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 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
698 * With new tasks being created, their initial util_avgs are extrapolated
699 * based on the cfs_rq's current util_avg:
701 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
703 * However, in many cases, the above util_avg does not give a desired
704 * value. Moreover, the sum of the util_avgs may be divergent, such
705 * as when the series is a harmonic series.
707 * To solve this problem, we also cap the util_avg of successive tasks to
708 * only 1/2 of the left utilization budget:
710 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
712 * where n denotes the nth task.
714 * For example, a simplest series from the beginning would be like:
716 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
717 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
719 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
720 * if util_avg > util_avg_cap.
722 void post_init_entity_util_avg(struct sched_entity *se)
724 struct cfs_rq *cfs_rq = cfs_rq_of(se);
725 struct sched_avg *sa = &se->avg;
726 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
727 u64 now = cfs_rq_clock_task(cfs_rq);
730 if (cfs_rq->avg.util_avg != 0) {
731 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
732 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
734 if (sa->util_avg > cap)
739 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
742 if (entity_is_task(se)) {
743 struct task_struct *p = task_of(se);
744 if (p->sched_class != &fair_sched_class) {
746 * For !fair tasks do:
748 update_cfs_rq_load_avg(now, cfs_rq, false);
749 attach_entity_load_avg(cfs_rq, se);
750 switched_from_fair(rq, p);
752 * such that the next switched_to_fair() has the
755 se->avg.last_update_time = now;
760 update_cfs_rq_load_avg(now, cfs_rq, false);
761 attach_entity_load_avg(cfs_rq, se);
764 #else /* !CONFIG_SMP */
765 void init_entity_runnable_average(struct sched_entity *se)
768 void post_init_entity_util_avg(struct sched_entity *se)
771 #endif /* CONFIG_SMP */
774 * Update the current task's runtime statistics.
776 static void update_curr(struct cfs_rq *cfs_rq)
778 struct sched_entity *curr = cfs_rq->curr;
779 u64 now = rq_clock_task(rq_of(cfs_rq));
785 delta_exec = now - curr->exec_start;
786 if (unlikely((s64)delta_exec <= 0))
789 curr->exec_start = now;
791 schedstat_set(curr->statistics.exec_max,
792 max(delta_exec, curr->statistics.exec_max));
794 curr->sum_exec_runtime += delta_exec;
795 schedstat_add(cfs_rq, exec_clock, delta_exec);
797 curr->vruntime += calc_delta_fair(delta_exec, curr);
798 update_min_vruntime(cfs_rq);
800 if (entity_is_task(curr)) {
801 struct task_struct *curtask = task_of(curr);
803 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
804 cpuacct_charge(curtask, delta_exec);
805 account_group_exec_runtime(curtask, delta_exec);
808 account_cfs_rq_runtime(cfs_rq, delta_exec);
811 static void update_curr_fair(struct rq *rq)
813 update_curr(cfs_rq_of(&rq->curr->se));
816 #ifdef CONFIG_SCHEDSTATS
818 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
820 u64 wait_start = rq_clock(rq_of(cfs_rq));
822 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
823 likely(wait_start > se->statistics.wait_start))
824 wait_start -= se->statistics.wait_start;
826 se->statistics.wait_start = wait_start;
830 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
832 struct task_struct *p;
835 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
837 if (entity_is_task(se)) {
839 if (task_on_rq_migrating(p)) {
841 * Preserve migrating task's wait time so wait_start
842 * time stamp can be adjusted to accumulate wait time
843 * prior to migration.
845 se->statistics.wait_start = delta;
848 trace_sched_stat_wait(p, delta);
851 se->statistics.wait_max = max(se->statistics.wait_max, delta);
852 se->statistics.wait_count++;
853 se->statistics.wait_sum += delta;
854 se->statistics.wait_start = 0;
858 * Task is being enqueued - update stats:
861 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
864 * Are we enqueueing a waiting task? (for current tasks
865 * a dequeue/enqueue event is a NOP)
867 if (se != cfs_rq->curr)
868 update_stats_wait_start(cfs_rq, se);
872 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
875 * Mark the end of the wait period if dequeueing a
878 if (se != cfs_rq->curr)
879 update_stats_wait_end(cfs_rq, se);
881 if (flags & DEQUEUE_SLEEP) {
882 if (entity_is_task(se)) {
883 struct task_struct *tsk = task_of(se);
885 if (tsk->state & TASK_INTERRUPTIBLE)
886 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
887 if (tsk->state & TASK_UNINTERRUPTIBLE)
888 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
895 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
900 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
905 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
910 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
916 * We are picking a new current task - update its stats:
919 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
922 * We are starting a new run period:
924 se->exec_start = rq_clock_task(rq_of(cfs_rq));
927 /**************************************************
928 * Scheduling class queueing methods:
931 #ifdef CONFIG_NUMA_BALANCING
933 * Approximate time to scan a full NUMA task in ms. The task scan period is
934 * calculated based on the tasks virtual memory size and
935 * numa_balancing_scan_size.
937 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
938 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
940 /* Portion of address space to scan in MB */
941 unsigned int sysctl_numa_balancing_scan_size = 256;
943 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
944 unsigned int sysctl_numa_balancing_scan_delay = 1000;
946 static unsigned int task_nr_scan_windows(struct task_struct *p)
948 unsigned long rss = 0;
949 unsigned long nr_scan_pages;
952 * Calculations based on RSS as non-present and empty pages are skipped
953 * by the PTE scanner and NUMA hinting faults should be trapped based
956 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
957 rss = get_mm_rss(p->mm);
961 rss = round_up(rss, nr_scan_pages);
962 return rss / nr_scan_pages;
965 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
966 #define MAX_SCAN_WINDOW 2560
968 static unsigned int task_scan_min(struct task_struct *p)
970 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
971 unsigned int scan, floor;
972 unsigned int windows = 1;
974 if (scan_size < MAX_SCAN_WINDOW)
975 windows = MAX_SCAN_WINDOW / scan_size;
976 floor = 1000 / windows;
978 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
979 return max_t(unsigned int, floor, scan);
982 static unsigned int task_scan_max(struct task_struct *p)
984 unsigned int smin = task_scan_min(p);
987 /* Watch for min being lower than max due to floor calculations */
988 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
989 return max(smin, smax);
992 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
994 rq->nr_numa_running += (p->numa_preferred_nid != -1);
995 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
998 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1000 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1001 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1007 spinlock_t lock; /* nr_tasks, tasks */
1012 struct rcu_head rcu;
1013 unsigned long total_faults;
1014 unsigned long max_faults_cpu;
1016 * Faults_cpu is used to decide whether memory should move
1017 * towards the CPU. As a consequence, these stats are weighted
1018 * more by CPU use than by memory faults.
1020 unsigned long *faults_cpu;
1021 unsigned long faults[0];
1024 /* Shared or private faults. */
1025 #define NR_NUMA_HINT_FAULT_TYPES 2
1027 /* Memory and CPU locality */
1028 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1030 /* Averaged statistics, and temporary buffers. */
1031 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1033 pid_t task_numa_group_id(struct task_struct *p)
1035 return p->numa_group ? p->numa_group->gid : 0;
1039 * The averaged statistics, shared & private, memory & cpu,
1040 * occupy the first half of the array. The second half of the
1041 * array is for current counters, which are averaged into the
1042 * first set by task_numa_placement.
1044 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1046 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1049 static inline unsigned long task_faults(struct task_struct *p, int nid)
1051 if (!p->numa_faults)
1054 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1055 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1058 static inline unsigned long group_faults(struct task_struct *p, int nid)
1063 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1064 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1067 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1069 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1070 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1074 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1075 * considered part of a numa group's pseudo-interleaving set. Migrations
1076 * between these nodes are slowed down, to allow things to settle down.
1078 #define ACTIVE_NODE_FRACTION 3
1080 static bool numa_is_active_node(int nid, struct numa_group *ng)
1082 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1085 /* Handle placement on systems where not all nodes are directly connected. */
1086 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1087 int maxdist, bool task)
1089 unsigned long score = 0;
1093 * All nodes are directly connected, and the same distance
1094 * from each other. No need for fancy placement algorithms.
1096 if (sched_numa_topology_type == NUMA_DIRECT)
1100 * This code is called for each node, introducing N^2 complexity,
1101 * which should be ok given the number of nodes rarely exceeds 8.
1103 for_each_online_node(node) {
1104 unsigned long faults;
1105 int dist = node_distance(nid, node);
1108 * The furthest away nodes in the system are not interesting
1109 * for placement; nid was already counted.
1111 if (dist == sched_max_numa_distance || node == nid)
1115 * On systems with a backplane NUMA topology, compare groups
1116 * of nodes, and move tasks towards the group with the most
1117 * memory accesses. When comparing two nodes at distance
1118 * "hoplimit", only nodes closer by than "hoplimit" are part
1119 * of each group. Skip other nodes.
1121 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1125 /* Add up the faults from nearby nodes. */
1127 faults = task_faults(p, node);
1129 faults = group_faults(p, node);
1132 * On systems with a glueless mesh NUMA topology, there are
1133 * no fixed "groups of nodes". Instead, nodes that are not
1134 * directly connected bounce traffic through intermediate
1135 * nodes; a numa_group can occupy any set of nodes.
1136 * The further away a node is, the less the faults count.
1137 * This seems to result in good task placement.
1139 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1140 faults *= (sched_max_numa_distance - dist);
1141 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1151 * These return the fraction of accesses done by a particular task, or
1152 * task group, on a particular numa node. The group weight is given a
1153 * larger multiplier, in order to group tasks together that are almost
1154 * evenly spread out between numa nodes.
1156 static inline unsigned long task_weight(struct task_struct *p, int nid,
1159 unsigned long faults, total_faults;
1161 if (!p->numa_faults)
1164 total_faults = p->total_numa_faults;
1169 faults = task_faults(p, nid);
1170 faults += score_nearby_nodes(p, nid, dist, true);
1172 return 1000 * faults / total_faults;
1175 static inline unsigned long group_weight(struct task_struct *p, int nid,
1178 unsigned long faults, total_faults;
1183 total_faults = p->numa_group->total_faults;
1188 faults = group_faults(p, nid);
1189 faults += score_nearby_nodes(p, nid, dist, false);
1191 return 1000 * faults / total_faults;
1194 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1195 int src_nid, int dst_cpu)
1197 struct numa_group *ng = p->numa_group;
1198 int dst_nid = cpu_to_node(dst_cpu);
1199 int last_cpupid, this_cpupid;
1201 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1204 * Multi-stage node selection is used in conjunction with a periodic
1205 * migration fault to build a temporal task<->page relation. By using
1206 * a two-stage filter we remove short/unlikely relations.
1208 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1209 * a task's usage of a particular page (n_p) per total usage of this
1210 * page (n_t) (in a given time-span) to a probability.
1212 * Our periodic faults will sample this probability and getting the
1213 * same result twice in a row, given these samples are fully
1214 * independent, is then given by P(n)^2, provided our sample period
1215 * is sufficiently short compared to the usage pattern.
1217 * This quadric squishes small probabilities, making it less likely we
1218 * act on an unlikely task<->page relation.
1220 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1221 if (!cpupid_pid_unset(last_cpupid) &&
1222 cpupid_to_nid(last_cpupid) != dst_nid)
1225 /* Always allow migrate on private faults */
1226 if (cpupid_match_pid(p, last_cpupid))
1229 /* A shared fault, but p->numa_group has not been set up yet. */
1234 * Destination node is much more heavily used than the source
1235 * node? Allow migration.
1237 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1238 ACTIVE_NODE_FRACTION)
1242 * Distribute memory according to CPU & memory use on each node,
1243 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1245 * faults_cpu(dst) 3 faults_cpu(src)
1246 * --------------- * - > ---------------
1247 * faults_mem(dst) 4 faults_mem(src)
1249 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1250 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1253 static unsigned long weighted_cpuload(const int cpu);
1254 static unsigned long source_load(int cpu, int type);
1255 static unsigned long target_load(int cpu, int type);
1256 static unsigned long capacity_of(int cpu);
1257 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1259 /* Cached statistics for all CPUs within a node */
1261 unsigned long nr_running;
1264 /* Total compute capacity of CPUs on a node */
1265 unsigned long compute_capacity;
1267 /* Approximate capacity in terms of runnable tasks on a node */
1268 unsigned long task_capacity;
1269 int has_free_capacity;
1273 * XXX borrowed from update_sg_lb_stats
1275 static void update_numa_stats(struct numa_stats *ns, int nid)
1277 int smt, cpu, cpus = 0;
1278 unsigned long capacity;
1280 memset(ns, 0, sizeof(*ns));
1281 for_each_cpu(cpu, cpumask_of_node(nid)) {
1282 struct rq *rq = cpu_rq(cpu);
1284 ns->nr_running += rq->nr_running;
1285 ns->load += weighted_cpuload(cpu);
1286 ns->compute_capacity += capacity_of(cpu);
1292 * If we raced with hotplug and there are no CPUs left in our mask
1293 * the @ns structure is NULL'ed and task_numa_compare() will
1294 * not find this node attractive.
1296 * We'll either bail at !has_free_capacity, or we'll detect a huge
1297 * imbalance and bail there.
1302 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1303 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1304 capacity = cpus / smt; /* cores */
1306 ns->task_capacity = min_t(unsigned, capacity,
1307 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1308 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1311 struct task_numa_env {
1312 struct task_struct *p;
1314 int src_cpu, src_nid;
1315 int dst_cpu, dst_nid;
1317 struct numa_stats src_stats, dst_stats;
1322 struct task_struct *best_task;
1327 static void task_numa_assign(struct task_numa_env *env,
1328 struct task_struct *p, long imp)
1331 put_task_struct(env->best_task);
1336 env->best_imp = imp;
1337 env->best_cpu = env->dst_cpu;
1340 static bool load_too_imbalanced(long src_load, long dst_load,
1341 struct task_numa_env *env)
1344 long orig_src_load, orig_dst_load;
1345 long src_capacity, dst_capacity;
1348 * The load is corrected for the CPU capacity available on each node.
1351 * ------------ vs ---------
1352 * src_capacity dst_capacity
1354 src_capacity = env->src_stats.compute_capacity;
1355 dst_capacity = env->dst_stats.compute_capacity;
1357 /* We care about the slope of the imbalance, not the direction. */
1358 if (dst_load < src_load)
1359 swap(dst_load, src_load);
1361 /* Is the difference below the threshold? */
1362 imb = dst_load * src_capacity * 100 -
1363 src_load * dst_capacity * env->imbalance_pct;
1368 * The imbalance is above the allowed threshold.
1369 * Compare it with the old imbalance.
1371 orig_src_load = env->src_stats.load;
1372 orig_dst_load = env->dst_stats.load;
1374 if (orig_dst_load < orig_src_load)
1375 swap(orig_dst_load, orig_src_load);
1377 old_imb = orig_dst_load * src_capacity * 100 -
1378 orig_src_load * dst_capacity * env->imbalance_pct;
1380 /* Would this change make things worse? */
1381 return (imb > old_imb);
1385 * This checks if the overall compute and NUMA accesses of the system would
1386 * be improved if the source tasks was migrated to the target dst_cpu taking
1387 * into account that it might be best if task running on the dst_cpu should
1388 * be exchanged with the source task
1390 static void task_numa_compare(struct task_numa_env *env,
1391 long taskimp, long groupimp)
1393 struct rq *src_rq = cpu_rq(env->src_cpu);
1394 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1395 struct task_struct *cur;
1396 long src_load, dst_load;
1398 long imp = env->p->numa_group ? groupimp : taskimp;
1400 int dist = env->dist;
1403 cur = task_rcu_dereference(&dst_rq->curr);
1404 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1408 * Because we have preemption enabled we can get migrated around and
1409 * end try selecting ourselves (current == env->p) as a swap candidate.
1415 * "imp" is the fault differential for the source task between the
1416 * source and destination node. Calculate the total differential for
1417 * the source task and potential destination task. The more negative
1418 * the value is, the more rmeote accesses that would be expected to
1419 * be incurred if the tasks were swapped.
1422 /* Skip this swap candidate if cannot move to the source cpu */
1423 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1427 * If dst and source tasks are in the same NUMA group, or not
1428 * in any group then look only at task weights.
1430 if (cur->numa_group == env->p->numa_group) {
1431 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1432 task_weight(cur, env->dst_nid, dist);
1434 * Add some hysteresis to prevent swapping the
1435 * tasks within a group over tiny differences.
1437 if (cur->numa_group)
1441 * Compare the group weights. If a task is all by
1442 * itself (not part of a group), use the task weight
1445 if (cur->numa_group)
1446 imp += group_weight(cur, env->src_nid, dist) -
1447 group_weight(cur, env->dst_nid, dist);
1449 imp += task_weight(cur, env->src_nid, dist) -
1450 task_weight(cur, env->dst_nid, dist);
1454 if (imp <= env->best_imp && moveimp <= env->best_imp)
1458 /* Is there capacity at our destination? */
1459 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1460 !env->dst_stats.has_free_capacity)
1466 /* Balance doesn't matter much if we're running a task per cpu */
1467 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1468 dst_rq->nr_running == 1)
1472 * In the overloaded case, try and keep the load balanced.
1475 load = task_h_load(env->p);
1476 dst_load = env->dst_stats.load + load;
1477 src_load = env->src_stats.load - load;
1479 if (moveimp > imp && moveimp > env->best_imp) {
1481 * If the improvement from just moving env->p direction is
1482 * better than swapping tasks around, check if a move is
1483 * possible. Store a slightly smaller score than moveimp,
1484 * so an actually idle CPU will win.
1486 if (!load_too_imbalanced(src_load, dst_load, env)) {
1493 if (imp <= env->best_imp)
1497 load = task_h_load(cur);
1502 if (load_too_imbalanced(src_load, dst_load, env))
1506 * One idle CPU per node is evaluated for a task numa move.
1507 * Call select_idle_sibling to maybe find a better one.
1510 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1513 task_numa_assign(env, cur, imp);
1518 static void task_numa_find_cpu(struct task_numa_env *env,
1519 long taskimp, long groupimp)
1523 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1524 /* Skip this CPU if the source task cannot migrate */
1525 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1529 task_numa_compare(env, taskimp, groupimp);
1533 /* Only move tasks to a NUMA node less busy than the current node. */
1534 static bool numa_has_capacity(struct task_numa_env *env)
1536 struct numa_stats *src = &env->src_stats;
1537 struct numa_stats *dst = &env->dst_stats;
1539 if (src->has_free_capacity && !dst->has_free_capacity)
1543 * Only consider a task move if the source has a higher load
1544 * than the destination, corrected for CPU capacity on each node.
1546 * src->load dst->load
1547 * --------------------- vs ---------------------
1548 * src->compute_capacity dst->compute_capacity
1550 if (src->load * dst->compute_capacity * env->imbalance_pct >
1552 dst->load * src->compute_capacity * 100)
1558 static int task_numa_migrate(struct task_struct *p)
1560 struct task_numa_env env = {
1563 .src_cpu = task_cpu(p),
1564 .src_nid = task_node(p),
1566 .imbalance_pct = 112,
1572 struct sched_domain *sd;
1573 unsigned long taskweight, groupweight;
1575 long taskimp, groupimp;
1578 * Pick the lowest SD_NUMA domain, as that would have the smallest
1579 * imbalance and would be the first to start moving tasks about.
1581 * And we want to avoid any moving of tasks about, as that would create
1582 * random movement of tasks -- counter the numa conditions we're trying
1586 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1588 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1592 * Cpusets can break the scheduler domain tree into smaller
1593 * balance domains, some of which do not cross NUMA boundaries.
1594 * Tasks that are "trapped" in such domains cannot be migrated
1595 * elsewhere, so there is no point in (re)trying.
1597 if (unlikely(!sd)) {
1598 p->numa_preferred_nid = task_node(p);
1602 env.dst_nid = p->numa_preferred_nid;
1603 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1604 taskweight = task_weight(p, env.src_nid, dist);
1605 groupweight = group_weight(p, env.src_nid, dist);
1606 update_numa_stats(&env.src_stats, env.src_nid);
1607 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1608 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1609 update_numa_stats(&env.dst_stats, env.dst_nid);
1611 /* Try to find a spot on the preferred nid. */
1612 if (numa_has_capacity(&env))
1613 task_numa_find_cpu(&env, taskimp, groupimp);
1616 * Look at other nodes in these cases:
1617 * - there is no space available on the preferred_nid
1618 * - the task is part of a numa_group that is interleaved across
1619 * multiple NUMA nodes; in order to better consolidate the group,
1620 * we need to check other locations.
1622 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1623 for_each_online_node(nid) {
1624 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1627 dist = node_distance(env.src_nid, env.dst_nid);
1628 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1630 taskweight = task_weight(p, env.src_nid, dist);
1631 groupweight = group_weight(p, env.src_nid, dist);
1634 /* Only consider nodes where both task and groups benefit */
1635 taskimp = task_weight(p, nid, dist) - taskweight;
1636 groupimp = group_weight(p, nid, dist) - groupweight;
1637 if (taskimp < 0 && groupimp < 0)
1642 update_numa_stats(&env.dst_stats, env.dst_nid);
1643 if (numa_has_capacity(&env))
1644 task_numa_find_cpu(&env, taskimp, groupimp);
1649 * If the task is part of a workload that spans multiple NUMA nodes,
1650 * and is migrating into one of the workload's active nodes, remember
1651 * this node as the task's preferred numa node, so the workload can
1653 * A task that migrated to a second choice node will be better off
1654 * trying for a better one later. Do not set the preferred node here.
1656 if (p->numa_group) {
1657 struct numa_group *ng = p->numa_group;
1659 if (env.best_cpu == -1)
1664 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1665 sched_setnuma(p, env.dst_nid);
1668 /* No better CPU than the current one was found. */
1669 if (env.best_cpu == -1)
1673 * Reset the scan period if the task is being rescheduled on an
1674 * alternative node to recheck if the tasks is now properly placed.
1676 p->numa_scan_period = task_scan_min(p);
1678 if (env.best_task == NULL) {
1679 ret = migrate_task_to(p, env.best_cpu);
1681 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1685 ret = migrate_swap(p, env.best_task);
1687 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1688 put_task_struct(env.best_task);
1692 /* Attempt to migrate a task to a CPU on the preferred node. */
1693 static void numa_migrate_preferred(struct task_struct *p)
1695 unsigned long interval = HZ;
1697 /* This task has no NUMA fault statistics yet */
1698 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1701 /* Periodically retry migrating the task to the preferred node */
1702 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1703 p->numa_migrate_retry = jiffies + interval;
1705 /* Success if task is already running on preferred CPU */
1706 if (task_node(p) == p->numa_preferred_nid)
1709 /* Otherwise, try migrate to a CPU on the preferred node */
1710 task_numa_migrate(p);
1714 * Find out how many nodes on the workload is actively running on. Do this by
1715 * tracking the nodes from which NUMA hinting faults are triggered. This can
1716 * be different from the set of nodes where the workload's memory is currently
1719 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1721 unsigned long faults, max_faults = 0;
1722 int nid, active_nodes = 0;
1724 for_each_online_node(nid) {
1725 faults = group_faults_cpu(numa_group, nid);
1726 if (faults > max_faults)
1727 max_faults = faults;
1730 for_each_online_node(nid) {
1731 faults = group_faults_cpu(numa_group, nid);
1732 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1736 numa_group->max_faults_cpu = max_faults;
1737 numa_group->active_nodes = active_nodes;
1741 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1742 * increments. The more local the fault statistics are, the higher the scan
1743 * period will be for the next scan window. If local/(local+remote) ratio is
1744 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1745 * the scan period will decrease. Aim for 70% local accesses.
1747 #define NUMA_PERIOD_SLOTS 10
1748 #define NUMA_PERIOD_THRESHOLD 7
1751 * Increase the scan period (slow down scanning) if the majority of
1752 * our memory is already on our local node, or if the majority of
1753 * the page accesses are shared with other processes.
1754 * Otherwise, decrease the scan period.
1756 static void update_task_scan_period(struct task_struct *p,
1757 unsigned long shared, unsigned long private)
1759 unsigned int period_slot;
1763 unsigned long remote = p->numa_faults_locality[0];
1764 unsigned long local = p->numa_faults_locality[1];
1767 * If there were no record hinting faults then either the task is
1768 * completely idle or all activity is areas that are not of interest
1769 * to automatic numa balancing. Related to that, if there were failed
1770 * migration then it implies we are migrating too quickly or the local
1771 * node is overloaded. In either case, scan slower
1773 if (local + shared == 0 || p->numa_faults_locality[2]) {
1774 p->numa_scan_period = min(p->numa_scan_period_max,
1775 p->numa_scan_period << 1);
1777 p->mm->numa_next_scan = jiffies +
1778 msecs_to_jiffies(p->numa_scan_period);
1784 * Prepare to scale scan period relative to the current period.
1785 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1786 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1787 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1789 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1790 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1791 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1792 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1795 diff = slot * period_slot;
1797 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1800 * Scale scan rate increases based on sharing. There is an
1801 * inverse relationship between the degree of sharing and
1802 * the adjustment made to the scanning period. Broadly
1803 * speaking the intent is that there is little point
1804 * scanning faster if shared accesses dominate as it may
1805 * simply bounce migrations uselessly
1807 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1808 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1811 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1812 task_scan_min(p), task_scan_max(p));
1813 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1817 * Get the fraction of time the task has been running since the last
1818 * NUMA placement cycle. The scheduler keeps similar statistics, but
1819 * decays those on a 32ms period, which is orders of magnitude off
1820 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1821 * stats only if the task is so new there are no NUMA statistics yet.
1823 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1825 u64 runtime, delta, now;
1826 /* Use the start of this time slice to avoid calculations. */
1827 now = p->se.exec_start;
1828 runtime = p->se.sum_exec_runtime;
1830 if (p->last_task_numa_placement) {
1831 delta = runtime - p->last_sum_exec_runtime;
1832 *period = now - p->last_task_numa_placement;
1834 delta = p->se.avg.load_sum / p->se.load.weight;
1835 *period = LOAD_AVG_MAX;
1838 p->last_sum_exec_runtime = runtime;
1839 p->last_task_numa_placement = now;
1845 * Determine the preferred nid for a task in a numa_group. This needs to
1846 * be done in a way that produces consistent results with group_weight,
1847 * otherwise workloads might not converge.
1849 static int preferred_group_nid(struct task_struct *p, int nid)
1854 /* Direct connections between all NUMA nodes. */
1855 if (sched_numa_topology_type == NUMA_DIRECT)
1859 * On a system with glueless mesh NUMA topology, group_weight
1860 * scores nodes according to the number of NUMA hinting faults on
1861 * both the node itself, and on nearby nodes.
1863 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1864 unsigned long score, max_score = 0;
1865 int node, max_node = nid;
1867 dist = sched_max_numa_distance;
1869 for_each_online_node(node) {
1870 score = group_weight(p, node, dist);
1871 if (score > max_score) {
1880 * Finding the preferred nid in a system with NUMA backplane
1881 * interconnect topology is more involved. The goal is to locate
1882 * tasks from numa_groups near each other in the system, and
1883 * untangle workloads from different sides of the system. This requires
1884 * searching down the hierarchy of node groups, recursively searching
1885 * inside the highest scoring group of nodes. The nodemask tricks
1886 * keep the complexity of the search down.
1888 nodes = node_online_map;
1889 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1890 unsigned long max_faults = 0;
1891 nodemask_t max_group = NODE_MASK_NONE;
1894 /* Are there nodes at this distance from each other? */
1895 if (!find_numa_distance(dist))
1898 for_each_node_mask(a, nodes) {
1899 unsigned long faults = 0;
1900 nodemask_t this_group;
1901 nodes_clear(this_group);
1903 /* Sum group's NUMA faults; includes a==b case. */
1904 for_each_node_mask(b, nodes) {
1905 if (node_distance(a, b) < dist) {
1906 faults += group_faults(p, b);
1907 node_set(b, this_group);
1908 node_clear(b, nodes);
1912 /* Remember the top group. */
1913 if (faults > max_faults) {
1914 max_faults = faults;
1915 max_group = this_group;
1917 * subtle: at the smallest distance there is
1918 * just one node left in each "group", the
1919 * winner is the preferred nid.
1924 /* Next round, evaluate the nodes within max_group. */
1932 static void task_numa_placement(struct task_struct *p)
1934 int seq, nid, max_nid = -1, max_group_nid = -1;
1935 unsigned long max_faults = 0, max_group_faults = 0;
1936 unsigned long fault_types[2] = { 0, 0 };
1937 unsigned long total_faults;
1938 u64 runtime, period;
1939 spinlock_t *group_lock = NULL;
1942 * The p->mm->numa_scan_seq field gets updated without
1943 * exclusive access. Use READ_ONCE() here to ensure
1944 * that the field is read in a single access:
1946 seq = READ_ONCE(p->mm->numa_scan_seq);
1947 if (p->numa_scan_seq == seq)
1949 p->numa_scan_seq = seq;
1950 p->numa_scan_period_max = task_scan_max(p);
1952 total_faults = p->numa_faults_locality[0] +
1953 p->numa_faults_locality[1];
1954 runtime = numa_get_avg_runtime(p, &period);
1956 /* If the task is part of a group prevent parallel updates to group stats */
1957 if (p->numa_group) {
1958 group_lock = &p->numa_group->lock;
1959 spin_lock_irq(group_lock);
1962 /* Find the node with the highest number of faults */
1963 for_each_online_node(nid) {
1964 /* Keep track of the offsets in numa_faults array */
1965 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1966 unsigned long faults = 0, group_faults = 0;
1969 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1970 long diff, f_diff, f_weight;
1972 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1973 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1974 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1975 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1977 /* Decay existing window, copy faults since last scan */
1978 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1979 fault_types[priv] += p->numa_faults[membuf_idx];
1980 p->numa_faults[membuf_idx] = 0;
1983 * Normalize the faults_from, so all tasks in a group
1984 * count according to CPU use, instead of by the raw
1985 * number of faults. Tasks with little runtime have
1986 * little over-all impact on throughput, and thus their
1987 * faults are less important.
1989 f_weight = div64_u64(runtime << 16, period + 1);
1990 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1992 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1993 p->numa_faults[cpubuf_idx] = 0;
1995 p->numa_faults[mem_idx] += diff;
1996 p->numa_faults[cpu_idx] += f_diff;
1997 faults += p->numa_faults[mem_idx];
1998 p->total_numa_faults += diff;
1999 if (p->numa_group) {
2001 * safe because we can only change our own group
2003 * mem_idx represents the offset for a given
2004 * nid and priv in a specific region because it
2005 * is at the beginning of the numa_faults array.
2007 p->numa_group->faults[mem_idx] += diff;
2008 p->numa_group->faults_cpu[mem_idx] += f_diff;
2009 p->numa_group->total_faults += diff;
2010 group_faults += p->numa_group->faults[mem_idx];
2014 if (faults > max_faults) {
2015 max_faults = faults;
2019 if (group_faults > max_group_faults) {
2020 max_group_faults = group_faults;
2021 max_group_nid = nid;
2025 update_task_scan_period(p, fault_types[0], fault_types[1]);
2027 if (p->numa_group) {
2028 numa_group_count_active_nodes(p->numa_group);
2029 spin_unlock_irq(group_lock);
2030 max_nid = preferred_group_nid(p, max_group_nid);
2034 /* Set the new preferred node */
2035 if (max_nid != p->numa_preferred_nid)
2036 sched_setnuma(p, max_nid);
2038 if (task_node(p) != p->numa_preferred_nid)
2039 numa_migrate_preferred(p);
2043 static inline int get_numa_group(struct numa_group *grp)
2045 return atomic_inc_not_zero(&grp->refcount);
2048 static inline void put_numa_group(struct numa_group *grp)
2050 if (atomic_dec_and_test(&grp->refcount))
2051 kfree_rcu(grp, rcu);
2054 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2057 struct numa_group *grp, *my_grp;
2058 struct task_struct *tsk;
2060 int cpu = cpupid_to_cpu(cpupid);
2063 if (unlikely(!p->numa_group)) {
2064 unsigned int size = sizeof(struct numa_group) +
2065 4*nr_node_ids*sizeof(unsigned long);
2067 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2071 atomic_set(&grp->refcount, 1);
2072 grp->active_nodes = 1;
2073 grp->max_faults_cpu = 0;
2074 spin_lock_init(&grp->lock);
2076 /* Second half of the array tracks nids where faults happen */
2077 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2080 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2081 grp->faults[i] = p->numa_faults[i];
2083 grp->total_faults = p->total_numa_faults;
2086 rcu_assign_pointer(p->numa_group, grp);
2090 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2092 if (!cpupid_match_pid(tsk, cpupid))
2095 grp = rcu_dereference(tsk->numa_group);
2099 my_grp = p->numa_group;
2104 * Only join the other group if its bigger; if we're the bigger group,
2105 * the other task will join us.
2107 if (my_grp->nr_tasks > grp->nr_tasks)
2111 * Tie-break on the grp address.
2113 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2116 /* Always join threads in the same process. */
2117 if (tsk->mm == current->mm)
2120 /* Simple filter to avoid false positives due to PID collisions */
2121 if (flags & TNF_SHARED)
2124 /* Update priv based on whether false sharing was detected */
2127 if (join && !get_numa_group(grp))
2135 BUG_ON(irqs_disabled());
2136 double_lock_irq(&my_grp->lock, &grp->lock);
2138 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2139 my_grp->faults[i] -= p->numa_faults[i];
2140 grp->faults[i] += p->numa_faults[i];
2142 my_grp->total_faults -= p->total_numa_faults;
2143 grp->total_faults += p->total_numa_faults;
2148 spin_unlock(&my_grp->lock);
2149 spin_unlock_irq(&grp->lock);
2151 rcu_assign_pointer(p->numa_group, grp);
2153 put_numa_group(my_grp);
2161 void task_numa_free(struct task_struct *p)
2163 struct numa_group *grp = p->numa_group;
2164 void *numa_faults = p->numa_faults;
2165 unsigned long flags;
2169 spin_lock_irqsave(&grp->lock, flags);
2170 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2171 grp->faults[i] -= p->numa_faults[i];
2172 grp->total_faults -= p->total_numa_faults;
2175 spin_unlock_irqrestore(&grp->lock, flags);
2176 RCU_INIT_POINTER(p->numa_group, NULL);
2177 put_numa_group(grp);
2180 p->numa_faults = NULL;
2185 * Got a PROT_NONE fault for a page on @node.
2187 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2189 struct task_struct *p = current;
2190 bool migrated = flags & TNF_MIGRATED;
2191 int cpu_node = task_node(current);
2192 int local = !!(flags & TNF_FAULT_LOCAL);
2193 struct numa_group *ng;
2196 if (!static_branch_likely(&sched_numa_balancing))
2199 /* for example, ksmd faulting in a user's mm */
2203 /* Allocate buffer to track faults on a per-node basis */
2204 if (unlikely(!p->numa_faults)) {
2205 int size = sizeof(*p->numa_faults) *
2206 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2208 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2209 if (!p->numa_faults)
2212 p->total_numa_faults = 0;
2213 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2217 * First accesses are treated as private, otherwise consider accesses
2218 * to be private if the accessing pid has not changed
2220 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2223 priv = cpupid_match_pid(p, last_cpupid);
2224 if (!priv && !(flags & TNF_NO_GROUP))
2225 task_numa_group(p, last_cpupid, flags, &priv);
2229 * If a workload spans multiple NUMA nodes, a shared fault that
2230 * occurs wholly within the set of nodes that the workload is
2231 * actively using should be counted as local. This allows the
2232 * scan rate to slow down when a workload has settled down.
2235 if (!priv && !local && ng && ng->active_nodes > 1 &&
2236 numa_is_active_node(cpu_node, ng) &&
2237 numa_is_active_node(mem_node, ng))
2240 task_numa_placement(p);
2243 * Retry task to preferred node migration periodically, in case it
2244 * case it previously failed, or the scheduler moved us.
2246 if (time_after(jiffies, p->numa_migrate_retry))
2247 numa_migrate_preferred(p);
2250 p->numa_pages_migrated += pages;
2251 if (flags & TNF_MIGRATE_FAIL)
2252 p->numa_faults_locality[2] += pages;
2254 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2255 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2256 p->numa_faults_locality[local] += pages;
2259 static void reset_ptenuma_scan(struct task_struct *p)
2262 * We only did a read acquisition of the mmap sem, so
2263 * p->mm->numa_scan_seq is written to without exclusive access
2264 * and the update is not guaranteed to be atomic. That's not
2265 * much of an issue though, since this is just used for
2266 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2267 * expensive, to avoid any form of compiler optimizations:
2269 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2270 p->mm->numa_scan_offset = 0;
2274 * The expensive part of numa migration is done from task_work context.
2275 * Triggered from task_tick_numa().
2277 void task_numa_work(struct callback_head *work)
2279 unsigned long migrate, next_scan, now = jiffies;
2280 struct task_struct *p = current;
2281 struct mm_struct *mm = p->mm;
2282 u64 runtime = p->se.sum_exec_runtime;
2283 struct vm_area_struct *vma;
2284 unsigned long start, end;
2285 unsigned long nr_pte_updates = 0;
2286 long pages, virtpages;
2288 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2290 work->next = work; /* protect against double add */
2292 * Who cares about NUMA placement when they're dying.
2294 * NOTE: make sure not to dereference p->mm before this check,
2295 * exit_task_work() happens _after_ exit_mm() so we could be called
2296 * without p->mm even though we still had it when we enqueued this
2299 if (p->flags & PF_EXITING)
2302 if (!mm->numa_next_scan) {
2303 mm->numa_next_scan = now +
2304 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2308 * Enforce maximal scan/migration frequency..
2310 migrate = mm->numa_next_scan;
2311 if (time_before(now, migrate))
2314 if (p->numa_scan_period == 0) {
2315 p->numa_scan_period_max = task_scan_max(p);
2316 p->numa_scan_period = task_scan_min(p);
2319 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2320 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2324 * Delay this task enough that another task of this mm will likely win
2325 * the next time around.
2327 p->node_stamp += 2 * TICK_NSEC;
2329 start = mm->numa_scan_offset;
2330 pages = sysctl_numa_balancing_scan_size;
2331 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2332 virtpages = pages * 8; /* Scan up to this much virtual space */
2337 down_read(&mm->mmap_sem);
2338 vma = find_vma(mm, start);
2340 reset_ptenuma_scan(p);
2344 for (; vma; vma = vma->vm_next) {
2345 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2346 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2351 * Shared library pages mapped by multiple processes are not
2352 * migrated as it is expected they are cache replicated. Avoid
2353 * hinting faults in read-only file-backed mappings or the vdso
2354 * as migrating the pages will be of marginal benefit.
2357 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2361 * Skip inaccessible VMAs to avoid any confusion between
2362 * PROT_NONE and NUMA hinting ptes
2364 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2368 start = max(start, vma->vm_start);
2369 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2370 end = min(end, vma->vm_end);
2371 nr_pte_updates = change_prot_numa(vma, start, end);
2374 * Try to scan sysctl_numa_balancing_size worth of
2375 * hpages that have at least one present PTE that
2376 * is not already pte-numa. If the VMA contains
2377 * areas that are unused or already full of prot_numa
2378 * PTEs, scan up to virtpages, to skip through those
2382 pages -= (end - start) >> PAGE_SHIFT;
2383 virtpages -= (end - start) >> PAGE_SHIFT;
2386 if (pages <= 0 || virtpages <= 0)
2390 } while (end != vma->vm_end);
2395 * It is possible to reach the end of the VMA list but the last few
2396 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2397 * would find the !migratable VMA on the next scan but not reset the
2398 * scanner to the start so check it now.
2401 mm->numa_scan_offset = start;
2403 reset_ptenuma_scan(p);
2404 up_read(&mm->mmap_sem);
2407 * Make sure tasks use at least 32x as much time to run other code
2408 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2409 * Usually update_task_scan_period slows down scanning enough; on an
2410 * overloaded system we need to limit overhead on a per task basis.
2412 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2413 u64 diff = p->se.sum_exec_runtime - runtime;
2414 p->node_stamp += 32 * diff;
2419 * Drive the periodic memory faults..
2421 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2423 struct callback_head *work = &curr->numa_work;
2427 * We don't care about NUMA placement if we don't have memory.
2429 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2433 * Using runtime rather than walltime has the dual advantage that
2434 * we (mostly) drive the selection from busy threads and that the
2435 * task needs to have done some actual work before we bother with
2438 now = curr->se.sum_exec_runtime;
2439 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2441 if (now > curr->node_stamp + period) {
2442 if (!curr->node_stamp)
2443 curr->numa_scan_period = task_scan_min(curr);
2444 curr->node_stamp += period;
2446 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2447 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2448 task_work_add(curr, work, true);
2453 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2457 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2461 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2464 #endif /* CONFIG_NUMA_BALANCING */
2467 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2469 update_load_add(&cfs_rq->load, se->load.weight);
2470 if (!parent_entity(se))
2471 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2473 if (entity_is_task(se)) {
2474 struct rq *rq = rq_of(cfs_rq);
2476 account_numa_enqueue(rq, task_of(se));
2477 list_add(&se->group_node, &rq->cfs_tasks);
2480 cfs_rq->nr_running++;
2484 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2486 update_load_sub(&cfs_rq->load, se->load.weight);
2487 if (!parent_entity(se))
2488 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2490 if (entity_is_task(se)) {
2491 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2492 list_del_init(&se->group_node);
2495 cfs_rq->nr_running--;
2498 #ifdef CONFIG_FAIR_GROUP_SCHED
2500 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2502 long tg_weight, load, shares;
2505 * This really should be: cfs_rq->avg.load_avg, but instead we use
2506 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2507 * the shares for small weight interactive tasks.
2509 load = scale_load_down(cfs_rq->load.weight);
2511 tg_weight = atomic_long_read(&tg->load_avg);
2513 /* Ensure tg_weight >= load */
2514 tg_weight -= cfs_rq->tg_load_avg_contrib;
2517 shares = (tg->shares * load);
2519 shares /= tg_weight;
2521 if (shares < MIN_SHARES)
2522 shares = MIN_SHARES;
2523 if (shares > tg->shares)
2524 shares = tg->shares;
2528 # else /* CONFIG_SMP */
2529 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2533 # endif /* CONFIG_SMP */
2535 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2536 unsigned long weight)
2539 /* commit outstanding execution time */
2540 if (cfs_rq->curr == se)
2541 update_curr(cfs_rq);
2542 account_entity_dequeue(cfs_rq, se);
2545 update_load_set(&se->load, weight);
2548 account_entity_enqueue(cfs_rq, se);
2551 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2553 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2555 struct task_group *tg;
2556 struct sched_entity *se;
2560 se = tg->se[cpu_of(rq_of(cfs_rq))];
2561 if (!se || throttled_hierarchy(cfs_rq))
2564 if (likely(se->load.weight == tg->shares))
2567 shares = calc_cfs_shares(cfs_rq, tg);
2569 reweight_entity(cfs_rq_of(se), se, shares);
2571 #else /* CONFIG_FAIR_GROUP_SCHED */
2572 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2575 #endif /* CONFIG_FAIR_GROUP_SCHED */
2578 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2579 static const u32 runnable_avg_yN_inv[] = {
2580 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2581 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2582 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2583 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2584 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2585 0x85aac367, 0x82cd8698,
2589 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2590 * over-estimates when re-combining.
2592 static const u32 runnable_avg_yN_sum[] = {
2593 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2594 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2595 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2599 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2600 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2603 static const u32 __accumulated_sum_N32[] = {
2604 0, 23371, 35056, 40899, 43820, 45281,
2605 46011, 46376, 46559, 46650, 46696, 46719,
2610 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2612 static __always_inline u64 decay_load(u64 val, u64 n)
2614 unsigned int local_n;
2618 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2621 /* after bounds checking we can collapse to 32-bit */
2625 * As y^PERIOD = 1/2, we can combine
2626 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2627 * With a look-up table which covers y^n (n<PERIOD)
2629 * To achieve constant time decay_load.
2631 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2632 val >>= local_n / LOAD_AVG_PERIOD;
2633 local_n %= LOAD_AVG_PERIOD;
2636 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2641 * For updates fully spanning n periods, the contribution to runnable
2642 * average will be: \Sum 1024*y^n
2644 * We can compute this reasonably efficiently by combining:
2645 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2647 static u32 __compute_runnable_contrib(u64 n)
2651 if (likely(n <= LOAD_AVG_PERIOD))
2652 return runnable_avg_yN_sum[n];
2653 else if (unlikely(n >= LOAD_AVG_MAX_N))
2654 return LOAD_AVG_MAX;
2656 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2657 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2658 n %= LOAD_AVG_PERIOD;
2659 contrib = decay_load(contrib, n);
2660 return contrib + runnable_avg_yN_sum[n];
2663 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2666 * We can represent the historical contribution to runnable average as the
2667 * coefficients of a geometric series. To do this we sub-divide our runnable
2668 * history into segments of approximately 1ms (1024us); label the segment that
2669 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2671 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2673 * (now) (~1ms ago) (~2ms ago)
2675 * Let u_i denote the fraction of p_i that the entity was runnable.
2677 * We then designate the fractions u_i as our co-efficients, yielding the
2678 * following representation of historical load:
2679 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2681 * We choose y based on the with of a reasonably scheduling period, fixing:
2684 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2685 * approximately half as much as the contribution to load within the last ms
2688 * When a period "rolls over" and we have new u_0`, multiplying the previous
2689 * sum again by y is sufficient to update:
2690 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2691 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2693 static __always_inline int
2694 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2695 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2697 u64 delta, scaled_delta, periods;
2699 unsigned int delta_w, scaled_delta_w, decayed = 0;
2700 unsigned long scale_freq, scale_cpu;
2702 delta = now - sa->last_update_time;
2704 * This should only happen when time goes backwards, which it
2705 * unfortunately does during sched clock init when we swap over to TSC.
2707 if ((s64)delta < 0) {
2708 sa->last_update_time = now;
2713 * Use 1024ns as the unit of measurement since it's a reasonable
2714 * approximation of 1us and fast to compute.
2719 sa->last_update_time = now;
2721 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2722 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2724 /* delta_w is the amount already accumulated against our next period */
2725 delta_w = sa->period_contrib;
2726 if (delta + delta_w >= 1024) {
2729 /* how much left for next period will start over, we don't know yet */
2730 sa->period_contrib = 0;
2733 * Now that we know we're crossing a period boundary, figure
2734 * out how much from delta we need to complete the current
2735 * period and accrue it.
2737 delta_w = 1024 - delta_w;
2738 scaled_delta_w = cap_scale(delta_w, scale_freq);
2740 sa->load_sum += weight * scaled_delta_w;
2742 cfs_rq->runnable_load_sum +=
2743 weight * scaled_delta_w;
2747 sa->util_sum += scaled_delta_w * scale_cpu;
2751 /* Figure out how many additional periods this update spans */
2752 periods = delta / 1024;
2755 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2757 cfs_rq->runnable_load_sum =
2758 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2760 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2762 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2763 contrib = __compute_runnable_contrib(periods);
2764 contrib = cap_scale(contrib, scale_freq);
2766 sa->load_sum += weight * contrib;
2768 cfs_rq->runnable_load_sum += weight * contrib;
2771 sa->util_sum += contrib * scale_cpu;
2774 /* Remainder of delta accrued against u_0` */
2775 scaled_delta = cap_scale(delta, scale_freq);
2777 sa->load_sum += weight * scaled_delta;
2779 cfs_rq->runnable_load_sum += weight * scaled_delta;
2782 sa->util_sum += scaled_delta * scale_cpu;
2784 sa->period_contrib += delta;
2787 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2789 cfs_rq->runnable_load_avg =
2790 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2792 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2798 #ifdef CONFIG_FAIR_GROUP_SCHED
2800 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2801 * and effective_load (which is not done because it is too costly).
2803 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2805 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2808 * No need to update load_avg for root_task_group as it is not used.
2810 if (cfs_rq->tg == &root_task_group)
2813 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2814 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2815 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2820 * Called within set_task_rq() right before setting a task's cpu. The
2821 * caller only guarantees p->pi_lock is held; no other assumptions,
2822 * including the state of rq->lock, should be made.
2824 void set_task_rq_fair(struct sched_entity *se,
2825 struct cfs_rq *prev, struct cfs_rq *next)
2827 if (!sched_feat(ATTACH_AGE_LOAD))
2831 * We are supposed to update the task to "current" time, then its up to
2832 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2833 * getting what current time is, so simply throw away the out-of-date
2834 * time. This will result in the wakee task is less decayed, but giving
2835 * the wakee more load sounds not bad.
2837 if (se->avg.last_update_time && prev) {
2838 u64 p_last_update_time;
2839 u64 n_last_update_time;
2841 #ifndef CONFIG_64BIT
2842 u64 p_last_update_time_copy;
2843 u64 n_last_update_time_copy;
2846 p_last_update_time_copy = prev->load_last_update_time_copy;
2847 n_last_update_time_copy = next->load_last_update_time_copy;
2851 p_last_update_time = prev->avg.last_update_time;
2852 n_last_update_time = next->avg.last_update_time;
2854 } while (p_last_update_time != p_last_update_time_copy ||
2855 n_last_update_time != n_last_update_time_copy);
2857 p_last_update_time = prev->avg.last_update_time;
2858 n_last_update_time = next->avg.last_update_time;
2860 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2861 &se->avg, 0, 0, NULL);
2862 se->avg.last_update_time = n_last_update_time;
2865 #else /* CONFIG_FAIR_GROUP_SCHED */
2866 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2867 #endif /* CONFIG_FAIR_GROUP_SCHED */
2869 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2871 struct rq *rq = rq_of(cfs_rq);
2872 int cpu = cpu_of(rq);
2874 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2875 unsigned long max = rq->cpu_capacity_orig;
2878 * There are a few boundary cases this might miss but it should
2879 * get called often enough that that should (hopefully) not be
2880 * a real problem -- added to that it only calls on the local
2881 * CPU, so if we enqueue remotely we'll miss an update, but
2882 * the next tick/schedule should update.
2884 * It will not get called when we go idle, because the idle
2885 * thread is a different class (!fair), nor will the utilization
2886 * number include things like RT tasks.
2888 * As is, the util number is not freq-invariant (we'd have to
2889 * implement arch_scale_freq_capacity() for that).
2893 cpufreq_update_util(rq_clock(rq),
2894 min(cfs_rq->avg.util_avg, max), max);
2899 * Unsigned subtract and clamp on underflow.
2901 * Explicitly do a load-store to ensure the intermediate value never hits
2902 * memory. This allows lockless observations without ever seeing the negative
2905 #define sub_positive(_ptr, _val) do { \
2906 typeof(_ptr) ptr = (_ptr); \
2907 typeof(*ptr) val = (_val); \
2908 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2912 WRITE_ONCE(*ptr, res); \
2915 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2917 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2919 struct sched_avg *sa = &cfs_rq->avg;
2920 int decayed, removed_load = 0, removed_util = 0;
2922 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2923 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2924 sub_positive(&sa->load_avg, r);
2925 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2929 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2930 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2931 sub_positive(&sa->util_avg, r);
2932 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2936 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2937 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2939 #ifndef CONFIG_64BIT
2941 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2944 if (update_freq && (decayed || removed_util))
2945 cfs_rq_util_change(cfs_rq);
2947 return decayed || removed_load;
2950 /* Update task and its cfs_rq load average */
2951 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2953 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2954 u64 now = cfs_rq_clock_task(cfs_rq);
2955 struct rq *rq = rq_of(cfs_rq);
2956 int cpu = cpu_of(rq);
2959 * Track task load average for carrying it to new CPU after migrated, and
2960 * track group sched_entity load average for task_h_load calc in migration
2962 __update_load_avg(now, cpu, &se->avg,
2963 se->on_rq * scale_load_down(se->load.weight),
2964 cfs_rq->curr == se, NULL);
2966 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2967 update_tg_load_avg(cfs_rq, 0);
2970 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2972 if (!sched_feat(ATTACH_AGE_LOAD))
2976 * If we got migrated (either between CPUs or between cgroups) we'll
2977 * have aged the average right before clearing @last_update_time.
2979 * Or we're fresh through post_init_entity_util_avg().
2981 if (se->avg.last_update_time) {
2982 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2983 &se->avg, 0, 0, NULL);
2986 * XXX: we could have just aged the entire load away if we've been
2987 * absent from the fair class for too long.
2992 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2993 cfs_rq->avg.load_avg += se->avg.load_avg;
2994 cfs_rq->avg.load_sum += se->avg.load_sum;
2995 cfs_rq->avg.util_avg += se->avg.util_avg;
2996 cfs_rq->avg.util_sum += se->avg.util_sum;
2998 cfs_rq_util_change(cfs_rq);
3001 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3004 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3005 cfs_rq->curr == se, NULL);
3007 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3008 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3009 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3010 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3012 cfs_rq_util_change(cfs_rq);
3015 /* Add the load generated by se into cfs_rq's load average */
3017 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3019 struct sched_avg *sa = &se->avg;
3020 u64 now = cfs_rq_clock_task(cfs_rq);
3021 int migrated, decayed;
3023 migrated = !sa->last_update_time;
3025 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3026 se->on_rq * scale_load_down(se->load.weight),
3027 cfs_rq->curr == se, NULL);
3030 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3032 cfs_rq->runnable_load_avg += sa->load_avg;
3033 cfs_rq->runnable_load_sum += sa->load_sum;
3036 attach_entity_load_avg(cfs_rq, se);
3038 if (decayed || migrated)
3039 update_tg_load_avg(cfs_rq, 0);
3042 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3044 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3046 update_load_avg(se, 1);
3048 cfs_rq->runnable_load_avg =
3049 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3050 cfs_rq->runnable_load_sum =
3051 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3054 #ifndef CONFIG_64BIT
3055 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3057 u64 last_update_time_copy;
3058 u64 last_update_time;
3061 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3063 last_update_time = cfs_rq->avg.last_update_time;
3064 } while (last_update_time != last_update_time_copy);
3066 return last_update_time;
3069 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3071 return cfs_rq->avg.last_update_time;
3076 * Task first catches up with cfs_rq, and then subtract
3077 * itself from the cfs_rq (task must be off the queue now).
3079 void remove_entity_load_avg(struct sched_entity *se)
3081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3082 u64 last_update_time;
3085 * tasks cannot exit without having gone through wake_up_new_task() ->
3086 * post_init_entity_util_avg() which will have added things to the
3087 * cfs_rq, so we can remove unconditionally.
3089 * Similarly for groups, they will have passed through
3090 * post_init_entity_util_avg() before unregister_sched_fair_group()
3094 last_update_time = cfs_rq_last_update_time(cfs_rq);
3096 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3097 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3098 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3101 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3103 return cfs_rq->runnable_load_avg;
3106 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3108 return cfs_rq->avg.load_avg;
3111 static int idle_balance(struct rq *this_rq);
3113 #else /* CONFIG_SMP */
3116 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3121 static inline void update_load_avg(struct sched_entity *se, int not_used)
3123 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3124 struct rq *rq = rq_of(cfs_rq);
3126 cpufreq_trigger_update(rq_clock(rq));
3130 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3132 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3133 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3136 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3138 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3140 static inline int idle_balance(struct rq *rq)
3145 #endif /* CONFIG_SMP */
3147 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3149 #ifdef CONFIG_SCHEDSTATS
3150 struct task_struct *tsk = NULL;
3152 if (entity_is_task(se))
3155 if (se->statistics.sleep_start) {
3156 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3161 if (unlikely(delta > se->statistics.sleep_max))
3162 se->statistics.sleep_max = delta;
3164 se->statistics.sleep_start = 0;
3165 se->statistics.sum_sleep_runtime += delta;
3168 account_scheduler_latency(tsk, delta >> 10, 1);
3169 trace_sched_stat_sleep(tsk, delta);
3172 if (se->statistics.block_start) {
3173 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3178 if (unlikely(delta > se->statistics.block_max))
3179 se->statistics.block_max = delta;
3181 se->statistics.block_start = 0;
3182 se->statistics.sum_sleep_runtime += delta;
3185 if (tsk->in_iowait) {
3186 se->statistics.iowait_sum += delta;
3187 se->statistics.iowait_count++;
3188 trace_sched_stat_iowait(tsk, delta);
3191 trace_sched_stat_blocked(tsk, delta);
3194 * Blocking time is in units of nanosecs, so shift by
3195 * 20 to get a milliseconds-range estimation of the
3196 * amount of time that the task spent sleeping:
3198 if (unlikely(prof_on == SLEEP_PROFILING)) {
3199 profile_hits(SLEEP_PROFILING,
3200 (void *)get_wchan(tsk),
3203 account_scheduler_latency(tsk, delta >> 10, 0);
3209 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3211 #ifdef CONFIG_SCHED_DEBUG
3212 s64 d = se->vruntime - cfs_rq->min_vruntime;
3217 if (d > 3*sysctl_sched_latency)
3218 schedstat_inc(cfs_rq, nr_spread_over);
3223 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3225 u64 vruntime = cfs_rq->min_vruntime;
3228 * The 'current' period is already promised to the current tasks,
3229 * however the extra weight of the new task will slow them down a
3230 * little, place the new task so that it fits in the slot that
3231 * stays open at the end.
3233 if (initial && sched_feat(START_DEBIT))
3234 vruntime += sched_vslice(cfs_rq, se);
3236 /* sleeps up to a single latency don't count. */
3238 unsigned long thresh = sysctl_sched_latency;
3241 * Halve their sleep time's effect, to allow
3242 * for a gentler effect of sleepers:
3244 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3250 /* ensure we never gain time by being placed backwards. */
3251 se->vruntime = max_vruntime(se->vruntime, vruntime);
3254 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3256 static inline void check_schedstat_required(void)
3258 #ifdef CONFIG_SCHEDSTATS
3259 if (schedstat_enabled())
3262 /* Force schedstat enabled if a dependent tracepoint is active */
3263 if (trace_sched_stat_wait_enabled() ||
3264 trace_sched_stat_sleep_enabled() ||
3265 trace_sched_stat_iowait_enabled() ||
3266 trace_sched_stat_blocked_enabled() ||
3267 trace_sched_stat_runtime_enabled()) {
3268 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3269 "stat_blocked and stat_runtime require the "
3270 "kernel parameter schedstats=enabled or "
3271 "kernel.sched_schedstats=1\n");
3282 * update_min_vruntime()
3283 * vruntime -= min_vruntime
3287 * update_min_vruntime()
3288 * vruntime += min_vruntime
3290 * this way the vruntime transition between RQs is done when both
3291 * min_vruntime are up-to-date.
3295 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3296 * vruntime -= min_vruntime
3300 * update_min_vruntime()
3301 * vruntime += min_vruntime
3303 * this way we don't have the most up-to-date min_vruntime on the originating
3304 * CPU and an up-to-date min_vruntime on the destination CPU.
3308 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3310 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3311 bool curr = cfs_rq->curr == se;
3314 * If we're the current task, we must renormalise before calling
3318 se->vruntime += cfs_rq->min_vruntime;
3320 update_curr(cfs_rq);
3323 * Otherwise, renormalise after, such that we're placed at the current
3324 * moment in time, instead of some random moment in the past. Being
3325 * placed in the past could significantly boost this task to the
3326 * fairness detriment of existing tasks.
3328 if (renorm && !curr)
3329 se->vruntime += cfs_rq->min_vruntime;
3331 enqueue_entity_load_avg(cfs_rq, se);
3332 account_entity_enqueue(cfs_rq, se);
3333 update_cfs_shares(cfs_rq);
3335 if (flags & ENQUEUE_WAKEUP) {
3336 place_entity(cfs_rq, se, 0);
3337 if (schedstat_enabled())
3338 enqueue_sleeper(cfs_rq, se);
3341 check_schedstat_required();
3342 if (schedstat_enabled()) {
3343 update_stats_enqueue(cfs_rq, se);
3344 check_spread(cfs_rq, se);
3347 __enqueue_entity(cfs_rq, se);
3350 if (cfs_rq->nr_running == 1) {
3351 list_add_leaf_cfs_rq(cfs_rq);
3352 check_enqueue_throttle(cfs_rq);
3356 static void __clear_buddies_last(struct sched_entity *se)
3358 for_each_sched_entity(se) {
3359 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3360 if (cfs_rq->last != se)
3363 cfs_rq->last = NULL;
3367 static void __clear_buddies_next(struct sched_entity *se)
3369 for_each_sched_entity(se) {
3370 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3371 if (cfs_rq->next != se)
3374 cfs_rq->next = NULL;
3378 static void __clear_buddies_skip(struct sched_entity *se)
3380 for_each_sched_entity(se) {
3381 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3382 if (cfs_rq->skip != se)
3385 cfs_rq->skip = NULL;
3389 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3391 if (cfs_rq->last == se)
3392 __clear_buddies_last(se);
3394 if (cfs_rq->next == se)
3395 __clear_buddies_next(se);
3397 if (cfs_rq->skip == se)
3398 __clear_buddies_skip(se);
3401 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3404 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3407 * Update run-time statistics of the 'current'.
3409 update_curr(cfs_rq);
3410 dequeue_entity_load_avg(cfs_rq, se);
3412 if (schedstat_enabled())
3413 update_stats_dequeue(cfs_rq, se, flags);
3415 clear_buddies(cfs_rq, se);
3417 if (se != cfs_rq->curr)
3418 __dequeue_entity(cfs_rq, se);
3420 account_entity_dequeue(cfs_rq, se);
3423 * Normalize the entity after updating the min_vruntime because the
3424 * update can refer to the ->curr item and we need to reflect this
3425 * movement in our normalized position.
3427 if (!(flags & DEQUEUE_SLEEP))
3428 se->vruntime -= cfs_rq->min_vruntime;
3430 /* return excess runtime on last dequeue */
3431 return_cfs_rq_runtime(cfs_rq);
3433 update_min_vruntime(cfs_rq);
3434 update_cfs_shares(cfs_rq);
3438 * Preempt the current task with a newly woken task if needed:
3441 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3443 unsigned long ideal_runtime, delta_exec;
3444 struct sched_entity *se;
3447 ideal_runtime = sched_slice(cfs_rq, curr);
3448 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3449 if (delta_exec > ideal_runtime) {
3450 resched_curr(rq_of(cfs_rq));
3452 * The current task ran long enough, ensure it doesn't get
3453 * re-elected due to buddy favours.
3455 clear_buddies(cfs_rq, curr);
3460 * Ensure that a task that missed wakeup preemption by a
3461 * narrow margin doesn't have to wait for a full slice.
3462 * This also mitigates buddy induced latencies under load.
3464 if (delta_exec < sysctl_sched_min_granularity)
3467 se = __pick_first_entity(cfs_rq);
3468 delta = curr->vruntime - se->vruntime;
3473 if (delta > ideal_runtime)
3474 resched_curr(rq_of(cfs_rq));
3478 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3480 /* 'current' is not kept within the tree. */
3483 * Any task has to be enqueued before it get to execute on
3484 * a CPU. So account for the time it spent waiting on the
3487 if (schedstat_enabled())
3488 update_stats_wait_end(cfs_rq, se);
3489 __dequeue_entity(cfs_rq, se);
3490 update_load_avg(se, 1);
3493 update_stats_curr_start(cfs_rq, se);
3495 #ifdef CONFIG_SCHEDSTATS
3497 * Track our maximum slice length, if the CPU's load is at
3498 * least twice that of our own weight (i.e. dont track it
3499 * when there are only lesser-weight tasks around):
3501 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3502 se->statistics.slice_max = max(se->statistics.slice_max,
3503 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3506 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3510 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3513 * Pick the next process, keeping these things in mind, in this order:
3514 * 1) keep things fair between processes/task groups
3515 * 2) pick the "next" process, since someone really wants that to run
3516 * 3) pick the "last" process, for cache locality
3517 * 4) do not run the "skip" process, if something else is available
3519 static struct sched_entity *
3520 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3522 struct sched_entity *left = __pick_first_entity(cfs_rq);
3523 struct sched_entity *se;
3526 * If curr is set we have to see if its left of the leftmost entity
3527 * still in the tree, provided there was anything in the tree at all.
3529 if (!left || (curr && entity_before(curr, left)))
3532 se = left; /* ideally we run the leftmost entity */
3535 * Avoid running the skip buddy, if running something else can
3536 * be done without getting too unfair.
3538 if (cfs_rq->skip == se) {
3539 struct sched_entity *second;
3542 second = __pick_first_entity(cfs_rq);
3544 second = __pick_next_entity(se);
3545 if (!second || (curr && entity_before(curr, second)))
3549 if (second && wakeup_preempt_entity(second, left) < 1)
3554 * Prefer last buddy, try to return the CPU to a preempted task.
3556 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3560 * Someone really wants this to run. If it's not unfair, run it.
3562 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3565 clear_buddies(cfs_rq, se);
3570 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3572 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3575 * If still on the runqueue then deactivate_task()
3576 * was not called and update_curr() has to be done:
3579 update_curr(cfs_rq);
3581 /* throttle cfs_rqs exceeding runtime */
3582 check_cfs_rq_runtime(cfs_rq);
3584 if (schedstat_enabled()) {
3585 check_spread(cfs_rq, prev);
3587 update_stats_wait_start(cfs_rq, prev);
3591 /* Put 'current' back into the tree. */
3592 __enqueue_entity(cfs_rq, prev);
3593 /* in !on_rq case, update occurred at dequeue */
3594 update_load_avg(prev, 0);
3596 cfs_rq->curr = NULL;
3600 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3603 * Update run-time statistics of the 'current'.
3605 update_curr(cfs_rq);
3608 * Ensure that runnable average is periodically updated.
3610 update_load_avg(curr, 1);
3611 update_cfs_shares(cfs_rq);
3613 #ifdef CONFIG_SCHED_HRTICK
3615 * queued ticks are scheduled to match the slice, so don't bother
3616 * validating it and just reschedule.
3619 resched_curr(rq_of(cfs_rq));
3623 * don't let the period tick interfere with the hrtick preemption
3625 if (!sched_feat(DOUBLE_TICK) &&
3626 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3630 if (cfs_rq->nr_running > 1)
3631 check_preempt_tick(cfs_rq, curr);
3635 /**************************************************
3636 * CFS bandwidth control machinery
3639 #ifdef CONFIG_CFS_BANDWIDTH
3641 #ifdef HAVE_JUMP_LABEL
3642 static struct static_key __cfs_bandwidth_used;
3644 static inline bool cfs_bandwidth_used(void)
3646 return static_key_false(&__cfs_bandwidth_used);
3649 void cfs_bandwidth_usage_inc(void)
3651 static_key_slow_inc(&__cfs_bandwidth_used);
3654 void cfs_bandwidth_usage_dec(void)
3656 static_key_slow_dec(&__cfs_bandwidth_used);
3658 #else /* HAVE_JUMP_LABEL */
3659 static bool cfs_bandwidth_used(void)
3664 void cfs_bandwidth_usage_inc(void) {}
3665 void cfs_bandwidth_usage_dec(void) {}
3666 #endif /* HAVE_JUMP_LABEL */
3669 * default period for cfs group bandwidth.
3670 * default: 0.1s, units: nanoseconds
3672 static inline u64 default_cfs_period(void)
3674 return 100000000ULL;
3677 static inline u64 sched_cfs_bandwidth_slice(void)
3679 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3683 * Replenish runtime according to assigned quota and update expiration time.
3684 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3685 * additional synchronization around rq->lock.
3687 * requires cfs_b->lock
3689 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3693 if (cfs_b->quota == RUNTIME_INF)
3696 now = sched_clock_cpu(smp_processor_id());
3697 cfs_b->runtime = cfs_b->quota;
3698 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3701 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3703 return &tg->cfs_bandwidth;
3706 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3707 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3709 if (unlikely(cfs_rq->throttle_count))
3710 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3712 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3715 /* returns 0 on failure to allocate runtime */
3716 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3718 struct task_group *tg = cfs_rq->tg;
3719 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3720 u64 amount = 0, min_amount, expires;
3722 /* note: this is a positive sum as runtime_remaining <= 0 */
3723 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3725 raw_spin_lock(&cfs_b->lock);
3726 if (cfs_b->quota == RUNTIME_INF)
3727 amount = min_amount;
3729 start_cfs_bandwidth(cfs_b);
3731 if (cfs_b->runtime > 0) {
3732 amount = min(cfs_b->runtime, min_amount);
3733 cfs_b->runtime -= amount;
3737 expires = cfs_b->runtime_expires;
3738 raw_spin_unlock(&cfs_b->lock);
3740 cfs_rq->runtime_remaining += amount;
3742 * we may have advanced our local expiration to account for allowed
3743 * spread between our sched_clock and the one on which runtime was
3746 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3747 cfs_rq->runtime_expires = expires;
3749 return cfs_rq->runtime_remaining > 0;
3753 * Note: This depends on the synchronization provided by sched_clock and the
3754 * fact that rq->clock snapshots this value.
3756 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3758 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3760 /* if the deadline is ahead of our clock, nothing to do */
3761 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3764 if (cfs_rq->runtime_remaining < 0)
3768 * If the local deadline has passed we have to consider the
3769 * possibility that our sched_clock is 'fast' and the global deadline
3770 * has not truly expired.
3772 * Fortunately we can check determine whether this the case by checking
3773 * whether the global deadline has advanced. It is valid to compare
3774 * cfs_b->runtime_expires without any locks since we only care about
3775 * exact equality, so a partial write will still work.
3778 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3779 /* extend local deadline, drift is bounded above by 2 ticks */
3780 cfs_rq->runtime_expires += TICK_NSEC;
3782 /* global deadline is ahead, expiration has passed */
3783 cfs_rq->runtime_remaining = 0;
3787 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3789 /* dock delta_exec before expiring quota (as it could span periods) */
3790 cfs_rq->runtime_remaining -= delta_exec;
3791 expire_cfs_rq_runtime(cfs_rq);
3793 if (likely(cfs_rq->runtime_remaining > 0))
3797 * if we're unable to extend our runtime we resched so that the active
3798 * hierarchy can be throttled
3800 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3801 resched_curr(rq_of(cfs_rq));
3804 static __always_inline
3805 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3807 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3810 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3813 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3815 return cfs_bandwidth_used() && cfs_rq->throttled;
3818 /* check whether cfs_rq, or any parent, is throttled */
3819 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3821 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3825 * Ensure that neither of the group entities corresponding to src_cpu or
3826 * dest_cpu are members of a throttled hierarchy when performing group
3827 * load-balance operations.
3829 static inline int throttled_lb_pair(struct task_group *tg,
3830 int src_cpu, int dest_cpu)
3832 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3834 src_cfs_rq = tg->cfs_rq[src_cpu];
3835 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3837 return throttled_hierarchy(src_cfs_rq) ||
3838 throttled_hierarchy(dest_cfs_rq);
3841 /* updated child weight may affect parent so we have to do this bottom up */
3842 static int tg_unthrottle_up(struct task_group *tg, void *data)
3844 struct rq *rq = data;
3845 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3847 cfs_rq->throttle_count--;
3848 if (!cfs_rq->throttle_count) {
3849 /* adjust cfs_rq_clock_task() */
3850 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3851 cfs_rq->throttled_clock_task;
3857 static int tg_throttle_down(struct task_group *tg, void *data)
3859 struct rq *rq = data;
3860 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3862 /* group is entering throttled state, stop time */
3863 if (!cfs_rq->throttle_count)
3864 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3865 cfs_rq->throttle_count++;
3870 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3872 struct rq *rq = rq_of(cfs_rq);
3873 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3874 struct sched_entity *se;
3875 long task_delta, dequeue = 1;
3878 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3880 /* freeze hierarchy runnable averages while throttled */
3882 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3885 task_delta = cfs_rq->h_nr_running;
3886 for_each_sched_entity(se) {
3887 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3888 /* throttled entity or throttle-on-deactivate */
3893 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3894 qcfs_rq->h_nr_running -= task_delta;
3896 if (qcfs_rq->load.weight)
3901 sub_nr_running(rq, task_delta);
3903 cfs_rq->throttled = 1;
3904 cfs_rq->throttled_clock = rq_clock(rq);
3905 raw_spin_lock(&cfs_b->lock);
3906 empty = list_empty(&cfs_b->throttled_cfs_rq);
3909 * Add to the _head_ of the list, so that an already-started
3910 * distribute_cfs_runtime will not see us
3912 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3915 * If we're the first throttled task, make sure the bandwidth
3919 start_cfs_bandwidth(cfs_b);
3921 raw_spin_unlock(&cfs_b->lock);
3924 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3926 struct rq *rq = rq_of(cfs_rq);
3927 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3928 struct sched_entity *se;
3932 se = cfs_rq->tg->se[cpu_of(rq)];
3934 cfs_rq->throttled = 0;
3936 update_rq_clock(rq);
3938 raw_spin_lock(&cfs_b->lock);
3939 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3940 list_del_rcu(&cfs_rq->throttled_list);
3941 raw_spin_unlock(&cfs_b->lock);
3943 /* update hierarchical throttle state */
3944 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3946 if (!cfs_rq->load.weight)
3949 task_delta = cfs_rq->h_nr_running;
3950 for_each_sched_entity(se) {
3954 cfs_rq = cfs_rq_of(se);
3956 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3957 cfs_rq->h_nr_running += task_delta;
3959 if (cfs_rq_throttled(cfs_rq))
3964 add_nr_running(rq, task_delta);
3966 /* determine whether we need to wake up potentially idle cpu */
3967 if (rq->curr == rq->idle && rq->cfs.nr_running)
3971 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3972 u64 remaining, u64 expires)
3974 struct cfs_rq *cfs_rq;
3976 u64 starting_runtime = remaining;
3979 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3981 struct rq *rq = rq_of(cfs_rq);
3983 raw_spin_lock(&rq->lock);
3984 if (!cfs_rq_throttled(cfs_rq))
3987 runtime = -cfs_rq->runtime_remaining + 1;
3988 if (runtime > remaining)
3989 runtime = remaining;
3990 remaining -= runtime;
3992 cfs_rq->runtime_remaining += runtime;
3993 cfs_rq->runtime_expires = expires;
3995 /* we check whether we're throttled above */
3996 if (cfs_rq->runtime_remaining > 0)
3997 unthrottle_cfs_rq(cfs_rq);
4000 raw_spin_unlock(&rq->lock);
4007 return starting_runtime - remaining;
4011 * Responsible for refilling a task_group's bandwidth and unthrottling its
4012 * cfs_rqs as appropriate. If there has been no activity within the last
4013 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4014 * used to track this state.
4016 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4018 u64 runtime, runtime_expires;
4021 /* no need to continue the timer with no bandwidth constraint */
4022 if (cfs_b->quota == RUNTIME_INF)
4023 goto out_deactivate;
4025 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4026 cfs_b->nr_periods += overrun;
4029 * idle depends on !throttled (for the case of a large deficit), and if
4030 * we're going inactive then everything else can be deferred
4032 if (cfs_b->idle && !throttled)
4033 goto out_deactivate;
4035 __refill_cfs_bandwidth_runtime(cfs_b);
4038 /* mark as potentially idle for the upcoming period */
4043 /* account preceding periods in which throttling occurred */
4044 cfs_b->nr_throttled += overrun;
4046 runtime_expires = cfs_b->runtime_expires;
4049 * This check is repeated as we are holding onto the new bandwidth while
4050 * we unthrottle. This can potentially race with an unthrottled group
4051 * trying to acquire new bandwidth from the global pool. This can result
4052 * in us over-using our runtime if it is all used during this loop, but
4053 * only by limited amounts in that extreme case.
4055 while (throttled && cfs_b->runtime > 0) {
4056 runtime = cfs_b->runtime;
4057 raw_spin_unlock(&cfs_b->lock);
4058 /* we can't nest cfs_b->lock while distributing bandwidth */
4059 runtime = distribute_cfs_runtime(cfs_b, runtime,
4061 raw_spin_lock(&cfs_b->lock);
4063 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4065 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4069 * While we are ensured activity in the period following an
4070 * unthrottle, this also covers the case in which the new bandwidth is
4071 * insufficient to cover the existing bandwidth deficit. (Forcing the
4072 * timer to remain active while there are any throttled entities.)
4082 /* a cfs_rq won't donate quota below this amount */
4083 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4084 /* minimum remaining period time to redistribute slack quota */
4085 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4086 /* how long we wait to gather additional slack before distributing */
4087 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4090 * Are we near the end of the current quota period?
4092 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4093 * hrtimer base being cleared by hrtimer_start. In the case of
4094 * migrate_hrtimers, base is never cleared, so we are fine.
4096 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4098 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4101 /* if the call-back is running a quota refresh is already occurring */
4102 if (hrtimer_callback_running(refresh_timer))
4105 /* is a quota refresh about to occur? */
4106 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4107 if (remaining < min_expire)
4113 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4115 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4117 /* if there's a quota refresh soon don't bother with slack */
4118 if (runtime_refresh_within(cfs_b, min_left))
4121 hrtimer_start(&cfs_b->slack_timer,
4122 ns_to_ktime(cfs_bandwidth_slack_period),
4126 /* we know any runtime found here is valid as update_curr() precedes return */
4127 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4129 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4130 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4132 if (slack_runtime <= 0)
4135 raw_spin_lock(&cfs_b->lock);
4136 if (cfs_b->quota != RUNTIME_INF &&
4137 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4138 cfs_b->runtime += slack_runtime;
4140 /* we are under rq->lock, defer unthrottling using a timer */
4141 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4142 !list_empty(&cfs_b->throttled_cfs_rq))
4143 start_cfs_slack_bandwidth(cfs_b);
4145 raw_spin_unlock(&cfs_b->lock);
4147 /* even if it's not valid for return we don't want to try again */
4148 cfs_rq->runtime_remaining -= slack_runtime;
4151 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4153 if (!cfs_bandwidth_used())
4156 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4159 __return_cfs_rq_runtime(cfs_rq);
4163 * This is done with a timer (instead of inline with bandwidth return) since
4164 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4166 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4168 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4171 /* confirm we're still not at a refresh boundary */
4172 raw_spin_lock(&cfs_b->lock);
4173 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4174 raw_spin_unlock(&cfs_b->lock);
4178 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4179 runtime = cfs_b->runtime;
4181 expires = cfs_b->runtime_expires;
4182 raw_spin_unlock(&cfs_b->lock);
4187 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4189 raw_spin_lock(&cfs_b->lock);
4190 if (expires == cfs_b->runtime_expires)
4191 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4192 raw_spin_unlock(&cfs_b->lock);
4196 * When a group wakes up we want to make sure that its quota is not already
4197 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4198 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4200 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4202 if (!cfs_bandwidth_used())
4205 /* Synchronize hierarchical throttle counter: */
4206 if (unlikely(!cfs_rq->throttle_uptodate)) {
4207 struct rq *rq = rq_of(cfs_rq);
4208 struct cfs_rq *pcfs_rq;
4209 struct task_group *tg;
4211 cfs_rq->throttle_uptodate = 1;
4213 /* Get closest up-to-date node, because leaves go first: */
4214 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4215 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4216 if (pcfs_rq->throttle_uptodate)
4220 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4221 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4225 /* an active group must be handled by the update_curr()->put() path */
4226 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4229 /* ensure the group is not already throttled */
4230 if (cfs_rq_throttled(cfs_rq))
4233 /* update runtime allocation */
4234 account_cfs_rq_runtime(cfs_rq, 0);
4235 if (cfs_rq->runtime_remaining <= 0)
4236 throttle_cfs_rq(cfs_rq);
4239 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4240 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4242 if (!cfs_bandwidth_used())
4245 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4249 * it's possible for a throttled entity to be forced into a running
4250 * state (e.g. set_curr_task), in this case we're finished.
4252 if (cfs_rq_throttled(cfs_rq))
4255 throttle_cfs_rq(cfs_rq);
4259 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4261 struct cfs_bandwidth *cfs_b =
4262 container_of(timer, struct cfs_bandwidth, slack_timer);
4264 do_sched_cfs_slack_timer(cfs_b);
4266 return HRTIMER_NORESTART;
4269 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4271 struct cfs_bandwidth *cfs_b =
4272 container_of(timer, struct cfs_bandwidth, period_timer);
4276 raw_spin_lock(&cfs_b->lock);
4278 overrun = hrtimer_forward_now(timer, cfs_b->period);
4282 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4285 cfs_b->period_active = 0;
4286 raw_spin_unlock(&cfs_b->lock);
4288 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4291 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4293 raw_spin_lock_init(&cfs_b->lock);
4295 cfs_b->quota = RUNTIME_INF;
4296 cfs_b->period = ns_to_ktime(default_cfs_period());
4298 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4299 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4300 cfs_b->period_timer.function = sched_cfs_period_timer;
4301 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4302 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4305 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4307 cfs_rq->runtime_enabled = 0;
4308 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4311 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4313 lockdep_assert_held(&cfs_b->lock);
4315 if (!cfs_b->period_active) {
4316 cfs_b->period_active = 1;
4317 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4318 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4322 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4324 /* init_cfs_bandwidth() was not called */
4325 if (!cfs_b->throttled_cfs_rq.next)
4328 hrtimer_cancel(&cfs_b->period_timer);
4329 hrtimer_cancel(&cfs_b->slack_timer);
4332 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4334 struct cfs_rq *cfs_rq;
4336 for_each_leaf_cfs_rq(rq, cfs_rq) {
4337 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4339 raw_spin_lock(&cfs_b->lock);
4340 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4341 raw_spin_unlock(&cfs_b->lock);
4345 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4347 struct cfs_rq *cfs_rq;
4349 for_each_leaf_cfs_rq(rq, cfs_rq) {
4350 if (!cfs_rq->runtime_enabled)
4354 * clock_task is not advancing so we just need to make sure
4355 * there's some valid quota amount
4357 cfs_rq->runtime_remaining = 1;
4359 * Offline rq is schedulable till cpu is completely disabled
4360 * in take_cpu_down(), so we prevent new cfs throttling here.
4362 cfs_rq->runtime_enabled = 0;
4364 if (cfs_rq_throttled(cfs_rq))
4365 unthrottle_cfs_rq(cfs_rq);
4369 #else /* CONFIG_CFS_BANDWIDTH */
4370 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4372 return rq_clock_task(rq_of(cfs_rq));
4375 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4376 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4377 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4378 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4380 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4385 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4390 static inline int throttled_lb_pair(struct task_group *tg,
4391 int src_cpu, int dest_cpu)
4396 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4398 #ifdef CONFIG_FAIR_GROUP_SCHED
4399 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4402 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4406 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4407 static inline void update_runtime_enabled(struct rq *rq) {}
4408 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4410 #endif /* CONFIG_CFS_BANDWIDTH */
4412 /**************************************************
4413 * CFS operations on tasks:
4416 #ifdef CONFIG_SCHED_HRTICK
4417 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4419 struct sched_entity *se = &p->se;
4420 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4422 WARN_ON(task_rq(p) != rq);
4424 if (cfs_rq->nr_running > 1) {
4425 u64 slice = sched_slice(cfs_rq, se);
4426 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4427 s64 delta = slice - ran;
4434 hrtick_start(rq, delta);
4439 * called from enqueue/dequeue and updates the hrtick when the
4440 * current task is from our class and nr_running is low enough
4443 static void hrtick_update(struct rq *rq)
4445 struct task_struct *curr = rq->curr;
4447 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4450 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4451 hrtick_start_fair(rq, curr);
4453 #else /* !CONFIG_SCHED_HRTICK */
4455 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4459 static inline void hrtick_update(struct rq *rq)
4465 * The enqueue_task method is called before nr_running is
4466 * increased. Here we update the fair scheduling stats and
4467 * then put the task into the rbtree:
4470 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4472 struct cfs_rq *cfs_rq;
4473 struct sched_entity *se = &p->se;
4475 for_each_sched_entity(se) {
4478 cfs_rq = cfs_rq_of(se);
4479 enqueue_entity(cfs_rq, se, flags);
4482 * end evaluation on encountering a throttled cfs_rq
4484 * note: in the case of encountering a throttled cfs_rq we will
4485 * post the final h_nr_running increment below.
4487 if (cfs_rq_throttled(cfs_rq))
4489 cfs_rq->h_nr_running++;
4491 flags = ENQUEUE_WAKEUP;
4494 for_each_sched_entity(se) {
4495 cfs_rq = cfs_rq_of(se);
4496 cfs_rq->h_nr_running++;
4498 if (cfs_rq_throttled(cfs_rq))
4501 update_load_avg(se, 1);
4502 update_cfs_shares(cfs_rq);
4506 add_nr_running(rq, 1);
4511 static void set_next_buddy(struct sched_entity *se);
4514 * The dequeue_task method is called before nr_running is
4515 * decreased. We remove the task from the rbtree and
4516 * update the fair scheduling stats:
4518 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4520 struct cfs_rq *cfs_rq;
4521 struct sched_entity *se = &p->se;
4522 int task_sleep = flags & DEQUEUE_SLEEP;
4524 for_each_sched_entity(se) {
4525 cfs_rq = cfs_rq_of(se);
4526 dequeue_entity(cfs_rq, se, flags);
4529 * end evaluation on encountering a throttled cfs_rq
4531 * note: in the case of encountering a throttled cfs_rq we will
4532 * post the final h_nr_running decrement below.
4534 if (cfs_rq_throttled(cfs_rq))
4536 cfs_rq->h_nr_running--;
4538 /* Don't dequeue parent if it has other entities besides us */
4539 if (cfs_rq->load.weight) {
4540 /* Avoid re-evaluating load for this entity: */
4541 se = parent_entity(se);
4543 * Bias pick_next to pick a task from this cfs_rq, as
4544 * p is sleeping when it is within its sched_slice.
4546 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4550 flags |= DEQUEUE_SLEEP;
4553 for_each_sched_entity(se) {
4554 cfs_rq = cfs_rq_of(se);
4555 cfs_rq->h_nr_running--;
4557 if (cfs_rq_throttled(cfs_rq))
4560 update_load_avg(se, 1);
4561 update_cfs_shares(cfs_rq);
4565 sub_nr_running(rq, 1);
4571 #ifdef CONFIG_NO_HZ_COMMON
4573 * per rq 'load' arrray crap; XXX kill this.
4577 * The exact cpuload calculated at every tick would be:
4579 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4581 * If a cpu misses updates for n ticks (as it was idle) and update gets
4582 * called on the n+1-th tick when cpu may be busy, then we have:
4584 * load_n = (1 - 1/2^i)^n * load_0
4585 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4587 * decay_load_missed() below does efficient calculation of
4589 * load' = (1 - 1/2^i)^n * load
4591 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4592 * This allows us to precompute the above in said factors, thereby allowing the
4593 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4594 * fixed_power_int())
4596 * The calculation is approximated on a 128 point scale.
4598 #define DEGRADE_SHIFT 7
4600 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4601 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4602 { 0, 0, 0, 0, 0, 0, 0, 0 },
4603 { 64, 32, 8, 0, 0, 0, 0, 0 },
4604 { 96, 72, 40, 12, 1, 0, 0, 0 },
4605 { 112, 98, 75, 43, 15, 1, 0, 0 },
4606 { 120, 112, 98, 76, 45, 16, 2, 0 }
4610 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4611 * would be when CPU is idle and so we just decay the old load without
4612 * adding any new load.
4614 static unsigned long
4615 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4619 if (!missed_updates)
4622 if (missed_updates >= degrade_zero_ticks[idx])
4626 return load >> missed_updates;
4628 while (missed_updates) {
4629 if (missed_updates % 2)
4630 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4632 missed_updates >>= 1;
4637 #endif /* CONFIG_NO_HZ_COMMON */
4640 * __cpu_load_update - update the rq->cpu_load[] statistics
4641 * @this_rq: The rq to update statistics for
4642 * @this_load: The current load
4643 * @pending_updates: The number of missed updates
4645 * Update rq->cpu_load[] statistics. This function is usually called every
4646 * scheduler tick (TICK_NSEC).
4648 * This function computes a decaying average:
4650 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4652 * Because of NOHZ it might not get called on every tick which gives need for
4653 * the @pending_updates argument.
4655 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4656 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4657 * = A * (A * load[i]_n-2 + B) + B
4658 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4659 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4660 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4661 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4662 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4664 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4665 * any change in load would have resulted in the tick being turned back on.
4667 * For regular NOHZ, this reduces to:
4669 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4671 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4674 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4675 unsigned long pending_updates)
4677 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4680 this_rq->nr_load_updates++;
4682 /* Update our load: */
4683 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4684 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4685 unsigned long old_load, new_load;
4687 /* scale is effectively 1 << i now, and >> i divides by scale */
4689 old_load = this_rq->cpu_load[i];
4690 #ifdef CONFIG_NO_HZ_COMMON
4691 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4692 if (tickless_load) {
4693 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4695 * old_load can never be a negative value because a
4696 * decayed tickless_load cannot be greater than the
4697 * original tickless_load.
4699 old_load += tickless_load;
4702 new_load = this_load;
4704 * Round up the averaging division if load is increasing. This
4705 * prevents us from getting stuck on 9 if the load is 10, for
4708 if (new_load > old_load)
4709 new_load += scale - 1;
4711 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4714 sched_avg_update(this_rq);
4717 /* Used instead of source_load when we know the type == 0 */
4718 static unsigned long weighted_cpuload(const int cpu)
4720 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4723 #ifdef CONFIG_NO_HZ_COMMON
4725 * There is no sane way to deal with nohz on smp when using jiffies because the
4726 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4727 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4729 * Therefore we need to avoid the delta approach from the regular tick when
4730 * possible since that would seriously skew the load calculation. This is why we
4731 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4732 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4733 * loop exit, nohz_idle_balance, nohz full exit...)
4735 * This means we might still be one tick off for nohz periods.
4738 static void cpu_load_update_nohz(struct rq *this_rq,
4739 unsigned long curr_jiffies,
4742 unsigned long pending_updates;
4744 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4745 if (pending_updates) {
4746 this_rq->last_load_update_tick = curr_jiffies;
4748 * In the regular NOHZ case, we were idle, this means load 0.
4749 * In the NOHZ_FULL case, we were non-idle, we should consider
4750 * its weighted load.
4752 cpu_load_update(this_rq, load, pending_updates);
4757 * Called from nohz_idle_balance() to update the load ratings before doing the
4760 static void cpu_load_update_idle(struct rq *this_rq)
4763 * bail if there's load or we're actually up-to-date.
4765 if (weighted_cpuload(cpu_of(this_rq)))
4768 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4772 * Record CPU load on nohz entry so we know the tickless load to account
4773 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4774 * than other cpu_load[idx] but it should be fine as cpu_load readers
4775 * shouldn't rely into synchronized cpu_load[*] updates.
4777 void cpu_load_update_nohz_start(void)
4779 struct rq *this_rq = this_rq();
4782 * This is all lockless but should be fine. If weighted_cpuload changes
4783 * concurrently we'll exit nohz. And cpu_load write can race with
4784 * cpu_load_update_idle() but both updater would be writing the same.
4786 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4790 * Account the tickless load in the end of a nohz frame.
4792 void cpu_load_update_nohz_stop(void)
4794 unsigned long curr_jiffies = READ_ONCE(jiffies);
4795 struct rq *this_rq = this_rq();
4798 if (curr_jiffies == this_rq->last_load_update_tick)
4801 load = weighted_cpuload(cpu_of(this_rq));
4802 raw_spin_lock(&this_rq->lock);
4803 update_rq_clock(this_rq);
4804 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4805 raw_spin_unlock(&this_rq->lock);
4807 #else /* !CONFIG_NO_HZ_COMMON */
4808 static inline void cpu_load_update_nohz(struct rq *this_rq,
4809 unsigned long curr_jiffies,
4810 unsigned long load) { }
4811 #endif /* CONFIG_NO_HZ_COMMON */
4813 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4815 #ifdef CONFIG_NO_HZ_COMMON
4816 /* See the mess around cpu_load_update_nohz(). */
4817 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4819 cpu_load_update(this_rq, load, 1);
4823 * Called from scheduler_tick()
4825 void cpu_load_update_active(struct rq *this_rq)
4827 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4829 if (tick_nohz_tick_stopped())
4830 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4832 cpu_load_update_periodic(this_rq, load);
4836 * Return a low guess at the load of a migration-source cpu weighted
4837 * according to the scheduling class and "nice" value.
4839 * We want to under-estimate the load of migration sources, to
4840 * balance conservatively.
4842 static unsigned long source_load(int cpu, int type)
4844 struct rq *rq = cpu_rq(cpu);
4845 unsigned long total = weighted_cpuload(cpu);
4847 if (type == 0 || !sched_feat(LB_BIAS))
4850 return min(rq->cpu_load[type-1], total);
4854 * Return a high guess at the load of a migration-target cpu weighted
4855 * according to the scheduling class and "nice" value.
4857 static unsigned long target_load(int cpu, int type)
4859 struct rq *rq = cpu_rq(cpu);
4860 unsigned long total = weighted_cpuload(cpu);
4862 if (type == 0 || !sched_feat(LB_BIAS))
4865 return max(rq->cpu_load[type-1], total);
4868 static unsigned long capacity_of(int cpu)
4870 return cpu_rq(cpu)->cpu_capacity;
4873 static unsigned long capacity_orig_of(int cpu)
4875 return cpu_rq(cpu)->cpu_capacity_orig;
4878 static unsigned long cpu_avg_load_per_task(int cpu)
4880 struct rq *rq = cpu_rq(cpu);
4881 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4882 unsigned long load_avg = weighted_cpuload(cpu);
4885 return load_avg / nr_running;
4890 #ifdef CONFIG_FAIR_GROUP_SCHED
4892 * effective_load() calculates the load change as seen from the root_task_group
4894 * Adding load to a group doesn't make a group heavier, but can cause movement
4895 * of group shares between cpus. Assuming the shares were perfectly aligned one
4896 * can calculate the shift in shares.
4898 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4899 * on this @cpu and results in a total addition (subtraction) of @wg to the
4900 * total group weight.
4902 * Given a runqueue weight distribution (rw_i) we can compute a shares
4903 * distribution (s_i) using:
4905 * s_i = rw_i / \Sum rw_j (1)
4907 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4908 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4909 * shares distribution (s_i):
4911 * rw_i = { 2, 4, 1, 0 }
4912 * s_i = { 2/7, 4/7, 1/7, 0 }
4914 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4915 * task used to run on and the CPU the waker is running on), we need to
4916 * compute the effect of waking a task on either CPU and, in case of a sync
4917 * wakeup, compute the effect of the current task going to sleep.
4919 * So for a change of @wl to the local @cpu with an overall group weight change
4920 * of @wl we can compute the new shares distribution (s'_i) using:
4922 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4924 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4925 * differences in waking a task to CPU 0. The additional task changes the
4926 * weight and shares distributions like:
4928 * rw'_i = { 3, 4, 1, 0 }
4929 * s'_i = { 3/8, 4/8, 1/8, 0 }
4931 * We can then compute the difference in effective weight by using:
4933 * dw_i = S * (s'_i - s_i) (3)
4935 * Where 'S' is the group weight as seen by its parent.
4937 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4938 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4939 * 4/7) times the weight of the group.
4941 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4943 struct sched_entity *se = tg->se[cpu];
4945 if (!tg->parent) /* the trivial, non-cgroup case */
4948 for_each_sched_entity(se) {
4949 struct cfs_rq *cfs_rq = se->my_q;
4950 long W, w = cfs_rq_load_avg(cfs_rq);
4955 * W = @wg + \Sum rw_j
4957 W = wg + atomic_long_read(&tg->load_avg);
4959 /* Ensure \Sum rw_j >= rw_i */
4960 W -= cfs_rq->tg_load_avg_contrib;
4969 * wl = S * s'_i; see (2)
4972 wl = (w * (long)tg->shares) / W;
4977 * Per the above, wl is the new se->load.weight value; since
4978 * those are clipped to [MIN_SHARES, ...) do so now. See
4979 * calc_cfs_shares().
4981 if (wl < MIN_SHARES)
4985 * wl = dw_i = S * (s'_i - s_i); see (3)
4987 wl -= se->avg.load_avg;
4990 * Recursively apply this logic to all parent groups to compute
4991 * the final effective load change on the root group. Since
4992 * only the @tg group gets extra weight, all parent groups can
4993 * only redistribute existing shares. @wl is the shift in shares
4994 * resulting from this level per the above.
5003 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5010 static void record_wakee(struct task_struct *p)
5013 * Only decay a single time; tasks that have less then 1 wakeup per
5014 * jiffy will not have built up many flips.
5016 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5017 current->wakee_flips >>= 1;
5018 current->wakee_flip_decay_ts = jiffies;
5021 if (current->last_wakee != p) {
5022 current->last_wakee = p;
5023 current->wakee_flips++;
5028 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5030 * A waker of many should wake a different task than the one last awakened
5031 * at a frequency roughly N times higher than one of its wakees.
5033 * In order to determine whether we should let the load spread vs consolidating
5034 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5035 * partner, and a factor of lls_size higher frequency in the other.
5037 * With both conditions met, we can be relatively sure that the relationship is
5038 * non-monogamous, with partner count exceeding socket size.
5040 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5041 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5044 static int wake_wide(struct task_struct *p)
5046 unsigned int master = current->wakee_flips;
5047 unsigned int slave = p->wakee_flips;
5048 int factor = this_cpu_read(sd_llc_size);
5051 swap(master, slave);
5052 if (slave < factor || master < slave * factor)
5057 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5059 s64 this_load, load;
5060 s64 this_eff_load, prev_eff_load;
5061 int idx, this_cpu, prev_cpu;
5062 struct task_group *tg;
5063 unsigned long weight;
5067 this_cpu = smp_processor_id();
5068 prev_cpu = task_cpu(p);
5069 load = source_load(prev_cpu, idx);
5070 this_load = target_load(this_cpu, idx);
5073 * If sync wakeup then subtract the (maximum possible)
5074 * effect of the currently running task from the load
5075 * of the current CPU:
5078 tg = task_group(current);
5079 weight = current->se.avg.load_avg;
5081 this_load += effective_load(tg, this_cpu, -weight, -weight);
5082 load += effective_load(tg, prev_cpu, 0, -weight);
5086 weight = p->se.avg.load_avg;
5089 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5090 * due to the sync cause above having dropped this_load to 0, we'll
5091 * always have an imbalance, but there's really nothing you can do
5092 * about that, so that's good too.
5094 * Otherwise check if either cpus are near enough in load to allow this
5095 * task to be woken on this_cpu.
5097 this_eff_load = 100;
5098 this_eff_load *= capacity_of(prev_cpu);
5100 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5101 prev_eff_load *= capacity_of(this_cpu);
5103 if (this_load > 0) {
5104 this_eff_load *= this_load +
5105 effective_load(tg, this_cpu, weight, weight);
5107 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5110 balanced = this_eff_load <= prev_eff_load;
5112 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5117 schedstat_inc(sd, ttwu_move_affine);
5118 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5124 * find_idlest_group finds and returns the least busy CPU group within the
5127 static struct sched_group *
5128 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5129 int this_cpu, int sd_flag)
5131 struct sched_group *idlest = NULL, *group = sd->groups;
5132 unsigned long min_load = ULONG_MAX, this_load = 0;
5133 int load_idx = sd->forkexec_idx;
5134 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5136 if (sd_flag & SD_BALANCE_WAKE)
5137 load_idx = sd->wake_idx;
5140 unsigned long load, avg_load;
5144 /* Skip over this group if it has no CPUs allowed */
5145 if (!cpumask_intersects(sched_group_cpus(group),
5146 tsk_cpus_allowed(p)))
5149 local_group = cpumask_test_cpu(this_cpu,
5150 sched_group_cpus(group));
5152 /* Tally up the load of all CPUs in the group */
5155 for_each_cpu(i, sched_group_cpus(group)) {
5156 /* Bias balancing toward cpus of our domain */
5158 load = source_load(i, load_idx);
5160 load = target_load(i, load_idx);
5165 /* Adjust by relative CPU capacity of the group */
5166 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5169 this_load = avg_load;
5170 } else if (avg_load < min_load) {
5171 min_load = avg_load;
5174 } while (group = group->next, group != sd->groups);
5176 if (!idlest || 100*this_load < imbalance*min_load)
5182 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5185 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5187 unsigned long load, min_load = ULONG_MAX;
5188 unsigned int min_exit_latency = UINT_MAX;
5189 u64 latest_idle_timestamp = 0;
5190 int least_loaded_cpu = this_cpu;
5191 int shallowest_idle_cpu = -1;
5194 /* Traverse only the allowed CPUs */
5195 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5197 struct rq *rq = cpu_rq(i);
5198 struct cpuidle_state *idle = idle_get_state(rq);
5199 if (idle && idle->exit_latency < min_exit_latency) {
5201 * We give priority to a CPU whose idle state
5202 * has the smallest exit latency irrespective
5203 * of any idle timestamp.
5205 min_exit_latency = idle->exit_latency;
5206 latest_idle_timestamp = rq->idle_stamp;
5207 shallowest_idle_cpu = i;
5208 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5209 rq->idle_stamp > latest_idle_timestamp) {
5211 * If equal or no active idle state, then
5212 * the most recently idled CPU might have
5215 latest_idle_timestamp = rq->idle_stamp;
5216 shallowest_idle_cpu = i;
5218 } else if (shallowest_idle_cpu == -1) {
5219 load = weighted_cpuload(i);
5220 if (load < min_load || (load == min_load && i == this_cpu)) {
5222 least_loaded_cpu = i;
5227 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5231 * Try and locate an idle CPU in the sched_domain.
5233 static int select_idle_sibling(struct task_struct *p, int target)
5235 struct sched_domain *sd;
5236 struct sched_group *sg;
5237 int i = task_cpu(p);
5239 if (idle_cpu(target))
5243 * If the prevous cpu is cache affine and idle, don't be stupid.
5245 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5249 * Otherwise, iterate the domains and find an eligible idle cpu.
5251 * A completely idle sched group at higher domains is more
5252 * desirable than an idle group at a lower level, because lower
5253 * domains have smaller groups and usually share hardware
5254 * resources which causes tasks to contend on them, e.g. x86
5255 * hyperthread siblings in the lowest domain (SMT) can contend
5256 * on the shared cpu pipeline.
5258 * However, while we prefer idle groups at higher domains
5259 * finding an idle cpu at the lowest domain is still better than
5260 * returning 'target', which we've already established, isn't
5263 sd = rcu_dereference(per_cpu(sd_llc, target));
5264 for_each_lower_domain(sd) {
5267 if (!cpumask_intersects(sched_group_cpus(sg),
5268 tsk_cpus_allowed(p)))
5271 /* Ensure the entire group is idle */
5272 for_each_cpu(i, sched_group_cpus(sg)) {
5273 if (i == target || !idle_cpu(i))
5278 * It doesn't matter which cpu we pick, the
5279 * whole group is idle.
5281 target = cpumask_first_and(sched_group_cpus(sg),
5282 tsk_cpus_allowed(p));
5286 } while (sg != sd->groups);
5293 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5294 * tasks. The unit of the return value must be the one of capacity so we can
5295 * compare the utilization with the capacity of the CPU that is available for
5296 * CFS task (ie cpu_capacity).
5298 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5299 * recent utilization of currently non-runnable tasks on a CPU. It represents
5300 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5301 * capacity_orig is the cpu_capacity available at the highest frequency
5302 * (arch_scale_freq_capacity()).
5303 * The utilization of a CPU converges towards a sum equal to or less than the
5304 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5305 * the running time on this CPU scaled by capacity_curr.
5307 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5308 * higher than capacity_orig because of unfortunate rounding in
5309 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5310 * the average stabilizes with the new running time. We need to check that the
5311 * utilization stays within the range of [0..capacity_orig] and cap it if
5312 * necessary. Without utilization capping, a group could be seen as overloaded
5313 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5314 * available capacity. We allow utilization to overshoot capacity_curr (but not
5315 * capacity_orig) as it useful for predicting the capacity required after task
5316 * migrations (scheduler-driven DVFS).
5318 static int cpu_util(int cpu)
5320 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5321 unsigned long capacity = capacity_orig_of(cpu);
5323 return (util >= capacity) ? capacity : util;
5327 * select_task_rq_fair: Select target runqueue for the waking task in domains
5328 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5329 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5331 * Balances load by selecting the idlest cpu in the idlest group, or under
5332 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5334 * Returns the target cpu number.
5336 * preempt must be disabled.
5339 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5341 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5342 int cpu = smp_processor_id();
5343 int new_cpu = prev_cpu;
5344 int want_affine = 0;
5345 int sync = wake_flags & WF_SYNC;
5347 if (sd_flag & SD_BALANCE_WAKE) {
5349 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5353 for_each_domain(cpu, tmp) {
5354 if (!(tmp->flags & SD_LOAD_BALANCE))
5358 * If both cpu and prev_cpu are part of this domain,
5359 * cpu is a valid SD_WAKE_AFFINE target.
5361 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5362 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5367 if (tmp->flags & sd_flag)
5369 else if (!want_affine)
5374 sd = NULL; /* Prefer wake_affine over balance flags */
5375 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5380 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5381 new_cpu = select_idle_sibling(p, new_cpu);
5384 struct sched_group *group;
5387 if (!(sd->flags & sd_flag)) {
5392 group = find_idlest_group(sd, p, cpu, sd_flag);
5398 new_cpu = find_idlest_cpu(group, p, cpu);
5399 if (new_cpu == -1 || new_cpu == cpu) {
5400 /* Now try balancing at a lower domain level of cpu */
5405 /* Now try balancing at a lower domain level of new_cpu */
5407 weight = sd->span_weight;
5409 for_each_domain(cpu, tmp) {
5410 if (weight <= tmp->span_weight)
5412 if (tmp->flags & sd_flag)
5415 /* while loop will break here if sd == NULL */
5423 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5424 * cfs_rq_of(p) references at time of call are still valid and identify the
5425 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5427 static void migrate_task_rq_fair(struct task_struct *p)
5430 * As blocked tasks retain absolute vruntime the migration needs to
5431 * deal with this by subtracting the old and adding the new
5432 * min_vruntime -- the latter is done by enqueue_entity() when placing
5433 * the task on the new runqueue.
5435 if (p->state == TASK_WAKING) {
5436 struct sched_entity *se = &p->se;
5437 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5440 #ifndef CONFIG_64BIT
5441 u64 min_vruntime_copy;
5444 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5446 min_vruntime = cfs_rq->min_vruntime;
5447 } while (min_vruntime != min_vruntime_copy);
5449 min_vruntime = cfs_rq->min_vruntime;
5452 se->vruntime -= min_vruntime;
5456 * We are supposed to update the task to "current" time, then its up to date
5457 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5458 * what current time is, so simply throw away the out-of-date time. This
5459 * will result in the wakee task is less decayed, but giving the wakee more
5460 * load sounds not bad.
5462 remove_entity_load_avg(&p->se);
5464 /* Tell new CPU we are migrated */
5465 p->se.avg.last_update_time = 0;
5467 /* We have migrated, no longer consider this task hot */
5468 p->se.exec_start = 0;
5471 static void task_dead_fair(struct task_struct *p)
5473 remove_entity_load_avg(&p->se);
5475 #endif /* CONFIG_SMP */
5477 static unsigned long
5478 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5480 unsigned long gran = sysctl_sched_wakeup_granularity;
5483 * Since its curr running now, convert the gran from real-time
5484 * to virtual-time in his units.
5486 * By using 'se' instead of 'curr' we penalize light tasks, so
5487 * they get preempted easier. That is, if 'se' < 'curr' then
5488 * the resulting gran will be larger, therefore penalizing the
5489 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5490 * be smaller, again penalizing the lighter task.
5492 * This is especially important for buddies when the leftmost
5493 * task is higher priority than the buddy.
5495 return calc_delta_fair(gran, se);
5499 * Should 'se' preempt 'curr'.
5513 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5515 s64 gran, vdiff = curr->vruntime - se->vruntime;
5520 gran = wakeup_gran(curr, se);
5527 static void set_last_buddy(struct sched_entity *se)
5529 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5532 for_each_sched_entity(se)
5533 cfs_rq_of(se)->last = se;
5536 static void set_next_buddy(struct sched_entity *se)
5538 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5541 for_each_sched_entity(se)
5542 cfs_rq_of(se)->next = se;
5545 static void set_skip_buddy(struct sched_entity *se)
5547 for_each_sched_entity(se)
5548 cfs_rq_of(se)->skip = se;
5552 * Preempt the current task with a newly woken task if needed:
5554 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5556 struct task_struct *curr = rq->curr;
5557 struct sched_entity *se = &curr->se, *pse = &p->se;
5558 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5559 int scale = cfs_rq->nr_running >= sched_nr_latency;
5560 int next_buddy_marked = 0;
5562 if (unlikely(se == pse))
5566 * This is possible from callers such as attach_tasks(), in which we
5567 * unconditionally check_prempt_curr() after an enqueue (which may have
5568 * lead to a throttle). This both saves work and prevents false
5569 * next-buddy nomination below.
5571 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5574 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5575 set_next_buddy(pse);
5576 next_buddy_marked = 1;
5580 * We can come here with TIF_NEED_RESCHED already set from new task
5583 * Note: this also catches the edge-case of curr being in a throttled
5584 * group (e.g. via set_curr_task), since update_curr() (in the
5585 * enqueue of curr) will have resulted in resched being set. This
5586 * prevents us from potentially nominating it as a false LAST_BUDDY
5589 if (test_tsk_need_resched(curr))
5592 /* Idle tasks are by definition preempted by non-idle tasks. */
5593 if (unlikely(curr->policy == SCHED_IDLE) &&
5594 likely(p->policy != SCHED_IDLE))
5598 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5599 * is driven by the tick):
5601 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5604 find_matching_se(&se, &pse);
5605 update_curr(cfs_rq_of(se));
5607 if (wakeup_preempt_entity(se, pse) == 1) {
5609 * Bias pick_next to pick the sched entity that is
5610 * triggering this preemption.
5612 if (!next_buddy_marked)
5613 set_next_buddy(pse);
5622 * Only set the backward buddy when the current task is still
5623 * on the rq. This can happen when a wakeup gets interleaved
5624 * with schedule on the ->pre_schedule() or idle_balance()
5625 * point, either of which can * drop the rq lock.
5627 * Also, during early boot the idle thread is in the fair class,
5628 * for obvious reasons its a bad idea to schedule back to it.
5630 if (unlikely(!se->on_rq || curr == rq->idle))
5633 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5637 static struct task_struct *
5638 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5640 struct cfs_rq *cfs_rq = &rq->cfs;
5641 struct sched_entity *se;
5642 struct task_struct *p;
5646 #ifdef CONFIG_FAIR_GROUP_SCHED
5647 if (!cfs_rq->nr_running)
5650 if (prev->sched_class != &fair_sched_class)
5654 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5655 * likely that a next task is from the same cgroup as the current.
5657 * Therefore attempt to avoid putting and setting the entire cgroup
5658 * hierarchy, only change the part that actually changes.
5662 struct sched_entity *curr = cfs_rq->curr;
5665 * Since we got here without doing put_prev_entity() we also
5666 * have to consider cfs_rq->curr. If it is still a runnable
5667 * entity, update_curr() will update its vruntime, otherwise
5668 * forget we've ever seen it.
5672 update_curr(cfs_rq);
5677 * This call to check_cfs_rq_runtime() will do the
5678 * throttle and dequeue its entity in the parent(s).
5679 * Therefore the 'simple' nr_running test will indeed
5682 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5686 se = pick_next_entity(cfs_rq, curr);
5687 cfs_rq = group_cfs_rq(se);
5693 * Since we haven't yet done put_prev_entity and if the selected task
5694 * is a different task than we started out with, try and touch the
5695 * least amount of cfs_rqs.
5698 struct sched_entity *pse = &prev->se;
5700 while (!(cfs_rq = is_same_group(se, pse))) {
5701 int se_depth = se->depth;
5702 int pse_depth = pse->depth;
5704 if (se_depth <= pse_depth) {
5705 put_prev_entity(cfs_rq_of(pse), pse);
5706 pse = parent_entity(pse);
5708 if (se_depth >= pse_depth) {
5709 set_next_entity(cfs_rq_of(se), se);
5710 se = parent_entity(se);
5714 put_prev_entity(cfs_rq, pse);
5715 set_next_entity(cfs_rq, se);
5718 if (hrtick_enabled(rq))
5719 hrtick_start_fair(rq, p);
5726 if (!cfs_rq->nr_running)
5729 put_prev_task(rq, prev);
5732 se = pick_next_entity(cfs_rq, NULL);
5733 set_next_entity(cfs_rq, se);
5734 cfs_rq = group_cfs_rq(se);
5739 if (hrtick_enabled(rq))
5740 hrtick_start_fair(rq, p);
5746 * This is OK, because current is on_cpu, which avoids it being picked
5747 * for load-balance and preemption/IRQs are still disabled avoiding
5748 * further scheduler activity on it and we're being very careful to
5749 * re-start the picking loop.
5751 lockdep_unpin_lock(&rq->lock, cookie);
5752 new_tasks = idle_balance(rq);
5753 lockdep_repin_lock(&rq->lock, cookie);
5755 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5756 * possible for any higher priority task to appear. In that case we
5757 * must re-start the pick_next_entity() loop.
5769 * Account for a descheduled task:
5771 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5773 struct sched_entity *se = &prev->se;
5774 struct cfs_rq *cfs_rq;
5776 for_each_sched_entity(se) {
5777 cfs_rq = cfs_rq_of(se);
5778 put_prev_entity(cfs_rq, se);
5783 * sched_yield() is very simple
5785 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5787 static void yield_task_fair(struct rq *rq)
5789 struct task_struct *curr = rq->curr;
5790 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5791 struct sched_entity *se = &curr->se;
5794 * Are we the only task in the tree?
5796 if (unlikely(rq->nr_running == 1))
5799 clear_buddies(cfs_rq, se);
5801 if (curr->policy != SCHED_BATCH) {
5802 update_rq_clock(rq);
5804 * Update run-time statistics of the 'current'.
5806 update_curr(cfs_rq);
5808 * Tell update_rq_clock() that we've just updated,
5809 * so we don't do microscopic update in schedule()
5810 * and double the fastpath cost.
5812 rq_clock_skip_update(rq, true);
5818 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5820 struct sched_entity *se = &p->se;
5822 /* throttled hierarchies are not runnable */
5823 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5826 /* Tell the scheduler that we'd really like pse to run next. */
5829 yield_task_fair(rq);
5835 /**************************************************
5836 * Fair scheduling class load-balancing methods.
5840 * The purpose of load-balancing is to achieve the same basic fairness the
5841 * per-cpu scheduler provides, namely provide a proportional amount of compute
5842 * time to each task. This is expressed in the following equation:
5844 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5846 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5847 * W_i,0 is defined as:
5849 * W_i,0 = \Sum_j w_i,j (2)
5851 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5852 * is derived from the nice value as per sched_prio_to_weight[].
5854 * The weight average is an exponential decay average of the instantaneous
5857 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5859 * C_i is the compute capacity of cpu i, typically it is the
5860 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5861 * can also include other factors [XXX].
5863 * To achieve this balance we define a measure of imbalance which follows
5864 * directly from (1):
5866 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5868 * We them move tasks around to minimize the imbalance. In the continuous
5869 * function space it is obvious this converges, in the discrete case we get
5870 * a few fun cases generally called infeasible weight scenarios.
5873 * - infeasible weights;
5874 * - local vs global optima in the discrete case. ]
5879 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5880 * for all i,j solution, we create a tree of cpus that follows the hardware
5881 * topology where each level pairs two lower groups (or better). This results
5882 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5883 * tree to only the first of the previous level and we decrease the frequency
5884 * of load-balance at each level inv. proportional to the number of cpus in
5890 * \Sum { --- * --- * 2^i } = O(n) (5)
5892 * `- size of each group
5893 * | | `- number of cpus doing load-balance
5895 * `- sum over all levels
5897 * Coupled with a limit on how many tasks we can migrate every balance pass,
5898 * this makes (5) the runtime complexity of the balancer.
5900 * An important property here is that each CPU is still (indirectly) connected
5901 * to every other cpu in at most O(log n) steps:
5903 * The adjacency matrix of the resulting graph is given by:
5906 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5909 * And you'll find that:
5911 * A^(log_2 n)_i,j != 0 for all i,j (7)
5913 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5914 * The task movement gives a factor of O(m), giving a convergence complexity
5917 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5922 * In order to avoid CPUs going idle while there's still work to do, new idle
5923 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5924 * tree itself instead of relying on other CPUs to bring it work.
5926 * This adds some complexity to both (5) and (8) but it reduces the total idle
5934 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5937 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5942 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5944 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5946 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5949 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5950 * rewrite all of this once again.]
5953 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5955 enum fbq_type { regular, remote, all };
5957 #define LBF_ALL_PINNED 0x01
5958 #define LBF_NEED_BREAK 0x02
5959 #define LBF_DST_PINNED 0x04
5960 #define LBF_SOME_PINNED 0x08
5963 struct sched_domain *sd;
5971 struct cpumask *dst_grpmask;
5973 enum cpu_idle_type idle;
5975 /* The set of CPUs under consideration for load-balancing */
5976 struct cpumask *cpus;
5981 unsigned int loop_break;
5982 unsigned int loop_max;
5984 enum fbq_type fbq_type;
5985 struct list_head tasks;
5989 * Is this task likely cache-hot:
5991 static int task_hot(struct task_struct *p, struct lb_env *env)
5995 lockdep_assert_held(&env->src_rq->lock);
5997 if (p->sched_class != &fair_sched_class)
6000 if (unlikely(p->policy == SCHED_IDLE))
6004 * Buddy candidates are cache hot:
6006 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6007 (&p->se == cfs_rq_of(&p->se)->next ||
6008 &p->se == cfs_rq_of(&p->se)->last))
6011 if (sysctl_sched_migration_cost == -1)
6013 if (sysctl_sched_migration_cost == 0)
6016 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6018 return delta < (s64)sysctl_sched_migration_cost;
6021 #ifdef CONFIG_NUMA_BALANCING
6023 * Returns 1, if task migration degrades locality
6024 * Returns 0, if task migration improves locality i.e migration preferred.
6025 * Returns -1, if task migration is not affected by locality.
6027 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6029 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6030 unsigned long src_faults, dst_faults;
6031 int src_nid, dst_nid;
6033 if (!static_branch_likely(&sched_numa_balancing))
6036 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6039 src_nid = cpu_to_node(env->src_cpu);
6040 dst_nid = cpu_to_node(env->dst_cpu);
6042 if (src_nid == dst_nid)
6045 /* Migrating away from the preferred node is always bad. */
6046 if (src_nid == p->numa_preferred_nid) {
6047 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6053 /* Encourage migration to the preferred node. */
6054 if (dst_nid == p->numa_preferred_nid)
6058 src_faults = group_faults(p, src_nid);
6059 dst_faults = group_faults(p, dst_nid);
6061 src_faults = task_faults(p, src_nid);
6062 dst_faults = task_faults(p, dst_nid);
6065 return dst_faults < src_faults;
6069 static inline int migrate_degrades_locality(struct task_struct *p,
6077 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6080 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6084 lockdep_assert_held(&env->src_rq->lock);
6087 * We do not migrate tasks that are:
6088 * 1) throttled_lb_pair, or
6089 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6090 * 3) running (obviously), or
6091 * 4) are cache-hot on their current CPU.
6093 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6096 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6099 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6101 env->flags |= LBF_SOME_PINNED;
6104 * Remember if this task can be migrated to any other cpu in
6105 * our sched_group. We may want to revisit it if we couldn't
6106 * meet load balance goals by pulling other tasks on src_cpu.
6108 * Also avoid computing new_dst_cpu if we have already computed
6109 * one in current iteration.
6111 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6114 /* Prevent to re-select dst_cpu via env's cpus */
6115 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6116 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6117 env->flags |= LBF_DST_PINNED;
6118 env->new_dst_cpu = cpu;
6126 /* Record that we found atleast one task that could run on dst_cpu */
6127 env->flags &= ~LBF_ALL_PINNED;
6129 if (task_running(env->src_rq, p)) {
6130 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6135 * Aggressive migration if:
6136 * 1) destination numa is preferred
6137 * 2) task is cache cold, or
6138 * 3) too many balance attempts have failed.
6140 tsk_cache_hot = migrate_degrades_locality(p, env);
6141 if (tsk_cache_hot == -1)
6142 tsk_cache_hot = task_hot(p, env);
6144 if (tsk_cache_hot <= 0 ||
6145 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6146 if (tsk_cache_hot == 1) {
6147 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6148 schedstat_inc(p, se.statistics.nr_forced_migrations);
6153 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6158 * detach_task() -- detach the task for the migration specified in env
6160 static void detach_task(struct task_struct *p, struct lb_env *env)
6162 lockdep_assert_held(&env->src_rq->lock);
6164 p->on_rq = TASK_ON_RQ_MIGRATING;
6165 deactivate_task(env->src_rq, p, 0);
6166 set_task_cpu(p, env->dst_cpu);
6170 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6171 * part of active balancing operations within "domain".
6173 * Returns a task if successful and NULL otherwise.
6175 static struct task_struct *detach_one_task(struct lb_env *env)
6177 struct task_struct *p, *n;
6179 lockdep_assert_held(&env->src_rq->lock);
6181 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6182 if (!can_migrate_task(p, env))
6185 detach_task(p, env);
6188 * Right now, this is only the second place where
6189 * lb_gained[env->idle] is updated (other is detach_tasks)
6190 * so we can safely collect stats here rather than
6191 * inside detach_tasks().
6193 schedstat_inc(env->sd, lb_gained[env->idle]);
6199 static const unsigned int sched_nr_migrate_break = 32;
6202 * detach_tasks() -- tries to detach up to imbalance weighted load from
6203 * busiest_rq, as part of a balancing operation within domain "sd".
6205 * Returns number of detached tasks if successful and 0 otherwise.
6207 static int detach_tasks(struct lb_env *env)
6209 struct list_head *tasks = &env->src_rq->cfs_tasks;
6210 struct task_struct *p;
6214 lockdep_assert_held(&env->src_rq->lock);
6216 if (env->imbalance <= 0)
6219 while (!list_empty(tasks)) {
6221 * We don't want to steal all, otherwise we may be treated likewise,
6222 * which could at worst lead to a livelock crash.
6224 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6227 p = list_first_entry(tasks, struct task_struct, se.group_node);
6230 /* We've more or less seen every task there is, call it quits */
6231 if (env->loop > env->loop_max)
6234 /* take a breather every nr_migrate tasks */
6235 if (env->loop > env->loop_break) {
6236 env->loop_break += sched_nr_migrate_break;
6237 env->flags |= LBF_NEED_BREAK;
6241 if (!can_migrate_task(p, env))
6244 load = task_h_load(p);
6246 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6249 if ((load / 2) > env->imbalance)
6252 detach_task(p, env);
6253 list_add(&p->se.group_node, &env->tasks);
6256 env->imbalance -= load;
6258 #ifdef CONFIG_PREEMPT
6260 * NEWIDLE balancing is a source of latency, so preemptible
6261 * kernels will stop after the first task is detached to minimize
6262 * the critical section.
6264 if (env->idle == CPU_NEWLY_IDLE)
6269 * We only want to steal up to the prescribed amount of
6272 if (env->imbalance <= 0)
6277 list_move_tail(&p->se.group_node, tasks);
6281 * Right now, this is one of only two places we collect this stat
6282 * so we can safely collect detach_one_task() stats here rather
6283 * than inside detach_one_task().
6285 schedstat_add(env->sd, lb_gained[env->idle], detached);
6291 * attach_task() -- attach the task detached by detach_task() to its new rq.
6293 static void attach_task(struct rq *rq, struct task_struct *p)
6295 lockdep_assert_held(&rq->lock);
6297 BUG_ON(task_rq(p) != rq);
6298 activate_task(rq, p, 0);
6299 p->on_rq = TASK_ON_RQ_QUEUED;
6300 check_preempt_curr(rq, p, 0);
6304 * attach_one_task() -- attaches the task returned from detach_one_task() to
6307 static void attach_one_task(struct rq *rq, struct task_struct *p)
6309 raw_spin_lock(&rq->lock);
6311 raw_spin_unlock(&rq->lock);
6315 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6318 static void attach_tasks(struct lb_env *env)
6320 struct list_head *tasks = &env->tasks;
6321 struct task_struct *p;
6323 raw_spin_lock(&env->dst_rq->lock);
6325 while (!list_empty(tasks)) {
6326 p = list_first_entry(tasks, struct task_struct, se.group_node);
6327 list_del_init(&p->se.group_node);
6329 attach_task(env->dst_rq, p);
6332 raw_spin_unlock(&env->dst_rq->lock);
6335 #ifdef CONFIG_FAIR_GROUP_SCHED
6336 static void update_blocked_averages(int cpu)
6338 struct rq *rq = cpu_rq(cpu);
6339 struct cfs_rq *cfs_rq;
6340 unsigned long flags;
6342 raw_spin_lock_irqsave(&rq->lock, flags);
6343 update_rq_clock(rq);
6346 * Iterates the task_group tree in a bottom up fashion, see
6347 * list_add_leaf_cfs_rq() for details.
6349 for_each_leaf_cfs_rq(rq, cfs_rq) {
6350 /* throttled entities do not contribute to load */
6351 if (throttled_hierarchy(cfs_rq))
6354 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6355 update_tg_load_avg(cfs_rq, 0);
6357 raw_spin_unlock_irqrestore(&rq->lock, flags);
6361 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6362 * This needs to be done in a top-down fashion because the load of a child
6363 * group is a fraction of its parents load.
6365 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6367 struct rq *rq = rq_of(cfs_rq);
6368 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6369 unsigned long now = jiffies;
6372 if (cfs_rq->last_h_load_update == now)
6375 cfs_rq->h_load_next = NULL;
6376 for_each_sched_entity(se) {
6377 cfs_rq = cfs_rq_of(se);
6378 cfs_rq->h_load_next = se;
6379 if (cfs_rq->last_h_load_update == now)
6384 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6385 cfs_rq->last_h_load_update = now;
6388 while ((se = cfs_rq->h_load_next) != NULL) {
6389 load = cfs_rq->h_load;
6390 load = div64_ul(load * se->avg.load_avg,
6391 cfs_rq_load_avg(cfs_rq) + 1);
6392 cfs_rq = group_cfs_rq(se);
6393 cfs_rq->h_load = load;
6394 cfs_rq->last_h_load_update = now;
6398 static unsigned long task_h_load(struct task_struct *p)
6400 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6402 update_cfs_rq_h_load(cfs_rq);
6403 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6404 cfs_rq_load_avg(cfs_rq) + 1);
6407 static inline void update_blocked_averages(int cpu)
6409 struct rq *rq = cpu_rq(cpu);
6410 struct cfs_rq *cfs_rq = &rq->cfs;
6411 unsigned long flags;
6413 raw_spin_lock_irqsave(&rq->lock, flags);
6414 update_rq_clock(rq);
6415 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6416 raw_spin_unlock_irqrestore(&rq->lock, flags);
6419 static unsigned long task_h_load(struct task_struct *p)
6421 return p->se.avg.load_avg;
6425 /********** Helpers for find_busiest_group ************************/
6434 * sg_lb_stats - stats of a sched_group required for load_balancing
6436 struct sg_lb_stats {
6437 unsigned long avg_load; /*Avg load across the CPUs of the group */
6438 unsigned long group_load; /* Total load over the CPUs of the group */
6439 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6440 unsigned long load_per_task;
6441 unsigned long group_capacity;
6442 unsigned long group_util; /* Total utilization of the group */
6443 unsigned int sum_nr_running; /* Nr tasks running in the group */
6444 unsigned int idle_cpus;
6445 unsigned int group_weight;
6446 enum group_type group_type;
6447 int group_no_capacity;
6448 #ifdef CONFIG_NUMA_BALANCING
6449 unsigned int nr_numa_running;
6450 unsigned int nr_preferred_running;
6455 * sd_lb_stats - Structure to store the statistics of a sched_domain
6456 * during load balancing.
6458 struct sd_lb_stats {
6459 struct sched_group *busiest; /* Busiest group in this sd */
6460 struct sched_group *local; /* Local group in this sd */
6461 unsigned long total_load; /* Total load of all groups in sd */
6462 unsigned long total_capacity; /* Total capacity of all groups in sd */
6463 unsigned long avg_load; /* Average load across all groups in sd */
6465 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6466 struct sg_lb_stats local_stat; /* Statistics of the local group */
6469 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6472 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6473 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6474 * We must however clear busiest_stat::avg_load because
6475 * update_sd_pick_busiest() reads this before assignment.
6477 *sds = (struct sd_lb_stats){
6481 .total_capacity = 0UL,
6484 .sum_nr_running = 0,
6485 .group_type = group_other,
6491 * get_sd_load_idx - Obtain the load index for a given sched domain.
6492 * @sd: The sched_domain whose load_idx is to be obtained.
6493 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6495 * Return: The load index.
6497 static inline int get_sd_load_idx(struct sched_domain *sd,
6498 enum cpu_idle_type idle)
6504 load_idx = sd->busy_idx;
6507 case CPU_NEWLY_IDLE:
6508 load_idx = sd->newidle_idx;
6511 load_idx = sd->idle_idx;
6518 static unsigned long scale_rt_capacity(int cpu)
6520 struct rq *rq = cpu_rq(cpu);
6521 u64 total, used, age_stamp, avg;
6525 * Since we're reading these variables without serialization make sure
6526 * we read them once before doing sanity checks on them.
6528 age_stamp = READ_ONCE(rq->age_stamp);
6529 avg = READ_ONCE(rq->rt_avg);
6530 delta = __rq_clock_broken(rq) - age_stamp;
6532 if (unlikely(delta < 0))
6535 total = sched_avg_period() + delta;
6537 used = div_u64(avg, total);
6539 if (likely(used < SCHED_CAPACITY_SCALE))
6540 return SCHED_CAPACITY_SCALE - used;
6545 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6547 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6548 struct sched_group *sdg = sd->groups;
6550 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6552 capacity *= scale_rt_capacity(cpu);
6553 capacity >>= SCHED_CAPACITY_SHIFT;
6558 cpu_rq(cpu)->cpu_capacity = capacity;
6559 sdg->sgc->capacity = capacity;
6562 void update_group_capacity(struct sched_domain *sd, int cpu)
6564 struct sched_domain *child = sd->child;
6565 struct sched_group *group, *sdg = sd->groups;
6566 unsigned long capacity;
6567 unsigned long interval;
6569 interval = msecs_to_jiffies(sd->balance_interval);
6570 interval = clamp(interval, 1UL, max_load_balance_interval);
6571 sdg->sgc->next_update = jiffies + interval;
6574 update_cpu_capacity(sd, cpu);
6580 if (child->flags & SD_OVERLAP) {
6582 * SD_OVERLAP domains cannot assume that child groups
6583 * span the current group.
6586 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6587 struct sched_group_capacity *sgc;
6588 struct rq *rq = cpu_rq(cpu);
6591 * build_sched_domains() -> init_sched_groups_capacity()
6592 * gets here before we've attached the domains to the
6595 * Use capacity_of(), which is set irrespective of domains
6596 * in update_cpu_capacity().
6598 * This avoids capacity from being 0 and
6599 * causing divide-by-zero issues on boot.
6601 if (unlikely(!rq->sd)) {
6602 capacity += capacity_of(cpu);
6606 sgc = rq->sd->groups->sgc;
6607 capacity += sgc->capacity;
6611 * !SD_OVERLAP domains can assume that child groups
6612 * span the current group.
6615 group = child->groups;
6617 capacity += group->sgc->capacity;
6618 group = group->next;
6619 } while (group != child->groups);
6622 sdg->sgc->capacity = capacity;
6626 * Check whether the capacity of the rq has been noticeably reduced by side
6627 * activity. The imbalance_pct is used for the threshold.
6628 * Return true is the capacity is reduced
6631 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6633 return ((rq->cpu_capacity * sd->imbalance_pct) <
6634 (rq->cpu_capacity_orig * 100));
6638 * Group imbalance indicates (and tries to solve) the problem where balancing
6639 * groups is inadequate due to tsk_cpus_allowed() constraints.
6641 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6642 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6645 * { 0 1 2 3 } { 4 5 6 7 }
6648 * If we were to balance group-wise we'd place two tasks in the first group and
6649 * two tasks in the second group. Clearly this is undesired as it will overload
6650 * cpu 3 and leave one of the cpus in the second group unused.
6652 * The current solution to this issue is detecting the skew in the first group
6653 * by noticing the lower domain failed to reach balance and had difficulty
6654 * moving tasks due to affinity constraints.
6656 * When this is so detected; this group becomes a candidate for busiest; see
6657 * update_sd_pick_busiest(). And calculate_imbalance() and
6658 * find_busiest_group() avoid some of the usual balance conditions to allow it
6659 * to create an effective group imbalance.
6661 * This is a somewhat tricky proposition since the next run might not find the
6662 * group imbalance and decide the groups need to be balanced again. A most
6663 * subtle and fragile situation.
6666 static inline int sg_imbalanced(struct sched_group *group)
6668 return group->sgc->imbalance;
6672 * group_has_capacity returns true if the group has spare capacity that could
6673 * be used by some tasks.
6674 * We consider that a group has spare capacity if the * number of task is
6675 * smaller than the number of CPUs or if the utilization is lower than the
6676 * available capacity for CFS tasks.
6677 * For the latter, we use a threshold to stabilize the state, to take into
6678 * account the variance of the tasks' load and to return true if the available
6679 * capacity in meaningful for the load balancer.
6680 * As an example, an available capacity of 1% can appear but it doesn't make
6681 * any benefit for the load balance.
6684 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6686 if (sgs->sum_nr_running < sgs->group_weight)
6689 if ((sgs->group_capacity * 100) >
6690 (sgs->group_util * env->sd->imbalance_pct))
6697 * group_is_overloaded returns true if the group has more tasks than it can
6699 * group_is_overloaded is not equals to !group_has_capacity because a group
6700 * with the exact right number of tasks, has no more spare capacity but is not
6701 * overloaded so both group_has_capacity and group_is_overloaded return
6705 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6707 if (sgs->sum_nr_running <= sgs->group_weight)
6710 if ((sgs->group_capacity * 100) <
6711 (sgs->group_util * env->sd->imbalance_pct))
6718 group_type group_classify(struct sched_group *group,
6719 struct sg_lb_stats *sgs)
6721 if (sgs->group_no_capacity)
6722 return group_overloaded;
6724 if (sg_imbalanced(group))
6725 return group_imbalanced;
6731 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6732 * @env: The load balancing environment.
6733 * @group: sched_group whose statistics are to be updated.
6734 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6735 * @local_group: Does group contain this_cpu.
6736 * @sgs: variable to hold the statistics for this group.
6737 * @overload: Indicate more than one runnable task for any CPU.
6739 static inline void update_sg_lb_stats(struct lb_env *env,
6740 struct sched_group *group, int load_idx,
6741 int local_group, struct sg_lb_stats *sgs,
6747 memset(sgs, 0, sizeof(*sgs));
6749 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6750 struct rq *rq = cpu_rq(i);
6752 /* Bias balancing toward cpus of our domain */
6754 load = target_load(i, load_idx);
6756 load = source_load(i, load_idx);
6758 sgs->group_load += load;
6759 sgs->group_util += cpu_util(i);
6760 sgs->sum_nr_running += rq->cfs.h_nr_running;
6762 nr_running = rq->nr_running;
6766 #ifdef CONFIG_NUMA_BALANCING
6767 sgs->nr_numa_running += rq->nr_numa_running;
6768 sgs->nr_preferred_running += rq->nr_preferred_running;
6770 sgs->sum_weighted_load += weighted_cpuload(i);
6772 * No need to call idle_cpu() if nr_running is not 0
6774 if (!nr_running && idle_cpu(i))
6778 /* Adjust by relative CPU capacity of the group */
6779 sgs->group_capacity = group->sgc->capacity;
6780 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6782 if (sgs->sum_nr_running)
6783 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6785 sgs->group_weight = group->group_weight;
6787 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6788 sgs->group_type = group_classify(group, sgs);
6792 * update_sd_pick_busiest - return 1 on busiest group
6793 * @env: The load balancing environment.
6794 * @sds: sched_domain statistics
6795 * @sg: sched_group candidate to be checked for being the busiest
6796 * @sgs: sched_group statistics
6798 * Determine if @sg is a busier group than the previously selected
6801 * Return: %true if @sg is a busier group than the previously selected
6802 * busiest group. %false otherwise.
6804 static bool update_sd_pick_busiest(struct lb_env *env,
6805 struct sd_lb_stats *sds,
6806 struct sched_group *sg,
6807 struct sg_lb_stats *sgs)
6809 struct sg_lb_stats *busiest = &sds->busiest_stat;
6811 if (sgs->group_type > busiest->group_type)
6814 if (sgs->group_type < busiest->group_type)
6817 if (sgs->avg_load <= busiest->avg_load)
6820 /* This is the busiest node in its class. */
6821 if (!(env->sd->flags & SD_ASYM_PACKING))
6824 /* No ASYM_PACKING if target cpu is already busy */
6825 if (env->idle == CPU_NOT_IDLE)
6828 * ASYM_PACKING needs to move all the work to the lowest
6829 * numbered CPUs in the group, therefore mark all groups
6830 * higher than ourself as busy.
6832 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6836 /* Prefer to move from highest possible cpu's work */
6837 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6844 #ifdef CONFIG_NUMA_BALANCING
6845 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6847 if (sgs->sum_nr_running > sgs->nr_numa_running)
6849 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6854 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6856 if (rq->nr_running > rq->nr_numa_running)
6858 if (rq->nr_running > rq->nr_preferred_running)
6863 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6868 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6872 #endif /* CONFIG_NUMA_BALANCING */
6875 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6876 * @env: The load balancing environment.
6877 * @sds: variable to hold the statistics for this sched_domain.
6879 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6881 struct sched_domain *child = env->sd->child;
6882 struct sched_group *sg = env->sd->groups;
6883 struct sg_lb_stats tmp_sgs;
6884 int load_idx, prefer_sibling = 0;
6885 bool overload = false;
6887 if (child && child->flags & SD_PREFER_SIBLING)
6890 load_idx = get_sd_load_idx(env->sd, env->idle);
6893 struct sg_lb_stats *sgs = &tmp_sgs;
6896 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6899 sgs = &sds->local_stat;
6901 if (env->idle != CPU_NEWLY_IDLE ||
6902 time_after_eq(jiffies, sg->sgc->next_update))
6903 update_group_capacity(env->sd, env->dst_cpu);
6906 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6913 * In case the child domain prefers tasks go to siblings
6914 * first, lower the sg capacity so that we'll try
6915 * and move all the excess tasks away. We lower the capacity
6916 * of a group only if the local group has the capacity to fit
6917 * these excess tasks. The extra check prevents the case where
6918 * you always pull from the heaviest group when it is already
6919 * under-utilized (possible with a large weight task outweighs
6920 * the tasks on the system).
6922 if (prefer_sibling && sds->local &&
6923 group_has_capacity(env, &sds->local_stat) &&
6924 (sgs->sum_nr_running > 1)) {
6925 sgs->group_no_capacity = 1;
6926 sgs->group_type = group_classify(sg, sgs);
6929 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6931 sds->busiest_stat = *sgs;
6935 /* Now, start updating sd_lb_stats */
6936 sds->total_load += sgs->group_load;
6937 sds->total_capacity += sgs->group_capacity;
6940 } while (sg != env->sd->groups);
6942 if (env->sd->flags & SD_NUMA)
6943 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6945 if (!env->sd->parent) {
6946 /* update overload indicator if we are at root domain */
6947 if (env->dst_rq->rd->overload != overload)
6948 env->dst_rq->rd->overload = overload;
6954 * check_asym_packing - Check to see if the group is packed into the
6957 * This is primarily intended to used at the sibling level. Some
6958 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6959 * case of POWER7, it can move to lower SMT modes only when higher
6960 * threads are idle. When in lower SMT modes, the threads will
6961 * perform better since they share less core resources. Hence when we
6962 * have idle threads, we want them to be the higher ones.
6964 * This packing function is run on idle threads. It checks to see if
6965 * the busiest CPU in this domain (core in the P7 case) has a higher
6966 * CPU number than the packing function is being run on. Here we are
6967 * assuming lower CPU number will be equivalent to lower a SMT thread
6970 * Return: 1 when packing is required and a task should be moved to
6971 * this CPU. The amount of the imbalance is returned in *imbalance.
6973 * @env: The load balancing environment.
6974 * @sds: Statistics of the sched_domain which is to be packed
6976 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6980 if (!(env->sd->flags & SD_ASYM_PACKING))
6983 if (env->idle == CPU_NOT_IDLE)
6989 busiest_cpu = group_first_cpu(sds->busiest);
6990 if (env->dst_cpu > busiest_cpu)
6993 env->imbalance = DIV_ROUND_CLOSEST(
6994 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6995 SCHED_CAPACITY_SCALE);
7001 * fix_small_imbalance - Calculate the minor imbalance that exists
7002 * amongst the groups of a sched_domain, during
7004 * @env: The load balancing environment.
7005 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7008 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7010 unsigned long tmp, capa_now = 0, capa_move = 0;
7011 unsigned int imbn = 2;
7012 unsigned long scaled_busy_load_per_task;
7013 struct sg_lb_stats *local, *busiest;
7015 local = &sds->local_stat;
7016 busiest = &sds->busiest_stat;
7018 if (!local->sum_nr_running)
7019 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7020 else if (busiest->load_per_task > local->load_per_task)
7023 scaled_busy_load_per_task =
7024 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7025 busiest->group_capacity;
7027 if (busiest->avg_load + scaled_busy_load_per_task >=
7028 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7029 env->imbalance = busiest->load_per_task;
7034 * OK, we don't have enough imbalance to justify moving tasks,
7035 * however we may be able to increase total CPU capacity used by
7039 capa_now += busiest->group_capacity *
7040 min(busiest->load_per_task, busiest->avg_load);
7041 capa_now += local->group_capacity *
7042 min(local->load_per_task, local->avg_load);
7043 capa_now /= SCHED_CAPACITY_SCALE;
7045 /* Amount of load we'd subtract */
7046 if (busiest->avg_load > scaled_busy_load_per_task) {
7047 capa_move += busiest->group_capacity *
7048 min(busiest->load_per_task,
7049 busiest->avg_load - scaled_busy_load_per_task);
7052 /* Amount of load we'd add */
7053 if (busiest->avg_load * busiest->group_capacity <
7054 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7055 tmp = (busiest->avg_load * busiest->group_capacity) /
7056 local->group_capacity;
7058 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7059 local->group_capacity;
7061 capa_move += local->group_capacity *
7062 min(local->load_per_task, local->avg_load + tmp);
7063 capa_move /= SCHED_CAPACITY_SCALE;
7065 /* Move if we gain throughput */
7066 if (capa_move > capa_now)
7067 env->imbalance = busiest->load_per_task;
7071 * calculate_imbalance - Calculate the amount of imbalance present within the
7072 * groups of a given sched_domain during load balance.
7073 * @env: load balance environment
7074 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7076 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7078 unsigned long max_pull, load_above_capacity = ~0UL;
7079 struct sg_lb_stats *local, *busiest;
7081 local = &sds->local_stat;
7082 busiest = &sds->busiest_stat;
7084 if (busiest->group_type == group_imbalanced) {
7086 * In the group_imb case we cannot rely on group-wide averages
7087 * to ensure cpu-load equilibrium, look at wider averages. XXX
7089 busiest->load_per_task =
7090 min(busiest->load_per_task, sds->avg_load);
7094 * Avg load of busiest sg can be less and avg load of local sg can
7095 * be greater than avg load across all sgs of sd because avg load
7096 * factors in sg capacity and sgs with smaller group_type are
7097 * skipped when updating the busiest sg:
7099 if (busiest->avg_load <= sds->avg_load ||
7100 local->avg_load >= sds->avg_load) {
7102 return fix_small_imbalance(env, sds);
7106 * If there aren't any idle cpus, avoid creating some.
7108 if (busiest->group_type == group_overloaded &&
7109 local->group_type == group_overloaded) {
7110 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7111 if (load_above_capacity > busiest->group_capacity) {
7112 load_above_capacity -= busiest->group_capacity;
7113 load_above_capacity *= NICE_0_LOAD;
7114 load_above_capacity /= busiest->group_capacity;
7116 load_above_capacity = ~0UL;
7120 * We're trying to get all the cpus to the average_load, so we don't
7121 * want to push ourselves above the average load, nor do we wish to
7122 * reduce the max loaded cpu below the average load. At the same time,
7123 * we also don't want to reduce the group load below the group
7124 * capacity. Thus we look for the minimum possible imbalance.
7126 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7128 /* How much load to actually move to equalise the imbalance */
7129 env->imbalance = min(
7130 max_pull * busiest->group_capacity,
7131 (sds->avg_load - local->avg_load) * local->group_capacity
7132 ) / SCHED_CAPACITY_SCALE;
7135 * if *imbalance is less than the average load per runnable task
7136 * there is no guarantee that any tasks will be moved so we'll have
7137 * a think about bumping its value to force at least one task to be
7140 if (env->imbalance < busiest->load_per_task)
7141 return fix_small_imbalance(env, sds);
7144 /******* find_busiest_group() helpers end here *********************/
7147 * find_busiest_group - Returns the busiest group within the sched_domain
7148 * if there is an imbalance.
7150 * Also calculates the amount of weighted load which should be moved
7151 * to restore balance.
7153 * @env: The load balancing environment.
7155 * Return: - The busiest group if imbalance exists.
7157 static struct sched_group *find_busiest_group(struct lb_env *env)
7159 struct sg_lb_stats *local, *busiest;
7160 struct sd_lb_stats sds;
7162 init_sd_lb_stats(&sds);
7165 * Compute the various statistics relavent for load balancing at
7168 update_sd_lb_stats(env, &sds);
7169 local = &sds.local_stat;
7170 busiest = &sds.busiest_stat;
7172 /* ASYM feature bypasses nice load balance check */
7173 if (check_asym_packing(env, &sds))
7176 /* There is no busy sibling group to pull tasks from */
7177 if (!sds.busiest || busiest->sum_nr_running == 0)
7180 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7181 / sds.total_capacity;
7184 * If the busiest group is imbalanced the below checks don't
7185 * work because they assume all things are equal, which typically
7186 * isn't true due to cpus_allowed constraints and the like.
7188 if (busiest->group_type == group_imbalanced)
7191 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7192 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7193 busiest->group_no_capacity)
7197 * If the local group is busier than the selected busiest group
7198 * don't try and pull any tasks.
7200 if (local->avg_load >= busiest->avg_load)
7204 * Don't pull any tasks if this group is already above the domain
7207 if (local->avg_load >= sds.avg_load)
7210 if (env->idle == CPU_IDLE) {
7212 * This cpu is idle. If the busiest group is not overloaded
7213 * and there is no imbalance between this and busiest group
7214 * wrt idle cpus, it is balanced. The imbalance becomes
7215 * significant if the diff is greater than 1 otherwise we
7216 * might end up to just move the imbalance on another group
7218 if ((busiest->group_type != group_overloaded) &&
7219 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7223 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7224 * imbalance_pct to be conservative.
7226 if (100 * busiest->avg_load <=
7227 env->sd->imbalance_pct * local->avg_load)
7232 /* Looks like there is an imbalance. Compute it */
7233 calculate_imbalance(env, &sds);
7242 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7244 static struct rq *find_busiest_queue(struct lb_env *env,
7245 struct sched_group *group)
7247 struct rq *busiest = NULL, *rq;
7248 unsigned long busiest_load = 0, busiest_capacity = 1;
7251 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7252 unsigned long capacity, wl;
7256 rt = fbq_classify_rq(rq);
7259 * We classify groups/runqueues into three groups:
7260 * - regular: there are !numa tasks
7261 * - remote: there are numa tasks that run on the 'wrong' node
7262 * - all: there is no distinction
7264 * In order to avoid migrating ideally placed numa tasks,
7265 * ignore those when there's better options.
7267 * If we ignore the actual busiest queue to migrate another
7268 * task, the next balance pass can still reduce the busiest
7269 * queue by moving tasks around inside the node.
7271 * If we cannot move enough load due to this classification
7272 * the next pass will adjust the group classification and
7273 * allow migration of more tasks.
7275 * Both cases only affect the total convergence complexity.
7277 if (rt > env->fbq_type)
7280 capacity = capacity_of(i);
7282 wl = weighted_cpuload(i);
7285 * When comparing with imbalance, use weighted_cpuload()
7286 * which is not scaled with the cpu capacity.
7289 if (rq->nr_running == 1 && wl > env->imbalance &&
7290 !check_cpu_capacity(rq, env->sd))
7294 * For the load comparisons with the other cpu's, consider
7295 * the weighted_cpuload() scaled with the cpu capacity, so
7296 * that the load can be moved away from the cpu that is
7297 * potentially running at a lower capacity.
7299 * Thus we're looking for max(wl_i / capacity_i), crosswise
7300 * multiplication to rid ourselves of the division works out
7301 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7302 * our previous maximum.
7304 if (wl * busiest_capacity > busiest_load * capacity) {
7306 busiest_capacity = capacity;
7315 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7316 * so long as it is large enough.
7318 #define MAX_PINNED_INTERVAL 512
7320 /* Working cpumask for load_balance and load_balance_newidle. */
7321 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7323 static int need_active_balance(struct lb_env *env)
7325 struct sched_domain *sd = env->sd;
7327 if (env->idle == CPU_NEWLY_IDLE) {
7330 * ASYM_PACKING needs to force migrate tasks from busy but
7331 * higher numbered CPUs in order to pack all tasks in the
7332 * lowest numbered CPUs.
7334 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7339 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7340 * It's worth migrating the task if the src_cpu's capacity is reduced
7341 * because of other sched_class or IRQs if more capacity stays
7342 * available on dst_cpu.
7344 if ((env->idle != CPU_NOT_IDLE) &&
7345 (env->src_rq->cfs.h_nr_running == 1)) {
7346 if ((check_cpu_capacity(env->src_rq, sd)) &&
7347 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7351 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7354 static int active_load_balance_cpu_stop(void *data);
7356 static int should_we_balance(struct lb_env *env)
7358 struct sched_group *sg = env->sd->groups;
7359 struct cpumask *sg_cpus, *sg_mask;
7360 int cpu, balance_cpu = -1;
7363 * In the newly idle case, we will allow all the cpu's
7364 * to do the newly idle load balance.
7366 if (env->idle == CPU_NEWLY_IDLE)
7369 sg_cpus = sched_group_cpus(sg);
7370 sg_mask = sched_group_mask(sg);
7371 /* Try to find first idle cpu */
7372 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7373 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7380 if (balance_cpu == -1)
7381 balance_cpu = group_balance_cpu(sg);
7384 * First idle cpu or the first cpu(busiest) in this sched group
7385 * is eligible for doing load balancing at this and above domains.
7387 return balance_cpu == env->dst_cpu;
7391 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7392 * tasks if there is an imbalance.
7394 static int load_balance(int this_cpu, struct rq *this_rq,
7395 struct sched_domain *sd, enum cpu_idle_type idle,
7396 int *continue_balancing)
7398 int ld_moved, cur_ld_moved, active_balance = 0;
7399 struct sched_domain *sd_parent = sd->parent;
7400 struct sched_group *group;
7402 unsigned long flags;
7403 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7405 struct lb_env env = {
7407 .dst_cpu = this_cpu,
7409 .dst_grpmask = sched_group_cpus(sd->groups),
7411 .loop_break = sched_nr_migrate_break,
7414 .tasks = LIST_HEAD_INIT(env.tasks),
7418 * For NEWLY_IDLE load_balancing, we don't need to consider
7419 * other cpus in our group
7421 if (idle == CPU_NEWLY_IDLE)
7422 env.dst_grpmask = NULL;
7424 cpumask_copy(cpus, cpu_active_mask);
7426 schedstat_inc(sd, lb_count[idle]);
7429 if (!should_we_balance(&env)) {
7430 *continue_balancing = 0;
7434 group = find_busiest_group(&env);
7436 schedstat_inc(sd, lb_nobusyg[idle]);
7440 busiest = find_busiest_queue(&env, group);
7442 schedstat_inc(sd, lb_nobusyq[idle]);
7446 BUG_ON(busiest == env.dst_rq);
7448 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7450 env.src_cpu = busiest->cpu;
7451 env.src_rq = busiest;
7454 if (busiest->nr_running > 1) {
7456 * Attempt to move tasks. If find_busiest_group has found
7457 * an imbalance but busiest->nr_running <= 1, the group is
7458 * still unbalanced. ld_moved simply stays zero, so it is
7459 * correctly treated as an imbalance.
7461 env.flags |= LBF_ALL_PINNED;
7462 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7465 raw_spin_lock_irqsave(&busiest->lock, flags);
7468 * cur_ld_moved - load moved in current iteration
7469 * ld_moved - cumulative load moved across iterations
7471 cur_ld_moved = detach_tasks(&env);
7474 * We've detached some tasks from busiest_rq. Every
7475 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7476 * unlock busiest->lock, and we are able to be sure
7477 * that nobody can manipulate the tasks in parallel.
7478 * See task_rq_lock() family for the details.
7481 raw_spin_unlock(&busiest->lock);
7485 ld_moved += cur_ld_moved;
7488 local_irq_restore(flags);
7490 if (env.flags & LBF_NEED_BREAK) {
7491 env.flags &= ~LBF_NEED_BREAK;
7496 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7497 * us and move them to an alternate dst_cpu in our sched_group
7498 * where they can run. The upper limit on how many times we
7499 * iterate on same src_cpu is dependent on number of cpus in our
7502 * This changes load balance semantics a bit on who can move
7503 * load to a given_cpu. In addition to the given_cpu itself
7504 * (or a ilb_cpu acting on its behalf where given_cpu is
7505 * nohz-idle), we now have balance_cpu in a position to move
7506 * load to given_cpu. In rare situations, this may cause
7507 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7508 * _independently_ and at _same_ time to move some load to
7509 * given_cpu) causing exceess load to be moved to given_cpu.
7510 * This however should not happen so much in practice and
7511 * moreover subsequent load balance cycles should correct the
7512 * excess load moved.
7514 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7516 /* Prevent to re-select dst_cpu via env's cpus */
7517 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7519 env.dst_rq = cpu_rq(env.new_dst_cpu);
7520 env.dst_cpu = env.new_dst_cpu;
7521 env.flags &= ~LBF_DST_PINNED;
7523 env.loop_break = sched_nr_migrate_break;
7526 * Go back to "more_balance" rather than "redo" since we
7527 * need to continue with same src_cpu.
7533 * We failed to reach balance because of affinity.
7536 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7538 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7539 *group_imbalance = 1;
7542 /* All tasks on this runqueue were pinned by CPU affinity */
7543 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7544 cpumask_clear_cpu(cpu_of(busiest), cpus);
7545 if (!cpumask_empty(cpus)) {
7547 env.loop_break = sched_nr_migrate_break;
7550 goto out_all_pinned;
7555 schedstat_inc(sd, lb_failed[idle]);
7557 * Increment the failure counter only on periodic balance.
7558 * We do not want newidle balance, which can be very
7559 * frequent, pollute the failure counter causing
7560 * excessive cache_hot migrations and active balances.
7562 if (idle != CPU_NEWLY_IDLE)
7563 sd->nr_balance_failed++;
7565 if (need_active_balance(&env)) {
7566 raw_spin_lock_irqsave(&busiest->lock, flags);
7568 /* don't kick the active_load_balance_cpu_stop,
7569 * if the curr task on busiest cpu can't be
7572 if (!cpumask_test_cpu(this_cpu,
7573 tsk_cpus_allowed(busiest->curr))) {
7574 raw_spin_unlock_irqrestore(&busiest->lock,
7576 env.flags |= LBF_ALL_PINNED;
7577 goto out_one_pinned;
7581 * ->active_balance synchronizes accesses to
7582 * ->active_balance_work. Once set, it's cleared
7583 * only after active load balance is finished.
7585 if (!busiest->active_balance) {
7586 busiest->active_balance = 1;
7587 busiest->push_cpu = this_cpu;
7590 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7592 if (active_balance) {
7593 stop_one_cpu_nowait(cpu_of(busiest),
7594 active_load_balance_cpu_stop, busiest,
7595 &busiest->active_balance_work);
7598 /* We've kicked active balancing, force task migration. */
7599 sd->nr_balance_failed = sd->cache_nice_tries+1;
7602 sd->nr_balance_failed = 0;
7604 if (likely(!active_balance)) {
7605 /* We were unbalanced, so reset the balancing interval */
7606 sd->balance_interval = sd->min_interval;
7609 * If we've begun active balancing, start to back off. This
7610 * case may not be covered by the all_pinned logic if there
7611 * is only 1 task on the busy runqueue (because we don't call
7614 if (sd->balance_interval < sd->max_interval)
7615 sd->balance_interval *= 2;
7622 * We reach balance although we may have faced some affinity
7623 * constraints. Clear the imbalance flag if it was set.
7626 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7628 if (*group_imbalance)
7629 *group_imbalance = 0;
7634 * We reach balance because all tasks are pinned at this level so
7635 * we can't migrate them. Let the imbalance flag set so parent level
7636 * can try to migrate them.
7638 schedstat_inc(sd, lb_balanced[idle]);
7640 sd->nr_balance_failed = 0;
7643 /* tune up the balancing interval */
7644 if (((env.flags & LBF_ALL_PINNED) &&
7645 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7646 (sd->balance_interval < sd->max_interval))
7647 sd->balance_interval *= 2;
7654 static inline unsigned long
7655 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7657 unsigned long interval = sd->balance_interval;
7660 interval *= sd->busy_factor;
7662 /* scale ms to jiffies */
7663 interval = msecs_to_jiffies(interval);
7664 interval = clamp(interval, 1UL, max_load_balance_interval);
7670 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7672 unsigned long interval, next;
7674 interval = get_sd_balance_interval(sd, cpu_busy);
7675 next = sd->last_balance + interval;
7677 if (time_after(*next_balance, next))
7678 *next_balance = next;
7682 * idle_balance is called by schedule() if this_cpu is about to become
7683 * idle. Attempts to pull tasks from other CPUs.
7685 static int idle_balance(struct rq *this_rq)
7687 unsigned long next_balance = jiffies + HZ;
7688 int this_cpu = this_rq->cpu;
7689 struct sched_domain *sd;
7690 int pulled_task = 0;
7694 * We must set idle_stamp _before_ calling idle_balance(), such that we
7695 * measure the duration of idle_balance() as idle time.
7697 this_rq->idle_stamp = rq_clock(this_rq);
7699 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7700 !this_rq->rd->overload) {
7702 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7704 update_next_balance(sd, 0, &next_balance);
7710 raw_spin_unlock(&this_rq->lock);
7712 update_blocked_averages(this_cpu);
7714 for_each_domain(this_cpu, sd) {
7715 int continue_balancing = 1;
7716 u64 t0, domain_cost;
7718 if (!(sd->flags & SD_LOAD_BALANCE))
7721 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7722 update_next_balance(sd, 0, &next_balance);
7726 if (sd->flags & SD_BALANCE_NEWIDLE) {
7727 t0 = sched_clock_cpu(this_cpu);
7729 pulled_task = load_balance(this_cpu, this_rq,
7731 &continue_balancing);
7733 domain_cost = sched_clock_cpu(this_cpu) - t0;
7734 if (domain_cost > sd->max_newidle_lb_cost)
7735 sd->max_newidle_lb_cost = domain_cost;
7737 curr_cost += domain_cost;
7740 update_next_balance(sd, 0, &next_balance);
7743 * Stop searching for tasks to pull if there are
7744 * now runnable tasks on this rq.
7746 if (pulled_task || this_rq->nr_running > 0)
7751 raw_spin_lock(&this_rq->lock);
7753 if (curr_cost > this_rq->max_idle_balance_cost)
7754 this_rq->max_idle_balance_cost = curr_cost;
7757 * While browsing the domains, we released the rq lock, a task could
7758 * have been enqueued in the meantime. Since we're not going idle,
7759 * pretend we pulled a task.
7761 if (this_rq->cfs.h_nr_running && !pulled_task)
7765 /* Move the next balance forward */
7766 if (time_after(this_rq->next_balance, next_balance))
7767 this_rq->next_balance = next_balance;
7769 /* Is there a task of a high priority class? */
7770 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7774 this_rq->idle_stamp = 0;
7780 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7781 * running tasks off the busiest CPU onto idle CPUs. It requires at
7782 * least 1 task to be running on each physical CPU where possible, and
7783 * avoids physical / logical imbalances.
7785 static int active_load_balance_cpu_stop(void *data)
7787 struct rq *busiest_rq = data;
7788 int busiest_cpu = cpu_of(busiest_rq);
7789 int target_cpu = busiest_rq->push_cpu;
7790 struct rq *target_rq = cpu_rq(target_cpu);
7791 struct sched_domain *sd;
7792 struct task_struct *p = NULL;
7794 raw_spin_lock_irq(&busiest_rq->lock);
7796 /* make sure the requested cpu hasn't gone down in the meantime */
7797 if (unlikely(busiest_cpu != smp_processor_id() ||
7798 !busiest_rq->active_balance))
7801 /* Is there any task to move? */
7802 if (busiest_rq->nr_running <= 1)
7806 * This condition is "impossible", if it occurs
7807 * we need to fix it. Originally reported by
7808 * Bjorn Helgaas on a 128-cpu setup.
7810 BUG_ON(busiest_rq == target_rq);
7812 /* Search for an sd spanning us and the target CPU. */
7814 for_each_domain(target_cpu, sd) {
7815 if ((sd->flags & SD_LOAD_BALANCE) &&
7816 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7821 struct lb_env env = {
7823 .dst_cpu = target_cpu,
7824 .dst_rq = target_rq,
7825 .src_cpu = busiest_rq->cpu,
7826 .src_rq = busiest_rq,
7830 schedstat_inc(sd, alb_count);
7832 p = detach_one_task(&env);
7834 schedstat_inc(sd, alb_pushed);
7835 /* Active balancing done, reset the failure counter. */
7836 sd->nr_balance_failed = 0;
7838 schedstat_inc(sd, alb_failed);
7843 busiest_rq->active_balance = 0;
7844 raw_spin_unlock(&busiest_rq->lock);
7847 attach_one_task(target_rq, p);
7854 static inline int on_null_domain(struct rq *rq)
7856 return unlikely(!rcu_dereference_sched(rq->sd));
7859 #ifdef CONFIG_NO_HZ_COMMON
7861 * idle load balancing details
7862 * - When one of the busy CPUs notice that there may be an idle rebalancing
7863 * needed, they will kick the idle load balancer, which then does idle
7864 * load balancing for all the idle CPUs.
7867 cpumask_var_t idle_cpus_mask;
7869 unsigned long next_balance; /* in jiffy units */
7870 } nohz ____cacheline_aligned;
7872 static inline int find_new_ilb(void)
7874 int ilb = cpumask_first(nohz.idle_cpus_mask);
7876 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7883 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7884 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7885 * CPU (if there is one).
7887 static void nohz_balancer_kick(void)
7891 nohz.next_balance++;
7893 ilb_cpu = find_new_ilb();
7895 if (ilb_cpu >= nr_cpu_ids)
7898 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7901 * Use smp_send_reschedule() instead of resched_cpu().
7902 * This way we generate a sched IPI on the target cpu which
7903 * is idle. And the softirq performing nohz idle load balance
7904 * will be run before returning from the IPI.
7906 smp_send_reschedule(ilb_cpu);
7910 void nohz_balance_exit_idle(unsigned int cpu)
7912 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7914 * Completely isolated CPUs don't ever set, so we must test.
7916 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7917 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7918 atomic_dec(&nohz.nr_cpus);
7920 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7924 static inline void set_cpu_sd_state_busy(void)
7926 struct sched_domain *sd;
7927 int cpu = smp_processor_id();
7930 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7932 if (!sd || !sd->nohz_idle)
7936 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7941 void set_cpu_sd_state_idle(void)
7943 struct sched_domain *sd;
7944 int cpu = smp_processor_id();
7947 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7949 if (!sd || sd->nohz_idle)
7953 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7959 * This routine will record that the cpu is going idle with tick stopped.
7960 * This info will be used in performing idle load balancing in the future.
7962 void nohz_balance_enter_idle(int cpu)
7965 * If this cpu is going down, then nothing needs to be done.
7967 if (!cpu_active(cpu))
7970 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7974 * If we're a completely isolated CPU, we don't play.
7976 if (on_null_domain(cpu_rq(cpu)))
7979 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7980 atomic_inc(&nohz.nr_cpus);
7981 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7985 static DEFINE_SPINLOCK(balancing);
7988 * Scale the max load_balance interval with the number of CPUs in the system.
7989 * This trades load-balance latency on larger machines for less cross talk.
7991 void update_max_interval(void)
7993 max_load_balance_interval = HZ*num_online_cpus()/10;
7997 * It checks each scheduling domain to see if it is due to be balanced,
7998 * and initiates a balancing operation if so.
8000 * Balancing parameters are set up in init_sched_domains.
8002 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8004 int continue_balancing = 1;
8006 unsigned long interval;
8007 struct sched_domain *sd;
8008 /* Earliest time when we have to do rebalance again */
8009 unsigned long next_balance = jiffies + 60*HZ;
8010 int update_next_balance = 0;
8011 int need_serialize, need_decay = 0;
8014 update_blocked_averages(cpu);
8017 for_each_domain(cpu, sd) {
8019 * Decay the newidle max times here because this is a regular
8020 * visit to all the domains. Decay ~1% per second.
8022 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8023 sd->max_newidle_lb_cost =
8024 (sd->max_newidle_lb_cost * 253) / 256;
8025 sd->next_decay_max_lb_cost = jiffies + HZ;
8028 max_cost += sd->max_newidle_lb_cost;
8030 if (!(sd->flags & SD_LOAD_BALANCE))
8034 * Stop the load balance at this level. There is another
8035 * CPU in our sched group which is doing load balancing more
8038 if (!continue_balancing) {
8044 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8046 need_serialize = sd->flags & SD_SERIALIZE;
8047 if (need_serialize) {
8048 if (!spin_trylock(&balancing))
8052 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8053 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8055 * The LBF_DST_PINNED logic could have changed
8056 * env->dst_cpu, so we can't know our idle
8057 * state even if we migrated tasks. Update it.
8059 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8061 sd->last_balance = jiffies;
8062 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8065 spin_unlock(&balancing);
8067 if (time_after(next_balance, sd->last_balance + interval)) {
8068 next_balance = sd->last_balance + interval;
8069 update_next_balance = 1;
8074 * Ensure the rq-wide value also decays but keep it at a
8075 * reasonable floor to avoid funnies with rq->avg_idle.
8077 rq->max_idle_balance_cost =
8078 max((u64)sysctl_sched_migration_cost, max_cost);
8083 * next_balance will be updated only when there is a need.
8084 * When the cpu is attached to null domain for ex, it will not be
8087 if (likely(update_next_balance)) {
8088 rq->next_balance = next_balance;
8090 #ifdef CONFIG_NO_HZ_COMMON
8092 * If this CPU has been elected to perform the nohz idle
8093 * balance. Other idle CPUs have already rebalanced with
8094 * nohz_idle_balance() and nohz.next_balance has been
8095 * updated accordingly. This CPU is now running the idle load
8096 * balance for itself and we need to update the
8097 * nohz.next_balance accordingly.
8099 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8100 nohz.next_balance = rq->next_balance;
8105 #ifdef CONFIG_NO_HZ_COMMON
8107 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8108 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8110 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8112 int this_cpu = this_rq->cpu;
8115 /* Earliest time when we have to do rebalance again */
8116 unsigned long next_balance = jiffies + 60*HZ;
8117 int update_next_balance = 0;
8119 if (idle != CPU_IDLE ||
8120 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8123 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8124 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8128 * If this cpu gets work to do, stop the load balancing
8129 * work being done for other cpus. Next load
8130 * balancing owner will pick it up.
8135 rq = cpu_rq(balance_cpu);
8138 * If time for next balance is due,
8141 if (time_after_eq(jiffies, rq->next_balance)) {
8142 raw_spin_lock_irq(&rq->lock);
8143 update_rq_clock(rq);
8144 cpu_load_update_idle(rq);
8145 raw_spin_unlock_irq(&rq->lock);
8146 rebalance_domains(rq, CPU_IDLE);
8149 if (time_after(next_balance, rq->next_balance)) {
8150 next_balance = rq->next_balance;
8151 update_next_balance = 1;
8156 * next_balance will be updated only when there is a need.
8157 * When the CPU is attached to null domain for ex, it will not be
8160 if (likely(update_next_balance))
8161 nohz.next_balance = next_balance;
8163 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8167 * Current heuristic for kicking the idle load balancer in the presence
8168 * of an idle cpu in the system.
8169 * - This rq has more than one task.
8170 * - This rq has at least one CFS task and the capacity of the CPU is
8171 * significantly reduced because of RT tasks or IRQs.
8172 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8173 * multiple busy cpu.
8174 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8175 * domain span are idle.
8177 static inline bool nohz_kick_needed(struct rq *rq)
8179 unsigned long now = jiffies;
8180 struct sched_domain *sd;
8181 struct sched_group_capacity *sgc;
8182 int nr_busy, cpu = rq->cpu;
8185 if (unlikely(rq->idle_balance))
8189 * We may be recently in ticked or tickless idle mode. At the first
8190 * busy tick after returning from idle, we will update the busy stats.
8192 set_cpu_sd_state_busy();
8193 nohz_balance_exit_idle(cpu);
8196 * None are in tickless mode and hence no need for NOHZ idle load
8199 if (likely(!atomic_read(&nohz.nr_cpus)))
8202 if (time_before(now, nohz.next_balance))
8205 if (rq->nr_running >= 2)
8209 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8211 sgc = sd->groups->sgc;
8212 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8221 sd = rcu_dereference(rq->sd);
8223 if ((rq->cfs.h_nr_running >= 1) &&
8224 check_cpu_capacity(rq, sd)) {
8230 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8231 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8232 sched_domain_span(sd)) < cpu)) {
8242 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8246 * run_rebalance_domains is triggered when needed from the scheduler tick.
8247 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8249 static void run_rebalance_domains(struct softirq_action *h)
8251 struct rq *this_rq = this_rq();
8252 enum cpu_idle_type idle = this_rq->idle_balance ?
8253 CPU_IDLE : CPU_NOT_IDLE;
8256 * If this cpu has a pending nohz_balance_kick, then do the
8257 * balancing on behalf of the other idle cpus whose ticks are
8258 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8259 * give the idle cpus a chance to load balance. Else we may
8260 * load balance only within the local sched_domain hierarchy
8261 * and abort nohz_idle_balance altogether if we pull some load.
8263 nohz_idle_balance(this_rq, idle);
8264 rebalance_domains(this_rq, idle);
8268 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8270 void trigger_load_balance(struct rq *rq)
8272 /* Don't need to rebalance while attached to NULL domain */
8273 if (unlikely(on_null_domain(rq)))
8276 if (time_after_eq(jiffies, rq->next_balance))
8277 raise_softirq(SCHED_SOFTIRQ);
8278 #ifdef CONFIG_NO_HZ_COMMON
8279 if (nohz_kick_needed(rq))
8280 nohz_balancer_kick();
8284 static void rq_online_fair(struct rq *rq)
8288 update_runtime_enabled(rq);
8291 static void rq_offline_fair(struct rq *rq)
8295 /* Ensure any throttled groups are reachable by pick_next_task */
8296 unthrottle_offline_cfs_rqs(rq);
8299 #endif /* CONFIG_SMP */
8302 * scheduler tick hitting a task of our scheduling class:
8304 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8306 struct cfs_rq *cfs_rq;
8307 struct sched_entity *se = &curr->se;
8309 for_each_sched_entity(se) {
8310 cfs_rq = cfs_rq_of(se);
8311 entity_tick(cfs_rq, se, queued);
8314 if (static_branch_unlikely(&sched_numa_balancing))
8315 task_tick_numa(rq, curr);
8319 * called on fork with the child task as argument from the parent's context
8320 * - child not yet on the tasklist
8321 * - preemption disabled
8323 static void task_fork_fair(struct task_struct *p)
8325 struct cfs_rq *cfs_rq;
8326 struct sched_entity *se = &p->se, *curr;
8327 struct rq *rq = this_rq();
8329 raw_spin_lock(&rq->lock);
8330 update_rq_clock(rq);
8332 cfs_rq = task_cfs_rq(current);
8333 curr = cfs_rq->curr;
8335 update_curr(cfs_rq);
8336 se->vruntime = curr->vruntime;
8338 place_entity(cfs_rq, se, 1);
8340 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8342 * Upon rescheduling, sched_class::put_prev_task() will place
8343 * 'current' within the tree based on its new key value.
8345 swap(curr->vruntime, se->vruntime);
8349 se->vruntime -= cfs_rq->min_vruntime;
8350 raw_spin_unlock(&rq->lock);
8354 * Priority of the task has changed. Check to see if we preempt
8358 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8360 if (!task_on_rq_queued(p))
8364 * Reschedule if we are currently running on this runqueue and
8365 * our priority decreased, or if we are not currently running on
8366 * this runqueue and our priority is higher than the current's
8368 if (rq->curr == p) {
8369 if (p->prio > oldprio)
8372 check_preempt_curr(rq, p, 0);
8375 static inline bool vruntime_normalized(struct task_struct *p)
8377 struct sched_entity *se = &p->se;
8380 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8381 * the dequeue_entity(.flags=0) will already have normalized the
8388 * When !on_rq, vruntime of the task has usually NOT been normalized.
8389 * But there are some cases where it has already been normalized:
8391 * - A forked child which is waiting for being woken up by
8392 * wake_up_new_task().
8393 * - A task which has been woken up by try_to_wake_up() and
8394 * waiting for actually being woken up by sched_ttwu_pending().
8396 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8402 static void detach_task_cfs_rq(struct task_struct *p)
8404 struct sched_entity *se = &p->se;
8405 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8406 u64 now = cfs_rq_clock_task(cfs_rq);
8408 if (!vruntime_normalized(p)) {
8410 * Fix up our vruntime so that the current sleep doesn't
8411 * cause 'unlimited' sleep bonus.
8413 place_entity(cfs_rq, se, 0);
8414 se->vruntime -= cfs_rq->min_vruntime;
8417 /* Catch up with the cfs_rq and remove our load when we leave */
8418 update_cfs_rq_load_avg(now, cfs_rq, false);
8419 detach_entity_load_avg(cfs_rq, se);
8422 static void attach_task_cfs_rq(struct task_struct *p)
8424 struct sched_entity *se = &p->se;
8425 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8426 u64 now = cfs_rq_clock_task(cfs_rq);
8428 #ifdef CONFIG_FAIR_GROUP_SCHED
8430 * Since the real-depth could have been changed (only FAIR
8431 * class maintain depth value), reset depth properly.
8433 se->depth = se->parent ? se->parent->depth + 1 : 0;
8436 /* Synchronize task with its cfs_rq */
8437 update_cfs_rq_load_avg(now, cfs_rq, false);
8438 attach_entity_load_avg(cfs_rq, se);
8440 if (!vruntime_normalized(p))
8441 se->vruntime += cfs_rq->min_vruntime;
8444 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8446 detach_task_cfs_rq(p);
8449 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8451 attach_task_cfs_rq(p);
8453 if (task_on_rq_queued(p)) {
8455 * We were most likely switched from sched_rt, so
8456 * kick off the schedule if running, otherwise just see
8457 * if we can still preempt the current task.
8462 check_preempt_curr(rq, p, 0);
8466 /* Account for a task changing its policy or group.
8468 * This routine is mostly called to set cfs_rq->curr field when a task
8469 * migrates between groups/classes.
8471 static void set_curr_task_fair(struct rq *rq)
8473 struct sched_entity *se = &rq->curr->se;
8475 for_each_sched_entity(se) {
8476 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8478 set_next_entity(cfs_rq, se);
8479 /* ensure bandwidth has been allocated on our new cfs_rq */
8480 account_cfs_rq_runtime(cfs_rq, 0);
8484 void init_cfs_rq(struct cfs_rq *cfs_rq)
8486 cfs_rq->tasks_timeline = RB_ROOT;
8487 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8488 #ifndef CONFIG_64BIT
8489 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8492 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8493 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8497 #ifdef CONFIG_FAIR_GROUP_SCHED
8498 static void task_set_group_fair(struct task_struct *p)
8500 struct sched_entity *se = &p->se;
8502 set_task_rq(p, task_cpu(p));
8503 se->depth = se->parent ? se->parent->depth + 1 : 0;
8506 static void task_move_group_fair(struct task_struct *p)
8508 detach_task_cfs_rq(p);
8509 set_task_rq(p, task_cpu(p));
8512 /* Tell se's cfs_rq has been changed -- migrated */
8513 p->se.avg.last_update_time = 0;
8515 attach_task_cfs_rq(p);
8518 static void task_change_group_fair(struct task_struct *p, int type)
8521 case TASK_SET_GROUP:
8522 task_set_group_fair(p);
8525 case TASK_MOVE_GROUP:
8526 task_move_group_fair(p);
8531 void free_fair_sched_group(struct task_group *tg)
8535 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8537 for_each_possible_cpu(i) {
8539 kfree(tg->cfs_rq[i]);
8548 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8550 struct sched_entity *se;
8551 struct cfs_rq *cfs_rq;
8555 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8558 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8562 tg->shares = NICE_0_LOAD;
8564 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8566 for_each_possible_cpu(i) {
8569 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8570 GFP_KERNEL, cpu_to_node(i));
8574 se = kzalloc_node(sizeof(struct sched_entity),
8575 GFP_KERNEL, cpu_to_node(i));
8579 init_cfs_rq(cfs_rq);
8580 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8581 init_entity_runnable_average(se);
8583 raw_spin_lock_irq(&rq->lock);
8584 post_init_entity_util_avg(se);
8585 raw_spin_unlock_irq(&rq->lock);
8596 void unregister_fair_sched_group(struct task_group *tg)
8598 unsigned long flags;
8602 for_each_possible_cpu(cpu) {
8604 remove_entity_load_avg(tg->se[cpu]);
8607 * Only empty task groups can be destroyed; so we can speculatively
8608 * check on_list without danger of it being re-added.
8610 if (!tg->cfs_rq[cpu]->on_list)
8615 raw_spin_lock_irqsave(&rq->lock, flags);
8616 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8617 raw_spin_unlock_irqrestore(&rq->lock, flags);
8621 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8622 struct sched_entity *se, int cpu,
8623 struct sched_entity *parent)
8625 struct rq *rq = cpu_rq(cpu);
8629 init_cfs_rq_runtime(cfs_rq);
8631 tg->cfs_rq[cpu] = cfs_rq;
8634 /* se could be NULL for root_task_group */
8639 se->cfs_rq = &rq->cfs;
8642 se->cfs_rq = parent->my_q;
8643 se->depth = parent->depth + 1;
8647 /* guarantee group entities always have weight */
8648 update_load_set(&se->load, NICE_0_LOAD);
8649 se->parent = parent;
8652 static DEFINE_MUTEX(shares_mutex);
8654 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8657 unsigned long flags;
8660 * We can't change the weight of the root cgroup.
8665 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8667 mutex_lock(&shares_mutex);
8668 if (tg->shares == shares)
8671 tg->shares = shares;
8672 for_each_possible_cpu(i) {
8673 struct rq *rq = cpu_rq(i);
8674 struct sched_entity *se;
8677 /* Propagate contribution to hierarchy */
8678 raw_spin_lock_irqsave(&rq->lock, flags);
8680 /* Possible calls to update_curr() need rq clock */
8681 update_rq_clock(rq);
8682 for_each_sched_entity(se)
8683 update_cfs_shares(group_cfs_rq(se));
8684 raw_spin_unlock_irqrestore(&rq->lock, flags);
8688 mutex_unlock(&shares_mutex);
8691 #else /* CONFIG_FAIR_GROUP_SCHED */
8693 void free_fair_sched_group(struct task_group *tg) { }
8695 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8700 void unregister_fair_sched_group(struct task_group *tg) { }
8702 #endif /* CONFIG_FAIR_GROUP_SCHED */
8705 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8707 struct sched_entity *se = &task->se;
8708 unsigned int rr_interval = 0;
8711 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8714 if (rq->cfs.load.weight)
8715 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8721 * All the scheduling class methods:
8723 const struct sched_class fair_sched_class = {
8724 .next = &idle_sched_class,
8725 .enqueue_task = enqueue_task_fair,
8726 .dequeue_task = dequeue_task_fair,
8727 .yield_task = yield_task_fair,
8728 .yield_to_task = yield_to_task_fair,
8730 .check_preempt_curr = check_preempt_wakeup,
8732 .pick_next_task = pick_next_task_fair,
8733 .put_prev_task = put_prev_task_fair,
8736 .select_task_rq = select_task_rq_fair,
8737 .migrate_task_rq = migrate_task_rq_fair,
8739 .rq_online = rq_online_fair,
8740 .rq_offline = rq_offline_fair,
8742 .task_dead = task_dead_fair,
8743 .set_cpus_allowed = set_cpus_allowed_common,
8746 .set_curr_task = set_curr_task_fair,
8747 .task_tick = task_tick_fair,
8748 .task_fork = task_fork_fair,
8750 .prio_changed = prio_changed_fair,
8751 .switched_from = switched_from_fair,
8752 .switched_to = switched_to_fair,
8754 .get_rr_interval = get_rr_interval_fair,
8756 .update_curr = update_curr_fair,
8758 #ifdef CONFIG_FAIR_GROUP_SCHED
8759 .task_change_group = task_change_group_fair,
8763 #ifdef CONFIG_SCHED_DEBUG
8764 void print_cfs_stats(struct seq_file *m, int cpu)
8766 struct cfs_rq *cfs_rq;
8769 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8770 print_cfs_rq(m, cpu, cfs_rq);
8774 #ifdef CONFIG_NUMA_BALANCING
8775 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8778 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8780 for_each_online_node(node) {
8781 if (p->numa_faults) {
8782 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8783 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8785 if (p->numa_group) {
8786 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8787 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8789 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8792 #endif /* CONFIG_NUMA_BALANCING */
8793 #endif /* CONFIG_SCHED_DEBUG */
8795 __init void init_sched_fair_class(void)
8798 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8800 #ifdef CONFIG_NO_HZ_COMMON
8801 nohz.next_balance = jiffies;
8802 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);