to equalise the imbalance */
7393 env->imbalance = min(
7394 max_pull * busiest->group_capacity,
7395 (sds->avg_load - local->avg_load) * local->group_capacity
7396 ) / SCHED_CAPACITY_SCALE;
7399 * if *imbalance is less than the average load per runnable task
7400 * there is no guarantee that any tasks will be moved so we'll have
7401 * a think about bumping its value to force at least one task to be
7404 if (env->imbalance < busiest->load_per_task)
7405 return fix_small_imbalance(env, sds);
7408 /******* find_busiest_group() helpers end here *********************/
7411 * find_busiest_group - Returns the busiest group within the sched_domain
7412 * if there is an imbalance.
7414 * Also calculates the amount of weighted load which should be moved
7415 * to restore balance.
7417 * @env: The load balancing environment.
7419 * Return: - The busiest group if imbalance exists.
7421 static struct sched_group *find_busiest_group(struct lb_env *env)
7423 struct sg_lb_stats *local, *busiest;
7424 struct sd_lb_stats sds;
7426 init_sd_lb_stats(&sds);
7429 * Compute the various statistics relavent for load balancing at
7432 update_sd_lb_stats(env, &sds);
7433 local = &sds.local_stat;
7434 busiest = &sds.busiest_stat;
7436 /* ASYM feature bypasses nice load balance check */
7437 if (check_asym_packing(env, &sds))
7440 /* There is no busy sibling group to pull tasks from */
7441 if (!sds.busiest || busiest->sum_nr_running == 0)
7444 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7445 / sds.total_capacity;
7448 * If the busiest group is imbalanced the below checks don't
7449 * work because they assume all things are equal, which typically
7450 * isn't true due to cpus_allowed constraints and the like.
7452 if (busiest->group_type == group_imbalanced)
7455 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7456 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7457 busiest->group_no_capacity)
7461 * If the local group is busier than the selected busiest group
7462 * don't try and pull any tasks.
7464 if (local->avg_load >= busiest->avg_load)
7468 * Don't pull any tasks if this group is already above the domain
7471 if (local->avg_load >= sds.avg_load)
7474 if (env->idle == CPU_IDLE) {
7476 * This cpu is idle. If the busiest group is not overloaded
7477 * and there is no imbalance between this and busiest group
7478 * wrt idle cpus, it is balanced. The imbalance becomes
7479 * significant if the diff is greater than 1 otherwise we
7480 * might end up to just move the imbalance on another group
7482 if ((busiest->group_type != group_overloaded) &&
7483 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7487 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7488 * imbalance_pct to be conservative.
7490 if (100 * busiest->avg_load <=
7491 env->sd->imbalance_pct * local->avg_load)
7496 /* Looks like there is an imbalance. Compute it */
7497 calculate_imbalance(env, &sds);
7506 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7508 static struct rq *find_busiest_queue(struct lb_env *env,
7509 struct sched_group *group)
7511 struct rq *busiest = NULL, *rq;
7512 unsigned long busiest_load = 0, busiest_capacity = 1;
7515 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7516 unsigned long capacity, wl;
7520 rt = fbq_classify_rq(rq);
7523 * We classify groups/runqueues into three groups:
7524 * - regular: there are !numa tasks
7525 * - remote: there are numa tasks that run on the 'wrong' node
7526 * - all: there is no distinction
7528 * In order to avoid migrating ideally placed numa tasks,
7529 * ignore those when there's better options.
7531 * If we ignore the actual busiest queue to migrate another
7532 * task, the next balance pass can still reduce the busiest
7533 * queue by moving tasks around inside the node.
7535 * If we cannot move enough load due to this classification
7536 * the next pass will adjust the group classification and
7537 * allow migration of more tasks.
7539 * Both cases only affect the total convergence complexity.
7541 if (rt > env->fbq_type)
7544 capacity = capacity_of(i);
7546 wl = weighted_cpuload(i);
7549 * When comparing with imbalance, use weighted_cpuload()
7550 * which is not scaled with the cpu capacity.
7553 if (rq->nr_running == 1 && wl > env->imbalance &&
7554 !check_cpu_capacity(rq, env->sd))
7558 * For the load comparisons with the other cpu's, consider
7559 * the weighted_cpuload() scaled with the cpu capacity, so
7560 * that the load can be moved away from the cpu that is
7561 * potentially running at a lower capacity.
7563 * Thus we're looking for max(wl_i / capacity_i), crosswise
7564 * multiplication to rid ourselves of the division works out
7565 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7566 * our previous maximum.
7568 if (wl * busiest_capacity > busiest_load * capacity) {
7570 busiest_capacity = capacity;
7579 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7580 * so long as it is large enough.
7582 #define MAX_PINNED_INTERVAL 512
7584 static int need_active_balance(struct lb_env *env)
7586 struct sched_domain *sd = env->sd;
7588 if (env->idle == CPU_NEWLY_IDLE) {
7591 * ASYM_PACKING needs to force migrate tasks from busy but
7592 * higher numbered CPUs in order to pack all tasks in the
7593 * lowest numbered CPUs.
7595 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7600 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7601 * It's worth migrating the task if the src_cpu's capacity is reduced
7602 * because of other sched_class or IRQs if more capacity stays
7603 * available on dst_cpu.
7605 if ((env->idle != CPU_NOT_IDLE) &&
7606 (env->src_rq->cfs.h_nr_running == 1)) {
7607 if ((check_cpu_capacity(env->src_rq, sd)) &&
7608 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7612 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7615 static int active_load_balance_cpu_stop(void *data);
7617 static int should_we_balance(struct lb_env *env)
7619 struct sched_group *sg = env->sd->groups;
7620 struct cpumask *sg_cpus, *sg_mask;
7621 int cpu, balance_cpu = -1;
7624 * In the newly idle case, we will allow all the cpu's
7625 * to do the newly idle load balance.
7627 if (env->idle == CPU_NEWLY_IDLE)
7630 sg_cpus = sched_group_cpus(sg);
7631 sg_mask = sched_group_mask(sg);
7632 /* Try to find first idle cpu */
7633 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7634 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7641 if (balance_cpu == -1)
7642 balance_cpu = group_balance_cpu(sg);
7645 * First idle cpu or the first cpu(busiest) in this sched group
7646 * is eligible for doing load balancing at this and above domains.
7648 return balance_cpu == env->dst_cpu;
7652 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7653 * tasks if there is an imbalance.
7655 static int load_balance(int this_cpu, struct rq *this_rq,
7656 struct sched_domain *sd, enum cpu_idle_type idle,
7657 int *continue_balancing)
7659 int ld_moved, cur_ld_moved, active_balance = 0;
7660 struct sched_domain *sd_parent = sd->parent;
7661 struct sched_group *group;
7663 unsigned long flags;
7664 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7666 struct lb_env env = {
7668 .dst_cpu = this_cpu,
7670 .dst_grpmask = sched_group_cpus(sd->groups),
7672 .loop_break = sched_nr_migrate_break,
7675 .tasks = LIST_HEAD_INIT(env.tasks),
7679 * For NEWLY_IDLE load_balancing, we don't need to consider
7680 * other cpus in our group
7682 if (idle == CPU_NEWLY_IDLE)
7683 env.dst_grpmask = NULL;
7685 cpumask_copy(cpus, cpu_active_mask);
7687 schedstat_inc(sd->lb_count[idle]);
7690 if (!should_we_balance(&env)) {
7691 *continue_balancing = 0;
7695 group = find_busiest_group(&env);
7697 schedstat_inc(sd->lb_nobusyg[idle]);
7701 busiest = find_busiest_queue(&env, group);
7703 schedstat_inc(sd->lb_nobusyq[idle]);
7707 BUG_ON(busiest == env.dst_rq);
7709 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
7711 env.src_cpu = busiest->cpu;
7712 env.src_rq = busiest;
7715 if (busiest->nr_running > 1) {
7717 * Attempt to move tasks. If find_busiest_group has found
7718 * an imbalance but busiest->nr_running <= 1, the group is
7719 * still unbalanced. ld_moved simply stays zero, so it is
7720 * correctly treated as an imbalance.
7722 env.flags |= LBF_ALL_PINNED;
7723 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7726 raw_spin_lock_irqsave(&busiest->lock, flags);
7729 * cur_ld_moved - load moved in current iteration
7730 * ld_moved - cumulative load moved across iterations
7732 cur_ld_moved = detach_tasks(&env);
7735 * We've detached some tasks from busiest_rq. Every
7736 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7737 * unlock busiest->lock, and we are able to be sure
7738 * that nobody can manipulate the tasks in parallel.
7739 * See task_rq_lock() family for the details.
7742 raw_spin_unlock(&busiest->lock);
7746 ld_moved += cur_ld_moved;
7749 local_irq_restore(flags);
7751 if (env.flags & LBF_NEED_BREAK) {
7752 env.flags &= ~LBF_NEED_BREAK;
7757 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7758 * us and move them to an alternate dst_cpu in our sched_group
7759 * where they can run. The upper limit on how many times we
7760 * iterate on same src_cpu is dependent on number of cpus in our
7763 * This changes load balance semantics a bit on who can move
7764 * load to a given_cpu. In addition to the given_cpu itself
7765 * (or a ilb_cpu acting on its behalf where given_cpu is
7766 * nohz-idle), we now have balance_cpu in a position to move
7767 * load to given_cpu. In rare situations, this may cause
7768 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7769 * _independently_ and at _same_ time to move some load to
7770 * given_cpu) causing exceess load to be moved to given_cpu.
7771 * This however should not happen so much in practice and
7772 * moreover subsequent load balance cycles should correct the
7773 * excess load moved.
7775 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7777 /* Prevent to re-select dst_cpu via env's cpus */
7778 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7780 env.dst_rq = cpu_rq(env.new_dst_cpu);
7781 env.dst_cpu = env.new_dst_cpu;
7782 env.flags &= ~LBF_DST_PINNED;
7784 env.loop_break = sched_nr_migrate_break;
7787 * Go back to "more_balance" rather than "redo" since we
7788 * need to continue with same src_cpu.
7794 * We failed to reach balance because of affinity.
7797 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7799 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7800 *group_imbalance = 1;
7803 /* All tasks on this runqueue were pinned by CPU affinity */
7804 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7805 cpumask_clear_cpu(cpu_of(busiest), cpus);
7806 if (!cpumask_empty(cpus)) {
7808 env.loop_break = sched_nr_migrate_break;
7811 goto out_all_pinned;
7816 schedstat_inc(sd->lb_failed[idle]);
7818 * Increment the failure counter only on periodic balance.
7819 * We do not want newidle balance, which can be very
7820 * frequent, pollute the failure counter causing
7821 * excessive cache_hot migrations and active balances.
7823 if (idle != CPU_NEWLY_IDLE)
7824 sd->nr_balance_failed++;
7826 if (need_active_balance(&env)) {
7827 raw_spin_lock_irqsave(&busiest->lock, flags);
7829 /* don't kick the active_load_balance_cpu_stop,
7830 * if the curr task on busiest cpu can't be
7833 if (!cpumask_test_cpu(this_cpu,
7834 tsk_cpus_allowed(busiest->curr))) {
7835 raw_spin_unlock_irqrestore(&busiest->lock,
7837 env.flags |= LBF_ALL_PINNED;
7838 goto out_one_pinned;
7842 * ->active_balance synchronizes accesses to
7843 * ->active_balance_work. Once set, it's cleared
7844 * only after active load balance is finished.
7846 if (!busiest->active_balance) {
7847 busiest->active_balance = 1;
7848 busiest->push_cpu = this_cpu;
7851 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7853 if (active_balance) {
7854 stop_one_cpu_nowait(cpu_of(busiest),
7855 active_load_balance_cpu_stop, busiest,
7856 &busiest->active_balance_work);
7859 /* We've kicked active balancing, force task migration. */
7860 sd->nr_balance_failed = sd->cache_nice_tries+1;
7863 sd->nr_balance_failed = 0;
7865 if (likely(!active_balance)) {
7866 /* We were unbalanced, so reset the balancing interval */
7867 sd->balance_interval = sd->min_interval;
7870 * If we've begun active balancing, start to back off. This
7871 * case may not be covered by the all_pinned logic if there
7872 * is only 1 task on the busy runqueue (because we don't call
7875 if (sd->balance_interval < sd->max_interval)
7876 sd->balance_interval *= 2;
7883 * We reach balance although we may have faced some affinity
7884 * constraints. Clear the imbalance flag if it was set.
7887 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7889 if (*group_imbalance)
7890 *group_imbalance = 0;
7895 * We reach balance because all tasks are pinned at this level so
7896 * we can't migrate them. Let the imbalance flag set so parent level
7897 * can try to migrate them.
7899 schedstat_inc(sd->lb_balanced[idle]);
7901 sd->nr_balance_failed = 0;
7904 /* tune up the balancing interval */
7905 if (((env.flags & LBF_ALL_PINNED) &&
7906 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7907 (sd->balance_interval < sd->max_interval))
7908 sd->balance_interval *= 2;
7915 static inline unsigned long
7916 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7918 unsigned long interval = sd->balance_interval;
7921 interval *= sd->busy_factor;
7923 /* scale ms to jiffies */
7924 interval = msecs_to_jiffies(interval);
7925 interval = clamp(interval, 1UL, max_load_balance_interval);
7931 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
7933 unsigned long interval, next;
7935 /* used by idle balance, so cpu_busy = 0 */
7936 interval = get_sd_balance_interval(sd, 0);
7937 next = sd->last_balance + interval;
7939 if (time_after(*next_balance, next))
7940 *next_balance = next;
7944 * idle_balance is called by schedule() if this_cpu is about to become
7945 * idle. Attempts to pull tasks from other CPUs.
7947 static int idle_balance(struct rq *this_rq)
7949 unsigned long next_balance = jiffies + HZ;
7950 int this_cpu = this_rq->cpu;
7951 struct sched_domain *sd;
7952 int pulled_task = 0;
7956 * We must set idle_stamp _before_ calling idle_balance(), such that we
7957 * measure the duration of idle_balance() as idle time.
7959 this_rq->idle_stamp = rq_clock(this_rq);
7961 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7962 !this_rq->rd->overload) {
7964 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7966 update_next_balance(sd, &next_balance);
7972 raw_spin_unlock(&this_rq->lock);
7974 update_blocked_averages(this_cpu);
7976 for_each_domain(this_cpu, sd) {
7977 int continue_balancing = 1;
7978 u64 t0, domain_cost;
7980 if (!(sd->flags & SD_LOAD_BALANCE))
7983 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7984 update_next_balance(sd, &next_balance);
7988 if (sd->flags & SD_BALANCE_NEWIDLE) {
7989 t0 = sched_clock_cpu(this_cpu);
7991 pulled_task = load_balance(this_cpu, this_rq,
7993 &continue_balancing);
7995 domain_cost = sched_clock_cpu(this_cpu) - t0;
7996 if (domain_cost > sd->max_newidle_lb_cost)
7997 sd->max_newidle_lb_cost = domain_cost;
7999 curr_cost += domain_cost;
8002 update_next_balance(sd, &next_balance);
8005 * Stop searching for tasks to pull if there are
8006 * now runnable tasks on this rq.
8008 if (pulled_task || this_rq->nr_running > 0)
8013 raw_spin_lock(&this_rq->lock);
8015 if (curr_cost > this_rq->max_idle_balance_cost)
8016 this_rq->max_idle_balance_cost = curr_cost;
8019 * While browsing the domains, we released the rq lock, a task could
8020 * have been enqueued in the meantime. Since we're not going idle,
8021 * pretend we pulled a task.
8023 if (this_rq->cfs.h_nr_running && !pulled_task)
8027 /* Move the next balance forward */
8028 if (time_after(this_rq->next_balance, next_balance))
8029 this_rq->next_balance = next_balance;
8031 /* Is there a task of a high priority class? */
8032 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8036 this_rq->idle_stamp = 0;
8042 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8043 * running tasks off the busiest CPU onto idle CPUs. It requires at
8044 * least 1 task to be running on each physical CPU where possible, and
8045 * avoids physical / logical imbalances.
8047 static int active_load_balance_cpu_stop(void *data)
8049 struct rq *busiest_rq = data;
8050 int busiest_cpu = cpu_of(busiest_rq);
8051 int target_cpu = busiest_rq->push_cpu;
8052 struct rq *target_rq = cpu_rq(target_cpu);
8053 struct sched_domain *sd;
8054 struct task_struct *p = NULL;
8056 raw_spin_lock_irq(&busiest_rq->lock);
8058 /* make sure the requested cpu hasn't gone down in the meantime */
8059 if (unlikely(busiest_cpu != smp_processor_id() ||
8060 !busiest_rq->active_balance))
8063 /* Is there any task to move? */
8064 if (busiest_rq->nr_running <= 1)
8068 * This condition is "impossible", if it occurs
8069 * we need to fix it. Originally reported by
8070 * Bjorn Helgaas on a 128-cpu setup.
8072 BUG_ON(busiest_rq == target_rq);
8074 /* Search for an sd spanning us and the target CPU. */
8076 for_each_domain(target_cpu, sd) {
8077 if ((sd->flags & SD_LOAD_BALANCE) &&
8078 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8083 struct lb_env env = {
8085 .dst_cpu = target_cpu,
8086 .dst_rq = target_rq,
8087 .src_cpu = busiest_rq->cpu,
8088 .src_rq = busiest_rq,
8092 schedstat_inc(sd->alb_count);
8094 p = detach_one_task(&env);
8096 schedstat_inc(sd->alb_pushed);
8097 /* Active balancing done, reset the failure counter. */
8098 sd->nr_balance_failed = 0;
8100 schedstat_inc(sd->alb_failed);
8105 busiest_rq->active_balance = 0;
8106 raw_spin_unlock(&busiest_rq->lock);
8109 attach_one_task(target_rq, p);
8116 static inline int on_null_domain(struct rq *rq)
8118 return unlikely(!rcu_dereference_sched(rq->sd));
8121 #ifdef CONFIG_NO_HZ_COMMON
8123 * idle load balancing details
8124 * - When one of the busy CPUs notice that there may be an idle rebalancing
8125 * needed, they will kick the idle load balancer, which then does idle
8126 * load balancing for all the idle CPUs.
8129 cpumask_var_t idle_cpus_mask;
8131 unsigned long next_balance; /* in jiffy units */
8132 } nohz ____cacheline_aligned;
8134 static inline int find_new_ilb(void)
8136 int ilb = cpumask_first(nohz.idle_cpus_mask);
8138 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8145 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8146 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8147 * CPU (if there is one).
8149 static void nohz_balancer_kick(void)
8153 nohz.next_balance++;
8155 ilb_cpu = find_new_ilb();
8157 if (ilb_cpu >= nr_cpu_ids)
8160 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8163 * Use smp_send_reschedule() instead of resched_cpu().
8164 * This way we generate a sched IPI on the target cpu which
8165 * is idle. And the softirq performing nohz idle load balance
8166 * will be run before returning from the IPI.
8168 smp_send_reschedule(ilb_cpu);
8172 void nohz_balance_exit_idle(unsigned int cpu)
8174 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8176 * Completely isolated CPUs don't ever set, so we must test.
8178 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8179 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8180 atomic_dec(&nohz.nr_cpus);
8182 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8186 static inline void set_cpu_sd_state_busy(void)
8188 struct sched_domain *sd;
8189 int cpu = smp_processor_id();
8192 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8194 if (!sd || !sd->nohz_idle)
8198 atomic_inc(&sd->shared->nr_busy_cpus);
8203 void set_cpu_sd_state_idle(void)
8205 struct sched_domain *sd;
8206 int cpu = smp_processor_id();
8209 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8211 if (!sd || sd->nohz_idle)
8215 atomic_dec(&sd->shared->nr_busy_cpus);
8221 * This routine will record that the cpu is going idle with tick stopped.
8222 * This info will be used in performing idle load balancing in the future.
8224 void nohz_balance_enter_idle(int cpu)
8227 * If this cpu is going down, then nothing needs to be done.
8229 if (!cpu_active(cpu))
8232 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8236 * If we're a completely isolated CPU, we don't play.
8238 if (on_null_domain(cpu_rq(cpu)))
8241 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8242 atomic_inc(&nohz.nr_cpus);
8243 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8247 static DEFINE_SPINLOCK(balancing);
8250 * Scale the max load_balance interval with the number of CPUs in the system.
8251 * This trades load-balance latency on larger machines for less cross talk.
8253 void update_max_interval(void)
8255 max_load_balance_interval = HZ*num_online_cpus()/10;
8259 * It checks each scheduling domain to see if it is due to be balanced,
8260 * and initiates a balancing operation if so.
8262 * Balancing parameters are set up in init_sched_domains.
8264 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8266 int continue_balancing = 1;
8268 unsigned long interval;
8269 struct sched_domain *sd;
8270 /* Earliest time when we have to do rebalance again */
8271 unsigned long next_balance = jiffies + 60*HZ;
8272 int update_next_balance = 0;
8273 int need_serialize, need_decay = 0;
8276 update_blocked_averages(cpu);
8279 for_each_domain(cpu, sd) {
8281 * Decay the newidle max times here because this is a regular
8282 * visit to all the domains. Decay ~1% per second.
8284 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8285 sd->max_newidle_lb_cost =
8286 (sd->max_newidle_lb_cost * 253) / 256;
8287 sd->next_decay_max_lb_cost = jiffies + HZ;
8290 max_cost += sd->max_newidle_lb_cost;
8292 if (!(sd->flags & SD_LOAD_BALANCE))
8296 * Stop the load balance at this level. There is another
8297 * CPU in our sched group which is doing load balancing more
8300 if (!continue_balancing) {
8306 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8308 need_serialize = sd->flags & SD_SERIALIZE;
8309 if (need_serialize) {
8310 if (!spin_trylock(&balancing))
8314 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8315 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8317 * The LBF_DST_PINNED logic could have changed
8318 * env->dst_cpu, so we can't know our idle
8319 * state even if we migrated tasks. Update it.
8321 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8323 sd->last_balance = jiffies;
8324 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8327 spin_unlock(&balancing);
8329 if (time_after(next_balance, sd->last_balance + interval)) {
8330 next_balance = sd->last_balance + interval;
8331 update_next_balance = 1;
8336 * Ensure the rq-wide value also decays but keep it at a
8337 * reasonable floor to avoid funnies with rq->avg_idle.
8339 rq->max_idle_balance_cost =
8340 max((u64)sysctl_sched_migration_cost, max_cost);
8345 * next_balance will be updated only when there is a need.
8346 * When the cpu is attached to null domain for ex, it will not be
8349 if (likely(update_next_balance)) {
8350 rq->next_balance = next_balance;
8352 #ifdef CONFIG_NO_HZ_COMMON
8354 * If this CPU has been elected to perform the nohz idle
8355 * balance. Other idle CPUs have already rebalanced with
8356 * nohz_idle_balance() and nohz.next_balance has been
8357 * updated accordingly. This CPU is now running the idle load
8358 * balance for itself and we need to update the
8359 * nohz.next_balance accordingly.
8361 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8362 nohz.next_balance = rq->next_balance;
8367 #ifdef CONFIG_NO_HZ_COMMON
8369 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8370 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8372 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8374 int this_cpu = this_rq->cpu;
8377 /* Earliest time when we have to do rebalance again */
8378 unsigned long next_balance = jiffies + 60*HZ;
8379 int update_next_balance = 0;
8381 if (idle != CPU_IDLE ||
8382 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8385 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8386 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8390 * If this cpu gets work to do, stop the load balancing
8391 * work being done for other cpus. Next load
8392 * balancing owner will pick it up.
8397 rq = cpu_rq(balance_cpu);
8400 * If time for next balance is due,
8403 if (time_after_eq(jiffies, rq->next_balance)) {
8404 raw_spin_lock_irq(&rq->lock);
8405 update_rq_clock(rq);
8406 cpu_load_update_idle(rq);
8407 raw_spin_unlock_irq(&rq->lock);
8408 rebalance_domains(rq, CPU_IDLE);
8411 if (time_after(next_balance, rq->next_balance)) {
8412 next_balance = rq->next_balance;
8413 update_next_balance = 1;
8418 * next_balance will be updated only when there is a need.
8419 * When the CPU is attached to null domain for ex, it will not be
8422 if (likely(update_next_balance))
8423 nohz.next_balance = next_balance;
8425 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8429 * Current heuristic for kicking the idle load balancer in the presence
8430 * of an idle cpu in the system.
8431 * - This rq has more than one task.
8432 * - This rq has at least one CFS task and the capacity of the CPU is
8433 * significantly reduced because of RT tasks or IRQs.
8434 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8435 * multiple busy cpu.
8436 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8437 * domain span are idle.
8439 static inline bool nohz_kick_needed(struct rq *rq)
8441 unsigned long now = jiffies;
8442 struct sched_domain_shared *sds;
8443 struct sched_domain *sd;
8444 int nr_busy, cpu = rq->cpu;
8447 if (unlikely(rq->idle_balance))
8451 * We may be recently in ticked or tickless idle mode. At the first
8452 * busy tick after returning from idle, we will update the busy stats.
8454 set_cpu_sd_state_busy();
8455 nohz_balance_exit_idle(cpu);
8458 * None are in tickless mode and hence no need for NOHZ idle load
8461 if (likely(!atomic_read(&nohz.nr_cpus)))
8464 if (time_before(now, nohz.next_balance))
8467 if (rq->nr_running >= 2)
8471 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8474 * XXX: write a coherent comment on why we do this.
8475 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8477 nr_busy = atomic_read(&sds->nr_busy_cpus);
8485 sd = rcu_dereference(rq->sd);
8487 if ((rq->cfs.h_nr_running >= 1) &&
8488 check_cpu_capacity(rq, sd)) {
8494 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8495 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8496 sched_domain_span(sd)) < cpu)) {
8506 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8510 * run_rebalance_domains is triggered when needed from the scheduler tick.
8511 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8513 static void run_rebalance_domains(struct softirq_action *h)
8515 struct rq *this_rq = this_rq();
8516 enum cpu_idle_type idle = this_rq->idle_balance ?
8517 CPU_IDLE : CPU_NOT_IDLE;
8520 * If this cpu has a pending nohz_balance_kick, then do the
8521 * balancing on behalf of the other idle cpus whose ticks are
8522 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8523 * give the idle cpus a chance to load balance. Else we may
8524 * load balance only within the local sched_domain hierarchy
8525 * and abort nohz_idle_balance altogether if we pull some load.
8527 nohz_idle_balance(this_rq, idle);
8528 rebalance_domains(this_rq, idle);
8532 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8534 void trigger_load_balance(struct rq *rq)
8536 /* Don't need to rebalance while attached to NULL domain */
8537 if (unlikely(on_null_domain(rq)))
8540 if (time_after_eq(jiffies, rq->next_balance))
8541 raise_softirq(SCHED_SOFTIRQ);
8542 #ifdef CONFIG_NO_HZ_COMMON
8543 if (nohz_kick_needed(rq))
8544 nohz_balancer_kick();
8548 static void rq_online_fair(struct rq *rq)
8552 update_runtime_enabled(rq);
8555 static void rq_offline_fair(struct rq *rq)
8559 /* Ensure any throttled groups are reachable by pick_next_task */
8560 unthrottle_offline_cfs_rqs(rq);
8563 #endif /* CONFIG_SMP */
8566 * scheduler tick hitting a task of our scheduling class:
8568 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8570 struct cfs_rq *cfs_rq;
8571 struct sched_entity *se = &curr->se;
8573 for_each_sched_entity(se) {
8574 cfs_rq = cfs_rq_of(se);
8575 entity_tick(cfs_rq, se, queued);
8578 if (static_branch_unlikely(&sched_numa_balancing))
8579 task_tick_numa(rq, curr);
8583 * called on fork with the child task as argument from the parent's context
8584 * - child not yet on the tasklist
8585 * - preemption disabled
8587 static void task_fork_fair(struct task_struct *p)
8589 struct cfs_rq *cfs_rq;
8590 struct sched_entity *se = &p->se, *curr;
8591 struct rq *rq = this_rq();
8593 raw_spin_lock(&rq->lock);
8594 update_rq_clock(rq);
8596 cfs_rq = task_cfs_rq(current);
8597 curr = cfs_rq->curr;
8599 update_curr(cfs_rq);
8600 se->vruntime = curr->vruntime;
8602 place_entity(cfs_rq, se, 1);
8604 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8606 * Upon rescheduling, sched_class::put_prev_task() will place
8607 * 'current' within the tree based on its new key value.
8609 swap(curr->vruntime, se->vruntime);
8613 se->vruntime -= cfs_rq->min_vruntime;
8614 raw_spin_unlock(&rq->lock);
8618 * Priority of the task has changed. Check to see if we preempt
8622 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8624 if (!task_on_rq_queued(p))
8628 * Reschedule if we are currently running on this runqueue and
8629 * our priority decreased, or if we are not currently running on
8630 * this runqueue and our priority is higher than the current's
8632 if (rq->curr == p) {
8633 if (p->prio > oldprio)
8636 check_preempt_curr(rq, p, 0);
8639 static inline bool vruntime_normalized(struct task_struct *p)
8641 struct sched_entity *se = &p->se;
8644 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8645 * the dequeue_entity(.flags=0) will already have normalized the
8652 * When !on_rq, vruntime of the task has usually NOT been normalized.
8653 * But there are some cases where it has already been normalized:
8655 * - A forked child which is waiting for being woken up by
8656 * wake_up_new_task().
8657 * - A task which has been woken up by try_to_wake_up() and
8658 * waiting for actually being woken up by sched_ttwu_pending().
8660 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8666 static void detach_task_cfs_rq(struct task_struct *p)
8668 struct sched_entity *se = &p->se;
8669 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8670 u64 now = cfs_rq_clock_task(cfs_rq);
8672 if (!vruntime_normalized(p)) {
8674 * Fix up our vruntime so that the current sleep doesn't
8675 * cause 'unlimited' sleep bonus.
8677 place_entity(cfs_rq, se, 0);
8678 se->vruntime -= cfs_rq->min_vruntime;
8681 /* Catch up with the cfs_rq and remove our load when we leave */
8682 update_cfs_rq_load_avg(now, cfs_rq, false);
8683 detach_entity_load_avg(cfs_rq, se);
8684 update_tg_load_avg(cfs_rq, false);
8687 static void attach_task_cfs_rq(struct task_struct *p)
8689 struct sched_entity *se = &p->se;
8690 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8691 u64 now = cfs_rq_clock_task(cfs_rq);
8693 #ifdef CONFIG_FAIR_GROUP_SCHED
8695 * Since the real-depth could have been changed (only FAIR
8696 * class maintain depth value), reset depth properly.
8698 se->depth = se->parent ? se->parent->depth + 1 : 0;
8701 /* Synchronize task with its cfs_rq */
8702 update_cfs_rq_load_avg(now, cfs_rq, false);
8703 attach_entity_load_avg(cfs_rq, se);
8704 update_tg_load_avg(cfs_rq, false);
8706 if (!vruntime_normalized(p))
8707 se->vruntime += cfs_rq->min_vruntime;
8710 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8712 detach_task_cfs_rq(p);
8715 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8717 attach_task_cfs_rq(p);
8719 if (task_on_rq_queued(p)) {
8721 * We were most likely switched from sched_rt, so
8722 * kick off the schedule if running, otherwise just see
8723 * if we can still preempt the current task.
8728 check_preempt_curr(rq, p, 0);
8732 /* Account for a task changing its policy or group.
8734 * This routine is mostly called to set cfs_rq->curr field when a task
8735 * migrates between groups/classes.
8737 static void set_curr_task_fair(struct rq *rq)
8739 struct sched_entity *se = &rq->curr->se;
8741 for_each_sched_entity(se) {
8742 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8744 set_next_entity(cfs_rq, se);
8745 /* ensure bandwidth has been allocated on our new cfs_rq */
8746 account_cfs_rq_runtime(cfs_rq, 0);
8750 void init_cfs_rq(struct cfs_rq *cfs_rq)
8752 cfs_rq->tasks_timeline = RB_ROOT;
8753 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8754 #ifndef CONFIG_64BIT
8755 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8758 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8759 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8763 #ifdef CONFIG_FAIR_GROUP_SCHED
8764 static void task_set_group_fair(struct task_struct *p)
8766 struct sched_entity *se = &p->se;
8768 set_task_rq(p, task_cpu(p));
8769 se->depth = se->parent ? se->parent->depth + 1 : 0;
8772 static void task_move_group_fair(struct task_struct *p)
8774 detach_task_cfs_rq(p);
8775 set_task_rq(p, task_cpu(p));
8778 /* Tell se's cfs_rq has been changed -- migrated */
8779 p->se.avg.last_update_time = 0;
8781 attach_task_cfs_rq(p);
8784 static void task_change_group_fair(struct task_struct *p, int type)
8787 case TASK_SET_GROUP:
8788 task_set_group_fair(p);
8791 case TASK_MOVE_GROUP:
8792 task_move_group_fair(p);
8797 void free_fair_sched_group(struct task_group *tg)
8801 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8803 for_each_possible_cpu(i) {
8805 kfree(tg->cfs_rq[i]);
8814 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8816 struct sched_entity *se;
8817 struct cfs_rq *cfs_rq;
8821 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8824 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8828 tg->shares = NICE_0_LOAD;
8830 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8832 for_each_possible_cpu(i) {
8835 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8836 GFP_KERNEL, cpu_to_node(i));
8840 se = kzalloc_node(sizeof(struct sched_entity),
8841 GFP_KERNEL, cpu_to_node(i));
8845 init_cfs_rq(cfs_rq);
8846 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8847 init_entity_runnable_average(se);
8858 void online_fair_sched_group(struct task_group *tg)
8860 struct sched_entity *se;
8864 for_each_possible_cpu(i) {
8868 raw_spin_lock_irq(&rq->lock);
8869 post_init_entity_util_avg(se);
8870 sync_throttle(tg, i);
8871 raw_spin_unlock_irq(&rq->lock);
8875 void unregister_fair_sched_group(struct task_group *tg)
8877 unsigned long flags;
8881 for_each_possible_cpu(cpu) {
8883 remove_entity_load_avg(tg->se[cpu]);
8886 * Only empty task groups can be destroyed; so we can speculatively
8887 * check on_list without danger of it being re-added.
8889 if (!tg->cfs_rq[cpu]->on_list)
8894 raw_spin_lock_irqsave(&rq->lock, flags);
8895 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8896 raw_spin_unlock_irqrestore(&rq->lock, flags);
8900 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8901 struct sched_entity *se, int cpu,
8902 struct sched_entity *parent)
8904 struct rq *rq = cpu_rq(cpu);
8908 init_cfs_rq_runtime(cfs_rq);
8910 tg->cfs_rq[cpu] = cfs_rq;
8913 /* se could be NULL for root_task_group */
8918 se->cfs_rq = &rq->cfs;
8921 se->cfs_rq = parent->my_q;
8922 se->depth = parent->depth + 1;
8926 /* guarantee group entities always have weight */
8927 update_load_set(&se->load, NICE_0_LOAD);
8928 se->parent = parent;
8931 static DEFINE_MUTEX(shares_mutex);
8933 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8936 unsigned long flags;
8939 * We can't change the weight of the root cgroup.
8944 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8946 mutex_lock(&shares_mutex);
8947 if (tg->shares == shares)
8950 tg->shares = shares;
8951 for_each_possible_cpu(i) {
8952 struct rq *rq = cpu_rq(i);
8953 struct sched_entity *se;
8956 /* Propagate contribution to hierarchy */
8957 raw_spin_lock_irqsave(&rq->lock, flags);
8959 /* Possible calls to update_curr() need rq clock */
8960 update_rq_clock(rq);
8961 for_each_sched_entity(se)
8962 update_cfs_shares(group_cfs_rq(se));
8963 raw_spin_unlock_irqrestore(&rq->lock, flags);
8967 mutex_unlock(&shares_mutex);
8970 #else /* CONFIG_FAIR_GROUP_SCHED */
8972 void free_fair_sched_group(struct task_group *tg) { }
8974 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8979 void online_fair_sched_group(struct task_group *tg) { }
8981 void unregister_fair_sched_group(struct task_group *tg) { }
8983 #endif /* CONFIG_FAIR_GROUP_SCHED */
8986 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8988 struct sched_entity *se = &task->se;
8989 unsigned int rr_interval = 0;
8992 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8995 if (rq->cfs.load.weight)
8996 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9002 * All the scheduling class methods:
9004 const struct sched_class fair_sched_class = {
9005 .next = &idle_sched_class,
9006 .enqueue_task = enqueue_task_fair,
9007 .dequeue_task = dequeue_task_fair,
9008 .yield_task = yield_task_fair,
9009 .yield_to_task = yield_to_task_fair,
9011 .check_preempt_curr = check_preempt_wakeup,
9013 .pick_next_task = pick_next_task_fair,
9014 .put_prev_task = put_prev_task_fair,
9017 .select_task_rq = select_task_rq_fair,
9018 .migrate_task_rq = migrate_task_rq_fair,
9020 .rq_online = rq_online_fair,
9021 .rq_offline = rq_offline_fair,
9023 .task_dead = task_dead_fair,
9024 .set_cpus_allowed = set_cpus_allowed_common,
9027 .set_curr_task = set_curr_task_fair,
9028 .task_tick = task_tick_fair,
9029 .task_fork = task_fork_fair,
9031 .prio_changed = prio_changed_fair,
9032 .switched_from = switched_from_fair,
9033 .switched_to = switched_to_fair,
9035 .get_rr_interval = get_rr_interval_fair,
9037 .update_curr = update_curr_fair,
9039 #ifdef CONFIG_FAIR_GROUP_SCHED
9040 .task_change_group = task_change_group_fair,
9044 #ifdef CONFIG_SCHED_DEBUG
9045 void print_cfs_stats(struct seq_file *m, int cpu)
9047 struct cfs_rq *cfs_rq;
9050 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9051 print_cfs_rq(m, cpu, cfs_rq);
9055 #ifdef CONFIG_NUMA_BALANCING
9056 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9059 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9061 for_each_online_node(node) {
9062 if (p->numa_faults) {
9063 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9064 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9066 if (p->numa_group) {
9067 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9068 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9070 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9073 #endif /* CONFIG_NUMA_BALANCING */
9074 #endif /* CONFIG_SCHED_DEBUG */
9076 __init void init_sched_fair_class(void)
9079 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9081 #ifdef CONFIG_NO_HZ_COMMON
9082 nohz.next_balance = jiffies;
9083 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);