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;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 * At this point, util_avg won't be used in select_task_rq_fair anyway
690 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
693 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
694 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
695 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
696 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
699 * With new tasks being created, their initial util_avgs are extrapolated
700 * based on the cfs_rq's current util_avg:
702 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
704 * However, in many cases, the above util_avg does not give a desired
705 * value. Moreover, the sum of the util_avgs may be divergent, such
706 * as when the series is a harmonic series.
708 * To solve this problem, we also cap the util_avg of successive tasks to
709 * only 1/2 of the left utilization budget:
711 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
713 * where n denotes the nth task.
715 * For example, a simplest series from the beginning would be like:
717 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
718 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
720 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
721 * if util_avg > util_avg_cap.
723 void post_init_entity_util_avg(struct sched_entity *se)
725 struct cfs_rq *cfs_rq = cfs_rq_of(se);
726 struct sched_avg *sa = &se->avg;
727 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
728 u64 now = cfs_rq_clock_task(cfs_rq);
732 if (cfs_rq->avg.util_avg != 0) {
733 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
734 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
736 if (sa->util_avg > cap)
741 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
744 if (entity_is_task(se)) {
745 struct task_struct *p = task_of(se);
746 if (p->sched_class != &fair_sched_class) {
748 * For !fair tasks do:
750 update_cfs_rq_load_avg(now, cfs_rq, false);
751 attach_entity_load_avg(cfs_rq, se);
752 switched_from_fair(rq, p);
754 * such that the next switched_to_fair() has the
757 se->avg.last_update_time = now;
762 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
763 attach_entity_load_avg(cfs_rq, se);
765 update_tg_load_avg(cfs_rq, false);
768 #else /* !CONFIG_SMP */
769 void init_entity_runnable_average(struct sched_entity *se)
772 void post_init_entity_util_avg(struct sched_entity *se)
775 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
778 #endif /* CONFIG_SMP */
781 * Update the current task's runtime statistics.
783 static void update_curr(struct cfs_rq *cfs_rq)
785 struct sched_entity *curr = cfs_rq->curr;
786 u64 now = rq_clock_task(rq_of(cfs_rq));
792 delta_exec = now - curr->exec_start;
793 if (unlikely((s64)delta_exec <= 0))
796 curr->exec_start = now;
798 schedstat_set(curr->statistics.exec_max,
799 max(delta_exec, curr->statistics.exec_max));
801 curr->sum_exec_runtime += delta_exec;
802 schedstat_add(cfs_rq, exec_clock, delta_exec);
804 curr->vruntime += calc_delta_fair(delta_exec, curr);
805 update_min_vruntime(cfs_rq);
807 if (entity_is_task(curr)) {
808 struct task_struct *curtask = task_of(curr);
810 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
811 cpuacct_charge(curtask, delta_exec);
812 account_group_exec_runtime(curtask, delta_exec);
815 account_cfs_rq_runtime(cfs_rq, delta_exec);
818 static void update_curr_fair(struct rq *rq)
820 update_curr(cfs_rq_of(&rq->curr->se));
823 #ifdef CONFIG_SCHEDSTATS
825 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
827 u64 wait_start = rq_clock(rq_of(cfs_rq));
829 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
830 likely(wait_start > se->statistics.wait_start))
831 wait_start -= se->statistics.wait_start;
833 se->statistics.wait_start = wait_start;
837 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
839 struct task_struct *p;
842 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
844 if (entity_is_task(se)) {
846 if (task_on_rq_migrating(p)) {
848 * Preserve migrating task's wait time so wait_start
849 * time stamp can be adjusted to accumulate wait time
850 * prior to migration.
852 se->statistics.wait_start = delta;
855 trace_sched_stat_wait(p, delta);
858 se->statistics.wait_max = max(se->statistics.wait_max, delta);
859 se->statistics.wait_count++;
860 se->statistics.wait_sum += delta;
861 se->statistics.wait_start = 0;
865 * Task is being enqueued - update stats:
868 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
871 * Are we enqueueing a waiting task? (for current tasks
872 * a dequeue/enqueue event is a NOP)
874 if (se != cfs_rq->curr)
875 update_stats_wait_start(cfs_rq, se);
879 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
882 * Mark the end of the wait period if dequeueing a
885 if (se != cfs_rq->curr)
886 update_stats_wait_end(cfs_rq, se);
888 if (flags & DEQUEUE_SLEEP) {
889 if (entity_is_task(se)) {
890 struct task_struct *tsk = task_of(se);
892 if (tsk->state & TASK_INTERRUPTIBLE)
893 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
894 if (tsk->state & TASK_UNINTERRUPTIBLE)
895 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
902 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
907 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
917 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
923 * We are picking a new current task - update its stats:
926 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
929 * We are starting a new run period:
931 se->exec_start = rq_clock_task(rq_of(cfs_rq));
934 /**************************************************
935 * Scheduling class queueing methods:
938 #ifdef CONFIG_NUMA_BALANCING
940 * Approximate time to scan a full NUMA task in ms. The task scan period is
941 * calculated based on the tasks virtual memory size and
942 * numa_balancing_scan_size.
944 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
945 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
947 /* Portion of address space to scan in MB */
948 unsigned int sysctl_numa_balancing_scan_size = 256;
950 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
951 unsigned int sysctl_numa_balancing_scan_delay = 1000;
953 static unsigned int task_nr_scan_windows(struct task_struct *p)
955 unsigned long rss = 0;
956 unsigned long nr_scan_pages;
959 * Calculations based on RSS as non-present and empty pages are skipped
960 * by the PTE scanner and NUMA hinting faults should be trapped based
963 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
964 rss = get_mm_rss(p->mm);
968 rss = round_up(rss, nr_scan_pages);
969 return rss / nr_scan_pages;
972 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
973 #define MAX_SCAN_WINDOW 2560
975 static unsigned int task_scan_min(struct task_struct *p)
977 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
978 unsigned int scan, floor;
979 unsigned int windows = 1;
981 if (scan_size < MAX_SCAN_WINDOW)
982 windows = MAX_SCAN_WINDOW / scan_size;
983 floor = 1000 / windows;
985 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
986 return max_t(unsigned int, floor, scan);
989 static unsigned int task_scan_max(struct task_struct *p)
991 unsigned int smin = task_scan_min(p);
994 /* Watch for min being lower than max due to floor calculations */
995 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
996 return max(smin, smax);
999 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1001 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1002 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1005 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1007 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1008 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1014 spinlock_t lock; /* nr_tasks, tasks */
1019 struct rcu_head rcu;
1020 unsigned long total_faults;
1021 unsigned long max_faults_cpu;
1023 * Faults_cpu is used to decide whether memory should move
1024 * towards the CPU. As a consequence, these stats are weighted
1025 * more by CPU use than by memory faults.
1027 unsigned long *faults_cpu;
1028 unsigned long faults[0];
1031 /* Shared or private faults. */
1032 #define NR_NUMA_HINT_FAULT_TYPES 2
1034 /* Memory and CPU locality */
1035 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1037 /* Averaged statistics, and temporary buffers. */
1038 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1040 pid_t task_numa_group_id(struct task_struct *p)
1042 return p->numa_group ? p->numa_group->gid : 0;
1046 * The averaged statistics, shared & private, memory & cpu,
1047 * occupy the first half of the array. The second half of the
1048 * array is for current counters, which are averaged into the
1049 * first set by task_numa_placement.
1051 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1053 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1056 static inline unsigned long task_faults(struct task_struct *p, int nid)
1058 if (!p->numa_faults)
1061 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1062 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1065 static inline unsigned long group_faults(struct task_struct *p, int nid)
1070 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1071 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1074 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1076 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1077 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1081 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1082 * considered part of a numa group's pseudo-interleaving set. Migrations
1083 * between these nodes are slowed down, to allow things to settle down.
1085 #define ACTIVE_NODE_FRACTION 3
1087 static bool numa_is_active_node(int nid, struct numa_group *ng)
1089 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1092 /* Handle placement on systems where not all nodes are directly connected. */
1093 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1094 int maxdist, bool task)
1096 unsigned long score = 0;
1100 * All nodes are directly connected, and the same distance
1101 * from each other. No need for fancy placement algorithms.
1103 if (sched_numa_topology_type == NUMA_DIRECT)
1107 * This code is called for each node, introducing N^2 complexity,
1108 * which should be ok given the number of nodes rarely exceeds 8.
1110 for_each_online_node(node) {
1111 unsigned long faults;
1112 int dist = node_distance(nid, node);
1115 * The furthest away nodes in the system are not interesting
1116 * for placement; nid was already counted.
1118 if (dist == sched_max_numa_distance || node == nid)
1122 * On systems with a backplane NUMA topology, compare groups
1123 * of nodes, and move tasks towards the group with the most
1124 * memory accesses. When comparing two nodes at distance
1125 * "hoplimit", only nodes closer by than "hoplimit" are part
1126 * of each group. Skip other nodes.
1128 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1132 /* Add up the faults from nearby nodes. */
1134 faults = task_faults(p, node);
1136 faults = group_faults(p, node);
1139 * On systems with a glueless mesh NUMA topology, there are
1140 * no fixed "groups of nodes". Instead, nodes that are not
1141 * directly connected bounce traffic through intermediate
1142 * nodes; a numa_group can occupy any set of nodes.
1143 * The further away a node is, the less the faults count.
1144 * This seems to result in good task placement.
1146 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1147 faults *= (sched_max_numa_distance - dist);
1148 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1158 * These return the fraction of accesses done by a particular task, or
1159 * task group, on a particular numa node. The group weight is given a
1160 * larger multiplier, in order to group tasks together that are almost
1161 * evenly spread out between numa nodes.
1163 static inline unsigned long task_weight(struct task_struct *p, int nid,
1166 unsigned long faults, total_faults;
1168 if (!p->numa_faults)
1171 total_faults = p->total_numa_faults;
1176 faults = task_faults(p, nid);
1177 faults += score_nearby_nodes(p, nid, dist, true);
1179 return 1000 * faults / total_faults;
1182 static inline unsigned long group_weight(struct task_struct *p, int nid,
1185 unsigned long faults, total_faults;
1190 total_faults = p->numa_group->total_faults;
1195 faults = group_faults(p, nid);
1196 faults += score_nearby_nodes(p, nid, dist, false);
1198 return 1000 * faults / total_faults;
1201 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1202 int src_nid, int dst_cpu)
1204 struct numa_group *ng = p->numa_group;
1205 int dst_nid = cpu_to_node(dst_cpu);
1206 int last_cpupid, this_cpupid;
1208 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1211 * Multi-stage node selection is used in conjunction with a periodic
1212 * migration fault to build a temporal task<->page relation. By using
1213 * a two-stage filter we remove short/unlikely relations.
1215 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1216 * a task's usage of a particular page (n_p) per total usage of this
1217 * page (n_t) (in a given time-span) to a probability.
1219 * Our periodic faults will sample this probability and getting the
1220 * same result twice in a row, given these samples are fully
1221 * independent, is then given by P(n)^2, provided our sample period
1222 * is sufficiently short compared to the usage pattern.
1224 * This quadric squishes small probabilities, making it less likely we
1225 * act on an unlikely task<->page relation.
1227 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1228 if (!cpupid_pid_unset(last_cpupid) &&
1229 cpupid_to_nid(last_cpupid) != dst_nid)
1232 /* Always allow migrate on private faults */
1233 if (cpupid_match_pid(p, last_cpupid))
1236 /* A shared fault, but p->numa_group has not been set up yet. */
1241 * Destination node is much more heavily used than the source
1242 * node? Allow migration.
1244 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1245 ACTIVE_NODE_FRACTION)
1249 * Distribute memory according to CPU & memory use on each node,
1250 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1252 * faults_cpu(dst) 3 faults_cpu(src)
1253 * --------------- * - > ---------------
1254 * faults_mem(dst) 4 faults_mem(src)
1256 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1257 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1260 static unsigned long weighted_cpuload(const int cpu);
1261 static unsigned long source_load(int cpu, int type);
1262 static unsigned long target_load(int cpu, int type);
1263 static unsigned long capacity_of(int cpu);
1264 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1266 /* Cached statistics for all CPUs within a node */
1268 unsigned long nr_running;
1271 /* Total compute capacity of CPUs on a node */
1272 unsigned long compute_capacity;
1274 /* Approximate capacity in terms of runnable tasks on a node */
1275 unsigned long task_capacity;
1276 int has_free_capacity;
1280 * XXX borrowed from update_sg_lb_stats
1282 static void update_numa_stats(struct numa_stats *ns, int nid)
1284 int smt, cpu, cpus = 0;
1285 unsigned long capacity;
1287 memset(ns, 0, sizeof(*ns));
1288 for_each_cpu(cpu, cpumask_of_node(nid)) {
1289 struct rq *rq = cpu_rq(cpu);
1291 ns->nr_running += rq->nr_running;
1292 ns->load += weighted_cpuload(cpu);
1293 ns->compute_capacity += capacity_of(cpu);
1299 * If we raced with hotplug and there are no CPUs left in our mask
1300 * the @ns structure is NULL'ed and task_numa_compare() will
1301 * not find this node attractive.
1303 * We'll either bail at !has_free_capacity, or we'll detect a huge
1304 * imbalance and bail there.
1309 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1310 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1311 capacity = cpus / smt; /* cores */
1313 ns->task_capacity = min_t(unsigned, capacity,
1314 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1315 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1318 struct task_numa_env {
1319 struct task_struct *p;
1321 int src_cpu, src_nid;
1322 int dst_cpu, dst_nid;
1324 struct numa_stats src_stats, dst_stats;
1329 struct task_struct *best_task;
1334 static void task_numa_assign(struct task_numa_env *env,
1335 struct task_struct *p, long imp)
1338 put_task_struct(env->best_task);
1343 env->best_imp = imp;
1344 env->best_cpu = env->dst_cpu;
1347 static bool load_too_imbalanced(long src_load, long dst_load,
1348 struct task_numa_env *env)
1351 long orig_src_load, orig_dst_load;
1352 long src_capacity, dst_capacity;
1355 * The load is corrected for the CPU capacity available on each node.
1358 * ------------ vs ---------
1359 * src_capacity dst_capacity
1361 src_capacity = env->src_stats.compute_capacity;
1362 dst_capacity = env->dst_stats.compute_capacity;
1364 /* We care about the slope of the imbalance, not the direction. */
1365 if (dst_load < src_load)
1366 swap(dst_load, src_load);
1368 /* Is the difference below the threshold? */
1369 imb = dst_load * src_capacity * 100 -
1370 src_load * dst_capacity * env->imbalance_pct;
1375 * The imbalance is above the allowed threshold.
1376 * Compare it with the old imbalance.
1378 orig_src_load = env->src_stats.load;
1379 orig_dst_load = env->dst_stats.load;
1381 if (orig_dst_load < orig_src_load)
1382 swap(orig_dst_load, orig_src_load);
1384 old_imb = orig_dst_load * src_capacity * 100 -
1385 orig_src_load * dst_capacity * env->imbalance_pct;
1387 /* Would this change make things worse? */
1388 return (imb > old_imb);
1392 * This checks if the overall compute and NUMA accesses of the system would
1393 * be improved if the source tasks was migrated to the target dst_cpu taking
1394 * into account that it might be best if task running on the dst_cpu should
1395 * be exchanged with the source task
1397 static void task_numa_compare(struct task_numa_env *env,
1398 long taskimp, long groupimp)
1400 struct rq *src_rq = cpu_rq(env->src_cpu);
1401 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1402 struct task_struct *cur;
1403 long src_load, dst_load;
1405 long imp = env->p->numa_group ? groupimp : taskimp;
1407 int dist = env->dist;
1410 cur = task_rcu_dereference(&dst_rq->curr);
1411 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1415 * Because we have preemption enabled we can get migrated around and
1416 * end try selecting ourselves (current == env->p) as a swap candidate.
1422 * "imp" is the fault differential for the source task between the
1423 * source and destination node. Calculate the total differential for
1424 * the source task and potential destination task. The more negative
1425 * the value is, the more rmeote accesses that would be expected to
1426 * be incurred if the tasks were swapped.
1429 /* Skip this swap candidate if cannot move to the source cpu */
1430 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1434 * If dst and source tasks are in the same NUMA group, or not
1435 * in any group then look only at task weights.
1437 if (cur->numa_group == env->p->numa_group) {
1438 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1439 task_weight(cur, env->dst_nid, dist);
1441 * Add some hysteresis to prevent swapping the
1442 * tasks within a group over tiny differences.
1444 if (cur->numa_group)
1448 * Compare the group weights. If a task is all by
1449 * itself (not part of a group), use the task weight
1452 if (cur->numa_group)
1453 imp += group_weight(cur, env->src_nid, dist) -
1454 group_weight(cur, env->dst_nid, dist);
1456 imp += task_weight(cur, env->src_nid, dist) -
1457 task_weight(cur, env->dst_nid, dist);
1461 if (imp <= env->best_imp && moveimp <= env->best_imp)
1465 /* Is there capacity at our destination? */
1466 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1467 !env->dst_stats.has_free_capacity)
1473 /* Balance doesn't matter much if we're running a task per cpu */
1474 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1475 dst_rq->nr_running == 1)
1479 * In the overloaded case, try and keep the load balanced.
1482 load = task_h_load(env->p);
1483 dst_load = env->dst_stats.load + load;
1484 src_load = env->src_stats.load - load;
1486 if (moveimp > imp && moveimp > env->best_imp) {
1488 * If the improvement from just moving env->p direction is
1489 * better than swapping tasks around, check if a move is
1490 * possible. Store a slightly smaller score than moveimp,
1491 * so an actually idle CPU will win.
1493 if (!load_too_imbalanced(src_load, dst_load, env)) {
1500 if (imp <= env->best_imp)
1504 load = task_h_load(cur);
1509 if (load_too_imbalanced(src_load, dst_load, env))
1513 * One idle CPU per node is evaluated for a task numa move.
1514 * Call select_idle_sibling to maybe find a better one.
1517 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1520 task_numa_assign(env, cur, imp);
1525 static void task_numa_find_cpu(struct task_numa_env *env,
1526 long taskimp, long groupimp)
1530 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1531 /* Skip this CPU if the source task cannot migrate */
1532 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1536 task_numa_compare(env, taskimp, groupimp);
1540 /* Only move tasks to a NUMA node less busy than the current node. */
1541 static bool numa_has_capacity(struct task_numa_env *env)
1543 struct numa_stats *src = &env->src_stats;
1544 struct numa_stats *dst = &env->dst_stats;
1546 if (src->has_free_capacity && !dst->has_free_capacity)
1550 * Only consider a task move if the source has a higher load
1551 * than the destination, corrected for CPU capacity on each node.
1553 * src->load dst->load
1554 * --------------------- vs ---------------------
1555 * src->compute_capacity dst->compute_capacity
1557 if (src->load * dst->compute_capacity * env->imbalance_pct >
1559 dst->load * src->compute_capacity * 100)
1565 static int task_numa_migrate(struct task_struct *p)
1567 struct task_numa_env env = {
1570 .src_cpu = task_cpu(p),
1571 .src_nid = task_node(p),
1573 .imbalance_pct = 112,
1579 struct sched_domain *sd;
1580 unsigned long taskweight, groupweight;
1582 long taskimp, groupimp;
1585 * Pick the lowest SD_NUMA domain, as that would have the smallest
1586 * imbalance and would be the first to start moving tasks about.
1588 * And we want to avoid any moving of tasks about, as that would create
1589 * random movement of tasks -- counter the numa conditions we're trying
1593 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1595 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1599 * Cpusets can break the scheduler domain tree into smaller
1600 * balance domains, some of which do not cross NUMA boundaries.
1601 * Tasks that are "trapped" in such domains cannot be migrated
1602 * elsewhere, so there is no point in (re)trying.
1604 if (unlikely(!sd)) {
1605 p->numa_preferred_nid = task_node(p);
1609 env.dst_nid = p->numa_preferred_nid;
1610 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1611 taskweight = task_weight(p, env.src_nid, dist);
1612 groupweight = group_weight(p, env.src_nid, dist);
1613 update_numa_stats(&env.src_stats, env.src_nid);
1614 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1615 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1616 update_numa_stats(&env.dst_stats, env.dst_nid);
1618 /* Try to find a spot on the preferred nid. */
1619 if (numa_has_capacity(&env))
1620 task_numa_find_cpu(&env, taskimp, groupimp);
1623 * Look at other nodes in these cases:
1624 * - there is no space available on the preferred_nid
1625 * - the task is part of a numa_group that is interleaved across
1626 * multiple NUMA nodes; in order to better consolidate the group,
1627 * we need to check other locations.
1629 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1630 for_each_online_node(nid) {
1631 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1634 dist = node_distance(env.src_nid, env.dst_nid);
1635 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1637 taskweight = task_weight(p, env.src_nid, dist);
1638 groupweight = group_weight(p, env.src_nid, dist);
1641 /* Only consider nodes where both task and groups benefit */
1642 taskimp = task_weight(p, nid, dist) - taskweight;
1643 groupimp = group_weight(p, nid, dist) - groupweight;
1644 if (taskimp < 0 && groupimp < 0)
1649 update_numa_stats(&env.dst_stats, env.dst_nid);
1650 if (numa_has_capacity(&env))
1651 task_numa_find_cpu(&env, taskimp, groupimp);
1656 * If the task is part of a workload that spans multiple NUMA nodes,
1657 * and is migrating into one of the workload's active nodes, remember
1658 * this node as the task's preferred numa node, so the workload can
1660 * A task that migrated to a second choice node will be better off
1661 * trying for a better one later. Do not set the preferred node here.
1663 if (p->numa_group) {
1664 struct numa_group *ng = p->numa_group;
1666 if (env.best_cpu == -1)
1671 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1672 sched_setnuma(p, env.dst_nid);
1675 /* No better CPU than the current one was found. */
1676 if (env.best_cpu == -1)
1680 * Reset the scan period if the task is being rescheduled on an
1681 * alternative node to recheck if the tasks is now properly placed.
1683 p->numa_scan_period = task_scan_min(p);
1685 if (env.best_task == NULL) {
1686 ret = migrate_task_to(p, env.best_cpu);
1688 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1692 ret = migrate_swap(p, env.best_task);
1694 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1695 put_task_struct(env.best_task);
1699 /* Attempt to migrate a task to a CPU on the preferred node. */
1700 static void numa_migrate_preferred(struct task_struct *p)
1702 unsigned long interval = HZ;
1704 /* This task has no NUMA fault statistics yet */
1705 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1708 /* Periodically retry migrating the task to the preferred node */
1709 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1710 p->numa_migrate_retry = jiffies + interval;
1712 /* Success if task is already running on preferred CPU */
1713 if (task_node(p) == p->numa_preferred_nid)
1716 /* Otherwise, try migrate to a CPU on the preferred node */
1717 task_numa_migrate(p);
1721 * Find out how many nodes on the workload is actively running on. Do this by
1722 * tracking the nodes from which NUMA hinting faults are triggered. This can
1723 * be different from the set of nodes where the workload's memory is currently
1726 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1728 unsigned long faults, max_faults = 0;
1729 int nid, active_nodes = 0;
1731 for_each_online_node(nid) {
1732 faults = group_faults_cpu(numa_group, nid);
1733 if (faults > max_faults)
1734 max_faults = faults;
1737 for_each_online_node(nid) {
1738 faults = group_faults_cpu(numa_group, nid);
1739 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1743 numa_group->max_faults_cpu = max_faults;
1744 numa_group->active_nodes = active_nodes;
1748 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1749 * increments. The more local the fault statistics are, the higher the scan
1750 * period will be for the next scan window. If local/(local+remote) ratio is
1751 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1752 * the scan period will decrease. Aim for 70% local accesses.
1754 #define NUMA_PERIOD_SLOTS 10
1755 #define NUMA_PERIOD_THRESHOLD 7
1758 * Increase the scan period (slow down scanning) if the majority of
1759 * our memory is already on our local node, or if the majority of
1760 * the page accesses are shared with other processes.
1761 * Otherwise, decrease the scan period.
1763 static void update_task_scan_period(struct task_struct *p,
1764 unsigned long shared, unsigned long private)
1766 unsigned int period_slot;
1770 unsigned long remote = p->numa_faults_locality[0];
1771 unsigned long local = p->numa_faults_locality[1];
1774 * If there were no record hinting faults then either the task is
1775 * completely idle or all activity is areas that are not of interest
1776 * to automatic numa balancing. Related to that, if there were failed
1777 * migration then it implies we are migrating too quickly or the local
1778 * node is overloaded. In either case, scan slower
1780 if (local + shared == 0 || p->numa_faults_locality[2]) {
1781 p->numa_scan_period = min(p->numa_scan_period_max,
1782 p->numa_scan_period << 1);
1784 p->mm->numa_next_scan = jiffies +
1785 msecs_to_jiffies(p->numa_scan_period);
1791 * Prepare to scale scan period relative to the current period.
1792 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1793 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1794 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1796 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1797 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1798 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1799 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1802 diff = slot * period_slot;
1804 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1807 * Scale scan rate increases based on sharing. There is an
1808 * inverse relationship between the degree of sharing and
1809 * the adjustment made to the scanning period. Broadly
1810 * speaking the intent is that there is little point
1811 * scanning faster if shared accesses dominate as it may
1812 * simply bounce migrations uselessly
1814 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1815 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1818 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1819 task_scan_min(p), task_scan_max(p));
1820 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1824 * Get the fraction of time the task has been running since the last
1825 * NUMA placement cycle. The scheduler keeps similar statistics, but
1826 * decays those on a 32ms period, which is orders of magnitude off
1827 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1828 * stats only if the task is so new there are no NUMA statistics yet.
1830 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1832 u64 runtime, delta, now;
1833 /* Use the start of this time slice to avoid calculations. */
1834 now = p->se.exec_start;
1835 runtime = p->se.sum_exec_runtime;
1837 if (p->last_task_numa_placement) {
1838 delta = runtime - p->last_sum_exec_runtime;
1839 *period = now - p->last_task_numa_placement;
1841 delta = p->se.avg.load_sum / p->se.load.weight;
1842 *period = LOAD_AVG_MAX;
1845 p->last_sum_exec_runtime = runtime;
1846 p->last_task_numa_placement = now;
1852 * Determine the preferred nid for a task in a numa_group. This needs to
1853 * be done in a way that produces consistent results with group_weight,
1854 * otherwise workloads might not converge.
1856 static int preferred_group_nid(struct task_struct *p, int nid)
1861 /* Direct connections between all NUMA nodes. */
1862 if (sched_numa_topology_type == NUMA_DIRECT)
1866 * On a system with glueless mesh NUMA topology, group_weight
1867 * scores nodes according to the number of NUMA hinting faults on
1868 * both the node itself, and on nearby nodes.
1870 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1871 unsigned long score, max_score = 0;
1872 int node, max_node = nid;
1874 dist = sched_max_numa_distance;
1876 for_each_online_node(node) {
1877 score = group_weight(p, node, dist);
1878 if (score > max_score) {
1887 * Finding the preferred nid in a system with NUMA backplane
1888 * interconnect topology is more involved. The goal is to locate
1889 * tasks from numa_groups near each other in the system, and
1890 * untangle workloads from different sides of the system. This requires
1891 * searching down the hierarchy of node groups, recursively searching
1892 * inside the highest scoring group of nodes. The nodemask tricks
1893 * keep the complexity of the search down.
1895 nodes = node_online_map;
1896 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1897 unsigned long max_faults = 0;
1898 nodemask_t max_group = NODE_MASK_NONE;
1901 /* Are there nodes at this distance from each other? */
1902 if (!find_numa_distance(dist))
1905 for_each_node_mask(a, nodes) {
1906 unsigned long faults = 0;
1907 nodemask_t this_group;
1908 nodes_clear(this_group);
1910 /* Sum group's NUMA faults; includes a==b case. */
1911 for_each_node_mask(b, nodes) {
1912 if (node_distance(a, b) < dist) {
1913 faults += group_faults(p, b);
1914 node_set(b, this_group);
1915 node_clear(b, nodes);
1919 /* Remember the top group. */
1920 if (faults > max_faults) {
1921 max_faults = faults;
1922 max_group = this_group;
1924 * subtle: at the smallest distance there is
1925 * just one node left in each "group", the
1926 * winner is the preferred nid.
1931 /* Next round, evaluate the nodes within max_group. */
1939 static void task_numa_placement(struct task_struct *p)
1941 int seq, nid, max_nid = -1, max_group_nid = -1;
1942 unsigned long max_faults = 0, max_group_faults = 0;
1943 unsigned long fault_types[2] = { 0, 0 };
1944 unsigned long total_faults;
1945 u64 runtime, period;
1946 spinlock_t *group_lock = NULL;
1949 * The p->mm->numa_scan_seq field gets updated without
1950 * exclusive access. Use READ_ONCE() here to ensure
1951 * that the field is read in a single access:
1953 seq = READ_ONCE(p->mm->numa_scan_seq);
1954 if (p->numa_scan_seq == seq)
1956 p->numa_scan_seq = seq;
1957 p->numa_scan_period_max = task_scan_max(p);
1959 total_faults = p->numa_faults_locality[0] +
1960 p->numa_faults_locality[1];
1961 runtime = numa_get_avg_runtime(p, &period);
1963 /* If the task is part of a group prevent parallel updates to group stats */
1964 if (p->numa_group) {
1965 group_lock = &p->numa_group->lock;
1966 spin_lock_irq(group_lock);
1969 /* Find the node with the highest number of faults */
1970 for_each_online_node(nid) {
1971 /* Keep track of the offsets in numa_faults array */
1972 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1973 unsigned long faults = 0, group_faults = 0;
1976 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1977 long diff, f_diff, f_weight;
1979 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1980 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1981 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1982 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1984 /* Decay existing window, copy faults since last scan */
1985 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1986 fault_types[priv] += p->numa_faults[membuf_idx];
1987 p->numa_faults[membuf_idx] = 0;
1990 * Normalize the faults_from, so all tasks in a group
1991 * count according to CPU use, instead of by the raw
1992 * number of faults. Tasks with little runtime have
1993 * little over-all impact on throughput, and thus their
1994 * faults are less important.
1996 f_weight = div64_u64(runtime << 16, period + 1);
1997 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1999 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2000 p->numa_faults[cpubuf_idx] = 0;
2002 p->numa_faults[mem_idx] += diff;
2003 p->numa_faults[cpu_idx] += f_diff;
2004 faults += p->numa_faults[mem_idx];
2005 p->total_numa_faults += diff;
2006 if (p->numa_group) {
2008 * safe because we can only change our own group
2010 * mem_idx represents the offset for a given
2011 * nid and priv in a specific region because it
2012 * is at the beginning of the numa_faults array.
2014 p->numa_group->faults[mem_idx] += diff;
2015 p->numa_group->faults_cpu[mem_idx] += f_diff;
2016 p->numa_group->total_faults += diff;
2017 group_faults += p->numa_group->faults[mem_idx];
2021 if (faults > max_faults) {
2022 max_faults = faults;
2026 if (group_faults > max_group_faults) {
2027 max_group_faults = group_faults;
2028 max_group_nid = nid;
2032 update_task_scan_period(p, fault_types[0], fault_types[1]);
2034 if (p->numa_group) {
2035 numa_group_count_active_nodes(p->numa_group);
2036 spin_unlock_irq(group_lock);
2037 max_nid = preferred_group_nid(p, max_group_nid);
2041 /* Set the new preferred node */
2042 if (max_nid != p->numa_preferred_nid)
2043 sched_setnuma(p, max_nid);
2045 if (task_node(p) != p->numa_preferred_nid)
2046 numa_migrate_preferred(p);
2050 static inline int get_numa_group(struct numa_group *grp)
2052 return atomic_inc_not_zero(&grp->refcount);
2055 static inline void put_numa_group(struct numa_group *grp)
2057 if (atomic_dec_and_test(&grp->refcount))
2058 kfree_rcu(grp, rcu);
2061 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2064 struct numa_group *grp, *my_grp;
2065 struct task_struct *tsk;
2067 int cpu = cpupid_to_cpu(cpupid);
2070 if (unlikely(!p->numa_group)) {
2071 unsigned int size = sizeof(struct numa_group) +
2072 4*nr_node_ids*sizeof(unsigned long);
2074 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2078 atomic_set(&grp->refcount, 1);
2079 grp->active_nodes = 1;
2080 grp->max_faults_cpu = 0;
2081 spin_lock_init(&grp->lock);
2083 /* Second half of the array tracks nids where faults happen */
2084 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2087 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2088 grp->faults[i] = p->numa_faults[i];
2090 grp->total_faults = p->total_numa_faults;
2093 rcu_assign_pointer(p->numa_group, grp);
2097 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2099 if (!cpupid_match_pid(tsk, cpupid))
2102 grp = rcu_dereference(tsk->numa_group);
2106 my_grp = p->numa_group;
2111 * Only join the other group if its bigger; if we're the bigger group,
2112 * the other task will join us.
2114 if (my_grp->nr_tasks > grp->nr_tasks)
2118 * Tie-break on the grp address.
2120 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2123 /* Always join threads in the same process. */
2124 if (tsk->mm == current->mm)
2127 /* Simple filter to avoid false positives due to PID collisions */
2128 if (flags & TNF_SHARED)
2131 /* Update priv based on whether false sharing was detected */
2134 if (join && !get_numa_group(grp))
2142 BUG_ON(irqs_disabled());
2143 double_lock_irq(&my_grp->lock, &grp->lock);
2145 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2146 my_grp->faults[i] -= p->numa_faults[i];
2147 grp->faults[i] += p->numa_faults[i];
2149 my_grp->total_faults -= p->total_numa_faults;
2150 grp->total_faults += p->total_numa_faults;
2155 spin_unlock(&my_grp->lock);
2156 spin_unlock_irq(&grp->lock);
2158 rcu_assign_pointer(p->numa_group, grp);
2160 put_numa_group(my_grp);
2168 void task_numa_free(struct task_struct *p)
2170 struct numa_group *grp = p->numa_group;
2171 void *numa_faults = p->numa_faults;
2172 unsigned long flags;
2176 spin_lock_irqsave(&grp->lock, flags);
2177 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2178 grp->faults[i] -= p->numa_faults[i];
2179 grp->total_faults -= p->total_numa_faults;
2182 spin_unlock_irqrestore(&grp->lock, flags);
2183 RCU_INIT_POINTER(p->numa_group, NULL);
2184 put_numa_group(grp);
2187 p->numa_faults = NULL;
2192 * Got a PROT_NONE fault for a page on @node.
2194 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2196 struct task_struct *p = current;
2197 bool migrated = flags & TNF_MIGRATED;
2198 int cpu_node = task_node(current);
2199 int local = !!(flags & TNF_FAULT_LOCAL);
2200 struct numa_group *ng;
2203 if (!static_branch_likely(&sched_numa_balancing))
2206 /* for example, ksmd faulting in a user's mm */
2210 /* Allocate buffer to track faults on a per-node basis */
2211 if (unlikely(!p->numa_faults)) {
2212 int size = sizeof(*p->numa_faults) *
2213 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2215 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2216 if (!p->numa_faults)
2219 p->total_numa_faults = 0;
2220 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2224 * First accesses are treated as private, otherwise consider accesses
2225 * to be private if the accessing pid has not changed
2227 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2230 priv = cpupid_match_pid(p, last_cpupid);
2231 if (!priv && !(flags & TNF_NO_GROUP))
2232 task_numa_group(p, last_cpupid, flags, &priv);
2236 * If a workload spans multiple NUMA nodes, a shared fault that
2237 * occurs wholly within the set of nodes that the workload is
2238 * actively using should be counted as local. This allows the
2239 * scan rate to slow down when a workload has settled down.
2242 if (!priv && !local && ng && ng->active_nodes > 1 &&
2243 numa_is_active_node(cpu_node, ng) &&
2244 numa_is_active_node(mem_node, ng))
2247 task_numa_placement(p);
2250 * Retry task to preferred node migration periodically, in case it
2251 * case it previously failed, or the scheduler moved us.
2253 if (time_after(jiffies, p->numa_migrate_retry))
2254 numa_migrate_preferred(p);
2257 p->numa_pages_migrated += pages;
2258 if (flags & TNF_MIGRATE_FAIL)
2259 p->numa_faults_locality[2] += pages;
2261 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2262 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2263 p->numa_faults_locality[local] += pages;
2266 static void reset_ptenuma_scan(struct task_struct *p)
2269 * We only did a read acquisition of the mmap sem, so
2270 * p->mm->numa_scan_seq is written to without exclusive access
2271 * and the update is not guaranteed to be atomic. That's not
2272 * much of an issue though, since this is just used for
2273 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2274 * expensive, to avoid any form of compiler optimizations:
2276 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2277 p->mm->numa_scan_offset = 0;
2281 * The expensive part of numa migration is done from task_work context.
2282 * Triggered from task_tick_numa().
2284 void task_numa_work(struct callback_head *work)
2286 unsigned long migrate, next_scan, now = jiffies;
2287 struct task_struct *p = current;
2288 struct mm_struct *mm = p->mm;
2289 u64 runtime = p->se.sum_exec_runtime;
2290 struct vm_area_struct *vma;
2291 unsigned long start, end;
2292 unsigned long nr_pte_updates = 0;
2293 long pages, virtpages;
2295 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2297 work->next = work; /* protect against double add */
2299 * Who cares about NUMA placement when they're dying.
2301 * NOTE: make sure not to dereference p->mm before this check,
2302 * exit_task_work() happens _after_ exit_mm() so we could be called
2303 * without p->mm even though we still had it when we enqueued this
2306 if (p->flags & PF_EXITING)
2309 if (!mm->numa_next_scan) {
2310 mm->numa_next_scan = now +
2311 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2315 * Enforce maximal scan/migration frequency..
2317 migrate = mm->numa_next_scan;
2318 if (time_before(now, migrate))
2321 if (p->numa_scan_period == 0) {
2322 p->numa_scan_period_max = task_scan_max(p);
2323 p->numa_scan_period = task_scan_min(p);
2326 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2327 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2331 * Delay this task enough that another task of this mm will likely win
2332 * the next time around.
2334 p->node_stamp += 2 * TICK_NSEC;
2336 start = mm->numa_scan_offset;
2337 pages = sysctl_numa_balancing_scan_size;
2338 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2339 virtpages = pages * 8; /* Scan up to this much virtual space */
2344 down_read(&mm->mmap_sem);
2345 vma = find_vma(mm, start);
2347 reset_ptenuma_scan(p);
2351 for (; vma; vma = vma->vm_next) {
2352 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2353 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2358 * Shared library pages mapped by multiple processes are not
2359 * migrated as it is expected they are cache replicated. Avoid
2360 * hinting faults in read-only file-backed mappings or the vdso
2361 * as migrating the pages will be of marginal benefit.
2364 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2368 * Skip inaccessible VMAs to avoid any confusion between
2369 * PROT_NONE and NUMA hinting ptes
2371 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2375 start = max(start, vma->vm_start);
2376 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2377 end = min(end, vma->vm_end);
2378 nr_pte_updates = change_prot_numa(vma, start, end);
2381 * Try to scan sysctl_numa_balancing_size worth of
2382 * hpages that have at least one present PTE that
2383 * is not already pte-numa. If the VMA contains
2384 * areas that are unused or already full of prot_numa
2385 * PTEs, scan up to virtpages, to skip through those
2389 pages -= (end - start) >> PAGE_SHIFT;
2390 virtpages -= (end - start) >> PAGE_SHIFT;
2393 if (pages <= 0 || virtpages <= 0)
2397 } while (end != vma->vm_end);
2402 * It is possible to reach the end of the VMA list but the last few
2403 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2404 * would find the !migratable VMA on the next scan but not reset the
2405 * scanner to the start so check it now.
2408 mm->numa_scan_offset = start;
2410 reset_ptenuma_scan(p);
2411 up_read(&mm->mmap_sem);
2414 * Make sure tasks use at least 32x as much time to run other code
2415 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2416 * Usually update_task_scan_period slows down scanning enough; on an
2417 * overloaded system we need to limit overhead on a per task basis.
2419 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2420 u64 diff = p->se.sum_exec_runtime - runtime;
2421 p->node_stamp += 32 * diff;
2426 * Drive the periodic memory faults..
2428 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2430 struct callback_head *work = &curr->numa_work;
2434 * We don't care about NUMA placement if we don't have memory.
2436 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2440 * Using runtime rather than walltime has the dual advantage that
2441 * we (mostly) drive the selection from busy threads and that the
2442 * task needs to have done some actual work before we bother with
2445 now = curr->se.sum_exec_runtime;
2446 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2448 if (now > curr->node_stamp + period) {
2449 if (!curr->node_stamp)
2450 curr->numa_scan_period = task_scan_min(curr);
2451 curr->node_stamp += period;
2453 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2454 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2455 task_work_add(curr, work, true);
2460 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2464 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2468 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2471 #endif /* CONFIG_NUMA_BALANCING */
2474 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2476 update_load_add(&cfs_rq->load, se->load.weight);
2477 if (!parent_entity(se))
2478 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2480 if (entity_is_task(se)) {
2481 struct rq *rq = rq_of(cfs_rq);
2483 account_numa_enqueue(rq, task_of(se));
2484 list_add(&se->group_node, &rq->cfs_tasks);
2487 cfs_rq->nr_running++;
2491 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2493 update_load_sub(&cfs_rq->load, se->load.weight);
2494 if (!parent_entity(se))
2495 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2497 if (entity_is_task(se)) {
2498 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2499 list_del_init(&se->group_node);
2502 cfs_rq->nr_running--;
2505 #ifdef CONFIG_FAIR_GROUP_SCHED
2507 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2509 long tg_weight, load, shares;
2512 * This really should be: cfs_rq->avg.load_avg, but instead we use
2513 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2514 * the shares for small weight interactive tasks.
2516 load = scale_load_down(cfs_rq->load.weight);
2518 tg_weight = atomic_long_read(&tg->load_avg);
2520 /* Ensure tg_weight >= load */
2521 tg_weight -= cfs_rq->tg_load_avg_contrib;
2524 shares = (tg->shares * load);
2526 shares /= tg_weight;
2528 if (shares < MIN_SHARES)
2529 shares = MIN_SHARES;
2530 if (shares > tg->shares)
2531 shares = tg->shares;
2535 # else /* CONFIG_SMP */
2536 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2540 # endif /* CONFIG_SMP */
2542 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2543 unsigned long weight)
2546 /* commit outstanding execution time */
2547 if (cfs_rq->curr == se)
2548 update_curr(cfs_rq);
2549 account_entity_dequeue(cfs_rq, se);
2552 update_load_set(&se->load, weight);
2555 account_entity_enqueue(cfs_rq, se);
2558 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2560 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2562 struct task_group *tg;
2563 struct sched_entity *se;
2567 se = tg->se[cpu_of(rq_of(cfs_rq))];
2568 if (!se || throttled_hierarchy(cfs_rq))
2571 if (likely(se->load.weight == tg->shares))
2574 shares = calc_cfs_shares(cfs_rq, tg);
2576 reweight_entity(cfs_rq_of(se), se, shares);
2578 #else /* CONFIG_FAIR_GROUP_SCHED */
2579 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2582 #endif /* CONFIG_FAIR_GROUP_SCHED */
2585 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2586 static const u32 runnable_avg_yN_inv[] = {
2587 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2588 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2589 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2590 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2591 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2592 0x85aac367, 0x82cd8698,
2596 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2597 * over-estimates when re-combining.
2599 static const u32 runnable_avg_yN_sum[] = {
2600 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2601 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2602 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2606 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2607 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2610 static const u32 __accumulated_sum_N32[] = {
2611 0, 23371, 35056, 40899, 43820, 45281,
2612 46011, 46376, 46559, 46650, 46696, 46719,
2617 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2619 static __always_inline u64 decay_load(u64 val, u64 n)
2621 unsigned int local_n;
2625 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2628 /* after bounds checking we can collapse to 32-bit */
2632 * As y^PERIOD = 1/2, we can combine
2633 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2634 * With a look-up table which covers y^n (n<PERIOD)
2636 * To achieve constant time decay_load.
2638 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2639 val >>= local_n / LOAD_AVG_PERIOD;
2640 local_n %= LOAD_AVG_PERIOD;
2643 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2648 * For updates fully spanning n periods, the contribution to runnable
2649 * average will be: \Sum 1024*y^n
2651 * We can compute this reasonably efficiently by combining:
2652 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2654 static u32 __compute_runnable_contrib(u64 n)
2658 if (likely(n <= LOAD_AVG_PERIOD))
2659 return runnable_avg_yN_sum[n];
2660 else if (unlikely(n >= LOAD_AVG_MAX_N))
2661 return LOAD_AVG_MAX;
2663 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2664 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2665 n %= LOAD_AVG_PERIOD;
2666 contrib = decay_load(contrib, n);
2667 return contrib + runnable_avg_yN_sum[n];
2670 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2673 * We can represent the historical contribution to runnable average as the
2674 * coefficients of a geometric series. To do this we sub-divide our runnable
2675 * history into segments of approximately 1ms (1024us); label the segment that
2676 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2678 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2680 * (now) (~1ms ago) (~2ms ago)
2682 * Let u_i denote the fraction of p_i that the entity was runnable.
2684 * We then designate the fractions u_i as our co-efficients, yielding the
2685 * following representation of historical load:
2686 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2688 * We choose y based on the with of a reasonably scheduling period, fixing:
2691 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2692 * approximately half as much as the contribution to load within the last ms
2695 * When a period "rolls over" and we have new u_0`, multiplying the previous
2696 * sum again by y is sufficient to update:
2697 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2698 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2700 static __always_inline int
2701 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2702 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2704 u64 delta, scaled_delta, periods;
2706 unsigned int delta_w, scaled_delta_w, decayed = 0;
2707 unsigned long scale_freq, scale_cpu;
2709 delta = now - sa->last_update_time;
2711 * This should only happen when time goes backwards, which it
2712 * unfortunately does during sched clock init when we swap over to TSC.
2714 if ((s64)delta < 0) {
2715 sa->last_update_time = now;
2720 * Use 1024ns as the unit of measurement since it's a reasonable
2721 * approximation of 1us and fast to compute.
2726 sa->last_update_time = now;
2728 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2729 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2731 /* delta_w is the amount already accumulated against our next period */
2732 delta_w = sa->period_contrib;
2733 if (delta + delta_w >= 1024) {
2736 /* how much left for next period will start over, we don't know yet */
2737 sa->period_contrib = 0;
2740 * Now that we know we're crossing a period boundary, figure
2741 * out how much from delta we need to complete the current
2742 * period and accrue it.
2744 delta_w = 1024 - delta_w;
2745 scaled_delta_w = cap_scale(delta_w, scale_freq);
2747 sa->load_sum += weight * scaled_delta_w;
2749 cfs_rq->runnable_load_sum +=
2750 weight * scaled_delta_w;
2754 sa->util_sum += scaled_delta_w * scale_cpu;
2758 /* Figure out how many additional periods this update spans */
2759 periods = delta / 1024;
2762 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2764 cfs_rq->runnable_load_sum =
2765 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2767 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2769 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2770 contrib = __compute_runnable_contrib(periods);
2771 contrib = cap_scale(contrib, scale_freq);
2773 sa->load_sum += weight * contrib;
2775 cfs_rq->runnable_load_sum += weight * contrib;
2778 sa->util_sum += contrib * scale_cpu;
2781 /* Remainder of delta accrued against u_0` */
2782 scaled_delta = cap_scale(delta, scale_freq);
2784 sa->load_sum += weight * scaled_delta;
2786 cfs_rq->runnable_load_sum += weight * scaled_delta;
2789 sa->util_sum += scaled_delta * scale_cpu;
2791 sa->period_contrib += delta;
2794 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2796 cfs_rq->runnable_load_avg =
2797 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2799 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2805 #ifdef CONFIG_FAIR_GROUP_SCHED
2807 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2808 * and effective_load (which is not done because it is too costly).
2810 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2812 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2815 * No need to update load_avg for root_task_group as it is not used.
2817 if (cfs_rq->tg == &root_task_group)
2820 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2821 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2822 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2827 * Called within set_task_rq() right before setting a task's cpu. The
2828 * caller only guarantees p->pi_lock is held; no other assumptions,
2829 * including the state of rq->lock, should be made.
2831 void set_task_rq_fair(struct sched_entity *se,
2832 struct cfs_rq *prev, struct cfs_rq *next)
2834 if (!sched_feat(ATTACH_AGE_LOAD))
2838 * We are supposed to update the task to "current" time, then its up to
2839 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2840 * getting what current time is, so simply throw away the out-of-date
2841 * time. This will result in the wakee task is less decayed, but giving
2842 * the wakee more load sounds not bad.
2844 if (se->avg.last_update_time && prev) {
2845 u64 p_last_update_time;
2846 u64 n_last_update_time;
2848 #ifndef CONFIG_64BIT
2849 u64 p_last_update_time_copy;
2850 u64 n_last_update_time_copy;
2853 p_last_update_time_copy = prev->load_last_update_time_copy;
2854 n_last_update_time_copy = next->load_last_update_time_copy;
2858 p_last_update_time = prev->avg.last_update_time;
2859 n_last_update_time = next->avg.last_update_time;
2861 } while (p_last_update_time != p_last_update_time_copy ||
2862 n_last_update_time != n_last_update_time_copy);
2864 p_last_update_time = prev->avg.last_update_time;
2865 n_last_update_time = next->avg.last_update_time;
2867 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2868 &se->avg, 0, 0, NULL);
2869 se->avg.last_update_time = n_last_update_time;
2872 #else /* CONFIG_FAIR_GROUP_SCHED */
2873 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2874 #endif /* CONFIG_FAIR_GROUP_SCHED */
2876 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2878 struct rq *rq = rq_of(cfs_rq);
2879 int cpu = cpu_of(rq);
2881 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2882 unsigned long max = rq->cpu_capacity_orig;
2885 * There are a few boundary cases this might miss but it should
2886 * get called often enough that that should (hopefully) not be
2887 * a real problem -- added to that it only calls on the local
2888 * CPU, so if we enqueue remotely we'll miss an update, but
2889 * the next tick/schedule should update.
2891 * It will not get called when we go idle, because the idle
2892 * thread is a different class (!fair), nor will the utilization
2893 * number include things like RT tasks.
2895 * As is, the util number is not freq-invariant (we'd have to
2896 * implement arch_scale_freq_capacity() for that).
2900 cpufreq_update_util(rq_clock(rq),
2901 min(cfs_rq->avg.util_avg, max), max);
2906 * Unsigned subtract and clamp on underflow.
2908 * Explicitly do a load-store to ensure the intermediate value never hits
2909 * memory. This allows lockless observations without ever seeing the negative
2912 #define sub_positive(_ptr, _val) do { \
2913 typeof(_ptr) ptr = (_ptr); \
2914 typeof(*ptr) val = (_val); \
2915 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2919 WRITE_ONCE(*ptr, res); \
2923 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2924 * @now: current time, as per cfs_rq_clock_task()
2925 * @cfs_rq: cfs_rq to update
2926 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2928 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2929 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2930 * post_init_entity_util_avg().
2932 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2934 * Returns true if the load decayed or we removed utilization. It is expected
2935 * that one calls update_tg_load_avg() on this condition, but after you've
2936 * modified the cfs_rq avg (attach/detach), such that we propagate the new
2940 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2942 struct sched_avg *sa = &cfs_rq->avg;
2943 int decayed, removed_load = 0, removed_util = 0;
2945 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2946 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2947 sub_positive(&sa->load_avg, r);
2948 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2952 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2953 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2954 sub_positive(&sa->util_avg, r);
2955 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2959 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2960 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2962 #ifndef CONFIG_64BIT
2964 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2967 if (update_freq && (decayed || removed_util))
2968 cfs_rq_util_change(cfs_rq);
2970 return decayed || removed_load;
2973 /* Update task and its cfs_rq load average */
2974 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2976 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2977 u64 now = cfs_rq_clock_task(cfs_rq);
2978 struct rq *rq = rq_of(cfs_rq);
2979 int cpu = cpu_of(rq);
2982 * Track task load average for carrying it to new CPU after migrated, and
2983 * track group sched_entity load average for task_h_load calc in migration
2985 __update_load_avg(now, cpu, &se->avg,
2986 se->on_rq * scale_load_down(se->load.weight),
2987 cfs_rq->curr == se, NULL);
2989 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2990 update_tg_load_avg(cfs_rq, 0);
2994 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2995 * @cfs_rq: cfs_rq to attach to
2996 * @se: sched_entity to attach
2998 * Must call update_cfs_rq_load_avg() before this, since we rely on
2999 * cfs_rq->avg.last_update_time being current.
3001 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 if (!sched_feat(ATTACH_AGE_LOAD))
3007 * If we got migrated (either between CPUs or between cgroups) we'll
3008 * have aged the average right before clearing @last_update_time.
3010 * Or we're fresh through post_init_entity_util_avg().
3012 if (se->avg.last_update_time) {
3013 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3014 &se->avg, 0, 0, NULL);
3017 * XXX: we could have just aged the entire load away if we've been
3018 * absent from the fair class for too long.
3023 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3024 cfs_rq->avg.load_avg += se->avg.load_avg;
3025 cfs_rq->avg.load_sum += se->avg.load_sum;
3026 cfs_rq->avg.util_avg += se->avg.util_avg;
3027 cfs_rq->avg.util_sum += se->avg.util_sum;
3029 cfs_rq_util_change(cfs_rq);
3033 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3034 * @cfs_rq: cfs_rq to detach from
3035 * @se: sched_entity to detach
3037 * Must call update_cfs_rq_load_avg() before this, since we rely on
3038 * cfs_rq->avg.last_update_time being current.
3040 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3042 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3043 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3044 cfs_rq->curr == se, NULL);
3046 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3047 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3048 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3049 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3051 cfs_rq_util_change(cfs_rq);
3054 /* Add the load generated by se into cfs_rq's load average */
3056 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3058 struct sched_avg *sa = &se->avg;
3059 u64 now = cfs_rq_clock_task(cfs_rq);
3060 int migrated, decayed;
3062 migrated = !sa->last_update_time;
3064 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3065 se->on_rq * scale_load_down(se->load.weight),
3066 cfs_rq->curr == se, NULL);
3069 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3071 cfs_rq->runnable_load_avg += sa->load_avg;
3072 cfs_rq->runnable_load_sum += sa->load_sum;
3075 attach_entity_load_avg(cfs_rq, se);
3077 if (decayed || migrated)
3078 update_tg_load_avg(cfs_rq, 0);
3081 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3083 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3085 update_load_avg(se, 1);
3087 cfs_rq->runnable_load_avg =
3088 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3089 cfs_rq->runnable_load_sum =
3090 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3093 #ifndef CONFIG_64BIT
3094 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3096 u64 last_update_time_copy;
3097 u64 last_update_time;
3100 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3102 last_update_time = cfs_rq->avg.last_update_time;
3103 } while (last_update_time != last_update_time_copy);
3105 return last_update_time;
3108 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3110 return cfs_rq->avg.last_update_time;
3115 * Task first catches up with cfs_rq, and then subtract
3116 * itself from the cfs_rq (task must be off the queue now).
3118 void remove_entity_load_avg(struct sched_entity *se)
3120 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3121 u64 last_update_time;
3124 * tasks cannot exit without having gone through wake_up_new_task() ->
3125 * post_init_entity_util_avg() which will have added things to the
3126 * cfs_rq, so we can remove unconditionally.
3128 * Similarly for groups, they will have passed through
3129 * post_init_entity_util_avg() before unregister_sched_fair_group()
3133 last_update_time = cfs_rq_last_update_time(cfs_rq);
3135 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3136 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3137 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3140 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3142 return cfs_rq->runnable_load_avg;
3145 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3147 return cfs_rq->avg.load_avg;
3150 static int idle_balance(struct rq *this_rq);
3152 #else /* CONFIG_SMP */
3155 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3160 static inline void update_load_avg(struct sched_entity *se, int not_used)
3162 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3163 struct rq *rq = rq_of(cfs_rq);
3165 cpufreq_trigger_update(rq_clock(rq));
3169 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3171 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3172 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3175 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3177 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3179 static inline int idle_balance(struct rq *rq)
3184 #endif /* CONFIG_SMP */
3186 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3188 #ifdef CONFIG_SCHEDSTATS
3189 struct task_struct *tsk = NULL;
3191 if (entity_is_task(se))
3194 if (se->statistics.sleep_start) {
3195 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3200 if (unlikely(delta > se->statistics.sleep_max))
3201 se->statistics.sleep_max = delta;
3203 se->statistics.sleep_start = 0;
3204 se->statistics.sum_sleep_runtime += delta;
3207 account_scheduler_latency(tsk, delta >> 10, 1);
3208 trace_sched_stat_sleep(tsk, delta);
3211 if (se->statistics.block_start) {
3212 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3217 if (unlikely(delta > se->statistics.block_max))
3218 se->statistics.block_max = delta;
3220 se->statistics.block_start = 0;
3221 se->statistics.sum_sleep_runtime += delta;
3224 if (tsk->in_iowait) {
3225 se->statistics.iowait_sum += delta;
3226 se->statistics.iowait_count++;
3227 trace_sched_stat_iowait(tsk, delta);
3230 trace_sched_stat_blocked(tsk, delta);
3233 * Blocking time is in units of nanosecs, so shift by
3234 * 20 to get a milliseconds-range estimation of the
3235 * amount of time that the task spent sleeping:
3237 if (unlikely(prof_on == SLEEP_PROFILING)) {
3238 profile_hits(SLEEP_PROFILING,
3239 (void *)get_wchan(tsk),
3242 account_scheduler_latency(tsk, delta >> 10, 0);
3248 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3250 #ifdef CONFIG_SCHED_DEBUG
3251 s64 d = se->vruntime - cfs_rq->min_vruntime;
3256 if (d > 3*sysctl_sched_latency)
3257 schedstat_inc(cfs_rq, nr_spread_over);
3262 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3264 u64 vruntime = cfs_rq->min_vruntime;
3267 * The 'current' period is already promised to the current tasks,
3268 * however the extra weight of the new task will slow them down a
3269 * little, place the new task so that it fits in the slot that
3270 * stays open at the end.
3272 if (initial && sched_feat(START_DEBIT))
3273 vruntime += sched_vslice(cfs_rq, se);
3275 /* sleeps up to a single latency don't count. */
3277 unsigned long thresh = sysctl_sched_latency;
3280 * Halve their sleep time's effect, to allow
3281 * for a gentler effect of sleepers:
3283 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3289 /* ensure we never gain time by being placed backwards. */
3290 se->vruntime = max_vruntime(se->vruntime, vruntime);
3293 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3295 static inline void check_schedstat_required(void)
3297 #ifdef CONFIG_SCHEDSTATS
3298 if (schedstat_enabled())
3301 /* Force schedstat enabled if a dependent tracepoint is active */
3302 if (trace_sched_stat_wait_enabled() ||
3303 trace_sched_stat_sleep_enabled() ||
3304 trace_sched_stat_iowait_enabled() ||
3305 trace_sched_stat_blocked_enabled() ||
3306 trace_sched_stat_runtime_enabled()) {
3307 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3308 "stat_blocked and stat_runtime require the "
3309 "kernel parameter schedstats=enabled or "
3310 "kernel.sched_schedstats=1\n");
3321 * update_min_vruntime()
3322 * vruntime -= min_vruntime
3326 * update_min_vruntime()
3327 * vruntime += min_vruntime
3329 * this way the vruntime transition between RQs is done when both
3330 * min_vruntime are up-to-date.
3334 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3335 * vruntime -= min_vruntime
3339 * update_min_vruntime()
3340 * vruntime += min_vruntime
3342 * this way we don't have the most up-to-date min_vruntime on the originating
3343 * CPU and an up-to-date min_vruntime on the destination CPU.
3347 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3349 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3350 bool curr = cfs_rq->curr == se;
3353 * If we're the current task, we must renormalise before calling
3357 se->vruntime += cfs_rq->min_vruntime;
3359 update_curr(cfs_rq);
3362 * Otherwise, renormalise after, such that we're placed at the current
3363 * moment in time, instead of some random moment in the past. Being
3364 * placed in the past could significantly boost this task to the
3365 * fairness detriment of existing tasks.
3367 if (renorm && !curr)
3368 se->vruntime += cfs_rq->min_vruntime;
3370 enqueue_entity_load_avg(cfs_rq, se);
3371 account_entity_enqueue(cfs_rq, se);
3372 update_cfs_shares(cfs_rq);
3374 if (flags & ENQUEUE_WAKEUP) {
3375 place_entity(cfs_rq, se, 0);
3376 if (schedstat_enabled())
3377 enqueue_sleeper(cfs_rq, se);
3380 check_schedstat_required();
3381 if (schedstat_enabled()) {
3382 update_stats_enqueue(cfs_rq, se);
3383 check_spread(cfs_rq, se);
3386 __enqueue_entity(cfs_rq, se);
3389 if (cfs_rq->nr_running == 1) {
3390 list_add_leaf_cfs_rq(cfs_rq);
3391 check_enqueue_throttle(cfs_rq);
3395 static void __clear_buddies_last(struct sched_entity *se)
3397 for_each_sched_entity(se) {
3398 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3399 if (cfs_rq->last != se)
3402 cfs_rq->last = NULL;
3406 static void __clear_buddies_next(struct sched_entity *se)
3408 for_each_sched_entity(se) {
3409 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3410 if (cfs_rq->next != se)
3413 cfs_rq->next = NULL;
3417 static void __clear_buddies_skip(struct sched_entity *se)
3419 for_each_sched_entity(se) {
3420 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3421 if (cfs_rq->skip != se)
3424 cfs_rq->skip = NULL;
3428 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3430 if (cfs_rq->last == se)
3431 __clear_buddies_last(se);
3433 if (cfs_rq->next == se)
3434 __clear_buddies_next(se);
3436 if (cfs_rq->skip == se)
3437 __clear_buddies_skip(se);
3440 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3443 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3446 * Update run-time statistics of the 'current'.
3448 update_curr(cfs_rq);
3449 dequeue_entity_load_avg(cfs_rq, se);
3451 if (schedstat_enabled())
3452 update_stats_dequeue(cfs_rq, se, flags);
3454 clear_buddies(cfs_rq, se);
3456 if (se != cfs_rq->curr)
3457 __dequeue_entity(cfs_rq, se);
3459 account_entity_dequeue(cfs_rq, se);
3462 * Normalize the entity after updating the min_vruntime because the
3463 * update can refer to the ->curr item and we need to reflect this
3464 * movement in our normalized position.
3466 if (!(flags & DEQUEUE_SLEEP))
3467 se->vruntime -= cfs_rq->min_vruntime;
3469 /* return excess runtime on last dequeue */
3470 return_cfs_rq_runtime(cfs_rq);
3472 update_min_vruntime(cfs_rq);
3473 update_cfs_shares(cfs_rq);
3477 * Preempt the current task with a newly woken task if needed:
3480 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3482 unsigned long ideal_runtime, delta_exec;
3483 struct sched_entity *se;
3486 ideal_runtime = sched_slice(cfs_rq, curr);
3487 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3488 if (delta_exec > ideal_runtime) {
3489 resched_curr(rq_of(cfs_rq));
3491 * The current task ran long enough, ensure it doesn't get
3492 * re-elected due to buddy favours.
3494 clear_buddies(cfs_rq, curr);
3499 * Ensure that a task that missed wakeup preemption by a
3500 * narrow margin doesn't have to wait for a full slice.
3501 * This also mitigates buddy induced latencies under load.
3503 if (delta_exec < sysctl_sched_min_granularity)
3506 se = __pick_first_entity(cfs_rq);
3507 delta = curr->vruntime - se->vruntime;
3512 if (delta > ideal_runtime)
3513 resched_curr(rq_of(cfs_rq));
3517 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3519 /* 'current' is not kept within the tree. */
3522 * Any task has to be enqueued before it get to execute on
3523 * a CPU. So account for the time it spent waiting on the
3526 if (schedstat_enabled())
3527 update_stats_wait_end(cfs_rq, se);
3528 __dequeue_entity(cfs_rq, se);
3529 update_load_avg(se, 1);
3532 update_stats_curr_start(cfs_rq, se);
3534 #ifdef CONFIG_SCHEDSTATS
3536 * Track our maximum slice length, if the CPU's load is at
3537 * least twice that of our own weight (i.e. dont track it
3538 * when there are only lesser-weight tasks around):
3540 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3541 se->statistics.slice_max = max(se->statistics.slice_max,
3542 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3545 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3549 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3552 * Pick the next process, keeping these things in mind, in this order:
3553 * 1) keep things fair between processes/task groups
3554 * 2) pick the "next" process, since someone really wants that to run
3555 * 3) pick the "last" process, for cache locality
3556 * 4) do not run the "skip" process, if something else is available
3558 static struct sched_entity *
3559 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3561 struct sched_entity *left = __pick_first_entity(cfs_rq);
3562 struct sched_entity *se;
3565 * If curr is set we have to see if its left of the leftmost entity
3566 * still in the tree, provided there was anything in the tree at all.
3568 if (!left || (curr && entity_before(curr, left)))
3571 se = left; /* ideally we run the leftmost entity */
3574 * Avoid running the skip buddy, if running something else can
3575 * be done without getting too unfair.
3577 if (cfs_rq->skip == se) {
3578 struct sched_entity *second;
3581 second = __pick_first_entity(cfs_rq);
3583 second = __pick_next_entity(se);
3584 if (!second || (curr && entity_before(curr, second)))
3588 if (second && wakeup_preempt_entity(second, left) < 1)
3593 * Prefer last buddy, try to return the CPU to a preempted task.
3595 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3599 * Someone really wants this to run. If it's not unfair, run it.
3601 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3604 clear_buddies(cfs_rq, se);
3609 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3611 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3614 * If still on the runqueue then deactivate_task()
3615 * was not called and update_curr() has to be done:
3618 update_curr(cfs_rq);
3620 /* throttle cfs_rqs exceeding runtime */
3621 check_cfs_rq_runtime(cfs_rq);
3623 if (schedstat_enabled()) {
3624 check_spread(cfs_rq, prev);
3626 update_stats_wait_start(cfs_rq, prev);
3630 /* Put 'current' back into the tree. */
3631 __enqueue_entity(cfs_rq, prev);
3632 /* in !on_rq case, update occurred at dequeue */
3633 update_load_avg(prev, 0);
3635 cfs_rq->curr = NULL;
3639 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3642 * Update run-time statistics of the 'current'.
3644 update_curr(cfs_rq);
3647 * Ensure that runnable average is periodically updated.
3649 update_load_avg(curr, 1);
3650 update_cfs_shares(cfs_rq);
3652 #ifdef CONFIG_SCHED_HRTICK
3654 * queued ticks are scheduled to match the slice, so don't bother
3655 * validating it and just reschedule.
3658 resched_curr(rq_of(cfs_rq));
3662 * don't let the period tick interfere with the hrtick preemption
3664 if (!sched_feat(DOUBLE_TICK) &&
3665 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3669 if (cfs_rq->nr_running > 1)
3670 check_preempt_tick(cfs_rq, curr);
3674 /**************************************************
3675 * CFS bandwidth control machinery
3678 #ifdef CONFIG_CFS_BANDWIDTH
3680 #ifdef HAVE_JUMP_LABEL
3681 static struct static_key __cfs_bandwidth_used;
3683 static inline bool cfs_bandwidth_used(void)
3685 return static_key_false(&__cfs_bandwidth_used);
3688 void cfs_bandwidth_usage_inc(void)
3690 static_key_slow_inc(&__cfs_bandwidth_used);
3693 void cfs_bandwidth_usage_dec(void)
3695 static_key_slow_dec(&__cfs_bandwidth_used);
3697 #else /* HAVE_JUMP_LABEL */
3698 static bool cfs_bandwidth_used(void)
3703 void cfs_bandwidth_usage_inc(void) {}
3704 void cfs_bandwidth_usage_dec(void) {}
3705 #endif /* HAVE_JUMP_LABEL */
3708 * default period for cfs group bandwidth.
3709 * default: 0.1s, units: nanoseconds
3711 static inline u64 default_cfs_period(void)
3713 return 100000000ULL;
3716 static inline u64 sched_cfs_bandwidth_slice(void)
3718 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3722 * Replenish runtime according to assigned quota and update expiration time.
3723 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3724 * additional synchronization around rq->lock.
3726 * requires cfs_b->lock
3728 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3732 if (cfs_b->quota == RUNTIME_INF)
3735 now = sched_clock_cpu(smp_processor_id());
3736 cfs_b->runtime = cfs_b->quota;
3737 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3740 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3742 return &tg->cfs_bandwidth;
3745 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3746 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3748 if (unlikely(cfs_rq->throttle_count))
3749 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3751 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3754 /* returns 0 on failure to allocate runtime */
3755 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3757 struct task_group *tg = cfs_rq->tg;
3758 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3759 u64 amount = 0, min_amount, expires;
3761 /* note: this is a positive sum as runtime_remaining <= 0 */
3762 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3764 raw_spin_lock(&cfs_b->lock);
3765 if (cfs_b->quota == RUNTIME_INF)
3766 amount = min_amount;
3768 start_cfs_bandwidth(cfs_b);
3770 if (cfs_b->runtime > 0) {
3771 amount = min(cfs_b->runtime, min_amount);
3772 cfs_b->runtime -= amount;
3776 expires = cfs_b->runtime_expires;
3777 raw_spin_unlock(&cfs_b->lock);
3779 cfs_rq->runtime_remaining += amount;
3781 * we may have advanced our local expiration to account for allowed
3782 * spread between our sched_clock and the one on which runtime was
3785 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3786 cfs_rq->runtime_expires = expires;
3788 return cfs_rq->runtime_remaining > 0;
3792 * Note: This depends on the synchronization provided by sched_clock and the
3793 * fact that rq->clock snapshots this value.
3795 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3797 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3799 /* if the deadline is ahead of our clock, nothing to do */
3800 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3803 if (cfs_rq->runtime_remaining < 0)
3807 * If the local deadline has passed we have to consider the
3808 * possibility that our sched_clock is 'fast' and the global deadline
3809 * has not truly expired.
3811 * Fortunately we can check determine whether this the case by checking
3812 * whether the global deadline has advanced. It is valid to compare
3813 * cfs_b->runtime_expires without any locks since we only care about
3814 * exact equality, so a partial write will still work.
3817 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3818 /* extend local deadline, drift is bounded above by 2 ticks */
3819 cfs_rq->runtime_expires += TICK_NSEC;
3821 /* global deadline is ahead, expiration has passed */
3822 cfs_rq->runtime_remaining = 0;
3826 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3828 /* dock delta_exec before expiring quota (as it could span periods) */
3829 cfs_rq->runtime_remaining -= delta_exec;
3830 expire_cfs_rq_runtime(cfs_rq);
3832 if (likely(cfs_rq->runtime_remaining > 0))
3836 * if we're unable to extend our runtime we resched so that the active
3837 * hierarchy can be throttled
3839 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3840 resched_curr(rq_of(cfs_rq));
3843 static __always_inline
3844 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3846 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3849 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3852 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3854 return cfs_bandwidth_used() && cfs_rq->throttled;
3857 /* check whether cfs_rq, or any parent, is throttled */
3858 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3860 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3864 * Ensure that neither of the group entities corresponding to src_cpu or
3865 * dest_cpu are members of a throttled hierarchy when performing group
3866 * load-balance operations.
3868 static inline int throttled_lb_pair(struct task_group *tg,
3869 int src_cpu, int dest_cpu)
3871 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3873 src_cfs_rq = tg->cfs_rq[src_cpu];
3874 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3876 return throttled_hierarchy(src_cfs_rq) ||
3877 throttled_hierarchy(dest_cfs_rq);
3880 /* updated child weight may affect parent so we have to do this bottom up */
3881 static int tg_unthrottle_up(struct task_group *tg, void *data)
3883 struct rq *rq = data;
3884 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3886 cfs_rq->throttle_count--;
3887 if (!cfs_rq->throttle_count) {
3888 /* adjust cfs_rq_clock_task() */
3889 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3890 cfs_rq->throttled_clock_task;
3896 static int tg_throttle_down(struct task_group *tg, void *data)
3898 struct rq *rq = data;
3899 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3901 /* group is entering throttled state, stop time */
3902 if (!cfs_rq->throttle_count)
3903 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3904 cfs_rq->throttle_count++;
3909 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3911 struct rq *rq = rq_of(cfs_rq);
3912 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3913 struct sched_entity *se;
3914 long task_delta, dequeue = 1;
3917 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3919 /* freeze hierarchy runnable averages while throttled */
3921 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3924 task_delta = cfs_rq->h_nr_running;
3925 for_each_sched_entity(se) {
3926 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3927 /* throttled entity or throttle-on-deactivate */
3932 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3933 qcfs_rq->h_nr_running -= task_delta;
3935 if (qcfs_rq->load.weight)
3940 sub_nr_running(rq, task_delta);
3942 cfs_rq->throttled = 1;
3943 cfs_rq->throttled_clock = rq_clock(rq);
3944 raw_spin_lock(&cfs_b->lock);
3945 empty = list_empty(&cfs_b->throttled_cfs_rq);
3948 * Add to the _head_ of the list, so that an already-started
3949 * distribute_cfs_runtime will not see us
3951 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3954 * If we're the first throttled task, make sure the bandwidth
3958 start_cfs_bandwidth(cfs_b);
3960 raw_spin_unlock(&cfs_b->lock);
3963 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3965 struct rq *rq = rq_of(cfs_rq);
3966 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3967 struct sched_entity *se;
3971 se = cfs_rq->tg->se[cpu_of(rq)];
3973 cfs_rq->throttled = 0;
3975 update_rq_clock(rq);
3977 raw_spin_lock(&cfs_b->lock);
3978 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3979 list_del_rcu(&cfs_rq->throttled_list);
3980 raw_spin_unlock(&cfs_b->lock);
3982 /* update hierarchical throttle state */
3983 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3985 if (!cfs_rq->load.weight)
3988 task_delta = cfs_rq->h_nr_running;
3989 for_each_sched_entity(se) {
3993 cfs_rq = cfs_rq_of(se);
3995 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3996 cfs_rq->h_nr_running += task_delta;
3998 if (cfs_rq_throttled(cfs_rq))
4003 add_nr_running(rq, task_delta);
4005 /* determine whether we need to wake up potentially idle cpu */
4006 if (rq->curr == rq->idle && rq->cfs.nr_running)
4010 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4011 u64 remaining, u64 expires)
4013 struct cfs_rq *cfs_rq;
4015 u64 starting_runtime = remaining;
4018 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4020 struct rq *rq = rq_of(cfs_rq);
4022 raw_spin_lock(&rq->lock);
4023 if (!cfs_rq_throttled(cfs_rq))
4026 runtime = -cfs_rq->runtime_remaining + 1;
4027 if (runtime > remaining)
4028 runtime = remaining;
4029 remaining -= runtime;
4031 cfs_rq->runtime_remaining += runtime;
4032 cfs_rq->runtime_expires = expires;
4034 /* we check whether we're throttled above */
4035 if (cfs_rq->runtime_remaining > 0)
4036 unthrottle_cfs_rq(cfs_rq);
4039 raw_spin_unlock(&rq->lock);
4046 return starting_runtime - remaining;
4050 * Responsible for refilling a task_group's bandwidth and unthrottling its
4051 * cfs_rqs as appropriate. If there has been no activity within the last
4052 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4053 * used to track this state.
4055 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4057 u64 runtime, runtime_expires;
4060 /* no need to continue the timer with no bandwidth constraint */
4061 if (cfs_b->quota == RUNTIME_INF)
4062 goto out_deactivate;
4064 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4065 cfs_b->nr_periods += overrun;
4068 * idle depends on !throttled (for the case of a large deficit), and if
4069 * we're going inactive then everything else can be deferred
4071 if (cfs_b->idle && !throttled)
4072 goto out_deactivate;
4074 __refill_cfs_bandwidth_runtime(cfs_b);
4077 /* mark as potentially idle for the upcoming period */
4082 /* account preceding periods in which throttling occurred */
4083 cfs_b->nr_throttled += overrun;
4085 runtime_expires = cfs_b->runtime_expires;
4088 * This check is repeated as we are holding onto the new bandwidth while
4089 * we unthrottle. This can potentially race with an unthrottled group
4090 * trying to acquire new bandwidth from the global pool. This can result
4091 * in us over-using our runtime if it is all used during this loop, but
4092 * only by limited amounts in that extreme case.
4094 while (throttled && cfs_b->runtime > 0) {
4095 runtime = cfs_b->runtime;
4096 raw_spin_unlock(&cfs_b->lock);
4097 /* we can't nest cfs_b->lock while distributing bandwidth */
4098 runtime = distribute_cfs_runtime(cfs_b, runtime,
4100 raw_spin_lock(&cfs_b->lock);
4102 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4104 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4108 * While we are ensured activity in the period following an
4109 * unthrottle, this also covers the case in which the new bandwidth is
4110 * insufficient to cover the existing bandwidth deficit. (Forcing the
4111 * timer to remain active while there are any throttled entities.)
4121 /* a cfs_rq won't donate quota below this amount */
4122 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4123 /* minimum remaining period time to redistribute slack quota */
4124 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4125 /* how long we wait to gather additional slack before distributing */
4126 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4129 * Are we near the end of the current quota period?
4131 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4132 * hrtimer base being cleared by hrtimer_start. In the case of
4133 * migrate_hrtimers, base is never cleared, so we are fine.
4135 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4137 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4140 /* if the call-back is running a quota refresh is already occurring */
4141 if (hrtimer_callback_running(refresh_timer))
4144 /* is a quota refresh about to occur? */
4145 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4146 if (remaining < min_expire)
4152 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4154 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4156 /* if there's a quota refresh soon don't bother with slack */
4157 if (runtime_refresh_within(cfs_b, min_left))
4160 hrtimer_start(&cfs_b->slack_timer,
4161 ns_to_ktime(cfs_bandwidth_slack_period),
4165 /* we know any runtime found here is valid as update_curr() precedes return */
4166 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4168 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4169 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4171 if (slack_runtime <= 0)
4174 raw_spin_lock(&cfs_b->lock);
4175 if (cfs_b->quota != RUNTIME_INF &&
4176 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4177 cfs_b->runtime += slack_runtime;
4179 /* we are under rq->lock, defer unthrottling using a timer */
4180 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4181 !list_empty(&cfs_b->throttled_cfs_rq))
4182 start_cfs_slack_bandwidth(cfs_b);
4184 raw_spin_unlock(&cfs_b->lock);
4186 /* even if it's not valid for return we don't want to try again */
4187 cfs_rq->runtime_remaining -= slack_runtime;
4190 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4192 if (!cfs_bandwidth_used())
4195 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4198 __return_cfs_rq_runtime(cfs_rq);
4202 * This is done with a timer (instead of inline with bandwidth return) since
4203 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4205 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4207 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4210 /* confirm we're still not at a refresh boundary */
4211 raw_spin_lock(&cfs_b->lock);
4212 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4213 raw_spin_unlock(&cfs_b->lock);
4217 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4218 runtime = cfs_b->runtime;
4220 expires = cfs_b->runtime_expires;
4221 raw_spin_unlock(&cfs_b->lock);
4226 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4228 raw_spin_lock(&cfs_b->lock);
4229 if (expires == cfs_b->runtime_expires)
4230 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4231 raw_spin_unlock(&cfs_b->lock);
4235 * When a group wakes up we want to make sure that its quota is not already
4236 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4237 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4239 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4241 if (!cfs_bandwidth_used())
4244 /* Synchronize hierarchical throttle counter: */
4245 if (unlikely(!cfs_rq->throttle_uptodate)) {
4246 struct rq *rq = rq_of(cfs_rq);
4247 struct cfs_rq *pcfs_rq;
4248 struct task_group *tg;
4250 cfs_rq->throttle_uptodate = 1;
4252 /* Get closest up-to-date node, because leaves go first: */
4253 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4254 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4255 if (pcfs_rq->throttle_uptodate)
4259 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4260 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4264 /* an active group must be handled by the update_curr()->put() path */
4265 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4268 /* ensure the group is not already throttled */
4269 if (cfs_rq_throttled(cfs_rq))
4272 /* update runtime allocation */
4273 account_cfs_rq_runtime(cfs_rq, 0);
4274 if (cfs_rq->runtime_remaining <= 0)
4275 throttle_cfs_rq(cfs_rq);
4278 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4279 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4281 if (!cfs_bandwidth_used())
4284 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4288 * it's possible for a throttled entity to be forced into a running
4289 * state (e.g. set_curr_task), in this case we're finished.
4291 if (cfs_rq_throttled(cfs_rq))
4294 throttle_cfs_rq(cfs_rq);
4298 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4300 struct cfs_bandwidth *cfs_b =
4301 container_of(timer, struct cfs_bandwidth, slack_timer);
4303 do_sched_cfs_slack_timer(cfs_b);
4305 return HRTIMER_NORESTART;
4308 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4310 struct cfs_bandwidth *cfs_b =
4311 container_of(timer, struct cfs_bandwidth, period_timer);
4315 raw_spin_lock(&cfs_b->lock);
4317 overrun = hrtimer_forward_now(timer, cfs_b->period);
4321 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4324 cfs_b->period_active = 0;
4325 raw_spin_unlock(&cfs_b->lock);
4327 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4330 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4332 raw_spin_lock_init(&cfs_b->lock);
4334 cfs_b->quota = RUNTIME_INF;
4335 cfs_b->period = ns_to_ktime(default_cfs_period());
4337 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4338 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4339 cfs_b->period_timer.function = sched_cfs_period_timer;
4340 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4341 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4344 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4346 cfs_rq->runtime_enabled = 0;
4347 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4350 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4352 lockdep_assert_held(&cfs_b->lock);
4354 if (!cfs_b->period_active) {
4355 cfs_b->period_active = 1;
4356 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4357 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4361 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4363 /* init_cfs_bandwidth() was not called */
4364 if (!cfs_b->throttled_cfs_rq.next)
4367 hrtimer_cancel(&cfs_b->period_timer);
4368 hrtimer_cancel(&cfs_b->slack_timer);
4371 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4373 struct cfs_rq *cfs_rq;
4375 for_each_leaf_cfs_rq(rq, cfs_rq) {
4376 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4378 raw_spin_lock(&cfs_b->lock);
4379 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4380 raw_spin_unlock(&cfs_b->lock);
4384 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4386 struct cfs_rq *cfs_rq;
4388 for_each_leaf_cfs_rq(rq, cfs_rq) {
4389 if (!cfs_rq->runtime_enabled)
4393 * clock_task is not advancing so we just need to make sure
4394 * there's some valid quota amount
4396 cfs_rq->runtime_remaining = 1;
4398 * Offline rq is schedulable till cpu is completely disabled
4399 * in take_cpu_down(), so we prevent new cfs throttling here.
4401 cfs_rq->runtime_enabled = 0;
4403 if (cfs_rq_throttled(cfs_rq))
4404 unthrottle_cfs_rq(cfs_rq);
4408 #else /* CONFIG_CFS_BANDWIDTH */
4409 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4411 return rq_clock_task(rq_of(cfs_rq));
4414 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4415 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4416 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4417 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4419 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4424 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4429 static inline int throttled_lb_pair(struct task_group *tg,
4430 int src_cpu, int dest_cpu)
4435 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4437 #ifdef CONFIG_FAIR_GROUP_SCHED
4438 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4441 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4445 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4446 static inline void update_runtime_enabled(struct rq *rq) {}
4447 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4449 #endif /* CONFIG_CFS_BANDWIDTH */
4451 /**************************************************
4452 * CFS operations on tasks:
4455 #ifdef CONFIG_SCHED_HRTICK
4456 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4458 struct sched_entity *se = &p->se;
4459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4461 WARN_ON(task_rq(p) != rq);
4463 if (cfs_rq->nr_running > 1) {
4464 u64 slice = sched_slice(cfs_rq, se);
4465 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4466 s64 delta = slice - ran;
4473 hrtick_start(rq, delta);
4478 * called from enqueue/dequeue and updates the hrtick when the
4479 * current task is from our class and nr_running is low enough
4482 static void hrtick_update(struct rq *rq)
4484 struct task_struct *curr = rq->curr;
4486 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4489 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4490 hrtick_start_fair(rq, curr);
4492 #else /* !CONFIG_SCHED_HRTICK */
4494 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4498 static inline void hrtick_update(struct rq *rq)
4504 * The enqueue_task method is called before nr_running is
4505 * increased. Here we update the fair scheduling stats and
4506 * then put the task into the rbtree:
4509 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4511 struct cfs_rq *cfs_rq;
4512 struct sched_entity *se = &p->se;
4514 for_each_sched_entity(se) {
4517 cfs_rq = cfs_rq_of(se);
4518 enqueue_entity(cfs_rq, se, flags);
4521 * end evaluation on encountering a throttled cfs_rq
4523 * note: in the case of encountering a throttled cfs_rq we will
4524 * post the final h_nr_running increment below.
4526 if (cfs_rq_throttled(cfs_rq))
4528 cfs_rq->h_nr_running++;
4530 flags = ENQUEUE_WAKEUP;
4533 for_each_sched_entity(se) {
4534 cfs_rq = cfs_rq_of(se);
4535 cfs_rq->h_nr_running++;
4537 if (cfs_rq_throttled(cfs_rq))
4540 update_load_avg(se, 1);
4541 update_cfs_shares(cfs_rq);
4545 add_nr_running(rq, 1);
4550 static void set_next_buddy(struct sched_entity *se);
4553 * The dequeue_task method is called before nr_running is
4554 * decreased. We remove the task from the rbtree and
4555 * update the fair scheduling stats:
4557 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4559 struct cfs_rq *cfs_rq;
4560 struct sched_entity *se = &p->se;
4561 int task_sleep = flags & DEQUEUE_SLEEP;
4563 for_each_sched_entity(se) {
4564 cfs_rq = cfs_rq_of(se);
4565 dequeue_entity(cfs_rq, se, flags);
4568 * end evaluation on encountering a throttled cfs_rq
4570 * note: in the case of encountering a throttled cfs_rq we will
4571 * post the final h_nr_running decrement below.
4573 if (cfs_rq_throttled(cfs_rq))
4575 cfs_rq->h_nr_running--;
4577 /* Don't dequeue parent if it has other entities besides us */
4578 if (cfs_rq->load.weight) {
4579 /* Avoid re-evaluating load for this entity: */
4580 se = parent_entity(se);
4582 * Bias pick_next to pick a task from this cfs_rq, as
4583 * p is sleeping when it is within its sched_slice.
4585 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4589 flags |= DEQUEUE_SLEEP;
4592 for_each_sched_entity(se) {
4593 cfs_rq = cfs_rq_of(se);
4594 cfs_rq->h_nr_running--;
4596 if (cfs_rq_throttled(cfs_rq))
4599 update_load_avg(se, 1);
4600 update_cfs_shares(cfs_rq);
4604 sub_nr_running(rq, 1);
4610 #ifdef CONFIG_NO_HZ_COMMON
4612 * per rq 'load' arrray crap; XXX kill this.
4616 * The exact cpuload calculated at every tick would be:
4618 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4620 * If a cpu misses updates for n ticks (as it was idle) and update gets
4621 * called on the n+1-th tick when cpu may be busy, then we have:
4623 * load_n = (1 - 1/2^i)^n * load_0
4624 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4626 * decay_load_missed() below does efficient calculation of
4628 * load' = (1 - 1/2^i)^n * load
4630 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4631 * This allows us to precompute the above in said factors, thereby allowing the
4632 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4633 * fixed_power_int())
4635 * The calculation is approximated on a 128 point scale.
4637 #define DEGRADE_SHIFT 7
4639 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4640 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4641 { 0, 0, 0, 0, 0, 0, 0, 0 },
4642 { 64, 32, 8, 0, 0, 0, 0, 0 },
4643 { 96, 72, 40, 12, 1, 0, 0, 0 },
4644 { 112, 98, 75, 43, 15, 1, 0, 0 },
4645 { 120, 112, 98, 76, 45, 16, 2, 0 }
4649 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4650 * would be when CPU is idle and so we just decay the old load without
4651 * adding any new load.
4653 static unsigned long
4654 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4658 if (!missed_updates)
4661 if (missed_updates >= degrade_zero_ticks[idx])
4665 return load >> missed_updates;
4667 while (missed_updates) {
4668 if (missed_updates % 2)
4669 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4671 missed_updates >>= 1;
4676 #endif /* CONFIG_NO_HZ_COMMON */
4679 * __cpu_load_update - update the rq->cpu_load[] statistics
4680 * @this_rq: The rq to update statistics for
4681 * @this_load: The current load
4682 * @pending_updates: The number of missed updates
4684 * Update rq->cpu_load[] statistics. This function is usually called every
4685 * scheduler tick (TICK_NSEC).
4687 * This function computes a decaying average:
4689 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4691 * Because of NOHZ it might not get called on every tick which gives need for
4692 * the @pending_updates argument.
4694 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4695 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4696 * = A * (A * load[i]_n-2 + B) + B
4697 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4698 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4699 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4700 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4701 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4703 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4704 * any change in load would have resulted in the tick being turned back on.
4706 * For regular NOHZ, this reduces to:
4708 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4710 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4713 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4714 unsigned long pending_updates)
4716 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4719 this_rq->nr_load_updates++;
4721 /* Update our load: */
4722 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4723 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4724 unsigned long old_load, new_load;
4726 /* scale is effectively 1 << i now, and >> i divides by scale */
4728 old_load = this_rq->cpu_load[i];
4729 #ifdef CONFIG_NO_HZ_COMMON
4730 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4731 if (tickless_load) {
4732 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4734 * old_load can never be a negative value because a
4735 * decayed tickless_load cannot be greater than the
4736 * original tickless_load.
4738 old_load += tickless_load;
4741 new_load = this_load;
4743 * Round up the averaging division if load is increasing. This
4744 * prevents us from getting stuck on 9 if the load is 10, for
4747 if (new_load > old_load)
4748 new_load += scale - 1;
4750 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4753 sched_avg_update(this_rq);
4756 /* Used instead of source_load when we know the type == 0 */
4757 static unsigned long weighted_cpuload(const int cpu)
4759 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4762 #ifdef CONFIG_NO_HZ_COMMON
4764 * There is no sane way to deal with nohz on smp when using jiffies because the
4765 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4766 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4768 * Therefore we need to avoid the delta approach from the regular tick when
4769 * possible since that would seriously skew the load calculation. This is why we
4770 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4771 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4772 * loop exit, nohz_idle_balance, nohz full exit...)
4774 * This means we might still be one tick off for nohz periods.
4777 static void cpu_load_update_nohz(struct rq *this_rq,
4778 unsigned long curr_jiffies,
4781 unsigned long pending_updates;
4783 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4784 if (pending_updates) {
4785 this_rq->last_load_update_tick = curr_jiffies;
4787 * In the regular NOHZ case, we were idle, this means load 0.
4788 * In the NOHZ_FULL case, we were non-idle, we should consider
4789 * its weighted load.
4791 cpu_load_update(this_rq, load, pending_updates);
4796 * Called from nohz_idle_balance() to update the load ratings before doing the
4799 static void cpu_load_update_idle(struct rq *this_rq)
4802 * bail if there's load or we're actually up-to-date.
4804 if (weighted_cpuload(cpu_of(this_rq)))
4807 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4811 * Record CPU load on nohz entry so we know the tickless load to account
4812 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4813 * than other cpu_load[idx] but it should be fine as cpu_load readers
4814 * shouldn't rely into synchronized cpu_load[*] updates.
4816 void cpu_load_update_nohz_start(void)
4818 struct rq *this_rq = this_rq();
4821 * This is all lockless but should be fine. If weighted_cpuload changes
4822 * concurrently we'll exit nohz. And cpu_load write can race with
4823 * cpu_load_update_idle() but both updater would be writing the same.
4825 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4829 * Account the tickless load in the end of a nohz frame.
4831 void cpu_load_update_nohz_stop(void)
4833 unsigned long curr_jiffies = READ_ONCE(jiffies);
4834 struct rq *this_rq = this_rq();
4837 if (curr_jiffies == this_rq->last_load_update_tick)
4840 load = weighted_cpuload(cpu_of(this_rq));
4841 raw_spin_lock(&this_rq->lock);
4842 update_rq_clock(this_rq);
4843 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4844 raw_spin_unlock(&this_rq->lock);
4846 #else /* !CONFIG_NO_HZ_COMMON */
4847 static inline void cpu_load_update_nohz(struct rq *this_rq,
4848 unsigned long curr_jiffies,
4849 unsigned long load) { }
4850 #endif /* CONFIG_NO_HZ_COMMON */
4852 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4854 #ifdef CONFIG_NO_HZ_COMMON
4855 /* See the mess around cpu_load_update_nohz(). */
4856 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4858 cpu_load_update(this_rq, load, 1);
4862 * Called from scheduler_tick()
4864 void cpu_load_update_active(struct rq *this_rq)
4866 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4868 if (tick_nohz_tick_stopped())
4869 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4871 cpu_load_update_periodic(this_rq, load);
4875 * Return a low guess at the load of a migration-source cpu weighted
4876 * according to the scheduling class and "nice" value.
4878 * We want to under-estimate the load of migration sources, to
4879 * balance conservatively.
4881 static unsigned long source_load(int cpu, int type)
4883 struct rq *rq = cpu_rq(cpu);
4884 unsigned long total = weighted_cpuload(cpu);
4886 if (type == 0 || !sched_feat(LB_BIAS))
4889 return min(rq->cpu_load[type-1], total);
4893 * Return a high guess at the load of a migration-target cpu weighted
4894 * according to the scheduling class and "nice" value.
4896 static unsigned long target_load(int cpu, int type)
4898 struct rq *rq = cpu_rq(cpu);
4899 unsigned long total = weighted_cpuload(cpu);
4901 if (type == 0 || !sched_feat(LB_BIAS))
4904 return max(rq->cpu_load[type-1], total);
4907 static unsigned long capacity_of(int cpu)
4909 return cpu_rq(cpu)->cpu_capacity;
4912 static unsigned long capacity_orig_of(int cpu)
4914 return cpu_rq(cpu)->cpu_capacity_orig;
4917 static unsigned long cpu_avg_load_per_task(int cpu)
4919 struct rq *rq = cpu_rq(cpu);
4920 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4921 unsigned long load_avg = weighted_cpuload(cpu);
4924 return load_avg / nr_running;
4929 #ifdef CONFIG_FAIR_GROUP_SCHED
4931 * effective_load() calculates the load change as seen from the root_task_group
4933 * Adding load to a group doesn't make a group heavier, but can cause movement
4934 * of group shares between cpus. Assuming the shares were perfectly aligned one
4935 * can calculate the shift in shares.
4937 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4938 * on this @cpu and results in a total addition (subtraction) of @wg to the
4939 * total group weight.
4941 * Given a runqueue weight distribution (rw_i) we can compute a shares
4942 * distribution (s_i) using:
4944 * s_i = rw_i / \Sum rw_j (1)
4946 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4947 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4948 * shares distribution (s_i):
4950 * rw_i = { 2, 4, 1, 0 }
4951 * s_i = { 2/7, 4/7, 1/7, 0 }
4953 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4954 * task used to run on and the CPU the waker is running on), we need to
4955 * compute the effect of waking a task on either CPU and, in case of a sync
4956 * wakeup, compute the effect of the current task going to sleep.
4958 * So for a change of @wl to the local @cpu with an overall group weight change
4959 * of @wl we can compute the new shares distribution (s'_i) using:
4961 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4963 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4964 * differences in waking a task to CPU 0. The additional task changes the
4965 * weight and shares distributions like:
4967 * rw'_i = { 3, 4, 1, 0 }
4968 * s'_i = { 3/8, 4/8, 1/8, 0 }
4970 * We can then compute the difference in effective weight by using:
4972 * dw_i = S * (s'_i - s_i) (3)
4974 * Where 'S' is the group weight as seen by its parent.
4976 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4977 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4978 * 4/7) times the weight of the group.
4980 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4982 struct sched_entity *se = tg->se[cpu];
4984 if (!tg->parent) /* the trivial, non-cgroup case */
4987 for_each_sched_entity(se) {
4988 struct cfs_rq *cfs_rq = se->my_q;
4989 long W, w = cfs_rq_load_avg(cfs_rq);
4994 * W = @wg + \Sum rw_j
4996 W = wg + atomic_long_read(&tg->load_avg);
4998 /* Ensure \Sum rw_j >= rw_i */
4999 W -= cfs_rq->tg_load_avg_contrib;
5008 * wl = S * s'_i; see (2)
5011 wl = (w * (long)tg->shares) / W;
5016 * Per the above, wl is the new se->load.weight value; since
5017 * those are clipped to [MIN_SHARES, ...) do so now. See
5018 * calc_cfs_shares().
5020 if (wl < MIN_SHARES)
5024 * wl = dw_i = S * (s'_i - s_i); see (3)
5026 wl -= se->avg.load_avg;
5029 * Recursively apply this logic to all parent groups to compute
5030 * the final effective load change on the root group. Since
5031 * only the @tg group gets extra weight, all parent groups can
5032 * only redistribute existing shares. @wl is the shift in shares
5033 * resulting from this level per the above.
5042 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5049 static void record_wakee(struct task_struct *p)
5052 * Only decay a single time; tasks that have less then 1 wakeup per
5053 * jiffy will not have built up many flips.
5055 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5056 current->wakee_flips >>= 1;
5057 current->wakee_flip_decay_ts = jiffies;
5060 if (current->last_wakee != p) {
5061 current->last_wakee = p;
5062 current->wakee_flips++;
5067 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5069 * A waker of many should wake a different task than the one last awakened
5070 * at a frequency roughly N times higher than one of its wakees.
5072 * In order to determine whether we should let the load spread vs consolidating
5073 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5074 * partner, and a factor of lls_size higher frequency in the other.
5076 * With both conditions met, we can be relatively sure that the relationship is
5077 * non-monogamous, with partner count exceeding socket size.
5079 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5080 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5083 static int wake_wide(struct task_struct *p)
5085 unsigned int master = current->wakee_flips;
5086 unsigned int slave = p->wakee_flips;
5087 int factor = this_cpu_read(sd_llc_size);
5090 swap(master, slave);
5091 if (slave < factor || master < slave * factor)
5096 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5098 s64 this_load, load;
5099 s64 this_eff_load, prev_eff_load;
5100 int idx, this_cpu, prev_cpu;
5101 struct task_group *tg;
5102 unsigned long weight;
5106 this_cpu = smp_processor_id();
5107 prev_cpu = task_cpu(p);
5108 load = source_load(prev_cpu, idx);
5109 this_load = target_load(this_cpu, idx);
5112 * If sync wakeup then subtract the (maximum possible)
5113 * effect of the currently running task from the load
5114 * of the current CPU:
5117 tg = task_group(current);
5118 weight = current->se.avg.load_avg;
5120 this_load += effective_load(tg, this_cpu, -weight, -weight);
5121 load += effective_load(tg, prev_cpu, 0, -weight);
5125 weight = p->se.avg.load_avg;
5128 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5129 * due to the sync cause above having dropped this_load to 0, we'll
5130 * always have an imbalance, but there's really nothing you can do
5131 * about that, so that's good too.
5133 * Otherwise check if either cpus are near enough in load to allow this
5134 * task to be woken on this_cpu.
5136 this_eff_load = 100;
5137 this_eff_load *= capacity_of(prev_cpu);
5139 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5140 prev_eff_load *= capacity_of(this_cpu);
5142 if (this_load > 0) {
5143 this_eff_load *= this_load +
5144 effective_load(tg, this_cpu, weight, weight);
5146 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5149 balanced = this_eff_load <= prev_eff_load;
5151 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5156 schedstat_inc(sd, ttwu_move_affine);
5157 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5163 * find_idlest_group finds and returns the least busy CPU group within the
5166 static struct sched_group *
5167 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5168 int this_cpu, int sd_flag)
5170 struct sched_group *idlest = NULL, *group = sd->groups;
5171 unsigned long min_load = ULONG_MAX, this_load = 0;
5172 int load_idx = sd->forkexec_idx;
5173 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5175 if (sd_flag & SD_BALANCE_WAKE)
5176 load_idx = sd->wake_idx;
5179 unsigned long load, avg_load;
5183 /* Skip over this group if it has no CPUs allowed */
5184 if (!cpumask_intersects(sched_group_cpus(group),
5185 tsk_cpus_allowed(p)))
5188 local_group = cpumask_test_cpu(this_cpu,
5189 sched_group_cpus(group));
5191 /* Tally up the load of all CPUs in the group */
5194 for_each_cpu(i, sched_group_cpus(group)) {
5195 /* Bias balancing toward cpus of our domain */
5197 load = source_load(i, load_idx);
5199 load = target_load(i, load_idx);
5204 /* Adjust by relative CPU capacity of the group */
5205 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5208 this_load = avg_load;
5209 } else if (avg_load < min_load) {
5210 min_load = avg_load;
5213 } while (group = group->next, group != sd->groups);
5215 if (!idlest || 100*this_load < imbalance*min_load)
5221 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5224 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5226 unsigned long load, min_load = ULONG_MAX;
5227 unsigned int min_exit_latency = UINT_MAX;
5228 u64 latest_idle_timestamp = 0;
5229 int least_loaded_cpu = this_cpu;
5230 int shallowest_idle_cpu = -1;
5233 /* Traverse only the allowed CPUs */
5234 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5236 struct rq *rq = cpu_rq(i);
5237 struct cpuidle_state *idle = idle_get_state(rq);
5238 if (idle && idle->exit_latency < min_exit_latency) {
5240 * We give priority to a CPU whose idle state
5241 * has the smallest exit latency irrespective
5242 * of any idle timestamp.
5244 min_exit_latency = idle->exit_latency;
5245 latest_idle_timestamp = rq->idle_stamp;
5246 shallowest_idle_cpu = i;
5247 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5248 rq->idle_stamp > latest_idle_timestamp) {
5250 * If equal or no active idle state, then
5251 * the most recently idled CPU might have
5254 latest_idle_timestamp = rq->idle_stamp;
5255 shallowest_idle_cpu = i;
5257 } else if (shallowest_idle_cpu == -1) {
5258 load = weighted_cpuload(i);
5259 if (load < min_load || (load == min_load && i == this_cpu)) {
5261 least_loaded_cpu = i;
5266 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5270 * Try and locate an idle CPU in the sched_domain.
5272 static int select_idle_sibling(struct task_struct *p, int target)
5274 struct sched_domain *sd;
5275 struct sched_group *sg;
5276 int i = task_cpu(p);
5278 if (idle_cpu(target))
5282 * If the prevous cpu is cache affine and idle, don't be stupid.
5284 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5288 * Otherwise, iterate the domains and find an eligible idle cpu.
5290 * A completely idle sched group at higher domains is more
5291 * desirable than an idle group at a lower level, because lower
5292 * domains have smaller groups and usually share hardware
5293 * resources which causes tasks to contend on them, e.g. x86
5294 * hyperthread siblings in the lowest domain (SMT) can contend
5295 * on the shared cpu pipeline.
5297 * However, while we prefer idle groups at higher domains
5298 * finding an idle cpu at the lowest domain is still better than
5299 * returning 'target', which we've already established, isn't
5302 sd = rcu_dereference(per_cpu(sd_llc, target));
5303 for_each_lower_domain(sd) {
5306 if (!cpumask_intersects(sched_group_cpus(sg),
5307 tsk_cpus_allowed(p)))
5310 /* Ensure the entire group is idle */
5311 for_each_cpu(i, sched_group_cpus(sg)) {
5312 if (i == target || !idle_cpu(i))
5317 * It doesn't matter which cpu we pick, the
5318 * whole group is idle.
5320 target = cpumask_first_and(sched_group_cpus(sg),
5321 tsk_cpus_allowed(p));
5325 } while (sg != sd->groups);
5332 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5333 * tasks. The unit of the return value must be the one of capacity so we can
5334 * compare the utilization with the capacity of the CPU that is available for
5335 * CFS task (ie cpu_capacity).
5337 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5338 * recent utilization of currently non-runnable tasks on a CPU. It represents
5339 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5340 * capacity_orig is the cpu_capacity available at the highest frequency
5341 * (arch_scale_freq_capacity()).
5342 * The utilization of a CPU converges towards a sum equal to or less than the
5343 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5344 * the running time on this CPU scaled by capacity_curr.
5346 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5347 * higher than capacity_orig because of unfortunate rounding in
5348 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5349 * the average stabilizes with the new running time. We need to check that the
5350 * utilization stays within the range of [0..capacity_orig] and cap it if
5351 * necessary. Without utilization capping, a group could be seen as overloaded
5352 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5353 * available capacity. We allow utilization to overshoot capacity_curr (but not
5354 * capacity_orig) as it useful for predicting the capacity required after task
5355 * migrations (scheduler-driven DVFS).
5357 static int cpu_util(int cpu)
5359 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5360 unsigned long capacity = capacity_orig_of(cpu);
5362 return (util >= capacity) ? capacity : util;
5366 * select_task_rq_fair: Select target runqueue for the waking task in domains
5367 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5368 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5370 * Balances load by selecting the idlest cpu in the idlest group, or under
5371 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5373 * Returns the target cpu number.
5375 * preempt must be disabled.
5378 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5380 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5381 int cpu = smp_processor_id();
5382 int new_cpu = prev_cpu;
5383 int want_affine = 0;
5384 int sync = wake_flags & WF_SYNC;
5386 if (sd_flag & SD_BALANCE_WAKE) {
5388 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5392 for_each_domain(cpu, tmp) {
5393 if (!(tmp->flags & SD_LOAD_BALANCE))
5397 * If both cpu and prev_cpu are part of this domain,
5398 * cpu is a valid SD_WAKE_AFFINE target.
5400 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5401 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5406 if (tmp->flags & sd_flag)
5408 else if (!want_affine)
5413 sd = NULL; /* Prefer wake_affine over balance flags */
5414 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5419 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5420 new_cpu = select_idle_sibling(p, new_cpu);
5423 struct sched_group *group;
5426 if (!(sd->flags & sd_flag)) {
5431 group = find_idlest_group(sd, p, cpu, sd_flag);
5437 new_cpu = find_idlest_cpu(group, p, cpu);
5438 if (new_cpu == -1 || new_cpu == cpu) {
5439 /* Now try balancing at a lower domain level of cpu */
5444 /* Now try balancing at a lower domain level of new_cpu */
5446 weight = sd->span_weight;
5448 for_each_domain(cpu, tmp) {
5449 if (weight <= tmp->span_weight)
5451 if (tmp->flags & sd_flag)
5454 /* while loop will break here if sd == NULL */
5462 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5463 * cfs_rq_of(p) references at time of call are still valid and identify the
5464 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5466 static void migrate_task_rq_fair(struct task_struct *p)
5469 * As blocked tasks retain absolute vruntime the migration needs to
5470 * deal with this by subtracting the old and adding the new
5471 * min_vruntime -- the latter is done by enqueue_entity() when placing
5472 * the task on the new runqueue.
5474 if (p->state == TASK_WAKING) {
5475 struct sched_entity *se = &p->se;
5476 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5479 #ifndef CONFIG_64BIT
5480 u64 min_vruntime_copy;
5483 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5485 min_vruntime = cfs_rq->min_vruntime;
5486 } while (min_vruntime != min_vruntime_copy);
5488 min_vruntime = cfs_rq->min_vruntime;
5491 se->vruntime -= min_vruntime;
5495 * We are supposed to update the task to "current" time, then its up to date
5496 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5497 * what current time is, so simply throw away the out-of-date time. This
5498 * will result in the wakee task is less decayed, but giving the wakee more
5499 * load sounds not bad.
5501 remove_entity_load_avg(&p->se);
5503 /* Tell new CPU we are migrated */
5504 p->se.avg.last_update_time = 0;
5506 /* We have migrated, no longer consider this task hot */
5507 p->se.exec_start = 0;
5510 static void task_dead_fair(struct task_struct *p)
5512 remove_entity_load_avg(&p->se);
5514 #endif /* CONFIG_SMP */
5516 static unsigned long
5517 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5519 unsigned long gran = sysctl_sched_wakeup_granularity;
5522 * Since its curr running now, convert the gran from real-time
5523 * to virtual-time in his units.
5525 * By using 'se' instead of 'curr' we penalize light tasks, so
5526 * they get preempted easier. That is, if 'se' < 'curr' then
5527 * the resulting gran will be larger, therefore penalizing the
5528 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5529 * be smaller, again penalizing the lighter task.
5531 * This is especially important for buddies when the leftmost
5532 * task is higher priority than the buddy.
5534 return calc_delta_fair(gran, se);
5538 * Should 'se' preempt 'curr'.
5552 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5554 s64 gran, vdiff = curr->vruntime - se->vruntime;
5559 gran = wakeup_gran(curr, se);
5566 static void set_last_buddy(struct sched_entity *se)
5568 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5571 for_each_sched_entity(se)
5572 cfs_rq_of(se)->last = se;
5575 static void set_next_buddy(struct sched_entity *se)
5577 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5580 for_each_sched_entity(se)
5581 cfs_rq_of(se)->next = se;
5584 static void set_skip_buddy(struct sched_entity *se)
5586 for_each_sched_entity(se)
5587 cfs_rq_of(se)->skip = se;
5591 * Preempt the current task with a newly woken task if needed:
5593 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5595 struct task_struct *curr = rq->curr;
5596 struct sched_entity *se = &curr->se, *pse = &p->se;
5597 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5598 int scale = cfs_rq->nr_running >= sched_nr_latency;
5599 int next_buddy_marked = 0;
5601 if (unlikely(se == pse))
5605 * This is possible from callers such as attach_tasks(), in which we
5606 * unconditionally check_prempt_curr() after an enqueue (which may have
5607 * lead to a throttle). This both saves work and prevents false
5608 * next-buddy nomination below.
5610 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5613 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5614 set_next_buddy(pse);
5615 next_buddy_marked = 1;
5619 * We can come here with TIF_NEED_RESCHED already set from new task
5622 * Note: this also catches the edge-case of curr being in a throttled
5623 * group (e.g. via set_curr_task), since update_curr() (in the
5624 * enqueue of curr) will have resulted in resched being set. This
5625 * prevents us from potentially nominating it as a false LAST_BUDDY
5628 if (test_tsk_need_resched(curr))
5631 /* Idle tasks are by definition preempted by non-idle tasks. */
5632 if (unlikely(curr->policy == SCHED_IDLE) &&
5633 likely(p->policy != SCHED_IDLE))
5637 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5638 * is driven by the tick):
5640 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5643 find_matching_se(&se, &pse);
5644 update_curr(cfs_rq_of(se));
5646 if (wakeup_preempt_entity(se, pse) == 1) {
5648 * Bias pick_next to pick the sched entity that is
5649 * triggering this preemption.
5651 if (!next_buddy_marked)
5652 set_next_buddy(pse);
5661 * Only set the backward buddy when the current task is still
5662 * on the rq. This can happen when a wakeup gets interleaved
5663 * with schedule on the ->pre_schedule() or idle_balance()
5664 * point, either of which can * drop the rq lock.
5666 * Also, during early boot the idle thread is in the fair class,
5667 * for obvious reasons its a bad idea to schedule back to it.
5669 if (unlikely(!se->on_rq || curr == rq->idle))
5672 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5676 static struct task_struct *
5677 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5679 struct cfs_rq *cfs_rq = &rq->cfs;
5680 struct sched_entity *se;
5681 struct task_struct *p;
5685 #ifdef CONFIG_FAIR_GROUP_SCHED
5686 if (!cfs_rq->nr_running)
5689 if (prev->sched_class != &fair_sched_class)
5693 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5694 * likely that a next task is from the same cgroup as the current.
5696 * Therefore attempt to avoid putting and setting the entire cgroup
5697 * hierarchy, only change the part that actually changes.
5701 struct sched_entity *curr = cfs_rq->curr;
5704 * Since we got here without doing put_prev_entity() we also
5705 * have to consider cfs_rq->curr. If it is still a runnable
5706 * entity, update_curr() will update its vruntime, otherwise
5707 * forget we've ever seen it.
5711 update_curr(cfs_rq);
5716 * This call to check_cfs_rq_runtime() will do the
5717 * throttle and dequeue its entity in the parent(s).
5718 * Therefore the 'simple' nr_running test will indeed
5721 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5725 se = pick_next_entity(cfs_rq, curr);
5726 cfs_rq = group_cfs_rq(se);
5732 * Since we haven't yet done put_prev_entity and if the selected task
5733 * is a different task than we started out with, try and touch the
5734 * least amount of cfs_rqs.
5737 struct sched_entity *pse = &prev->se;
5739 while (!(cfs_rq = is_same_group(se, pse))) {
5740 int se_depth = se->depth;
5741 int pse_depth = pse->depth;
5743 if (se_depth <= pse_depth) {
5744 put_prev_entity(cfs_rq_of(pse), pse);
5745 pse = parent_entity(pse);
5747 if (se_depth >= pse_depth) {
5748 set_next_entity(cfs_rq_of(se), se);
5749 se = parent_entity(se);
5753 put_prev_entity(cfs_rq, pse);
5754 set_next_entity(cfs_rq, se);
5757 if (hrtick_enabled(rq))
5758 hrtick_start_fair(rq, p);
5765 if (!cfs_rq->nr_running)
5768 put_prev_task(rq, prev);
5771 se = pick_next_entity(cfs_rq, NULL);
5772 set_next_entity(cfs_rq, se);
5773 cfs_rq = group_cfs_rq(se);
5778 if (hrtick_enabled(rq))
5779 hrtick_start_fair(rq, p);
5785 * This is OK, because current is on_cpu, which avoids it being picked
5786 * for load-balance and preemption/IRQs are still disabled avoiding
5787 * further scheduler activity on it and we're being very careful to
5788 * re-start the picking loop.
5790 lockdep_unpin_lock(&rq->lock, cookie);
5791 new_tasks = idle_balance(rq);
5792 lockdep_repin_lock(&rq->lock, cookie);
5794 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5795 * possible for any higher priority task to appear. In that case we
5796 * must re-start the pick_next_entity() loop.
5808 * Account for a descheduled task:
5810 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5812 struct sched_entity *se = &prev->se;
5813 struct cfs_rq *cfs_rq;
5815 for_each_sched_entity(se) {
5816 cfs_rq = cfs_rq_of(se);
5817 put_prev_entity(cfs_rq, se);
5822 * sched_yield() is very simple
5824 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5826 static void yield_task_fair(struct rq *rq)
5828 struct task_struct *curr = rq->curr;
5829 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5830 struct sched_entity *se = &curr->se;
5833 * Are we the only task in the tree?
5835 if (unlikely(rq->nr_running == 1))
5838 clear_buddies(cfs_rq, se);
5840 if (curr->policy != SCHED_BATCH) {
5841 update_rq_clock(rq);
5843 * Update run-time statistics of the 'current'.
5845 update_curr(cfs_rq);
5847 * Tell update_rq_clock() that we've just updated,
5848 * so we don't do microscopic update in schedule()
5849 * and double the fastpath cost.
5851 rq_clock_skip_update(rq, true);
5857 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5859 struct sched_entity *se = &p->se;
5861 /* throttled hierarchies are not runnable */
5862 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5865 /* Tell the scheduler that we'd really like pse to run next. */
5868 yield_task_fair(rq);
5874 /**************************************************
5875 * Fair scheduling class load-balancing methods.
5879 * The purpose of load-balancing is to achieve the same basic fairness the
5880 * per-cpu scheduler provides, namely provide a proportional amount of compute
5881 * time to each task. This is expressed in the following equation:
5883 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5885 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5886 * W_i,0 is defined as:
5888 * W_i,0 = \Sum_j w_i,j (2)
5890 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5891 * is derived from the nice value as per sched_prio_to_weight[].
5893 * The weight average is an exponential decay average of the instantaneous
5896 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5898 * C_i is the compute capacity of cpu i, typically it is the
5899 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5900 * can also include other factors [XXX].
5902 * To achieve this balance we define a measure of imbalance which follows
5903 * directly from (1):
5905 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5907 * We them move tasks around to minimize the imbalance. In the continuous
5908 * function space it is obvious this converges, in the discrete case we get
5909 * a few fun cases generally called infeasible weight scenarios.
5912 * - infeasible weights;
5913 * - local vs global optima in the discrete case. ]
5918 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5919 * for all i,j solution, we create a tree of cpus that follows the hardware
5920 * topology where each level pairs two lower groups (or better). This results
5921 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5922 * tree to only the first of the previous level and we decrease the frequency
5923 * of load-balance at each level inv. proportional to the number of cpus in
5929 * \Sum { --- * --- * 2^i } = O(n) (5)
5931 * `- size of each group
5932 * | | `- number of cpus doing load-balance
5934 * `- sum over all levels
5936 * Coupled with a limit on how many tasks we can migrate every balance pass,
5937 * this makes (5) the runtime complexity of the balancer.
5939 * An important property here is that each CPU is still (indirectly) connected
5940 * to every other cpu in at most O(log n) steps:
5942 * The adjacency matrix of the resulting graph is given by:
5945 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5948 * And you'll find that:
5950 * A^(log_2 n)_i,j != 0 for all i,j (7)
5952 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5953 * The task movement gives a factor of O(m), giving a convergence complexity
5956 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5961 * In order to avoid CPUs going idle while there's still work to do, new idle
5962 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5963 * tree itself instead of relying on other CPUs to bring it work.
5965 * This adds some complexity to both (5) and (8) but it reduces the total idle
5973 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5976 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5981 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5983 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5985 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5988 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5989 * rewrite all of this once again.]
5992 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5994 enum fbq_type { regular, remote, all };
5996 #define LBF_ALL_PINNED 0x01
5997 #define LBF_NEED_BREAK 0x02
5998 #define LBF_DST_PINNED 0x04
5999 #define LBF_SOME_PINNED 0x08
6002 struct sched_domain *sd;
6010 struct cpumask *dst_grpmask;
6012 enum cpu_idle_type idle;
6014 /* The set of CPUs under consideration for load-balancing */
6015 struct cpumask *cpus;
6020 unsigned int loop_break;
6021 unsigned int loop_max;
6023 enum fbq_type fbq_type;
6024 struct list_head tasks;
6028 * Is this task likely cache-hot:
6030 static int task_hot(struct task_struct *p, struct lb_env *env)
6034 lockdep_assert_held(&env->src_rq->lock);
6036 if (p->sched_class != &fair_sched_class)
6039 if (unlikely(p->policy == SCHED_IDLE))
6043 * Buddy candidates are cache hot:
6045 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6046 (&p->se == cfs_rq_of(&p->se)->next ||
6047 &p->se == cfs_rq_of(&p->se)->last))
6050 if (sysctl_sched_migration_cost == -1)
6052 if (sysctl_sched_migration_cost == 0)
6055 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6057 return delta < (s64)sysctl_sched_migration_cost;
6060 #ifdef CONFIG_NUMA_BALANCING
6062 * Returns 1, if task migration degrades locality
6063 * Returns 0, if task migration improves locality i.e migration preferred.
6064 * Returns -1, if task migration is not affected by locality.
6066 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6068 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6069 unsigned long src_faults, dst_faults;
6070 int src_nid, dst_nid;
6072 if (!static_branch_likely(&sched_numa_balancing))
6075 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6078 src_nid = cpu_to_node(env->src_cpu);
6079 dst_nid = cpu_to_node(env->dst_cpu);
6081 if (src_nid == dst_nid)
6084 /* Migrating away from the preferred node is always bad. */
6085 if (src_nid == p->numa_preferred_nid) {
6086 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6092 /* Encourage migration to the preferred node. */
6093 if (dst_nid == p->numa_preferred_nid)
6097 src_faults = group_faults(p, src_nid);
6098 dst_faults = group_faults(p, dst_nid);
6100 src_faults = task_faults(p, src_nid);
6101 dst_faults = task_faults(p, dst_nid);
6104 return dst_faults < src_faults;
6108 static inline int migrate_degrades_locality(struct task_struct *p,
6116 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6119 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6123 lockdep_assert_held(&env->src_rq->lock);
6126 * We do not migrate tasks that are:
6127 * 1) throttled_lb_pair, or
6128 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6129 * 3) running (obviously), or
6130 * 4) are cache-hot on their current CPU.
6132 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6135 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6138 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6140 env->flags |= LBF_SOME_PINNED;
6143 * Remember if this task can be migrated to any other cpu in
6144 * our sched_group. We may want to revisit it if we couldn't
6145 * meet load balance goals by pulling other tasks on src_cpu.
6147 * Also avoid computing new_dst_cpu if we have already computed
6148 * one in current iteration.
6150 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6153 /* Prevent to re-select dst_cpu via env's cpus */
6154 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6155 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6156 env->flags |= LBF_DST_PINNED;
6157 env->new_dst_cpu = cpu;
6165 /* Record that we found atleast one task that could run on dst_cpu */
6166 env->flags &= ~LBF_ALL_PINNED;
6168 if (task_running(env->src_rq, p)) {
6169 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6174 * Aggressive migration if:
6175 * 1) destination numa is preferred
6176 * 2) task is cache cold, or
6177 * 3) too many balance attempts have failed.
6179 tsk_cache_hot = migrate_degrades_locality(p, env);
6180 if (tsk_cache_hot == -1)
6181 tsk_cache_hot = task_hot(p, env);
6183 if (tsk_cache_hot <= 0 ||
6184 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6185 if (tsk_cache_hot == 1) {
6186 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6187 schedstat_inc(p, se.statistics.nr_forced_migrations);
6192 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6197 * detach_task() -- detach the task for the migration specified in env
6199 static void detach_task(struct task_struct *p, struct lb_env *env)
6201 lockdep_assert_held(&env->src_rq->lock);
6203 p->on_rq = TASK_ON_RQ_MIGRATING;
6204 deactivate_task(env->src_rq, p, 0);
6205 set_task_cpu(p, env->dst_cpu);
6209 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6210 * part of active balancing operations within "domain".
6212 * Returns a task if successful and NULL otherwise.
6214 static struct task_struct *detach_one_task(struct lb_env *env)
6216 struct task_struct *p, *n;
6218 lockdep_assert_held(&env->src_rq->lock);
6220 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6221 if (!can_migrate_task(p, env))
6224 detach_task(p, env);
6227 * Right now, this is only the second place where
6228 * lb_gained[env->idle] is updated (other is detach_tasks)
6229 * so we can safely collect stats here rather than
6230 * inside detach_tasks().
6232 schedstat_inc(env->sd, lb_gained[env->idle]);
6238 static const unsigned int sched_nr_migrate_break = 32;
6241 * detach_tasks() -- tries to detach up to imbalance weighted load from
6242 * busiest_rq, as part of a balancing operation within domain "sd".
6244 * Returns number of detached tasks if successful and 0 otherwise.
6246 static int detach_tasks(struct lb_env *env)
6248 struct list_head *tasks = &env->src_rq->cfs_tasks;
6249 struct task_struct *p;
6253 lockdep_assert_held(&env->src_rq->lock);
6255 if (env->imbalance <= 0)
6258 while (!list_empty(tasks)) {
6260 * We don't want to steal all, otherwise we may be treated likewise,
6261 * which could at worst lead to a livelock crash.
6263 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6266 p = list_first_entry(tasks, struct task_struct, se.group_node);
6269 /* We've more or less seen every task there is, call it quits */
6270 if (env->loop > env->loop_max)
6273 /* take a breather every nr_migrate tasks */
6274 if (env->loop > env->loop_break) {
6275 env->loop_break += sched_nr_migrate_break;
6276 env->flags |= LBF_NEED_BREAK;
6280 if (!can_migrate_task(p, env))
6283 load = task_h_load(p);
6285 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6288 if ((load / 2) > env->imbalance)
6291 detach_task(p, env);
6292 list_add(&p->se.group_node, &env->tasks);
6295 env->imbalance -= load;
6297 #ifdef CONFIG_PREEMPT
6299 * NEWIDLE balancing is a source of latency, so preemptible
6300 * kernels will stop after the first task is detached to minimize
6301 * the critical section.
6303 if (env->idle == CPU_NEWLY_IDLE)
6308 * We only want to steal up to the prescribed amount of
6311 if (env->imbalance <= 0)
6316 list_move_tail(&p->se.group_node, tasks);
6320 * Right now, this is one of only two places we collect this stat
6321 * so we can safely collect detach_one_task() stats here rather
6322 * than inside detach_one_task().
6324 schedstat_add(env->sd, lb_gained[env->idle], detached);
6330 * attach_task() -- attach the task detached by detach_task() to its new rq.
6332 static void attach_task(struct rq *rq, struct task_struct *p)
6334 lockdep_assert_held(&rq->lock);
6336 BUG_ON(task_rq(p) != rq);
6337 activate_task(rq, p, 0);
6338 p->on_rq = TASK_ON_RQ_QUEUED;
6339 check_preempt_curr(rq, p, 0);
6343 * attach_one_task() -- attaches the task returned from detach_one_task() to
6346 static void attach_one_task(struct rq *rq, struct task_struct *p)
6348 raw_spin_lock(&rq->lock);
6350 raw_spin_unlock(&rq->lock);
6354 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6357 static void attach_tasks(struct lb_env *env)
6359 struct list_head *tasks = &env->tasks;
6360 struct task_struct *p;
6362 raw_spin_lock(&env->dst_rq->lock);
6364 while (!list_empty(tasks)) {
6365 p = list_first_entry(tasks, struct task_struct, se.group_node);
6366 list_del_init(&p->se.group_node);
6368 attach_task(env->dst_rq, p);
6371 raw_spin_unlock(&env->dst_rq->lock);
6374 #ifdef CONFIG_FAIR_GROUP_SCHED
6375 static void update_blocked_averages(int cpu)
6377 struct rq *rq = cpu_rq(cpu);
6378 struct cfs_rq *cfs_rq;
6379 unsigned long flags;
6381 raw_spin_lock_irqsave(&rq->lock, flags);
6382 update_rq_clock(rq);
6385 * Iterates the task_group tree in a bottom up fashion, see
6386 * list_add_leaf_cfs_rq() for details.
6388 for_each_leaf_cfs_rq(rq, cfs_rq) {
6389 /* throttled entities do not contribute to load */
6390 if (throttled_hierarchy(cfs_rq))
6393 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6394 update_tg_load_avg(cfs_rq, 0);
6396 raw_spin_unlock_irqrestore(&rq->lock, flags);
6400 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6401 * This needs to be done in a top-down fashion because the load of a child
6402 * group is a fraction of its parents load.
6404 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6406 struct rq *rq = rq_of(cfs_rq);
6407 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6408 unsigned long now = jiffies;
6411 if (cfs_rq->last_h_load_update == now)
6414 cfs_rq->h_load_next = NULL;
6415 for_each_sched_entity(se) {
6416 cfs_rq = cfs_rq_of(se);
6417 cfs_rq->h_load_next = se;
6418 if (cfs_rq->last_h_load_update == now)
6423 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6424 cfs_rq->last_h_load_update = now;
6427 while ((se = cfs_rq->h_load_next) != NULL) {
6428 load = cfs_rq->h_load;
6429 load = div64_ul(load * se->avg.load_avg,
6430 cfs_rq_load_avg(cfs_rq) + 1);
6431 cfs_rq = group_cfs_rq(se);
6432 cfs_rq->h_load = load;
6433 cfs_rq->last_h_load_update = now;
6437 static unsigned long task_h_load(struct task_struct *p)
6439 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6441 update_cfs_rq_h_load(cfs_rq);
6442 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6443 cfs_rq_load_avg(cfs_rq) + 1);
6446 static inline void update_blocked_averages(int cpu)
6448 struct rq *rq = cpu_rq(cpu);
6449 struct cfs_rq *cfs_rq = &rq->cfs;
6450 unsigned long flags;
6452 raw_spin_lock_irqsave(&rq->lock, flags);
6453 update_rq_clock(rq);
6454 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6455 raw_spin_unlock_irqrestore(&rq->lock, flags);
6458 static unsigned long task_h_load(struct task_struct *p)
6460 return p->se.avg.load_avg;
6464 /********** Helpers for find_busiest_group ************************/
6473 * sg_lb_stats - stats of a sched_group required for load_balancing
6475 struct sg_lb_stats {
6476 unsigned long avg_load; /*Avg load across the CPUs of the group */
6477 unsigned long group_load; /* Total load over the CPUs of the group */
6478 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6479 unsigned long load_per_task;
6480 unsigned long group_capacity;
6481 unsigned long group_util; /* Total utilization of the group */
6482 unsigned int sum_nr_running; /* Nr tasks running in the group */
6483 unsigned int idle_cpus;
6484 unsigned int group_weight;
6485 enum group_type group_type;
6486 int group_no_capacity;
6487 #ifdef CONFIG_NUMA_BALANCING
6488 unsigned int nr_numa_running;
6489 unsigned int nr_preferred_running;
6494 * sd_lb_stats - Structure to store the statistics of a sched_domain
6495 * during load balancing.
6497 struct sd_lb_stats {
6498 struct sched_group *busiest; /* Busiest group in this sd */
6499 struct sched_group *local; /* Local group in this sd */
6500 unsigned long total_load; /* Total load of all groups in sd */
6501 unsigned long total_capacity; /* Total capacity of all groups in sd */
6502 unsigned long avg_load; /* Average load across all groups in sd */
6504 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6505 struct sg_lb_stats local_stat; /* Statistics of the local group */
6508 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6511 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6512 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6513 * We must however clear busiest_stat::avg_load because
6514 * update_sd_pick_busiest() reads this before assignment.
6516 *sds = (struct sd_lb_stats){
6520 .total_capacity = 0UL,
6523 .sum_nr_running = 0,
6524 .group_type = group_other,
6530 * get_sd_load_idx - Obtain the load index for a given sched domain.
6531 * @sd: The sched_domain whose load_idx is to be obtained.
6532 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6534 * Return: The load index.
6536 static inline int get_sd_load_idx(struct sched_domain *sd,
6537 enum cpu_idle_type idle)
6543 load_idx = sd->busy_idx;
6546 case CPU_NEWLY_IDLE:
6547 load_idx = sd->newidle_idx;
6550 load_idx = sd->idle_idx;
6557 static unsigned long scale_rt_capacity(int cpu)
6559 struct rq *rq = cpu_rq(cpu);
6560 u64 total, used, age_stamp, avg;
6564 * Since we're reading these variables without serialization make sure
6565 * we read them once before doing sanity checks on them.
6567 age_stamp = READ_ONCE(rq->age_stamp);
6568 avg = READ_ONCE(rq->rt_avg);
6569 delta = __rq_clock_broken(rq) - age_stamp;
6571 if (unlikely(delta < 0))
6574 total = sched_avg_period() + delta;
6576 used = div_u64(avg, total);
6578 if (likely(used < SCHED_CAPACITY_SCALE))
6579 return SCHED_CAPACITY_SCALE - used;
6584 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6586 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6587 struct sched_group *sdg = sd->groups;
6589 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6591 capacity *= scale_rt_capacity(cpu);
6592 capacity >>= SCHED_CAPACITY_SHIFT;
6597 cpu_rq(cpu)->cpu_capacity = capacity;
6598 sdg->sgc->capacity = capacity;
6601 void update_group_capacity(struct sched_domain *sd, int cpu)
6603 struct sched_domain *child = sd->child;
6604 struct sched_group *group, *sdg = sd->groups;
6605 unsigned long capacity;
6606 unsigned long interval;
6608 interval = msecs_to_jiffies(sd->balance_interval);
6609 interval = clamp(interval, 1UL, max_load_balance_interval);
6610 sdg->sgc->next_update = jiffies + interval;
6613 update_cpu_capacity(sd, cpu);
6619 if (child->flags & SD_OVERLAP) {
6621 * SD_OVERLAP domains cannot assume that child groups
6622 * span the current group.
6625 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6626 struct sched_group_capacity *sgc;
6627 struct rq *rq = cpu_rq(cpu);
6630 * build_sched_domains() -> init_sched_groups_capacity()
6631 * gets here before we've attached the domains to the
6634 * Use capacity_of(), which is set irrespective of domains
6635 * in update_cpu_capacity().
6637 * This avoids capacity from being 0 and
6638 * causing divide-by-zero issues on boot.
6640 if (unlikely(!rq->sd)) {
6641 capacity += capacity_of(cpu);
6645 sgc = rq->sd->groups->sgc;
6646 capacity += sgc->capacity;
6650 * !SD_OVERLAP domains can assume that child groups
6651 * span the current group.
6654 group = child->groups;
6656 capacity += group->sgc->capacity;
6657 group = group->next;
6658 } while (group != child->groups);
6661 sdg->sgc->capacity = capacity;
6665 * Check whether the capacity of the rq has been noticeably reduced by side
6666 * activity. The imbalance_pct is used for the threshold.
6667 * Return true is the capacity is reduced
6670 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6672 return ((rq->cpu_capacity * sd->imbalance_pct) <
6673 (rq->cpu_capacity_orig * 100));
6677 * Group imbalance indicates (and tries to solve) the problem where balancing
6678 * groups is inadequate due to tsk_cpus_allowed() constraints.
6680 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6681 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6684 * { 0 1 2 3 } { 4 5 6 7 }
6687 * If we were to balance group-wise we'd place two tasks in the first group and
6688 * two tasks in the second group. Clearly this is undesired as it will overload
6689 * cpu 3 and leave one of the cpus in the second group unused.
6691 * The current solution to this issue is detecting the skew in the first group
6692 * by noticing the lower domain failed to reach balance and had difficulty
6693 * moving tasks due to affinity constraints.
6695 * When this is so detected; this group becomes a candidate for busiest; see
6696 * update_sd_pick_busiest(). And calculate_imbalance() and
6697 * find_busiest_group() avoid some of the usual balance conditions to allow it
6698 * to create an effective group imbalance.
6700 * This is a somewhat tricky proposition since the next run might not find the
6701 * group imbalance and decide the groups need to be balanced again. A most
6702 * subtle and fragile situation.
6705 static inline int sg_imbalanced(struct sched_group *group)
6707 return group->sgc->imbalance;
6711 * group_has_capacity returns true if the group has spare capacity that could
6712 * be used by some tasks.
6713 * We consider that a group has spare capacity if the * number of task is
6714 * smaller than the number of CPUs or if the utilization is lower than the
6715 * available capacity for CFS tasks.
6716 * For the latter, we use a threshold to stabilize the state, to take into
6717 * account the variance of the tasks' load and to return true if the available
6718 * capacity in meaningful for the load balancer.
6719 * As an example, an available capacity of 1% can appear but it doesn't make
6720 * any benefit for the load balance.
6723 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6725 if (sgs->sum_nr_running < sgs->group_weight)
6728 if ((sgs->group_capacity * 100) >
6729 (sgs->group_util * env->sd->imbalance_pct))
6736 * group_is_overloaded returns true if the group has more tasks than it can
6738 * group_is_overloaded is not equals to !group_has_capacity because a group
6739 * with the exact right number of tasks, has no more spare capacity but is not
6740 * overloaded so both group_has_capacity and group_is_overloaded return
6744 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6746 if (sgs->sum_nr_running <= sgs->group_weight)
6749 if ((sgs->group_capacity * 100) <
6750 (sgs->group_util * env->sd->imbalance_pct))
6757 group_type group_classify(struct sched_group *group,
6758 struct sg_lb_stats *sgs)
6760 if (sgs->group_no_capacity)
6761 return group_overloaded;
6763 if (sg_imbalanced(group))
6764 return group_imbalanced;
6770 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6771 * @env: The load balancing environment.
6772 * @group: sched_group whose statistics are to be updated.
6773 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6774 * @local_group: Does group contain this_cpu.
6775 * @sgs: variable to hold the statistics for this group.
6776 * @overload: Indicate more than one runnable task for any CPU.
6778 static inline void update_sg_lb_stats(struct lb_env *env,
6779 struct sched_group *group, int load_idx,
6780 int local_group, struct sg_lb_stats *sgs,
6786 memset(sgs, 0, sizeof(*sgs));
6788 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6789 struct rq *rq = cpu_rq(i);
6791 /* Bias balancing toward cpus of our domain */
6793 load = target_load(i, load_idx);
6795 load = source_load(i, load_idx);
6797 sgs->group_load += load;
6798 sgs->group_util += cpu_util(i);
6799 sgs->sum_nr_running += rq->cfs.h_nr_running;
6801 nr_running = rq->nr_running;
6805 #ifdef CONFIG_NUMA_BALANCING
6806 sgs->nr_numa_running += rq->nr_numa_running;
6807 sgs->nr_preferred_running += rq->nr_preferred_running;
6809 sgs->sum_weighted_load += weighted_cpuload(i);
6811 * No need to call idle_cpu() if nr_running is not 0
6813 if (!nr_running && idle_cpu(i))
6817 /* Adjust by relative CPU capacity of the group */
6818 sgs->group_capacity = group->sgc->capacity;
6819 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6821 if (sgs->sum_nr_running)
6822 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6824 sgs->group_weight = group->group_weight;
6826 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6827 sgs->group_type = group_classify(group, sgs);
6831 * update_sd_pick_busiest - return 1 on busiest group
6832 * @env: The load balancing environment.
6833 * @sds: sched_domain statistics
6834 * @sg: sched_group candidate to be checked for being the busiest
6835 * @sgs: sched_group statistics
6837 * Determine if @sg is a busier group than the previously selected
6840 * Return: %true if @sg is a busier group than the previously selected
6841 * busiest group. %false otherwise.
6843 static bool update_sd_pick_busiest(struct lb_env *env,
6844 struct sd_lb_stats *sds,
6845 struct sched_group *sg,
6846 struct sg_lb_stats *sgs)
6848 struct sg_lb_stats *busiest = &sds->busiest_stat;
6850 if (sgs->group_type > busiest->group_type)
6853 if (sgs->group_type < busiest->group_type)
6856 if (sgs->avg_load <= busiest->avg_load)
6859 /* This is the busiest node in its class. */
6860 if (!(env->sd->flags & SD_ASYM_PACKING))
6863 /* No ASYM_PACKING if target cpu is already busy */
6864 if (env->idle == CPU_NOT_IDLE)
6867 * ASYM_PACKING needs to move all the work to the lowest
6868 * numbered CPUs in the group, therefore mark all groups
6869 * higher than ourself as busy.
6871 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6875 /* Prefer to move from highest possible cpu's work */
6876 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6883 #ifdef CONFIG_NUMA_BALANCING
6884 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6886 if (sgs->sum_nr_running > sgs->nr_numa_running)
6888 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6893 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6895 if (rq->nr_running > rq->nr_numa_running)
6897 if (rq->nr_running > rq->nr_preferred_running)
6902 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6907 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6911 #endif /* CONFIG_NUMA_BALANCING */
6914 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6915 * @env: The load balancing environment.
6916 * @sds: variable to hold the statistics for this sched_domain.
6918 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6920 struct sched_domain *child = env->sd->child;
6921 struct sched_group *sg = env->sd->groups;
6922 struct sg_lb_stats tmp_sgs;
6923 int load_idx, prefer_sibling = 0;
6924 bool overload = false;
6926 if (child && child->flags & SD_PREFER_SIBLING)
6929 load_idx = get_sd_load_idx(env->sd, env->idle);
6932 struct sg_lb_stats *sgs = &tmp_sgs;
6935 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6938 sgs = &sds->local_stat;
6940 if (env->idle != CPU_NEWLY_IDLE ||
6941 time_after_eq(jiffies, sg->sgc->next_update))
6942 update_group_capacity(env->sd, env->dst_cpu);
6945 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6952 * In case the child domain prefers tasks go to siblings
6953 * first, lower the sg capacity so that we'll try
6954 * and move all the excess tasks away. We lower the capacity
6955 * of a group only if the local group has the capacity to fit
6956 * these excess tasks. The extra check prevents the case where
6957 * you always pull from the heaviest group when it is already
6958 * under-utilized (possible with a large weight task outweighs
6959 * the tasks on the system).
6961 if (prefer_sibling && sds->local &&
6962 group_has_capacity(env, &sds->local_stat) &&
6963 (sgs->sum_nr_running > 1)) {
6964 sgs->group_no_capacity = 1;
6965 sgs->group_type = group_classify(sg, sgs);
6968 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6970 sds->busiest_stat = *sgs;
6974 /* Now, start updating sd_lb_stats */
6975 sds->total_load += sgs->group_load;
6976 sds->total_capacity += sgs->group_capacity;
6979 } while (sg != env->sd->groups);
6981 if (env->sd->flags & SD_NUMA)
6982 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6984 if (!env->sd->parent) {
6985 /* update overload indicator if we are at root domain */
6986 if (env->dst_rq->rd->overload != overload)
6987 env->dst_rq->rd->overload = overload;
6993 * check_asym_packing - Check to see if the group is packed into the
6996 * This is primarily intended to used at the sibling level. Some
6997 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6998 * case of POWER7, it can move to lower SMT modes only when higher
6999 * threads are idle. When in lower SMT modes, the threads will
7000 * perform better since they share less core resources. Hence when we
7001 * have idle threads, we want them to be the higher ones.
7003 * This packing function is run on idle threads. It checks to see if
7004 * the busiest CPU in this domain (core in the P7 case) has a higher
7005 * CPU number than the packing function is being run on. Here we are
7006 * assuming lower CPU number will be equivalent to lower a SMT thread
7009 * Return: 1 when packing is required and a task should be moved to
7010 * this CPU. The amount of the imbalance is returned in *imbalance.
7012 * @env: The load balancing environment.
7013 * @sds: Statistics of the sched_domain which is to be packed
7015 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7019 if (!(env->sd->flags & SD_ASYM_PACKING))
7022 if (env->idle == CPU_NOT_IDLE)
7028 busiest_cpu = group_first_cpu(sds->busiest);
7029 if (env->dst_cpu > busiest_cpu)
7032 env->imbalance = DIV_ROUND_CLOSEST(
7033 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7034 SCHED_CAPACITY_SCALE);
7040 * fix_small_imbalance - Calculate the minor imbalance that exists
7041 * amongst the groups of a sched_domain, during
7043 * @env: The load balancing environment.
7044 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7047 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7049 unsigned long tmp, capa_now = 0, capa_move = 0;
7050 unsigned int imbn = 2;
7051 unsigned long scaled_busy_load_per_task;
7052 struct sg_lb_stats *local, *busiest;
7054 local = &sds->local_stat;
7055 busiest = &sds->busiest_stat;
7057 if (!local->sum_nr_running)
7058 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7059 else if (busiest->load_per_task > local->load_per_task)
7062 scaled_busy_load_per_task =
7063 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7064 busiest->group_capacity;
7066 if (busiest->avg_load + scaled_busy_load_per_task >=
7067 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7068 env->imbalance = busiest->load_per_task;
7073 * OK, we don't have enough imbalance to justify moving tasks,
7074 * however we may be able to increase total CPU capacity used by
7078 capa_now += busiest->group_capacity *
7079 min(busiest->load_per_task, busiest->avg_load);
7080 capa_now += local->group_capacity *
7081 min(local->load_per_task, local->avg_load);
7082 capa_now /= SCHED_CAPACITY_SCALE;
7084 /* Amount of load we'd subtract */
7085 if (busiest->avg_load > scaled_busy_load_per_task) {
7086 capa_move += busiest->group_capacity *
7087 min(busiest->load_per_task,
7088 busiest->avg_load - scaled_busy_load_per_task);
7091 /* Amount of load we'd add */
7092 if (busiest->avg_load * busiest->group_capacity <
7093 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7094 tmp = (busiest->avg_load * busiest->group_capacity) /
7095 local->group_capacity;
7097 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7098 local->group_capacity;
7100 capa_move += local->group_capacity *
7101 min(local->load_per_task, local->avg_load + tmp);
7102 capa_move /= SCHED_CAPACITY_SCALE;
7104 /* Move if we gain throughput */
7105 if (capa_move > capa_now)
7106 env->imbalance = busiest->load_per_task;
7110 * calculate_imbalance - Calculate the amount of imbalance present within the
7111 * groups of a given sched_domain during load balance.
7112 * @env: load balance environment
7113 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7115 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7117 unsigned long max_pull, load_above_capacity = ~0UL;
7118 struct sg_lb_stats *local, *busiest;
7120 local = &sds->local_stat;
7121 busiest = &sds->busiest_stat;
7123 if (busiest->group_type == group_imbalanced) {
7125 * In the group_imb case we cannot rely on group-wide averages
7126 * to ensure cpu-load equilibrium, look at wider averages. XXX
7128 busiest->load_per_task =
7129 min(busiest->load_per_task, sds->avg_load);
7133 * Avg load of busiest sg can be less and avg load of local sg can
7134 * be greater than avg load across all sgs of sd because avg load
7135 * factors in sg capacity and sgs with smaller group_type are
7136 * skipped when updating the busiest sg:
7138 if (busiest->avg_load <= sds->avg_load ||
7139 local->avg_load >= sds->avg_load) {
7141 return fix_small_imbalance(env, sds);
7145 * If there aren't any idle cpus, avoid creating some.
7147 if (busiest->group_type == group_overloaded &&
7148 local->group_type == group_overloaded) {
7149 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7150 if (load_above_capacity > busiest->group_capacity) {
7151 load_above_capacity -= busiest->group_capacity;
7152 load_above_capacity *= NICE_0_LOAD;
7153 load_above_capacity /= busiest->group_capacity;
7155 load_above_capacity = ~0UL;
7159 * We're trying to get all the cpus to the average_load, so we don't
7160 * want to push ourselves above the average load, nor do we wish to
7161 * reduce the max loaded cpu below the average load. At the same time,
7162 * we also don't want to reduce the group load below the group
7163 * capacity. Thus we look for the minimum possible imbalance.
7165 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7167 /* How much load to actually move to equalise the imbalance */
7168 env->imbalance = min(
7169 max_pull * busiest->group_capacity,
7170 (sds->avg_load - local->avg_load) * local->group_capacity
7171 ) / SCHED_CAPACITY_SCALE;
7174 * if *imbalance is less than the average load per runnable task
7175 * there is no guarantee that any tasks will be moved so we'll have
7176 * a think about bumping its value to force at least one task to be
7179 if (env->imbalance < busiest->load_per_task)
7180 return fix_small_imbalance(env, sds);
7183 /******* find_busiest_group() helpers end here *********************/
7186 * find_busiest_group - Returns the busiest group within the sched_domain
7187 * if there is an imbalance.
7189 * Also calculates the amount of weighted load which should be moved
7190 * to restore balance.
7192 * @env: The load balancing environment.
7194 * Return: - The busiest group if imbalance exists.
7196 static struct sched_group *find_busiest_group(struct lb_env *env)
7198 struct sg_lb_stats *local, *busiest;
7199 struct sd_lb_stats sds;
7201 init_sd_lb_stats(&sds);
7204 * Compute the various statistics relavent for load balancing at
7207 update_sd_lb_stats(env, &sds);
7208 local = &sds.local_stat;
7209 busiest = &sds.busiest_stat;
7211 /* ASYM feature bypasses nice load balance check */
7212 if (check_asym_packing(env, &sds))
7215 /* There is no busy sibling group to pull tasks from */
7216 if (!sds.busiest || busiest->sum_nr_running == 0)
7219 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7220 / sds.total_capacity;
7223 * If the busiest group is imbalanced the below checks don't
7224 * work because they assume all things are equal, which typically
7225 * isn't true due to cpus_allowed constraints and the like.
7227 if (busiest->group_type == group_imbalanced)
7230 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7231 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7232 busiest->group_no_capacity)
7236 * If the local group is busier than the selected busiest group
7237 * don't try and pull any tasks.
7239 if (local->avg_load >= busiest->avg_load)
7243 * Don't pull any tasks if this group is already above the domain
7246 if (local->avg_load >= sds.avg_load)
7249 if (env->idle == CPU_IDLE) {
7251 * This cpu is idle. If the busiest group is not overloaded
7252 * and there is no imbalance between this and busiest group
7253 * wrt idle cpus, it is balanced. The imbalance becomes
7254 * significant if the diff is greater than 1 otherwise we
7255 * might end up to just move the imbalance on another group
7257 if ((busiest->group_type != group_overloaded) &&
7258 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7262 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7263 * imbalance_pct to be conservative.
7265 if (100 * busiest->avg_load <=
7266 env->sd->imbalance_pct * local->avg_load)
7271 /* Looks like there is an imbalance. Compute it */
7272 calculate_imbalance(env, &sds);
7281 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7283 static struct rq *find_busiest_queue(struct lb_env *env,
7284 struct sched_group *group)
7286 struct rq *busiest = NULL, *rq;
7287 unsigned long busiest_load = 0, busiest_capacity = 1;
7290 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7291 unsigned long capacity, wl;
7295 rt = fbq_classify_rq(rq);
7298 * We classify groups/runqueues into three groups:
7299 * - regular: there are !numa tasks
7300 * - remote: there are numa tasks that run on the 'wrong' node
7301 * - all: there is no distinction
7303 * In order to avoid migrating ideally placed numa tasks,
7304 * ignore those when there's better options.
7306 * If we ignore the actual busiest queue to migrate another
7307 * task, the next balance pass can still reduce the busiest
7308 * queue by moving tasks around inside the node.
7310 * If we cannot move enough load due to this classification
7311 * the next pass will adjust the group classification and
7312 * allow migration of more tasks.
7314 * Both cases only affect the total convergence complexity.
7316 if (rt > env->fbq_type)
7319 capacity = capacity_of(i);
7321 wl = weighted_cpuload(i);
7324 * When comparing with imbalance, use weighted_cpuload()
7325 * which is not scaled with the cpu capacity.
7328 if (rq->nr_running == 1 && wl > env->imbalance &&
7329 !check_cpu_capacity(rq, env->sd))
7333 * For the load comparisons with the other cpu's, consider
7334 * the weighted_cpuload() scaled with the cpu capacity, so
7335 * that the load can be moved away from the cpu that is
7336 * potentially running at a lower capacity.
7338 * Thus we're looking for max(wl_i / capacity_i), crosswise
7339 * multiplication to rid ourselves of the division works out
7340 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7341 * our previous maximum.
7343 if (wl * busiest_capacity > busiest_load * capacity) {
7345 busiest_capacity = capacity;
7354 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7355 * so long as it is large enough.
7357 #define MAX_PINNED_INTERVAL 512
7359 /* Working cpumask for load_balance and load_balance_newidle. */
7360 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7362 static int need_active_balance(struct lb_env *env)
7364 struct sched_domain *sd = env->sd;
7366 if (env->idle == CPU_NEWLY_IDLE) {
7369 * ASYM_PACKING needs to force migrate tasks from busy but
7370 * higher numbered CPUs in order to pack all tasks in the
7371 * lowest numbered CPUs.
7373 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7378 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7379 * It's worth migrating the task if the src_cpu's capacity is reduced
7380 * because of other sched_class or IRQs if more capacity stays
7381 * available on dst_cpu.
7383 if ((env->idle != CPU_NOT_IDLE) &&
7384 (env->src_rq->cfs.h_nr_running == 1)) {
7385 if ((check_cpu_capacity(env->src_rq, sd)) &&
7386 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7390 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7393 static int active_load_balance_cpu_stop(void *data);
7395 static int should_we_balance(struct lb_env *env)
7397 struct sched_group *sg = env->sd->groups;
7398 struct cpumask *sg_cpus, *sg_mask;
7399 int cpu, balance_cpu = -1;
7402 * In the newly idle case, we will allow all the cpu's
7403 * to do the newly idle load balance.
7405 if (env->idle == CPU_NEWLY_IDLE)
7408 sg_cpus = sched_group_cpus(sg);
7409 sg_mask = sched_group_mask(sg);
7410 /* Try to find first idle cpu */
7411 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7412 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7419 if (balance_cpu == -1)
7420 balance_cpu = group_balance_cpu(sg);
7423 * First idle cpu or the first cpu(busiest) in this sched group
7424 * is eligible for doing load balancing at this and above domains.
7426 return balance_cpu == env->dst_cpu;
7430 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7431 * tasks if there is an imbalance.
7433 static int load_balance(int this_cpu, struct rq *this_rq,
7434 struct sched_domain *sd, enum cpu_idle_type idle,
7435 int *continue_balancing)
7437 int ld_moved, cur_ld_moved, active_balance = 0;
7438 struct sched_domain *sd_parent = sd->parent;
7439 struct sched_group *group;
7441 unsigned long flags;
7442 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7444 struct lb_env env = {
7446 .dst_cpu = this_cpu,
7448 .dst_grpmask = sched_group_cpus(sd->groups),
7450 .loop_break = sched_nr_migrate_break,
7453 .tasks = LIST_HEAD_INIT(env.tasks),
7457 * For NEWLY_IDLE load_balancing, we don't need to consider
7458 * other cpus in our group
7460 if (idle == CPU_NEWLY_IDLE)
7461 env.dst_grpmask = NULL;
7463 cpumask_copy(cpus, cpu_active_mask);
7465 schedstat_inc(sd, lb_count[idle]);
7468 if (!should_we_balance(&env)) {
7469 *continue_balancing = 0;
7473 group = find_busiest_group(&env);
7475 schedstat_inc(sd, lb_nobusyg[idle]);
7479 busiest = find_busiest_queue(&env, group);
7481 schedstat_inc(sd, lb_nobusyq[idle]);
7485 BUG_ON(busiest == env.dst_rq);
7487 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7489 env.src_cpu = busiest->cpu;
7490 env.src_rq = busiest;
7493 if (busiest->nr_running > 1) {
7495 * Attempt to move tasks. If find_busiest_group has found
7496 * an imbalance but busiest->nr_running <= 1, the group is
7497 * still unbalanced. ld_moved simply stays zero, so it is
7498 * correctly treated as an imbalance.
7500 env.flags |= LBF_ALL_PINNED;
7501 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7504 raw_spin_lock_irqsave(&busiest->lock, flags);
7507 * cur_ld_moved - load moved in current iteration
7508 * ld_moved - cumulative load moved across iterations
7510 cur_ld_moved = detach_tasks(&env);
7513 * We've detached some tasks from busiest_rq. Every
7514 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7515 * unlock busiest->lock, and we are able to be sure
7516 * that nobody can manipulate the tasks in parallel.
7517 * See task_rq_lock() family for the details.
7520 raw_spin_unlock(&busiest->lock);
7524 ld_moved += cur_ld_moved;
7527 local_irq_restore(flags);
7529 if (env.flags & LBF_NEED_BREAK) {
7530 env.flags &= ~LBF_NEED_BREAK;
7535 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7536 * us and move them to an alternate dst_cpu in our sched_group
7537 * where they can run. The upper limit on how many times we
7538 * iterate on same src_cpu is dependent on number of cpus in our
7541 * This changes load balance semantics a bit on who can move
7542 * load to a given_cpu. In addition to the given_cpu itself
7543 * (or a ilb_cpu acting on its behalf where given_cpu is
7544 * nohz-idle), we now have balance_cpu in a position to move
7545 * load to given_cpu. In rare situations, this may cause
7546 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7547 * _independently_ and at _same_ time to move some load to
7548 * given_cpu) causing exceess load to be moved to given_cpu.
7549 * This however should not happen so much in practice and
7550 * moreover subsequent load balance cycles should correct the
7551 * excess load moved.
7553 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7555 /* Prevent to re-select dst_cpu via env's cpus */
7556 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7558 env.dst_rq = cpu_rq(env.new_dst_cpu);
7559 env.dst_cpu = env.new_dst_cpu;
7560 env.flags &= ~LBF_DST_PINNED;
7562 env.loop_break = sched_nr_migrate_break;
7565 * Go back to "more_balance" rather than "redo" since we
7566 * need to continue with same src_cpu.
7572 * We failed to reach balance because of affinity.
7575 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7577 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7578 *group_imbalance = 1;
7581 /* All tasks on this runqueue were pinned by CPU affinity */
7582 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7583 cpumask_clear_cpu(cpu_of(busiest), cpus);
7584 if (!cpumask_empty(cpus)) {
7586 env.loop_break = sched_nr_migrate_break;
7589 goto out_all_pinned;
7594 schedstat_inc(sd, lb_failed[idle]);
7596 * Increment the failure counter only on periodic balance.
7597 * We do not want newidle balance, which can be very
7598 * frequent, pollute the failure counter causing
7599 * excessive cache_hot migrations and active balances.
7601 if (idle != CPU_NEWLY_IDLE)
7602 sd->nr_balance_failed++;
7604 if (need_active_balance(&env)) {
7605 raw_spin_lock_irqsave(&busiest->lock, flags);
7607 /* don't kick the active_load_balance_cpu_stop,
7608 * if the curr task on busiest cpu can't be
7611 if (!cpumask_test_cpu(this_cpu,
7612 tsk_cpus_allowed(busiest->curr))) {
7613 raw_spin_unlock_irqrestore(&busiest->lock,
7615 env.flags |= LBF_ALL_PINNED;
7616 goto out_one_pinned;
7620 * ->active_balance synchronizes accesses to
7621 * ->active_balance_work. Once set, it's cleared
7622 * only after active load balance is finished.
7624 if (!busiest->active_balance) {
7625 busiest->active_balance = 1;
7626 busiest->push_cpu = this_cpu;
7629 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7631 if (active_balance) {
7632 stop_one_cpu_nowait(cpu_of(busiest),
7633 active_load_balance_cpu_stop, busiest,
7634 &busiest->active_balance_work);
7637 /* We've kicked active balancing, force task migration. */
7638 sd->nr_balance_failed = sd->cache_nice_tries+1;
7641 sd->nr_balance_failed = 0;
7643 if (likely(!active_balance)) {
7644 /* We were unbalanced, so reset the balancing interval */
7645 sd->balance_interval = sd->min_interval;
7648 * If we've begun active balancing, start to back off. This
7649 * case may not be covered by the all_pinned logic if there
7650 * is only 1 task on the busy runqueue (because we don't call
7653 if (sd->balance_interval < sd->max_interval)
7654 sd->balance_interval *= 2;
7661 * We reach balance although we may have faced some affinity
7662 * constraints. Clear the imbalance flag if it was set.
7665 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7667 if (*group_imbalance)
7668 *group_imbalance = 0;
7673 * We reach balance because all tasks are pinned at this level so
7674 * we can't migrate them. Let the imbalance flag set so parent level
7675 * can try to migrate them.
7677 schedstat_inc(sd, lb_balanced[idle]);
7679 sd->nr_balance_failed = 0;
7682 /* tune up the balancing interval */
7683 if (((env.flags & LBF_ALL_PINNED) &&
7684 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7685 (sd->balance_interval < sd->max_interval))
7686 sd->balance_interval *= 2;
7693 static inline unsigned long
7694 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7696 unsigned long interval = sd->balance_interval;
7699 interval *= sd->busy_factor;
7701 /* scale ms to jiffies */
7702 interval = msecs_to_jiffies(interval);
7703 interval = clamp(interval, 1UL, max_load_balance_interval);
7709 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7711 unsigned long interval, next;
7713 interval = get_sd_balance_interval(sd, cpu_busy);
7714 next = sd->last_balance + interval;
7716 if (time_after(*next_balance, next))
7717 *next_balance = next;
7721 * idle_balance is called by schedule() if this_cpu is about to become
7722 * idle. Attempts to pull tasks from other CPUs.
7724 static int idle_balance(struct rq *this_rq)
7726 unsigned long next_balance = jiffies + HZ;
7727 int this_cpu = this_rq->cpu;
7728 struct sched_domain *sd;
7729 int pulled_task = 0;
7733 * We must set idle_stamp _before_ calling idle_balance(), such that we
7734 * measure the duration of idle_balance() as idle time.
7736 this_rq->idle_stamp = rq_clock(this_rq);
7738 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7739 !this_rq->rd->overload) {
7741 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7743 update_next_balance(sd, 0, &next_balance);
7749 raw_spin_unlock(&this_rq->lock);
7751 update_blocked_averages(this_cpu);
7753 for_each_domain(this_cpu, sd) {
7754 int continue_balancing = 1;
7755 u64 t0, domain_cost;
7757 if (!(sd->flags & SD_LOAD_BALANCE))
7760 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7761 update_next_balance(sd, 0, &next_balance);
7765 if (sd->flags & SD_BALANCE_NEWIDLE) {
7766 t0 = sched_clock_cpu(this_cpu);
7768 pulled_task = load_balance(this_cpu, this_rq,
7770 &continue_balancing);
7772 domain_cost = sched_clock_cpu(this_cpu) - t0;
7773 if (domain_cost > sd->max_newidle_lb_cost)
7774 sd->max_newidle_lb_cost = domain_cost;
7776 curr_cost += domain_cost;
7779 update_next_balance(sd, 0, &next_balance);
7782 * Stop searching for tasks to pull if there are
7783 * now runnable tasks on this rq.
7785 if (pulled_task || this_rq->nr_running > 0)
7790 raw_spin_lock(&this_rq->lock);
7792 if (curr_cost > this_rq->max_idle_balance_cost)
7793 this_rq->max_idle_balance_cost = curr_cost;
7796 * While browsing the domains, we released the rq lock, a task could
7797 * have been enqueued in the meantime. Since we're not going idle,
7798 * pretend we pulled a task.
7800 if (this_rq->cfs.h_nr_running && !pulled_task)
7804 /* Move the next balance forward */
7805 if (time_after(this_rq->next_balance, next_balance))
7806 this_rq->next_balance = next_balance;
7808 /* Is there a task of a high priority class? */
7809 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7813 this_rq->idle_stamp = 0;
7819 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7820 * running tasks off the busiest CPU onto idle CPUs. It requires at
7821 * least 1 task to be running on each physical CPU where possible, and
7822 * avoids physical / logical imbalances.
7824 static int active_load_balance_cpu_stop(void *data)
7826 struct rq *busiest_rq = data;
7827 int busiest_cpu = cpu_of(busiest_rq);
7828 int target_cpu = busiest_rq->push_cpu;
7829 struct rq *target_rq = cpu_rq(target_cpu);
7830 struct sched_domain *sd;
7831 struct task_struct *p = NULL;
7833 raw_spin_lock_irq(&busiest_rq->lock);
7835 /* make sure the requested cpu hasn't gone down in the meantime */
7836 if (unlikely(busiest_cpu != smp_processor_id() ||
7837 !busiest_rq->active_balance))
7840 /* Is there any task to move? */
7841 if (busiest_rq->nr_running <= 1)
7845 * This condition is "impossible", if it occurs
7846 * we need to fix it. Originally reported by
7847 * Bjorn Helgaas on a 128-cpu setup.
7849 BUG_ON(busiest_rq == target_rq);
7851 /* Search for an sd spanning us and the target CPU. */
7853 for_each_domain(target_cpu, sd) {
7854 if ((sd->flags & SD_LOAD_BALANCE) &&
7855 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7860 struct lb_env env = {
7862 .dst_cpu = target_cpu,
7863 .dst_rq = target_rq,
7864 .src_cpu = busiest_rq->cpu,
7865 .src_rq = busiest_rq,
7869 schedstat_inc(sd, alb_count);
7871 p = detach_one_task(&env);
7873 schedstat_inc(sd, alb_pushed);
7874 /* Active balancing done, reset the failure counter. */
7875 sd->nr_balance_failed = 0;
7877 schedstat_inc(sd, alb_failed);
7882 busiest_rq->active_balance = 0;
7883 raw_spin_unlock(&busiest_rq->lock);
7886 attach_one_task(target_rq, p);
7893 static inline int on_null_domain(struct rq *rq)
7895 return unlikely(!rcu_dereference_sched(rq->sd));
7898 #ifdef CONFIG_NO_HZ_COMMON
7900 * idle load balancing details
7901 * - When one of the busy CPUs notice that there may be an idle rebalancing
7902 * needed, they will kick the idle load balancer, which then does idle
7903 * load balancing for all the idle CPUs.
7906 cpumask_var_t idle_cpus_mask;
7908 unsigned long next_balance; /* in jiffy units */
7909 } nohz ____cacheline_aligned;
7911 static inline int find_new_ilb(void)
7913 int ilb = cpumask_first(nohz.idle_cpus_mask);
7915 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7922 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7923 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7924 * CPU (if there is one).
7926 static void nohz_balancer_kick(void)
7930 nohz.next_balance++;
7932 ilb_cpu = find_new_ilb();
7934 if (ilb_cpu >= nr_cpu_ids)
7937 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7940 * Use smp_send_reschedule() instead of resched_cpu().
7941 * This way we generate a sched IPI on the target cpu which
7942 * is idle. And the softirq performing nohz idle load balance
7943 * will be run before returning from the IPI.
7945 smp_send_reschedule(ilb_cpu);
7949 void nohz_balance_exit_idle(unsigned int cpu)
7951 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7953 * Completely isolated CPUs don't ever set, so we must test.
7955 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7956 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7957 atomic_dec(&nohz.nr_cpus);
7959 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7963 static inline void set_cpu_sd_state_busy(void)
7965 struct sched_domain *sd;
7966 int cpu = smp_processor_id();
7969 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7971 if (!sd || !sd->nohz_idle)
7975 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7980 void set_cpu_sd_state_idle(void)
7982 struct sched_domain *sd;
7983 int cpu = smp_processor_id();
7986 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7988 if (!sd || sd->nohz_idle)
7992 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7998 * This routine will record that the cpu is going idle with tick stopped.
7999 * This info will be used in performing idle load balancing in the future.
8001 void nohz_balance_enter_idle(int cpu)
8004 * If this cpu is going down, then nothing needs to be done.
8006 if (!cpu_active(cpu))
8009 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8013 * If we're a completely isolated CPU, we don't play.
8015 if (on_null_domain(cpu_rq(cpu)))
8018 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8019 atomic_inc(&nohz.nr_cpus);
8020 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8024 static DEFINE_SPINLOCK(balancing);
8027 * Scale the max load_balance interval with the number of CPUs in the system.
8028 * This trades load-balance latency on larger machines for less cross talk.
8030 void update_max_interval(void)
8032 max_load_balance_interval = HZ*num_online_cpus()/10;
8036 * It checks each scheduling domain to see if it is due to be balanced,
8037 * and initiates a balancing operation if so.
8039 * Balancing parameters are set up in init_sched_domains.
8041 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8043 int continue_balancing = 1;
8045 unsigned long interval;
8046 struct sched_domain *sd;
8047 /* Earliest time when we have to do rebalance again */
8048 unsigned long next_balance = jiffies + 60*HZ;
8049 int update_next_balance = 0;
8050 int need_serialize, need_decay = 0;
8053 update_blocked_averages(cpu);
8056 for_each_domain(cpu, sd) {
8058 * Decay the newidle max times here because this is a regular
8059 * visit to all the domains. Decay ~1% per second.
8061 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8062 sd->max_newidle_lb_cost =
8063 (sd->max_newidle_lb_cost * 253) / 256;
8064 sd->next_decay_max_lb_cost = jiffies + HZ;
8067 max_cost += sd->max_newidle_lb_cost;
8069 if (!(sd->flags & SD_LOAD_BALANCE))
8073 * Stop the load balance at this level. There is another
8074 * CPU in our sched group which is doing load balancing more
8077 if (!continue_balancing) {
8083 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8085 need_serialize = sd->flags & SD_SERIALIZE;
8086 if (need_serialize) {
8087 if (!spin_trylock(&balancing))
8091 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8092 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8094 * The LBF_DST_PINNED logic could have changed
8095 * env->dst_cpu, so we can't know our idle
8096 * state even if we migrated tasks. Update it.
8098 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8100 sd->last_balance = jiffies;
8101 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8104 spin_unlock(&balancing);
8106 if (time_after(next_balance, sd->last_balance + interval)) {
8107 next_balance = sd->last_balance + interval;
8108 update_next_balance = 1;
8113 * Ensure the rq-wide value also decays but keep it at a
8114 * reasonable floor to avoid funnies with rq->avg_idle.
8116 rq->max_idle_balance_cost =
8117 max((u64)sysctl_sched_migration_cost, max_cost);
8122 * next_balance will be updated only when there is a need.
8123 * When the cpu is attached to null domain for ex, it will not be
8126 if (likely(update_next_balance)) {
8127 rq->next_balance = next_balance;
8129 #ifdef CONFIG_NO_HZ_COMMON
8131 * If this CPU has been elected to perform the nohz idle
8132 * balance. Other idle CPUs have already rebalanced with
8133 * nohz_idle_balance() and nohz.next_balance has been
8134 * updated accordingly. This CPU is now running the idle load
8135 * balance for itself and we need to update the
8136 * nohz.next_balance accordingly.
8138 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8139 nohz.next_balance = rq->next_balance;
8144 #ifdef CONFIG_NO_HZ_COMMON
8146 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8147 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8149 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8151 int this_cpu = this_rq->cpu;
8154 /* Earliest time when we have to do rebalance again */
8155 unsigned long next_balance = jiffies + 60*HZ;
8156 int update_next_balance = 0;
8158 if (idle != CPU_IDLE ||
8159 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8162 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8163 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8167 * If this cpu gets work to do, stop the load balancing
8168 * work being done for other cpus. Next load
8169 * balancing owner will pick it up.
8174 rq = cpu_rq(balance_cpu);
8177 * If time for next balance is due,
8180 if (time_after_eq(jiffies, rq->next_balance)) {
8181 raw_spin_lock_irq(&rq->lock);
8182 update_rq_clock(rq);
8183 cpu_load_update_idle(rq);
8184 raw_spin_unlock_irq(&rq->lock);
8185 rebalance_domains(rq, CPU_IDLE);
8188 if (time_after(next_balance, rq->next_balance)) {
8189 next_balance = rq->next_balance;
8190 update_next_balance = 1;
8195 * next_balance will be updated only when there is a need.
8196 * When the CPU is attached to null domain for ex, it will not be
8199 if (likely(update_next_balance))
8200 nohz.next_balance = next_balance;
8202 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8206 * Current heuristic for kicking the idle load balancer in the presence
8207 * of an idle cpu in the system.
8208 * - This rq has more than one task.
8209 * - This rq has at least one CFS task and the capacity of the CPU is
8210 * significantly reduced because of RT tasks or IRQs.
8211 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8212 * multiple busy cpu.
8213 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8214 * domain span are idle.
8216 static inline bool nohz_kick_needed(struct rq *rq)
8218 unsigned long now = jiffies;
8219 struct sched_domain *sd;
8220 struct sched_group_capacity *sgc;
8221 int nr_busy, cpu = rq->cpu;
8224 if (unlikely(rq->idle_balance))
8228 * We may be recently in ticked or tickless idle mode. At the first
8229 * busy tick after returning from idle, we will update the busy stats.
8231 set_cpu_sd_state_busy();
8232 nohz_balance_exit_idle(cpu);
8235 * None are in tickless mode and hence no need for NOHZ idle load
8238 if (likely(!atomic_read(&nohz.nr_cpus)))
8241 if (time_before(now, nohz.next_balance))
8244 if (rq->nr_running >= 2)
8248 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8250 sgc = sd->groups->sgc;
8251 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8260 sd = rcu_dereference(rq->sd);
8262 if ((rq->cfs.h_nr_running >= 1) &&
8263 check_cpu_capacity(rq, sd)) {
8269 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8270 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8271 sched_domain_span(sd)) < cpu)) {
8281 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8285 * run_rebalance_domains is triggered when needed from the scheduler tick.
8286 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8288 static void run_rebalance_domains(struct softirq_action *h)
8290 struct rq *this_rq = this_rq();
8291 enum cpu_idle_type idle = this_rq->idle_balance ?
8292 CPU_IDLE : CPU_NOT_IDLE;
8295 * If this cpu has a pending nohz_balance_kick, then do the
8296 * balancing on behalf of the other idle cpus whose ticks are
8297 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8298 * give the idle cpus a chance to load balance. Else we may
8299 * load balance only within the local sched_domain hierarchy
8300 * and abort nohz_idle_balance altogether if we pull some load.
8302 nohz_idle_balance(this_rq, idle);
8303 rebalance_domains(this_rq, idle);
8307 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8309 void trigger_load_balance(struct rq *rq)
8311 /* Don't need to rebalance while attached to NULL domain */
8312 if (unlikely(on_null_domain(rq)))
8315 if (time_after_eq(jiffies, rq->next_balance))
8316 raise_softirq(SCHED_SOFTIRQ);
8317 #ifdef CONFIG_NO_HZ_COMMON
8318 if (nohz_kick_needed(rq))
8319 nohz_balancer_kick();
8323 static void rq_online_fair(struct rq *rq)
8327 update_runtime_enabled(rq);
8330 static void rq_offline_fair(struct rq *rq)
8334 /* Ensure any throttled groups are reachable by pick_next_task */
8335 unthrottle_offline_cfs_rqs(rq);
8338 #endif /* CONFIG_SMP */
8341 * scheduler tick hitting a task of our scheduling class:
8343 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8345 struct cfs_rq *cfs_rq;
8346 struct sched_entity *se = &curr->se;
8348 for_each_sched_entity(se) {
8349 cfs_rq = cfs_rq_of(se);
8350 entity_tick(cfs_rq, se, queued);
8353 if (static_branch_unlikely(&sched_numa_balancing))
8354 task_tick_numa(rq, curr);
8358 * called on fork with the child task as argument from the parent's context
8359 * - child not yet on the tasklist
8360 * - preemption disabled
8362 static void task_fork_fair(struct task_struct *p)
8364 struct cfs_rq *cfs_rq;
8365 struct sched_entity *se = &p->se, *curr;
8366 struct rq *rq = this_rq();
8368 raw_spin_lock(&rq->lock);
8369 update_rq_clock(rq);
8371 cfs_rq = task_cfs_rq(current);
8372 curr = cfs_rq->curr;
8374 update_curr(cfs_rq);
8375 se->vruntime = curr->vruntime;
8377 place_entity(cfs_rq, se, 1);
8379 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8381 * Upon rescheduling, sched_class::put_prev_task() will place
8382 * 'current' within the tree based on its new key value.
8384 swap(curr->vruntime, se->vruntime);
8388 se->vruntime -= cfs_rq->min_vruntime;
8389 raw_spin_unlock(&rq->lock);
8393 * Priority of the task has changed. Check to see if we preempt
8397 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8399 if (!task_on_rq_queued(p))
8403 * Reschedule if we are currently running on this runqueue and
8404 * our priority decreased, or if we are not currently running on
8405 * this runqueue and our priority is higher than the current's
8407 if (rq->curr == p) {
8408 if (p->prio > oldprio)
8411 check_preempt_curr(rq, p, 0);
8414 static inline bool vruntime_normalized(struct task_struct *p)
8416 struct sched_entity *se = &p->se;
8419 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8420 * the dequeue_entity(.flags=0) will already have normalized the
8427 * When !on_rq, vruntime of the task has usually NOT been normalized.
8428 * But there are some cases where it has already been normalized:
8430 * - A forked child which is waiting for being woken up by
8431 * wake_up_new_task().
8432 * - A task which has been woken up by try_to_wake_up() and
8433 * waiting for actually being woken up by sched_ttwu_pending().
8435 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8441 static void detach_task_cfs_rq(struct task_struct *p)
8443 struct sched_entity *se = &p->se;
8444 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8445 u64 now = cfs_rq_clock_task(cfs_rq);
8448 if (!vruntime_normalized(p)) {
8450 * Fix up our vruntime so that the current sleep doesn't
8451 * cause 'unlimited' sleep bonus.
8453 place_entity(cfs_rq, se, 0);
8454 se->vruntime -= cfs_rq->min_vruntime;
8457 /* Catch up with the cfs_rq and remove our load when we leave */
8458 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
8459 detach_entity_load_avg(cfs_rq, se);
8461 update_tg_load_avg(cfs_rq, false);
8464 static void attach_task_cfs_rq(struct task_struct *p)
8466 struct sched_entity *se = &p->se;
8467 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8468 u64 now = cfs_rq_clock_task(cfs_rq);
8471 #ifdef CONFIG_FAIR_GROUP_SCHED
8473 * Since the real-depth could have been changed (only FAIR
8474 * class maintain depth value), reset depth properly.
8476 se->depth = se->parent ? se->parent->depth + 1 : 0;
8479 /* Synchronize task with its cfs_rq */
8480 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
8481 attach_entity_load_avg(cfs_rq, se);
8483 update_tg_load_avg(cfs_rq, false);
8485 if (!vruntime_normalized(p))
8486 se->vruntime += cfs_rq->min_vruntime;
8489 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8491 detach_task_cfs_rq(p);
8494 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8496 attach_task_cfs_rq(p);
8498 if (task_on_rq_queued(p)) {
8500 * We were most likely switched from sched_rt, so
8501 * kick off the schedule if running, otherwise just see
8502 * if we can still preempt the current task.
8507 check_preempt_curr(rq, p, 0);
8511 /* Account for a task changing its policy or group.
8513 * This routine is mostly called to set cfs_rq->curr field when a task
8514 * migrates between groups/classes.
8516 static void set_curr_task_fair(struct rq *rq)
8518 struct sched_entity *se = &rq->curr->se;
8520 for_each_sched_entity(se) {
8521 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8523 set_next_entity(cfs_rq, se);
8524 /* ensure bandwidth has been allocated on our new cfs_rq */
8525 account_cfs_rq_runtime(cfs_rq, 0);
8529 void init_cfs_rq(struct cfs_rq *cfs_rq)
8531 cfs_rq->tasks_timeline = RB_ROOT;
8532 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8533 #ifndef CONFIG_64BIT
8534 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8537 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8538 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8542 #ifdef CONFIG_FAIR_GROUP_SCHED
8543 static void task_set_group_fair(struct task_struct *p)
8545 struct sched_entity *se = &p->se;
8547 set_task_rq(p, task_cpu(p));
8548 se->depth = se->parent ? se->parent->depth + 1 : 0;
8551 static void task_move_group_fair(struct task_struct *p)
8553 detach_task_cfs_rq(p);
8554 set_task_rq(p, task_cpu(p));
8557 /* Tell se's cfs_rq has been changed -- migrated */
8558 p->se.avg.last_update_time = 0;
8560 attach_task_cfs_rq(p);
8563 static void task_change_group_fair(struct task_struct *p, int type)
8566 case TASK_SET_GROUP:
8567 task_set_group_fair(p);
8570 case TASK_MOVE_GROUP:
8571 task_move_group_fair(p);
8576 void free_fair_sched_group(struct task_group *tg)
8580 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8582 for_each_possible_cpu(i) {
8584 kfree(tg->cfs_rq[i]);
8593 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8595 struct sched_entity *se;
8596 struct cfs_rq *cfs_rq;
8600 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8603 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8607 tg->shares = NICE_0_LOAD;
8609 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8611 for_each_possible_cpu(i) {
8614 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8615 GFP_KERNEL, cpu_to_node(i));
8619 se = kzalloc_node(sizeof(struct sched_entity),
8620 GFP_KERNEL, cpu_to_node(i));
8624 init_cfs_rq(cfs_rq);
8625 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8626 init_entity_runnable_average(se);
8628 raw_spin_lock_irq(&rq->lock);
8629 post_init_entity_util_avg(se);
8630 raw_spin_unlock_irq(&rq->lock);
8641 void unregister_fair_sched_group(struct task_group *tg)
8643 unsigned long flags;
8647 for_each_possible_cpu(cpu) {
8649 remove_entity_load_avg(tg->se[cpu]);
8652 * Only empty task groups can be destroyed; so we can speculatively
8653 * check on_list without danger of it being re-added.
8655 if (!tg->cfs_rq[cpu]->on_list)
8660 raw_spin_lock_irqsave(&rq->lock, flags);
8661 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8662 raw_spin_unlock_irqrestore(&rq->lock, flags);
8666 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8667 struct sched_entity *se, int cpu,
8668 struct sched_entity *parent)
8670 struct rq *rq = cpu_rq(cpu);
8674 init_cfs_rq_runtime(cfs_rq);
8676 tg->cfs_rq[cpu] = cfs_rq;
8679 /* se could be NULL for root_task_group */
8684 se->cfs_rq = &rq->cfs;
8687 se->cfs_rq = parent->my_q;
8688 se->depth = parent->depth + 1;
8692 /* guarantee group entities always have weight */
8693 update_load_set(&se->load, NICE_0_LOAD);
8694 se->parent = parent;
8697 static DEFINE_MUTEX(shares_mutex);
8699 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8702 unsigned long flags;
8705 * We can't change the weight of the root cgroup.
8710 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8712 mutex_lock(&shares_mutex);
8713 if (tg->shares == shares)
8716 tg->shares = shares;
8717 for_each_possible_cpu(i) {
8718 struct rq *rq = cpu_rq(i);
8719 struct sched_entity *se;
8722 /* Propagate contribution to hierarchy */
8723 raw_spin_lock_irqsave(&rq->lock, flags);
8725 /* Possible calls to update_curr() need rq clock */
8726 update_rq_clock(rq);
8727 for_each_sched_entity(se)
8728 update_cfs_shares(group_cfs_rq(se));
8729 raw_spin_unlock_irqrestore(&rq->lock, flags);
8733 mutex_unlock(&shares_mutex);
8736 #else /* CONFIG_FAIR_GROUP_SCHED */
8738 void free_fair_sched_group(struct task_group *tg) { }
8740 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8745 void unregister_fair_sched_group(struct task_group *tg) { }
8747 #endif /* CONFIG_FAIR_GROUP_SCHED */
8750 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8752 struct sched_entity *se = &task->se;
8753 unsigned int rr_interval = 0;
8756 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8759 if (rq->cfs.load.weight)
8760 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8766 * All the scheduling class methods:
8768 const struct sched_class fair_sched_class = {
8769 .next = &idle_sched_class,
8770 .enqueue_task = enqueue_task_fair,
8771 .dequeue_task = dequeue_task_fair,
8772 .yield_task = yield_task_fair,
8773 .yield_to_task = yield_to_task_fair,
8775 .check_preempt_curr = check_preempt_wakeup,
8777 .pick_next_task = pick_next_task_fair,
8778 .put_prev_task = put_prev_task_fair,
8781 .select_task_rq = select_task_rq_fair,
8782 .migrate_task_rq = migrate_task_rq_fair,
8784 .rq_online = rq_online_fair,
8785 .rq_offline = rq_offline_fair,
8787 .task_dead = task_dead_fair,
8788 .set_cpus_allowed = set_cpus_allowed_common,
8791 .set_curr_task = set_curr_task_fair,
8792 .task_tick = task_tick_fair,
8793 .task_fork = task_fork_fair,
8795 .prio_changed = prio_changed_fair,
8796 .switched_from = switched_from_fair,
8797 .switched_to = switched_to_fair,
8799 .get_rr_interval = get_rr_interval_fair,
8801 .update_curr = update_curr_fair,
8803 #ifdef CONFIG_FAIR_GROUP_SCHED
8804 .task_change_group = task_change_group_fair,
8808 #ifdef CONFIG_SCHED_DEBUG
8809 void print_cfs_stats(struct seq_file *m, int cpu)
8811 struct cfs_rq *cfs_rq;
8814 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8815 print_cfs_rq(m, cpu, cfs_rq);
8819 #ifdef CONFIG_NUMA_BALANCING
8820 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8823 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8825 for_each_online_node(node) {
8826 if (p->numa_faults) {
8827 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8828 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8830 if (p->numa_group) {
8831 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8832 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8834 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8837 #endif /* CONFIG_NUMA_BALANCING */
8838 #endif /* CONFIG_SCHED_DEBUG */
8840 __init void init_sched_fair_class(void)
8843 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8845 #ifdef CONFIG_NO_HZ_COMMON
8846 nohz.next_balance = jiffies;
8847 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);