2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched.h>
24 #include <linux/latencytop.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 * At this point, util_avg won't be used in select_task_rq_fair anyway
690 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
693 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
694 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
695 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
696 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
699 * With new tasks being created, their initial util_avgs are extrapolated
700 * based on the cfs_rq's current util_avg:
702 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
704 * However, in many cases, the above util_avg does not give a desired
705 * value. Moreover, the sum of the util_avgs may be divergent, such
706 * as when the series is a harmonic series.
708 * To solve this problem, we also cap the util_avg of successive tasks to
709 * only 1/2 of the left utilization budget:
711 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
713 * where n denotes the nth task.
715 * For example, a simplest series from the beginning would be like:
717 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
718 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
720 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
721 * if util_avg > util_avg_cap.
723 void post_init_entity_util_avg(struct sched_entity *se)
725 struct cfs_rq *cfs_rq = cfs_rq_of(se);
726 struct sched_avg *sa = &se->avg;
727 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
728 u64 now = cfs_rq_clock_task(cfs_rq);
732 if (cfs_rq->avg.util_avg != 0) {
733 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
734 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
736 if (sa->util_avg > cap)
741 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
744 if (entity_is_task(se)) {
745 struct task_struct *p = task_of(se);
746 if (p->sched_class != &fair_sched_class) {
748 * For !fair tasks do:
750 update_cfs_rq_load_avg(now, cfs_rq, false);
751 attach_entity_load_avg(cfs_rq, se);
752 switched_from_fair(rq, p);
754 * such that the next switched_to_fair() has the
757 se->avg.last_update_time = now;
762 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
763 attach_entity_load_avg(cfs_rq, se);
765 update_tg_load_avg(cfs_rq, false);
768 #else /* !CONFIG_SMP */
769 void init_entity_runnable_average(struct sched_entity *se)
772 void post_init_entity_util_avg(struct sched_entity *se)
775 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
778 #endif /* CONFIG_SMP */
781 * Update the current task's runtime statistics.
783 static void update_curr(struct cfs_rq *cfs_rq)
785 struct sched_entity *curr = cfs_rq->curr;
786 u64 now = rq_clock_task(rq_of(cfs_rq));
792 delta_exec = now - curr->exec_start;
793 if (unlikely((s64)delta_exec <= 0))
796 curr->exec_start = now;
798 schedstat_set(curr->statistics.exec_max,
799 max(delta_exec, curr->statistics.exec_max));
801 curr->sum_exec_runtime += delta_exec;
802 schedstat_add(cfs_rq, exec_clock, delta_exec);
804 curr->vruntime += calc_delta_fair(delta_exec, curr);
805 update_min_vruntime(cfs_rq);
807 if (entity_is_task(curr)) {
808 struct task_struct *curtask = task_of(curr);
810 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
811 cpuacct_charge(curtask, delta_exec);
812 account_group_exec_runtime(curtask, delta_exec);
815 account_cfs_rq_runtime(cfs_rq, delta_exec);
818 static void update_curr_fair(struct rq *rq)
820 update_curr(cfs_rq_of(&rq->curr->se));
823 #ifdef CONFIG_SCHEDSTATS
825 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
827 u64 wait_start = rq_clock(rq_of(cfs_rq));
829 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
830 likely(wait_start > se->statistics.wait_start))
831 wait_start -= se->statistics.wait_start;
833 se->statistics.wait_start = wait_start;
837 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
839 struct task_struct *p;
842 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
844 if (entity_is_task(se)) {
846 if (task_on_rq_migrating(p)) {
848 * Preserve migrating task's wait time so wait_start
849 * time stamp can be adjusted to accumulate wait time
850 * prior to migration.
852 se->statistics.wait_start = delta;
855 trace_sched_stat_wait(p, delta);
858 se->statistics.wait_max = max(se->statistics.wait_max, delta);
859 se->statistics.wait_count++;
860 se->statistics.wait_sum += delta;
861 se->statistics.wait_start = 0;
865 * Task is being enqueued - update stats:
868 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
871 * Are we enqueueing a waiting task? (for current tasks
872 * a dequeue/enqueue event is a NOP)
874 if (se != cfs_rq->curr)
875 update_stats_wait_start(cfs_rq, se);
879 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
882 * Mark the end of the wait period if dequeueing a
885 if (se != cfs_rq->curr)
886 update_stats_wait_end(cfs_rq, se);
888 if (flags & DEQUEUE_SLEEP) {
889 if (entity_is_task(se)) {
890 struct task_struct *tsk = task_of(se);
892 if (tsk->state & TASK_INTERRUPTIBLE)
893 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
894 if (tsk->state & TASK_UNINTERRUPTIBLE)
895 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
902 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
907 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
917 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
923 * We are picking a new current task - update its stats:
926 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
929 * We are starting a new run period:
931 se->exec_start = rq_clock_task(rq_of(cfs_rq));
934 /**************************************************
935 * Scheduling class queueing methods:
938 #ifdef CONFIG_NUMA_BALANCING
940 * Approximate time to scan a full NUMA task in ms. The task scan period is
941 * calculated based on the tasks virtual memory size and
942 * numa_balancing_scan_size.
944 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
945 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
947 /* Portion of address space to scan in MB */
948 unsigned int sysctl_numa_balancing_scan_size = 256;
950 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
951 unsigned int sysctl_numa_balancing_scan_delay = 1000;
953 static unsigned int task_nr_scan_windows(struct task_struct *p)
955 unsigned long rss = 0;
956 unsigned long nr_scan_pages;
959 * Calculations based on RSS as non-present and empty pages are skipped
960 * by the PTE scanner and NUMA hinting faults should be trapped based
963 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
964 rss = get_mm_rss(p->mm);
968 rss = round_up(rss, nr_scan_pages);
969 return rss / nr_scan_pages;
972 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
973 #define MAX_SCAN_WINDOW 2560
975 static unsigned int task_scan_min(struct task_struct *p)
977 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
978 unsigned int scan, floor;
979 unsigned int windows = 1;
981 if (scan_size < MAX_SCAN_WINDOW)
982 windows = MAX_SCAN_WINDOW / scan_size;
983 floor = 1000 / windows;
985 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
986 return max_t(unsigned int, floor, scan);
989 static unsigned int task_scan_max(struct task_struct *p)
991 unsigned int smin = task_scan_min(p);
994 /* Watch for min being lower than max due to floor calculations */
995 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
996 return max(smin, smax);
999 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1001 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1002 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1005 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1007 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1008 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1014 spinlock_t lock; /* nr_tasks, tasks */
1019 struct rcu_head rcu;
1020 unsigned long total_faults;
1021 unsigned long max_faults_cpu;
1023 * Faults_cpu is used to decide whether memory should move
1024 * towards the CPU. As a consequence, these stats are weighted
1025 * more by CPU use than by memory faults.
1027 unsigned long *faults_cpu;
1028 unsigned long faults[0];
1031 /* Shared or private faults. */
1032 #define NR_NUMA_HINT_FAULT_TYPES 2
1034 /* Memory and CPU locality */
1035 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1037 /* Averaged statistics, and temporary buffers. */
1038 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1040 pid_t task_numa_group_id(struct task_struct *p)
1042 return p->numa_group ? p->numa_group->gid : 0;
1046 * The averaged statistics, shared & private, memory & cpu,
1047 * occupy the first half of the array. The second half of the
1048 * array is for current counters, which are averaged into the
1049 * first set by task_numa_placement.
1051 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1053 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1056 static inline unsigned long task_faults(struct task_struct *p, int nid)
1058 if (!p->numa_faults)
1061 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1062 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1065 static inline unsigned long group_faults(struct task_struct *p, int nid)
1070 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1071 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1074 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1076 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1077 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1081 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1082 * considered part of a numa group's pseudo-interleaving set. Migrations
1083 * between these nodes are slowed down, to allow things to settle down.
1085 #define ACTIVE_NODE_FRACTION 3
1087 static bool numa_is_active_node(int nid, struct numa_group *ng)
1089 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1092 /* Handle placement on systems where not all nodes are directly connected. */
1093 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1094 int maxdist, bool task)
1096 unsigned long score = 0;
1100 * All nodes are directly connected, and the same distance
1101 * from each other. No need for fancy placement algorithms.
1103 if (sched_numa_topology_type == NUMA_DIRECT)
1107 * This code is called for each node, introducing N^2 complexity,
1108 * which should be ok given the number of nodes rarely exceeds 8.
1110 for_each_online_node(node) {
1111 unsigned long faults;
1112 int dist = node_distance(nid, node);
1115 * The furthest away nodes in the system are not interesting
1116 * for placement; nid was already counted.
1118 if (dist == sched_max_numa_distance || node == nid)
1122 * On systems with a backplane NUMA topology, compare groups
1123 * of nodes, and move tasks towards the group with the most
1124 * memory accesses. When comparing two nodes at distance
1125 * "hoplimit", only nodes closer by than "hoplimit" are part
1126 * of each group. Skip other nodes.
1128 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1132 /* Add up the faults from nearby nodes. */
1134 faults = task_faults(p, node);
1136 faults = group_faults(p, node);
1139 * On systems with a glueless mesh NUMA topology, there are
1140 * no fixed "groups of nodes". Instead, nodes that are not
1141 * directly connected bounce traffic through intermediate
1142 * nodes; a numa_group can occupy any set of nodes.
1143 * The further away a node is, the less the faults count.
1144 * This seems to result in good task placement.
1146 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1147 faults *= (sched_max_numa_distance - dist);
1148 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1158 * These return the fraction of accesses done by a particular task, or
1159 * task group, on a particular numa node. The group weight is given a
1160 * larger multiplier, in order to group tasks together that are almost
1161 * evenly spread out between numa nodes.
1163 static inline unsigned long task_weight(struct task_struct *p, int nid,
1166 unsigned long faults, total_faults;
1168 if (!p->numa_faults)
1171 total_faults = p->total_numa_faults;
1176 faults = task_faults(p, nid);
1177 faults += score_nearby_nodes(p, nid, dist, true);
1179 return 1000 * faults / total_faults;
1182 static inline unsigned long group_weight(struct task_struct *p, int nid,
1185 unsigned long faults, total_faults;
1190 total_faults = p->numa_group->total_faults;
1195 faults = group_faults(p, nid);
1196 faults += score_nearby_nodes(p, nid, dist, false);
1198 return 1000 * faults / total_faults;
1201 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1202 int src_nid, int dst_cpu)
1204 struct numa_group *ng = p->numa_group;
1205 int dst_nid = cpu_to_node(dst_cpu);
1206 int last_cpupid, this_cpupid;
1208 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1211 * Multi-stage node selection is used in conjunction with a periodic
1212 * migration fault to build a temporal task<->page relation. By using
1213 * a two-stage filter we remove short/unlikely relations.
1215 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1216 * a task's usage of a particular page (n_p) per total usage of this
1217 * page (n_t) (in a given time-span) to a probability.
1219 * Our periodic faults will sample this probability and getting the
1220 * same result twice in a row, given these samples are fully
1221 * independent, is then given by P(n)^2, provided our sample period
1222 * is sufficiently short compared to the usage pattern.
1224 * This quadric squishes small probabilities, making it less likely we
1225 * act on an unlikely task<->page relation.
1227 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1228 if (!cpupid_pid_unset(last_cpupid) &&
1229 cpupid_to_nid(last_cpupid) != dst_nid)
1232 /* Always allow migrate on private faults */
1233 if (cpupid_match_pid(p, last_cpupid))
1236 /* A shared fault, but p->numa_group has not been set up yet. */
1241 * Destination node is much more heavily used than the source
1242 * node? Allow migration.
1244 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1245 ACTIVE_NODE_FRACTION)
1249 * Distribute memory according to CPU & memory use on each node,
1250 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1252 * faults_cpu(dst) 3 faults_cpu(src)
1253 * --------------- * - > ---------------
1254 * faults_mem(dst) 4 faults_mem(src)
1256 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1257 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1260 static unsigned long weighted_cpuload(const int cpu);
1261 static unsigned long source_load(int cpu, int type);
1262 static unsigned long target_load(int cpu, int type);
1263 static unsigned long capacity_of(int cpu);
1264 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1266 /* Cached statistics for all CPUs within a node */
1268 unsigned long nr_running;
1271 /* Total compute capacity of CPUs on a node */
1272 unsigned long compute_capacity;
1274 /* Approximate capacity in terms of runnable tasks on a node */
1275 unsigned long task_capacity;
1276 int has_free_capacity;
1280 * XXX borrowed from update_sg_lb_stats
1282 static void update_numa_stats(struct numa_stats *ns, int nid)
1284 int smt, cpu, cpus = 0;
1285 unsigned long capacity;
1287 memset(ns, 0, sizeof(*ns));
1288 for_each_cpu(cpu, cpumask_of_node(nid)) {
1289 struct rq *rq = cpu_rq(cpu);
1291 ns->nr_running += rq->nr_running;
1292 ns->load += weighted_cpuload(cpu);
1293 ns->compute_capacity += capacity_of(cpu);
1299 * If we raced with hotplug and there are no CPUs left in our mask
1300 * the @ns structure is NULL'ed and task_numa_compare() will
1301 * not find this node attractive.
1303 * We'll either bail at !has_free_capacity, or we'll detect a huge
1304 * imbalance and bail there.
1309 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1310 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1311 capacity = cpus / smt; /* cores */
1313 ns->task_capacity = min_t(unsigned, capacity,
1314 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1315 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1318 struct task_numa_env {
1319 struct task_struct *p;
1321 int src_cpu, src_nid;
1322 int dst_cpu, dst_nid;
1324 struct numa_stats src_stats, dst_stats;
1329 struct task_struct *best_task;
1334 static void task_numa_assign(struct task_numa_env *env,
1335 struct task_struct *p, long imp)
1338 put_task_struct(env->best_task);
1343 env->best_imp = imp;
1344 env->best_cpu = env->dst_cpu;
1347 static bool load_too_imbalanced(long src_load, long dst_load,
1348 struct task_numa_env *env)
1351 long orig_src_load, orig_dst_load;
1352 long src_capacity, dst_capacity;
1355 * The load is corrected for the CPU capacity available on each node.
1358 * ------------ vs ---------
1359 * src_capacity dst_capacity
1361 src_capacity = env->src_stats.compute_capacity;
1362 dst_capacity = env->dst_stats.compute_capacity;
1364 /* We care about the slope of the imbalance, not the direction. */
1365 if (dst_load < src_load)
1366 swap(dst_load, src_load);
1368 /* Is the difference below the threshold? */
1369 imb = dst_load * src_capacity * 100 -
1370 src_load * dst_capacity * env->imbalance_pct;
1375 * The imbalance is above the allowed threshold.
1376 * Compare it with the old imbalance.
1378 orig_src_load = env->src_stats.load;
1379 orig_dst_load = env->dst_stats.load;
1381 if (orig_dst_load < orig_src_load)
1382 swap(orig_dst_load, orig_src_load);
1384 old_imb = orig_dst_load * src_capacity * 100 -
1385 orig_src_load * dst_capacity * env->imbalance_pct;
1387 /* Would this change make things worse? */
1388 return (imb > old_imb);
1392 * This checks if the overall compute and NUMA accesses of the system would
1393 * be improved if the source tasks was migrated to the target dst_cpu taking
1394 * into account that it might be best if task running on the dst_cpu should
1395 * be exchanged with the source task
1397 static void task_numa_compare(struct task_numa_env *env,
1398 long taskimp, long groupimp)
1400 struct rq *src_rq = cpu_rq(env->src_cpu);
1401 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1402 struct task_struct *cur;
1403 long src_load, dst_load;
1405 long imp = env->p->numa_group ? groupimp : taskimp;
1407 int dist = env->dist;
1410 cur = task_rcu_dereference(&dst_rq->curr);
1411 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1415 * Because we have preemption enabled we can get migrated around and
1416 * end try selecting ourselves (current == env->p) as a swap candidate.
1422 * "imp" is the fault differential for the source task between the
1423 * source and destination node. Calculate the total differential for
1424 * the source task and potential destination task. The more negative
1425 * the value is, the more rmeote accesses that would be expected to
1426 * be incurred if the tasks were swapped.
1429 /* Skip this swap candidate if cannot move to the source cpu */
1430 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1434 * If dst and source tasks are in the same NUMA group, or not
1435 * in any group then look only at task weights.
1437 if (cur->numa_group == env->p->numa_group) {
1438 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1439 task_weight(cur, env->dst_nid, dist);
1441 * Add some hysteresis to prevent swapping the
1442 * tasks within a group over tiny differences.
1444 if (cur->numa_group)
1448 * Compare the group weights. If a task is all by
1449 * itself (not part of a group), use the task weight
1452 if (cur->numa_group)
1453 imp += group_weight(cur, env->src_nid, dist) -
1454 group_weight(cur, env->dst_nid, dist);
1456 imp += task_weight(cur, env->src_nid, dist) -
1457 task_weight(cur, env->dst_nid, dist);
1461 if (imp <= env->best_imp && moveimp <= env->best_imp)
1465 /* Is there capacity at our destination? */
1466 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1467 !env->dst_stats.has_free_capacity)
1473 /* Balance doesn't matter much if we're running a task per cpu */
1474 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1475 dst_rq->nr_running == 1)
1479 * In the overloaded case, try and keep the load balanced.
1482 load = task_h_load(env->p);
1483 dst_load = env->dst_stats.load + load;
1484 src_load = env->src_stats.load - load;
1486 if (moveimp > imp && moveimp > env->best_imp) {
1488 * If the improvement from just moving env->p direction is
1489 * better than swapping tasks around, check if a move is
1490 * possible. Store a slightly smaller score than moveimp,
1491 * so an actually idle CPU will win.
1493 if (!load_too_imbalanced(src_load, dst_load, env)) {
1500 if (imp <= env->best_imp)
1504 load = task_h_load(cur);
1509 if (load_too_imbalanced(src_load, dst_load, env))
1513 * One idle CPU per node is evaluated for a task numa move.
1514 * Call select_idle_sibling to maybe find a better one.
1517 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1520 task_numa_assign(env, cur, imp);
1525 static void task_numa_find_cpu(struct task_numa_env *env,
1526 long taskimp, long groupimp)
1530 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1531 /* Skip this CPU if the source task cannot migrate */
1532 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1536 task_numa_compare(env, taskimp, groupimp);
1540 /* Only move tasks to a NUMA node less busy than the current node. */
1541 static bool numa_has_capacity(struct task_numa_env *env)
1543 struct numa_stats *src = &env->src_stats;
1544 struct numa_stats *dst = &env->dst_stats;
1546 if (src->has_free_capacity && !dst->has_free_capacity)
1550 * Only consider a task move if the source has a higher load
1551 * than the destination, corrected for CPU capacity on each node.
1553 * src->load dst->load
1554 * --------------------- vs ---------------------
1555 * src->compute_capacity dst->compute_capacity
1557 if (src->load * dst->compute_capacity * env->imbalance_pct >
1559 dst->load * src->compute_capacity * 100)
1565 static int task_numa_migrate(struct task_struct *p)
1567 struct task_numa_env env = {
1570 .src_cpu = task_cpu(p),
1571 .src_nid = task_node(p),
1573 .imbalance_pct = 112,
1579 struct sched_domain *sd;
1580 unsigned long taskweight, groupweight;
1582 long taskimp, groupimp;
1585 * Pick the lowest SD_NUMA domain, as that would have the smallest
1586 * imbalance and would be the first to start moving tasks about.
1588 * And we want to avoid any moving of tasks about, as that would create
1589 * random movement of tasks -- counter the numa conditions we're trying
1593 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1595 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1599 * Cpusets can break the scheduler domain tree into smaller
1600 * balance domains, some of which do not cross NUMA boundaries.
1601 * Tasks that are "trapped" in such domains cannot be migrated
1602 * elsewhere, so there is no point in (re)trying.
1604 if (unlikely(!sd)) {
1605 p->numa_preferred_nid = task_node(p);
1609 env.dst_nid = p->numa_preferred_nid;
1610 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1611 taskweight = task_weight(p, env.src_nid, dist);
1612 groupweight = group_weight(p, env.src_nid, dist);
1613 update_numa_stats(&env.src_stats, env.src_nid);
1614 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1615 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1616 update_numa_stats(&env.dst_stats, env.dst_nid);
1618 /* Try to find a spot on the preferred nid. */
1619 if (numa_has_capacity(&env))
1620 task_numa_find_cpu(&env, taskimp, groupimp);
1623 * Look at other nodes in these cases:
1624 * - there is no space available on the preferred_nid
1625 * - the task is part of a numa_group that is interleaved across
1626 * multiple NUMA nodes; in order to better consolidate the group,
1627 * we need to check other locations.
1629 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1630 for_each_online_node(nid) {
1631 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1634 dist = node_distance(env.src_nid, env.dst_nid);
1635 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1637 taskweight = task_weight(p, env.src_nid, dist);
1638 groupweight = group_weight(p, env.src_nid, dist);
1641 /* Only consider nodes where both task and groups benefit */
1642 taskimp = task_weight(p, nid, dist) - taskweight;
1643 groupimp = group_weight(p, nid, dist) - groupweight;
1644 if (taskimp < 0 && groupimp < 0)
1649 update_numa_stats(&env.dst_stats, env.dst_nid);
1650 if (numa_has_capacity(&env))
1651 task_numa_find_cpu(&env, taskimp, groupimp);
1656 * If the task is part of a workload that spans multiple NUMA nodes,
1657 * and is migrating into one of the workload's active nodes, remember
1658 * this node as the task's preferred numa node, so the workload can
1660 * A task that migrated to a second choice node will be better off
1661 * trying for a better one later. Do not set the preferred node here.
1663 if (p->numa_group) {
1664 struct numa_group *ng = p->numa_group;
1666 if (env.best_cpu == -1)
1671 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1672 sched_setnuma(p, env.dst_nid);
1675 /* No better CPU than the current one was found. */
1676 if (env.best_cpu == -1)
1680 * Reset the scan period if the task is being rescheduled on an
1681 * alternative node to recheck if the tasks is now properly placed.
1683 p->numa_scan_period = task_scan_min(p);
1685 if (env.best_task == NULL) {
1686 ret = migrate_task_to(p, env.best_cpu);
1688 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1692 ret = migrate_swap(p, env.best_task);
1694 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1695 put_task_struct(env.best_task);
1699 /* Attempt to migrate a task to a CPU on the preferred node. */
1700 static void numa_migrate_preferred(struct task_struct *p)
1702 unsigned long interval = HZ;
1704 /* This task has no NUMA fault statistics yet */
1705 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1708 /* Periodically retry migrating the task to the preferred node */
1709 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1710 p->numa_migrate_retry = jiffies + interval;
1712 /* Success if task is already running on preferred CPU */
1713 if (task_node(p) == p->numa_preferred_nid)
1716 /* Otherwise, try migrate to a CPU on the preferred node */
1717 task_numa_migrate(p);
1721 * Find out how many nodes on the workload is actively running on. Do this by
1722 * tracking the nodes from which NUMA hinting faults are triggered. This can
1723 * be different from the set of nodes where the workload's memory is currently
1726 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1728 unsigned long faults, max_faults = 0;
1729 int nid, active_nodes = 0;
1731 for_each_online_node(nid) {
1732 faults = group_faults_cpu(numa_group, nid);
1733 if (faults > max_faults)
1734 max_faults = faults;
1737 for_each_online_node(nid) {
1738 faults = group_faults_cpu(numa_group, nid);
1739 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1743 numa_group->max_faults_cpu = max_faults;
1744 numa_group->active_nodes = active_nodes;
1748 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1749 * increments. The more local the fault statistics are, the higher the scan
1750 * period will be for the next scan window. If local/(local+remote) ratio is
1751 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1752 * the scan period will decrease. Aim for 70% local accesses.
1754 #define NUMA_PERIOD_SLOTS 10
1755 #define NUMA_PERIOD_THRESHOLD 7
1758 * Increase the scan period (slow down scanning) if the majority of
1759 * our memory is already on our local node, or if the majority of
1760 * the page accesses are shared with other processes.
1761 * Otherwise, decrease the scan period.
1763 static void update_task_scan_period(struct task_struct *p,
1764 unsigned long shared, unsigned long private)
1766 unsigned int period_slot;
1770 unsigned long remote = p->numa_faults_locality[0];
1771 unsigned long local = p->numa_faults_locality[1];
1774 * If there were no record hinting faults then either the task is
1775 * completely idle or all activity is areas that are not of interest
1776 * to automatic numa balancing. Related to that, if there were failed
1777 * migration then it implies we are migrating too quickly or the local
1778 * node is overloaded. In either case, scan slower
1780 if (local + shared == 0 || p->numa_faults_locality[2]) {
1781 p->numa_scan_period = min(p->numa_scan_period_max,
1782 p->numa_scan_period << 1);
1784 p->mm->numa_next_scan = jiffies +
1785 msecs_to_jiffies(p->numa_scan_period);
1791 * Prepare to scale scan period relative to the current period.
1792 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1793 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1794 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1796 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1797 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1798 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1799 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1802 diff = slot * period_slot;
1804 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1807 * Scale scan rate increases based on sharing. There is an
1808 * inverse relationship between the degree of sharing and
1809 * the adjustment made to the scanning period. Broadly
1810 * speaking the intent is that there is little point
1811 * scanning faster if shared accesses dominate as it may
1812 * simply bounce migrations uselessly
1814 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1815 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1818 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1819 task_scan_min(p), task_scan_max(p));
1820 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1824 * Get the fraction of time the task has been running since the last
1825 * NUMA placement cycle. The scheduler keeps similar statistics, but
1826 * decays those on a 32ms period, which is orders of magnitude off
1827 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1828 * stats only if the task is so new there are no NUMA statistics yet.
1830 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1832 u64 runtime, delta, now;
1833 /* Use the start of this time slice to avoid calculations. */
1834 now = p->se.exec_start;
1835 runtime = p->se.sum_exec_runtime;
1837 if (p->last_task_numa_placement) {
1838 delta = runtime - p->last_sum_exec_runtime;
1839 *period = now - p->last_task_numa_placement;
1841 delta = p->se.avg.load_sum / p->se.load.weight;
1842 *period = LOAD_AVG_MAX;
1845 p->last_sum_exec_runtime = runtime;
1846 p->last_task_numa_placement = now;
1852 * Determine the preferred nid for a task in a numa_group. This needs to
1853 * be done in a way that produces consistent results with group_weight,
1854 * otherwise workloads might not converge.
1856 static int preferred_group_nid(struct task_struct *p, int nid)
1861 /* Direct connections between all NUMA nodes. */
1862 if (sched_numa_topology_type == NUMA_DIRECT)
1866 * On a system with glueless mesh NUMA topology, group_weight
1867 * scores nodes according to the number of NUMA hinting faults on
1868 * both the node itself, and on nearby nodes.
1870 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1871 unsigned long score, max_score = 0;
1872 int node, max_node = nid;
1874 dist = sched_max_numa_distance;
1876 for_each_online_node(node) {
1877 score = group_weight(p, node, dist);
1878 if (score > max_score) {
1887 * Finding the preferred nid in a system with NUMA backplane
1888 * interconnect topology is more involved. The goal is to locate
1889 * tasks from numa_groups near each other in the system, and
1890 * untangle workloads from different sides of the system. This requires
1891 * searching down the hierarchy of node groups, recursively searching
1892 * inside the highest scoring group of nodes. The nodemask tricks
1893 * keep the complexity of the search down.
1895 nodes = node_online_map;
1896 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1897 unsigned long max_faults = 0;
1898 nodemask_t max_group = NODE_MASK_NONE;
1901 /* Are there nodes at this distance from each other? */
1902 if (!find_numa_distance(dist))
1905 for_each_node_mask(a, nodes) {
1906 unsigned long faults = 0;
1907 nodemask_t this_group;
1908 nodes_clear(this_group);
1910 /* Sum group's NUMA faults; includes a==b case. */
1911 for_each_node_mask(b, nodes) {
1912 if (node_distance(a, b) < dist) {
1913 faults += group_faults(p, b);
1914 node_set(b, this_group);
1915 node_clear(b, nodes);
1919 /* Remember the top group. */
1920 if (faults > max_faults) {
1921 max_faults = faults;
1922 max_group = this_group;
1924 * subtle: at the smallest distance there is
1925 * just one node left in each "group", the
1926 * winner is the preferred nid.
1931 /* Next round, evaluate the nodes within max_group. */
1939 static void task_numa_placement(struct task_struct *p)
1941 int seq, nid, max_nid = -1, max_group_nid = -1;
1942 unsigned long max_faults = 0, max_group_faults = 0;
1943 unsigned long fault_types[2] = { 0, 0 };
1944 unsigned long total_faults;
1945 u64 runtime, period;
1946 spinlock_t *group_lock = NULL;
1949 * The p->mm->numa_scan_seq field gets updated without
1950 * exclusive access. Use READ_ONCE() here to ensure
1951 * that the field is read in a single access:
1953 seq = READ_ONCE(p->mm->numa_scan_seq);
1954 if (p->numa_scan_seq == seq)
1956 p->numa_scan_seq = seq;
1957 p->numa_scan_period_max = task_scan_max(p);
1959 total_faults = p->numa_faults_locality[0] +
1960 p->numa_faults_locality[1];
1961 runtime = numa_get_avg_runtime(p, &period);
1963 /* If the task is part of a group prevent parallel updates to group stats */
1964 if (p->numa_group) {
1965 group_lock = &p->numa_group->lock;
1966 spin_lock_irq(group_lock);
1969 /* Find the node with the highest number of faults */
1970 for_each_online_node(nid) {
1971 /* Keep track of the offsets in numa_faults array */
1972 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1973 unsigned long faults = 0, group_faults = 0;
1976 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1977 long diff, f_diff, f_weight;
1979 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1980 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1981 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1982 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1984 /* Decay existing window, copy faults since last scan */
1985 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1986 fault_types[priv] += p->numa_faults[membuf_idx];
1987 p->numa_faults[membuf_idx] = 0;
1990 * Normalize the faults_from, so all tasks in a group
1991 * count according to CPU use, instead of by the raw
1992 * number of faults. Tasks with little runtime have
1993 * little over-all impact on throughput, and thus their
1994 * faults are less important.
1996 f_weight = div64_u64(runtime << 16, period + 1);
1997 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1999 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2000 p->numa_faults[cpubuf_idx] = 0;
2002 p->numa_faults[mem_idx] += diff;
2003 p->numa_faults[cpu_idx] += f_diff;
2004 faults += p->numa_faults[mem_idx];
2005 p->total_numa_faults += diff;
2006 if (p->numa_group) {
2008 * safe because we can only change our own group
2010 * mem_idx represents the offset for a given
2011 * nid and priv in a specific region because it
2012 * is at the beginning of the numa_faults array.
2014 p->numa_group->faults[mem_idx] += diff;
2015 p->numa_group->faults_cpu[mem_idx] += f_diff;
2016 p->numa_group->total_faults += diff;
2017 group_faults += p->numa_group->faults[mem_idx];
2021 if (faults > max_faults) {
2022 max_faults = faults;
2026 if (group_faults > max_group_faults) {
2027 max_group_faults = group_faults;
2028 max_group_nid = nid;
2032 update_task_scan_period(p, fault_types[0], fault_types[1]);
2034 if (p->numa_group) {
2035 numa_group_count_active_nodes(p->numa_group);
2036 spin_unlock_irq(group_lock);
2037 max_nid = preferred_group_nid(p, max_group_nid);
2041 /* Set the new preferred node */
2042 if (max_nid != p->numa_preferred_nid)
2043 sched_setnuma(p, max_nid);
2045 if (task_node(p) != p->numa_preferred_nid)
2046 numa_migrate_preferred(p);
2050 static inline int get_numa_group(struct numa_group *grp)
2052 return atomic_inc_not_zero(&grp->refcount);
2055 static inline void put_numa_group(struct numa_group *grp)
2057 if (atomic_dec_and_test(&grp->refcount))
2058 kfree_rcu(grp, rcu);
2061 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2064 struct numa_group *grp, *my_grp;
2065 struct task_struct *tsk;
2067 int cpu = cpupid_to_cpu(cpupid);
2070 if (unlikely(!p->numa_group)) {
2071 unsigned int size = sizeof(struct numa_group) +
2072 4*nr_node_ids*sizeof(unsigned long);
2074 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2078 atomic_set(&grp->refcount, 1);
2079 grp->active_nodes = 1;
2080 grp->max_faults_cpu = 0;
2081 spin_lock_init(&grp->lock);
2083 /* Second half of the array tracks nids where faults happen */
2084 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2087 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2088 grp->faults[i] = p->numa_faults[i];
2090 grp->total_faults = p->total_numa_faults;
2093 rcu_assign_pointer(p->numa_group, grp);
2097 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2099 if (!cpupid_match_pid(tsk, cpupid))
2102 grp = rcu_dereference(tsk->numa_group);
2106 my_grp = p->numa_group;
2111 * Only join the other group if its bigger; if we're the bigger group,
2112 * the other task will join us.
2114 if (my_grp->nr_tasks > grp->nr_tasks)
2118 * Tie-break on the grp address.
2120 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2123 /* Always join threads in the same process. */
2124 if (tsk->mm == current->mm)
2127 /* Simple filter to avoid false positives due to PID collisions */
2128 if (flags & TNF_SHARED)
2131 /* Update priv based on whether false sharing was detected */
2134 if (join && !get_numa_group(grp))
2142 BUG_ON(irqs_disabled());
2143 double_lock_irq(&my_grp->lock, &grp->lock);
2145 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2146 my_grp->faults[i] -= p->numa_faults[i];
2147 grp->faults[i] += p->numa_faults[i];
2149 my_grp->total_faults -= p->total_numa_faults;
2150 grp->total_faults += p->total_numa_faults;
2155 spin_unlock(&my_grp->lock);
2156 spin_unlock_irq(&grp->lock);
2158 rcu_assign_pointer(p->numa_group, grp);
2160 put_numa_group(my_grp);
2168 void task_numa_free(struct task_struct *p)
2170 struct numa_group *grp = p->numa_group;
2171 void *numa_faults = p->numa_faults;
2172 unsigned long flags;
2176 spin_lock_irqsave(&grp->lock, flags);
2177 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2178 grp->faults[i] -= p->numa_faults[i];
2179 grp->total_faults -= p->total_numa_faults;
2182 spin_unlock_irqrestore(&grp->lock, flags);
2183 RCU_INIT_POINTER(p->numa_group, NULL);
2184 put_numa_group(grp);
2187 p->numa_faults = NULL;
2192 * Got a PROT_NONE fault for a page on @node.
2194 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2196 struct task_struct *p = current;
2197 bool migrated = flags & TNF_MIGRATED;
2198 int cpu_node = task_node(current);
2199 int local = !!(flags & TNF_FAULT_LOCAL);
2200 struct numa_group *ng;
2203 if (!static_branch_likely(&sched_numa_balancing))
2206 /* for example, ksmd faulting in a user's mm */
2210 /* Allocate buffer to track faults on a per-node basis */
2211 if (unlikely(!p->numa_faults)) {
2212 int size = sizeof(*p->numa_faults) *
2213 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2215 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2216 if (!p->numa_faults)
2219 p->total_numa_faults = 0;
2220 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2224 * First accesses are treated as private, otherwise consider accesses
2225 * to be private if the accessing pid has not changed
2227 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2230 priv = cpupid_match_pid(p, last_cpupid);
2231 if (!priv && !(flags & TNF_NO_GROUP))
2232 task_numa_group(p, last_cpupid, flags, &priv);
2236 * If a workload spans multiple NUMA nodes, a shared fault that
2237 * occurs wholly within the set of nodes that the workload is
2238 * actively using should be counted as local. This allows the
2239 * scan rate to slow down when a workload has settled down.
2242 if (!priv && !local && ng && ng->active_nodes > 1 &&
2243 numa_is_active_node(cpu_node, ng) &&
2244 numa_is_active_node(mem_node, ng))
2247 task_numa_placement(p);
2250 * Retry task to preferred node migration periodically, in case it
2251 * case it previously failed, or the scheduler moved us.
2253 if (time_after(jiffies, p->numa_migrate_retry))
2254 numa_migrate_preferred(p);
2257 p->numa_pages_migrated += pages;
2258 if (flags & TNF_MIGRATE_FAIL)
2259 p->numa_faults_locality[2] += pages;
2261 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2262 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2263 p->numa_faults_locality[local] += pages;
2266 static void reset_ptenuma_scan(struct task_struct *p)
2269 * We only did a read acquisition of the mmap sem, so
2270 * p->mm->numa_scan_seq is written to without exclusive access
2271 * and the update is not guaranteed to be atomic. That's not
2272 * much of an issue though, since this is just used for
2273 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2274 * expensive, to avoid any form of compiler optimizations:
2276 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2277 p->mm->numa_scan_offset = 0;
2281 * The expensive part of numa migration is done from task_work context.
2282 * Triggered from task_tick_numa().
2284 void task_numa_work(struct callback_head *work)
2286 unsigned long migrate, next_scan, now = jiffies;
2287 struct task_struct *p = current;
2288 struct mm_struct *mm = p->mm;
2289 u64 runtime = p->se.sum_exec_runtime;
2290 struct vm_area_struct *vma;
2291 unsigned long start, end;
2292 unsigned long nr_pte_updates = 0;
2293 long pages, virtpages;
2295 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2297 work->next = work; /* protect against double add */
2299 * Who cares about NUMA placement when they're dying.
2301 * NOTE: make sure not to dereference p->mm before this check,
2302 * exit_task_work() happens _after_ exit_mm() so we could be called
2303 * without p->mm even though we still had it when we enqueued this
2306 if (p->flags & PF_EXITING)
2309 if (!mm->numa_next_scan) {
2310 mm->numa_next_scan = now +
2311 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2315 * Enforce maximal scan/migration frequency..
2317 migrate = mm->numa_next_scan;
2318 if (time_before(now, migrate))
2321 if (p->numa_scan_period == 0) {
2322 p->numa_scan_period_max = task_scan_max(p);
2323 p->numa_scan_period = task_scan_min(p);
2326 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2327 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2331 * Delay this task enough that another task of this mm will likely win
2332 * the next time around.
2334 p->node_stamp += 2 * TICK_NSEC;
2336 start = mm->numa_scan_offset;
2337 pages = sysctl_numa_balancing_scan_size;
2338 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2339 virtpages = pages * 8; /* Scan up to this much virtual space */
2344 down_read(&mm->mmap_sem);
2345 vma = find_vma(mm, start);
2347 reset_ptenuma_scan(p);
2351 for (; vma; vma = vma->vm_next) {
2352 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2353 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2358 * Shared library pages mapped by multiple processes are not
2359 * migrated as it is expected they are cache replicated. Avoid
2360 * hinting faults in read-only file-backed mappings or the vdso
2361 * as migrating the pages will be of marginal benefit.
2364 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2368 * Skip inaccessible VMAs to avoid any confusion between
2369 * PROT_NONE and NUMA hinting ptes
2371 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2375 start = max(start, vma->vm_start);
2376 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2377 end = min(end, vma->vm_end);
2378 nr_pte_updates = change_prot_numa(vma, start, end);
2381 * Try to scan sysctl_numa_balancing_size worth of
2382 * hpages that have at least one present PTE that
2383 * is not already pte-numa. If the VMA contains
2384 * areas that are unused or already full of prot_numa
2385 * PTEs, scan up to virtpages, to skip through those
2389 pages -= (end - start) >> PAGE_SHIFT;
2390 virtpages -= (end - start) >> PAGE_SHIFT;
2393 if (pages <= 0 || virtpages <= 0)
2397 } while (end != vma->vm_end);
2402 * It is possible to reach the end of the VMA list but the last few
2403 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2404 * would find the !migratable VMA on the next scan but not reset the
2405 * scanner to the start so check it now.
2408 mm->numa_scan_offset = start;
2410 reset_ptenuma_scan(p);
2411 up_read(&mm->mmap_sem);
2414 * Make sure tasks use at least 32x as much time to run other code
2415 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2416 * Usually update_task_scan_period slows down scanning enough; on an
2417 * overloaded system we need to limit overhead on a per task basis.
2419 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2420 u64 diff = p->se.sum_exec_runtime - runtime;
2421 p->node_stamp += 32 * diff;
2426 * Drive the periodic memory faults..
2428 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2430 struct callback_head *work = &curr->numa_work;
2434 * We don't care about NUMA placement if we don't have memory.
2436 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2440 * Using runtime rather than walltime has the dual advantage that
2441 * we (mostly) drive the selection from busy threads and that the
2442 * task needs to have done some actual work before we bother with
2445 now = curr->se.sum_exec_runtime;
2446 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2448 if (now > curr->node_stamp + period) {
2449 if (!curr->node_stamp)
2450 curr->numa_scan_period = task_scan_min(curr);
2451 curr->node_stamp += period;
2453 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2454 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2455 task_work_add(curr, work, true);
2460 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2464 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2468 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2471 #endif /* CONFIG_NUMA_BALANCING */
2474 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2476 update_load_add(&cfs_rq->load, se->load.weight);
2477 if (!parent_entity(se))
2478 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2480 if (entity_is_task(se)) {
2481 struct rq *rq = rq_of(cfs_rq);
2483 account_numa_enqueue(rq, task_of(se));
2484 list_add(&se->group_node, &rq->cfs_tasks);
2487 cfs_rq->nr_running++;
2491 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2493 update_load_sub(&cfs_rq->load, se->load.weight);
2494 if (!parent_entity(se))
2495 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2497 if (entity_is_task(se)) {
2498 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2499 list_del_init(&se->group_node);
2502 cfs_rq->nr_running--;
2505 #ifdef CONFIG_FAIR_GROUP_SCHED
2507 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2509 long tg_weight, load, shares;
2512 * This really should be: cfs_rq->avg.load_avg, but instead we use
2513 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2514 * the shares for small weight interactive tasks.
2516 load = scale_load_down(cfs_rq->load.weight);
2518 tg_weight = atomic_long_read(&tg->load_avg);
2520 /* Ensure tg_weight >= load */
2521 tg_weight -= cfs_rq->tg_load_avg_contrib;
2524 shares = (tg->shares * load);
2526 shares /= tg_weight;
2528 if (shares < MIN_SHARES)
2529 shares = MIN_SHARES;
2530 if (shares > tg->shares)
2531 shares = tg->shares;
2535 # else /* CONFIG_SMP */
2536 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2540 # endif /* CONFIG_SMP */
2542 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2543 unsigned long weight)
2546 /* commit outstanding execution time */
2547 if (cfs_rq->curr == se)
2548 update_curr(cfs_rq);
2549 account_entity_dequeue(cfs_rq, se);
2552 update_load_set(&se->load, weight);
2555 account_entity_enqueue(cfs_rq, se);
2558 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2560 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2562 struct task_group *tg;
2563 struct sched_entity *se;
2567 se = tg->se[cpu_of(rq_of(cfs_rq))];
2568 if (!se || throttled_hierarchy(cfs_rq))
2571 if (likely(se->load.weight == tg->shares))
2574 shares = calc_cfs_shares(cfs_rq, tg);
2576 reweight_entity(cfs_rq_of(se), se, shares);
2578 #else /* CONFIG_FAIR_GROUP_SCHED */
2579 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2582 #endif /* CONFIG_FAIR_GROUP_SCHED */
2585 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2586 static const u32 runnable_avg_yN_inv[] = {
2587 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2588 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2589 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2590 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2591 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2592 0x85aac367, 0x82cd8698,
2596 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2597 * over-estimates when re-combining.
2599 static const u32 runnable_avg_yN_sum[] = {
2600 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2601 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2602 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2606 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2607 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2610 static const u32 __accumulated_sum_N32[] = {
2611 0, 23371, 35056, 40899, 43820, 45281,
2612 46011, 46376, 46559, 46650, 46696, 46719,
2617 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2619 static __always_inline u64 decay_load(u64 val, u64 n)
2621 unsigned int local_n;
2625 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2628 /* after bounds checking we can collapse to 32-bit */
2632 * As y^PERIOD = 1/2, we can combine
2633 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2634 * With a look-up table which covers y^n (n<PERIOD)
2636 * To achieve constant time decay_load.
2638 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2639 val >>= local_n / LOAD_AVG_PERIOD;
2640 local_n %= LOAD_AVG_PERIOD;
2643 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2648 * For updates fully spanning n periods, the contribution to runnable
2649 * average will be: \Sum 1024*y^n
2651 * We can compute this reasonably efficiently by combining:
2652 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2654 static u32 __compute_runnable_contrib(u64 n)
2658 if (likely(n <= LOAD_AVG_PERIOD))
2659 return runnable_avg_yN_sum[n];
2660 else if (unlikely(n >= LOAD_AVG_MAX_N))
2661 return LOAD_AVG_MAX;
2663 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2664 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2665 n %= LOAD_AVG_PERIOD;
2666 contrib = decay_load(contrib, n);
2667 return contrib + runnable_avg_yN_sum[n];
2670 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2673 * We can represent the historical contribution to runnable average as the
2674 * coefficients of a geometric series. To do this we sub-divide our runnable
2675 * history into segments of approximately 1ms (1024us); label the segment that
2676 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2678 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2680 * (now) (~1ms ago) (~2ms ago)
2682 * Let u_i denote the fraction of p_i that the entity was runnable.
2684 * We then designate the fractions u_i as our co-efficients, yielding the
2685 * following representation of historical load:
2686 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2688 * We choose y based on the with of a reasonably scheduling period, fixing:
2691 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2692 * approximately half as much as the contribution to load within the last ms
2695 * When a period "rolls over" and we have new u_0`, multiplying the previous
2696 * sum again by y is sufficient to update:
2697 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2698 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2700 static __always_inline int
2701 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2702 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2704 u64 delta, scaled_delta, periods;
2706 unsigned int delta_w, scaled_delta_w, decayed = 0;
2707 unsigned long scale_freq, scale_cpu;
2709 delta = now - sa->last_update_time;
2711 * This should only happen when time goes backwards, which it
2712 * unfortunately does during sched clock init when we swap over to TSC.
2714 if ((s64)delta < 0) {
2715 sa->last_update_time = now;
2720 * Use 1024ns as the unit of measurement since it's a reasonable
2721 * approximation of 1us and fast to compute.
2726 sa->last_update_time = now;
2728 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2729 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2731 /* delta_w is the amount already accumulated against our next period */
2732 delta_w = sa->period_contrib;
2733 if (delta + delta_w >= 1024) {
2736 /* how much left for next period will start over, we don't know yet */
2737 sa->period_contrib = 0;
2740 * Now that we know we're crossing a period boundary, figure
2741 * out how much from delta we need to complete the current
2742 * period and accrue it.
2744 delta_w = 1024 - delta_w;
2745 scaled_delta_w = cap_scale(delta_w, scale_freq);
2747 sa->load_sum += weight * scaled_delta_w;
2749 cfs_rq->runnable_load_sum +=
2750 weight * scaled_delta_w;
2754 sa->util_sum += scaled_delta_w * scale_cpu;
2758 /* Figure out how many additional periods this update spans */
2759 periods = delta / 1024;
2762 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2764 cfs_rq->runnable_load_sum =
2765 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2767 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2769 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2770 contrib = __compute_runnable_contrib(periods);
2771 contrib = cap_scale(contrib, scale_freq);
2773 sa->load_sum += weight * contrib;
2775 cfs_rq->runnable_load_sum += weight * contrib;
2778 sa->util_sum += contrib * scale_cpu;
2781 /* Remainder of delta accrued against u_0` */
2782 scaled_delta = cap_scale(delta, scale_freq);
2784 sa->load_sum += weight * scaled_delta;
2786 cfs_rq->runnable_load_sum += weight * scaled_delta;
2789 sa->util_sum += scaled_delta * scale_cpu;
2791 sa->period_contrib += delta;
2794 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2796 cfs_rq->runnable_load_avg =
2797 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2799 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2805 #ifdef CONFIG_FAIR_GROUP_SCHED
2807 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2808 * and effective_load (which is not done because it is too costly).
2810 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2812 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2815 * No need to update load_avg for root_task_group as it is not used.
2817 if (cfs_rq->tg == &root_task_group)
2820 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2821 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2822 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2827 * Called within set_task_rq() right before setting a task's cpu. The
2828 * caller only guarantees p->pi_lock is held; no other assumptions,
2829 * including the state of rq->lock, should be made.
2831 void set_task_rq_fair(struct sched_entity *se,
2832 struct cfs_rq *prev, struct cfs_rq *next)
2834 if (!sched_feat(ATTACH_AGE_LOAD))
2838 * We are supposed to update the task to "current" time, then its up to
2839 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2840 * getting what current time is, so simply throw away the out-of-date
2841 * time. This will result in the wakee task is less decayed, but giving
2842 * the wakee more load sounds not bad.
2844 if (se->avg.last_update_time && prev) {
2845 u64 p_last_update_time;
2846 u64 n_last_update_time;
2848 #ifndef CONFIG_64BIT
2849 u64 p_last_update_time_copy;
2850 u64 n_last_update_time_copy;
2853 p_last_update_time_copy = prev->load_last_update_time_copy;
2854 n_last_update_time_copy = next->load_last_update_time_copy;
2858 p_last_update_time = prev->avg.last_update_time;
2859 n_last_update_time = next->avg.last_update_time;
2861 } while (p_last_update_time != p_last_update_time_copy ||
2862 n_last_update_time != n_last_update_time_copy);
2864 p_last_update_time = prev->avg.last_update_time;
2865 n_last_update_time = next->avg.last_update_time;
2867 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2868 &se->avg, 0, 0, NULL);
2869 se->avg.last_update_time = n_last_update_time;
2872 #else /* CONFIG_FAIR_GROUP_SCHED */
2873 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2874 #endif /* CONFIG_FAIR_GROUP_SCHED */
2876 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2878 if (&this_rq()->cfs == cfs_rq) {
2880 * There are a few boundary cases this might miss but it should
2881 * get called often enough that that should (hopefully) not be
2882 * a real problem -- added to that it only calls on the local
2883 * CPU, so if we enqueue remotely we'll miss an update, but
2884 * the next tick/schedule should update.
2886 * It will not get called when we go idle, because the idle
2887 * thread is a different class (!fair), nor will the utilization
2888 * number include things like RT tasks.
2890 * As is, the util number is not freq-invariant (we'd have to
2891 * implement arch_scale_freq_capacity() for that).
2895 cpufreq_update_util(rq_of(cfs_rq), 0);
2900 * Unsigned subtract and clamp on underflow.
2902 * Explicitly do a load-store to ensure the intermediate value never hits
2903 * memory. This allows lockless observations without ever seeing the negative
2906 #define sub_positive(_ptr, _val) do { \
2907 typeof(_ptr) ptr = (_ptr); \
2908 typeof(*ptr) val = (_val); \
2909 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2913 WRITE_ONCE(*ptr, res); \
2917 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2918 * @now: current time, as per cfs_rq_clock_task()
2919 * @cfs_rq: cfs_rq to update
2920 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2922 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2923 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2924 * post_init_entity_util_avg().
2926 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2928 * Returns true if the load decayed or we removed utilization. It is expected
2929 * that one calls update_tg_load_avg() on this condition, but after you've
2930 * modified the cfs_rq avg (attach/detach), such that we propagate the new
2934 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2936 struct sched_avg *sa = &cfs_rq->avg;
2937 int decayed, removed_load = 0, removed_util = 0;
2939 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2940 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2941 sub_positive(&sa->load_avg, r);
2942 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2946 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2947 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2948 sub_positive(&sa->util_avg, r);
2949 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2953 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2954 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2956 #ifndef CONFIG_64BIT
2958 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2961 if (update_freq && (decayed || removed_util))
2962 cfs_rq_util_change(cfs_rq);
2964 return decayed || removed_load;
2967 /* Update task and its cfs_rq load average */
2968 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2970 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2971 u64 now = cfs_rq_clock_task(cfs_rq);
2972 struct rq *rq = rq_of(cfs_rq);
2973 int cpu = cpu_of(rq);
2976 * Track task load average for carrying it to new CPU after migrated, and
2977 * track group sched_entity load average for task_h_load calc in migration
2979 __update_load_avg(now, cpu, &se->avg,
2980 se->on_rq * scale_load_down(se->load.weight),
2981 cfs_rq->curr == se, NULL);
2983 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2984 update_tg_load_avg(cfs_rq, 0);
2988 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2989 * @cfs_rq: cfs_rq to attach to
2990 * @se: sched_entity to attach
2992 * Must call update_cfs_rq_load_avg() before this, since we rely on
2993 * cfs_rq->avg.last_update_time being current.
2995 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2997 if (!sched_feat(ATTACH_AGE_LOAD))
3001 * If we got migrated (either between CPUs or between cgroups) we'll
3002 * have aged the average right before clearing @last_update_time.
3004 * Or we're fresh through post_init_entity_util_avg().
3006 if (se->avg.last_update_time) {
3007 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3008 &se->avg, 0, 0, NULL);
3011 * XXX: we could have just aged the entire load away if we've been
3012 * absent from the fair class for too long.
3017 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3018 cfs_rq->avg.load_avg += se->avg.load_avg;
3019 cfs_rq->avg.load_sum += se->avg.load_sum;
3020 cfs_rq->avg.util_avg += se->avg.util_avg;
3021 cfs_rq->avg.util_sum += se->avg.util_sum;
3023 cfs_rq_util_change(cfs_rq);
3027 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3028 * @cfs_rq: cfs_rq to detach from
3029 * @se: sched_entity to detach
3031 * Must call update_cfs_rq_load_avg() before this, since we rely on
3032 * cfs_rq->avg.last_update_time being current.
3034 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3036 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3037 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3038 cfs_rq->curr == se, NULL);
3040 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3041 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3042 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3043 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3045 cfs_rq_util_change(cfs_rq);
3048 /* Add the load generated by se into cfs_rq's load average */
3050 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3052 struct sched_avg *sa = &se->avg;
3053 u64 now = cfs_rq_clock_task(cfs_rq);
3054 int migrated, decayed;
3056 migrated = !sa->last_update_time;
3058 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3059 se->on_rq * scale_load_down(se->load.weight),
3060 cfs_rq->curr == se, NULL);
3063 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3065 cfs_rq->runnable_load_avg += sa->load_avg;
3066 cfs_rq->runnable_load_sum += sa->load_sum;
3069 attach_entity_load_avg(cfs_rq, se);
3071 if (decayed || migrated)
3072 update_tg_load_avg(cfs_rq, 0);
3075 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3077 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3079 update_load_avg(se, 1);
3081 cfs_rq->runnable_load_avg =
3082 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3083 cfs_rq->runnable_load_sum =
3084 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3087 #ifndef CONFIG_64BIT
3088 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3090 u64 last_update_time_copy;
3091 u64 last_update_time;
3094 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3096 last_update_time = cfs_rq->avg.last_update_time;
3097 } while (last_update_time != last_update_time_copy);
3099 return last_update_time;
3102 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3104 return cfs_rq->avg.last_update_time;
3109 * Task first catches up with cfs_rq, and then subtract
3110 * itself from the cfs_rq (task must be off the queue now).
3112 void remove_entity_load_avg(struct sched_entity *se)
3114 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3115 u64 last_update_time;
3118 * tasks cannot exit without having gone through wake_up_new_task() ->
3119 * post_init_entity_util_avg() which will have added things to the
3120 * cfs_rq, so we can remove unconditionally.
3122 * Similarly for groups, they will have passed through
3123 * post_init_entity_util_avg() before unregister_sched_fair_group()
3127 last_update_time = cfs_rq_last_update_time(cfs_rq);
3129 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3130 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3131 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3134 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3136 return cfs_rq->runnable_load_avg;
3139 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3141 return cfs_rq->avg.load_avg;
3144 static int idle_balance(struct rq *this_rq);
3146 #else /* CONFIG_SMP */
3149 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3154 static inline void update_load_avg(struct sched_entity *se, int not_used)
3156 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3160 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3162 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3163 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3166 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3168 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3170 static inline int idle_balance(struct rq *rq)
3175 #endif /* CONFIG_SMP */
3177 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3179 #ifdef CONFIG_SCHEDSTATS
3180 struct task_struct *tsk = NULL;
3182 if (entity_is_task(se))
3185 if (se->statistics.sleep_start) {
3186 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3191 if (unlikely(delta > se->statistics.sleep_max))
3192 se->statistics.sleep_max = delta;
3194 se->statistics.sleep_start = 0;
3195 se->statistics.sum_sleep_runtime += delta;
3198 account_scheduler_latency(tsk, delta >> 10, 1);
3199 trace_sched_stat_sleep(tsk, delta);
3202 if (se->statistics.block_start) {
3203 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3208 if (unlikely(delta > se->statistics.block_max))
3209 se->statistics.block_max = delta;
3211 se->statistics.block_start = 0;
3212 se->statistics.sum_sleep_runtime += delta;
3215 if (tsk->in_iowait) {
3216 se->statistics.iowait_sum += delta;
3217 se->statistics.iowait_count++;
3218 trace_sched_stat_iowait(tsk, delta);
3221 trace_sched_stat_blocked(tsk, delta);
3224 * Blocking time is in units of nanosecs, so shift by
3225 * 20 to get a milliseconds-range estimation of the
3226 * amount of time that the task spent sleeping:
3228 if (unlikely(prof_on == SLEEP_PROFILING)) {
3229 profile_hits(SLEEP_PROFILING,
3230 (void *)get_wchan(tsk),
3233 account_scheduler_latency(tsk, delta >> 10, 0);
3239 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3241 #ifdef CONFIG_SCHED_DEBUG
3242 s64 d = se->vruntime - cfs_rq->min_vruntime;
3247 if (d > 3*sysctl_sched_latency)
3248 schedstat_inc(cfs_rq, nr_spread_over);
3253 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3255 u64 vruntime = cfs_rq->min_vruntime;
3258 * The 'current' period is already promised to the current tasks,
3259 * however the extra weight of the new task will slow them down a
3260 * little, place the new task so that it fits in the slot that
3261 * stays open at the end.
3263 if (initial && sched_feat(START_DEBIT))
3264 vruntime += sched_vslice(cfs_rq, se);
3266 /* sleeps up to a single latency don't count. */
3268 unsigned long thresh = sysctl_sched_latency;
3271 * Halve their sleep time's effect, to allow
3272 * for a gentler effect of sleepers:
3274 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3280 /* ensure we never gain time by being placed backwards. */
3281 se->vruntime = max_vruntime(se->vruntime, vruntime);
3284 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3286 static inline void check_schedstat_required(void)
3288 #ifdef CONFIG_SCHEDSTATS
3289 if (schedstat_enabled())
3292 /* Force schedstat enabled if a dependent tracepoint is active */
3293 if (trace_sched_stat_wait_enabled() ||
3294 trace_sched_stat_sleep_enabled() ||
3295 trace_sched_stat_iowait_enabled() ||
3296 trace_sched_stat_blocked_enabled() ||
3297 trace_sched_stat_runtime_enabled()) {
3298 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3299 "stat_blocked and stat_runtime require the "
3300 "kernel parameter schedstats=enabled or "
3301 "kernel.sched_schedstats=1\n");
3312 * update_min_vruntime()
3313 * vruntime -= min_vruntime
3317 * update_min_vruntime()
3318 * vruntime += min_vruntime
3320 * this way the vruntime transition between RQs is done when both
3321 * min_vruntime are up-to-date.
3325 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3326 * vruntime -= min_vruntime
3330 * update_min_vruntime()
3331 * vruntime += min_vruntime
3333 * this way we don't have the most up-to-date min_vruntime on the originating
3334 * CPU and an up-to-date min_vruntime on the destination CPU.
3338 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3340 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3341 bool curr = cfs_rq->curr == se;
3344 * If we're the current task, we must renormalise before calling
3348 se->vruntime += cfs_rq->min_vruntime;
3350 update_curr(cfs_rq);
3353 * Otherwise, renormalise after, such that we're placed at the current
3354 * moment in time, instead of some random moment in the past. Being
3355 * placed in the past could significantly boost this task to the
3356 * fairness detriment of existing tasks.
3358 if (renorm && !curr)
3359 se->vruntime += cfs_rq->min_vruntime;
3361 enqueue_entity_load_avg(cfs_rq, se);
3362 account_entity_enqueue(cfs_rq, se);
3363 update_cfs_shares(cfs_rq);
3365 if (flags & ENQUEUE_WAKEUP) {
3366 place_entity(cfs_rq, se, 0);
3367 if (schedstat_enabled())
3368 enqueue_sleeper(cfs_rq, se);
3371 check_schedstat_required();
3372 if (schedstat_enabled()) {
3373 update_stats_enqueue(cfs_rq, se);
3374 check_spread(cfs_rq, se);
3377 __enqueue_entity(cfs_rq, se);
3380 if (cfs_rq->nr_running == 1) {
3381 list_add_leaf_cfs_rq(cfs_rq);
3382 check_enqueue_throttle(cfs_rq);
3386 static void __clear_buddies_last(struct sched_entity *se)
3388 for_each_sched_entity(se) {
3389 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3390 if (cfs_rq->last != se)
3393 cfs_rq->last = NULL;
3397 static void __clear_buddies_next(struct sched_entity *se)
3399 for_each_sched_entity(se) {
3400 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3401 if (cfs_rq->next != se)
3404 cfs_rq->next = NULL;
3408 static void __clear_buddies_skip(struct sched_entity *se)
3410 for_each_sched_entity(se) {
3411 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3412 if (cfs_rq->skip != se)
3415 cfs_rq->skip = NULL;
3419 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3421 if (cfs_rq->last == se)
3422 __clear_buddies_last(se);
3424 if (cfs_rq->next == se)
3425 __clear_buddies_next(se);
3427 if (cfs_rq->skip == se)
3428 __clear_buddies_skip(se);
3431 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3434 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3437 * Update run-time statistics of the 'current'.
3439 update_curr(cfs_rq);
3440 dequeue_entity_load_avg(cfs_rq, se);
3442 if (schedstat_enabled())
3443 update_stats_dequeue(cfs_rq, se, flags);
3445 clear_buddies(cfs_rq, se);
3447 if (se != cfs_rq->curr)
3448 __dequeue_entity(cfs_rq, se);
3450 account_entity_dequeue(cfs_rq, se);
3453 * Normalize the entity after updating the min_vruntime because the
3454 * update can refer to the ->curr item and we need to reflect this
3455 * movement in our normalized position.
3457 if (!(flags & DEQUEUE_SLEEP))
3458 se->vruntime -= cfs_rq->min_vruntime;
3460 /* return excess runtime on last dequeue */
3461 return_cfs_rq_runtime(cfs_rq);
3463 update_min_vruntime(cfs_rq);
3464 update_cfs_shares(cfs_rq);
3468 * Preempt the current task with a newly woken task if needed:
3471 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3473 unsigned long ideal_runtime, delta_exec;
3474 struct sched_entity *se;
3477 ideal_runtime = sched_slice(cfs_rq, curr);
3478 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3479 if (delta_exec > ideal_runtime) {
3480 resched_curr(rq_of(cfs_rq));
3482 * The current task ran long enough, ensure it doesn't get
3483 * re-elected due to buddy favours.
3485 clear_buddies(cfs_rq, curr);
3490 * Ensure that a task that missed wakeup preemption by a
3491 * narrow margin doesn't have to wait for a full slice.
3492 * This also mitigates buddy induced latencies under load.
3494 if (delta_exec < sysctl_sched_min_granularity)
3497 se = __pick_first_entity(cfs_rq);
3498 delta = curr->vruntime - se->vruntime;
3503 if (delta > ideal_runtime)
3504 resched_curr(rq_of(cfs_rq));
3508 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3510 /* 'current' is not kept within the tree. */
3513 * Any task has to be enqueued before it get to execute on
3514 * a CPU. So account for the time it spent waiting on the
3517 if (schedstat_enabled())
3518 update_stats_wait_end(cfs_rq, se);
3519 __dequeue_entity(cfs_rq, se);
3520 update_load_avg(se, 1);
3523 update_stats_curr_start(cfs_rq, se);
3525 #ifdef CONFIG_SCHEDSTATS
3527 * Track our maximum slice length, if the CPU's load is at
3528 * least twice that of our own weight (i.e. dont track it
3529 * when there are only lesser-weight tasks around):
3531 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3532 se->statistics.slice_max = max(se->statistics.slice_max,
3533 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3536 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3540 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3543 * Pick the next process, keeping these things in mind, in this order:
3544 * 1) keep things fair between processes/task groups
3545 * 2) pick the "next" process, since someone really wants that to run
3546 * 3) pick the "last" process, for cache locality
3547 * 4) do not run the "skip" process, if something else is available
3549 static struct sched_entity *
3550 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3552 struct sched_entity *left = __pick_first_entity(cfs_rq);
3553 struct sched_entity *se;
3556 * If curr is set we have to see if its left of the leftmost entity
3557 * still in the tree, provided there was anything in the tree at all.
3559 if (!left || (curr && entity_before(curr, left)))
3562 se = left; /* ideally we run the leftmost entity */
3565 * Avoid running the skip buddy, if running something else can
3566 * be done without getting too unfair.
3568 if (cfs_rq->skip == se) {
3569 struct sched_entity *second;
3572 second = __pick_first_entity(cfs_rq);
3574 second = __pick_next_entity(se);
3575 if (!second || (curr && entity_before(curr, second)))
3579 if (second && wakeup_preempt_entity(second, left) < 1)
3584 * Prefer last buddy, try to return the CPU to a preempted task.
3586 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3590 * Someone really wants this to run. If it's not unfair, run it.
3592 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3595 clear_buddies(cfs_rq, se);
3600 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3602 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3605 * If still on the runqueue then deactivate_task()
3606 * was not called and update_curr() has to be done:
3609 update_curr(cfs_rq);
3611 /* throttle cfs_rqs exceeding runtime */
3612 check_cfs_rq_runtime(cfs_rq);
3614 if (schedstat_enabled()) {
3615 check_spread(cfs_rq, prev);
3617 update_stats_wait_start(cfs_rq, prev);
3621 /* Put 'current' back into the tree. */
3622 __enqueue_entity(cfs_rq, prev);
3623 /* in !on_rq case, update occurred at dequeue */
3624 update_load_avg(prev, 0);
3626 cfs_rq->curr = NULL;
3630 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3633 * Update run-time statistics of the 'current'.
3635 update_curr(cfs_rq);
3638 * Ensure that runnable average is periodically updated.
3640 update_load_avg(curr, 1);
3641 update_cfs_shares(cfs_rq);
3643 #ifdef CONFIG_SCHED_HRTICK
3645 * queued ticks are scheduled to match the slice, so don't bother
3646 * validating it and just reschedule.
3649 resched_curr(rq_of(cfs_rq));
3653 * don't let the period tick interfere with the hrtick preemption
3655 if (!sched_feat(DOUBLE_TICK) &&
3656 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3660 if (cfs_rq->nr_running > 1)
3661 check_preempt_tick(cfs_rq, curr);
3665 /**************************************************
3666 * CFS bandwidth control machinery
3669 #ifdef CONFIG_CFS_BANDWIDTH
3671 #ifdef HAVE_JUMP_LABEL
3672 static struct static_key __cfs_bandwidth_used;
3674 static inline bool cfs_bandwidth_used(void)
3676 return static_key_false(&__cfs_bandwidth_used);
3679 void cfs_bandwidth_usage_inc(void)
3681 static_key_slow_inc(&__cfs_bandwidth_used);
3684 void cfs_bandwidth_usage_dec(void)
3686 static_key_slow_dec(&__cfs_bandwidth_used);
3688 #else /* HAVE_JUMP_LABEL */
3689 static bool cfs_bandwidth_used(void)
3694 void cfs_bandwidth_usage_inc(void) {}
3695 void cfs_bandwidth_usage_dec(void) {}
3696 #endif /* HAVE_JUMP_LABEL */
3699 * default period for cfs group bandwidth.
3700 * default: 0.1s, units: nanoseconds
3702 static inline u64 default_cfs_period(void)
3704 return 100000000ULL;
3707 static inline u64 sched_cfs_bandwidth_slice(void)
3709 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3713 * Replenish runtime according to assigned quota and update expiration time.
3714 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3715 * additional synchronization around rq->lock.
3717 * requires cfs_b->lock
3719 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3723 if (cfs_b->quota == RUNTIME_INF)
3726 now = sched_clock_cpu(smp_processor_id());
3727 cfs_b->runtime = cfs_b->quota;
3728 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3731 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3733 return &tg->cfs_bandwidth;
3736 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3737 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3739 if (unlikely(cfs_rq->throttle_count))
3740 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3742 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3745 /* returns 0 on failure to allocate runtime */
3746 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3748 struct task_group *tg = cfs_rq->tg;
3749 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3750 u64 amount = 0, min_amount, expires;
3752 /* note: this is a positive sum as runtime_remaining <= 0 */
3753 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3755 raw_spin_lock(&cfs_b->lock);
3756 if (cfs_b->quota == RUNTIME_INF)
3757 amount = min_amount;
3759 start_cfs_bandwidth(cfs_b);
3761 if (cfs_b->runtime > 0) {
3762 amount = min(cfs_b->runtime, min_amount);
3763 cfs_b->runtime -= amount;
3767 expires = cfs_b->runtime_expires;
3768 raw_spin_unlock(&cfs_b->lock);
3770 cfs_rq->runtime_remaining += amount;
3772 * we may have advanced our local expiration to account for allowed
3773 * spread between our sched_clock and the one on which runtime was
3776 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3777 cfs_rq->runtime_expires = expires;
3779 return cfs_rq->runtime_remaining > 0;
3783 * Note: This depends on the synchronization provided by sched_clock and the
3784 * fact that rq->clock snapshots this value.
3786 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3788 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3790 /* if the deadline is ahead of our clock, nothing to do */
3791 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3794 if (cfs_rq->runtime_remaining < 0)
3798 * If the local deadline has passed we have to consider the
3799 * possibility that our sched_clock is 'fast' and the global deadline
3800 * has not truly expired.
3802 * Fortunately we can check determine whether this the case by checking
3803 * whether the global deadline has advanced. It is valid to compare
3804 * cfs_b->runtime_expires without any locks since we only care about
3805 * exact equality, so a partial write will still work.
3808 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3809 /* extend local deadline, drift is bounded above by 2 ticks */
3810 cfs_rq->runtime_expires += TICK_NSEC;
3812 /* global deadline is ahead, expiration has passed */
3813 cfs_rq->runtime_remaining = 0;
3817 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3819 /* dock delta_exec before expiring quota (as it could span periods) */
3820 cfs_rq->runtime_remaining -= delta_exec;
3821 expire_cfs_rq_runtime(cfs_rq);
3823 if (likely(cfs_rq->runtime_remaining > 0))
3827 * if we're unable to extend our runtime we resched so that the active
3828 * hierarchy can be throttled
3830 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3831 resched_curr(rq_of(cfs_rq));
3834 static __always_inline
3835 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3837 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3840 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3843 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3845 return cfs_bandwidth_used() && cfs_rq->throttled;
3848 /* check whether cfs_rq, or any parent, is throttled */
3849 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3851 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3855 * Ensure that neither of the group entities corresponding to src_cpu or
3856 * dest_cpu are members of a throttled hierarchy when performing group
3857 * load-balance operations.
3859 static inline int throttled_lb_pair(struct task_group *tg,
3860 int src_cpu, int dest_cpu)
3862 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3864 src_cfs_rq = tg->cfs_rq[src_cpu];
3865 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3867 return throttled_hierarchy(src_cfs_rq) ||
3868 throttled_hierarchy(dest_cfs_rq);
3871 /* updated child weight may affect parent so we have to do this bottom up */
3872 static int tg_unthrottle_up(struct task_group *tg, void *data)
3874 struct rq *rq = data;
3875 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3877 cfs_rq->throttle_count--;
3878 if (!cfs_rq->throttle_count) {
3879 /* adjust cfs_rq_clock_task() */
3880 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3881 cfs_rq->throttled_clock_task;
3887 static int tg_throttle_down(struct task_group *tg, void *data)
3889 struct rq *rq = data;
3890 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3892 /* group is entering throttled state, stop time */
3893 if (!cfs_rq->throttle_count)
3894 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3895 cfs_rq->throttle_count++;
3900 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3902 struct rq *rq = rq_of(cfs_rq);
3903 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3904 struct sched_entity *se;
3905 long task_delta, dequeue = 1;
3908 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3910 /* freeze hierarchy runnable averages while throttled */
3912 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3915 task_delta = cfs_rq->h_nr_running;
3916 for_each_sched_entity(se) {
3917 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3918 /* throttled entity or throttle-on-deactivate */
3923 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3924 qcfs_rq->h_nr_running -= task_delta;
3926 if (qcfs_rq->load.weight)
3931 sub_nr_running(rq, task_delta);
3933 cfs_rq->throttled = 1;
3934 cfs_rq->throttled_clock = rq_clock(rq);
3935 raw_spin_lock(&cfs_b->lock);
3936 empty = list_empty(&cfs_b->throttled_cfs_rq);
3939 * Add to the _head_ of the list, so that an already-started
3940 * distribute_cfs_runtime will not see us
3942 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3945 * If we're the first throttled task, make sure the bandwidth
3949 start_cfs_bandwidth(cfs_b);
3951 raw_spin_unlock(&cfs_b->lock);
3954 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3956 struct rq *rq = rq_of(cfs_rq);
3957 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3958 struct sched_entity *se;
3962 se = cfs_rq->tg->se[cpu_of(rq)];
3964 cfs_rq->throttled = 0;
3966 update_rq_clock(rq);
3968 raw_spin_lock(&cfs_b->lock);
3969 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3970 list_del_rcu(&cfs_rq->throttled_list);
3971 raw_spin_unlock(&cfs_b->lock);
3973 /* update hierarchical throttle state */
3974 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3976 if (!cfs_rq->load.weight)
3979 task_delta = cfs_rq->h_nr_running;
3980 for_each_sched_entity(se) {
3984 cfs_rq = cfs_rq_of(se);
3986 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3987 cfs_rq->h_nr_running += task_delta;
3989 if (cfs_rq_throttled(cfs_rq))
3994 add_nr_running(rq, task_delta);
3996 /* determine whether we need to wake up potentially idle cpu */
3997 if (rq->curr == rq->idle && rq->cfs.nr_running)
4001 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4002 u64 remaining, u64 expires)
4004 struct cfs_rq *cfs_rq;
4006 u64 starting_runtime = remaining;
4009 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4011 struct rq *rq = rq_of(cfs_rq);
4013 raw_spin_lock(&rq->lock);
4014 if (!cfs_rq_throttled(cfs_rq))
4017 runtime = -cfs_rq->runtime_remaining + 1;
4018 if (runtime > remaining)
4019 runtime = remaining;
4020 remaining -= runtime;
4022 cfs_rq->runtime_remaining += runtime;
4023 cfs_rq->runtime_expires = expires;
4025 /* we check whether we're throttled above */
4026 if (cfs_rq->runtime_remaining > 0)
4027 unthrottle_cfs_rq(cfs_rq);
4030 raw_spin_unlock(&rq->lock);
4037 return starting_runtime - remaining;
4041 * Responsible for refilling a task_group's bandwidth and unthrottling its
4042 * cfs_rqs as appropriate. If there has been no activity within the last
4043 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4044 * used to track this state.
4046 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4048 u64 runtime, runtime_expires;
4051 /* no need to continue the timer with no bandwidth constraint */
4052 if (cfs_b->quota == RUNTIME_INF)
4053 goto out_deactivate;
4055 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4056 cfs_b->nr_periods += overrun;
4059 * idle depends on !throttled (for the case of a large deficit), and if
4060 * we're going inactive then everything else can be deferred
4062 if (cfs_b->idle && !throttled)
4063 goto out_deactivate;
4065 __refill_cfs_bandwidth_runtime(cfs_b);
4068 /* mark as potentially idle for the upcoming period */
4073 /* account preceding periods in which throttling occurred */
4074 cfs_b->nr_throttled += overrun;
4076 runtime_expires = cfs_b->runtime_expires;
4079 * This check is repeated as we are holding onto the new bandwidth while
4080 * we unthrottle. This can potentially race with an unthrottled group
4081 * trying to acquire new bandwidth from the global pool. This can result
4082 * in us over-using our runtime if it is all used during this loop, but
4083 * only by limited amounts in that extreme case.
4085 while (throttled && cfs_b->runtime > 0) {
4086 runtime = cfs_b->runtime;
4087 raw_spin_unlock(&cfs_b->lock);
4088 /* we can't nest cfs_b->lock while distributing bandwidth */
4089 runtime = distribute_cfs_runtime(cfs_b, runtime,
4091 raw_spin_lock(&cfs_b->lock);
4093 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4095 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4099 * While we are ensured activity in the period following an
4100 * unthrottle, this also covers the case in which the new bandwidth is
4101 * insufficient to cover the existing bandwidth deficit. (Forcing the
4102 * timer to remain active while there are any throttled entities.)
4112 /* a cfs_rq won't donate quota below this amount */
4113 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4114 /* minimum remaining period time to redistribute slack quota */
4115 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4116 /* how long we wait to gather additional slack before distributing */
4117 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4120 * Are we near the end of the current quota period?
4122 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4123 * hrtimer base being cleared by hrtimer_start. In the case of
4124 * migrate_hrtimers, base is never cleared, so we are fine.
4126 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4128 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4131 /* if the call-back is running a quota refresh is already occurring */
4132 if (hrtimer_callback_running(refresh_timer))
4135 /* is a quota refresh about to occur? */
4136 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4137 if (remaining < min_expire)
4143 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4145 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4147 /* if there's a quota refresh soon don't bother with slack */
4148 if (runtime_refresh_within(cfs_b, min_left))
4151 hrtimer_start(&cfs_b->slack_timer,
4152 ns_to_ktime(cfs_bandwidth_slack_period),
4156 /* we know any runtime found here is valid as update_curr() precedes return */
4157 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4159 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4160 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4162 if (slack_runtime <= 0)
4165 raw_spin_lock(&cfs_b->lock);
4166 if (cfs_b->quota != RUNTIME_INF &&
4167 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4168 cfs_b->runtime += slack_runtime;
4170 /* we are under rq->lock, defer unthrottling using a timer */
4171 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4172 !list_empty(&cfs_b->throttled_cfs_rq))
4173 start_cfs_slack_bandwidth(cfs_b);
4175 raw_spin_unlock(&cfs_b->lock);
4177 /* even if it's not valid for return we don't want to try again */
4178 cfs_rq->runtime_remaining -= slack_runtime;
4181 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4183 if (!cfs_bandwidth_used())
4186 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4189 __return_cfs_rq_runtime(cfs_rq);
4193 * This is done with a timer (instead of inline with bandwidth return) since
4194 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4196 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4198 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4201 /* confirm we're still not at a refresh boundary */
4202 raw_spin_lock(&cfs_b->lock);
4203 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4204 raw_spin_unlock(&cfs_b->lock);
4208 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4209 runtime = cfs_b->runtime;
4211 expires = cfs_b->runtime_expires;
4212 raw_spin_unlock(&cfs_b->lock);
4217 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4219 raw_spin_lock(&cfs_b->lock);
4220 if (expires == cfs_b->runtime_expires)
4221 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4222 raw_spin_unlock(&cfs_b->lock);
4226 * When a group wakes up we want to make sure that its quota is not already
4227 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4228 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4230 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4232 if (!cfs_bandwidth_used())
4235 /* an active group must be handled by the update_curr()->put() path */
4236 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4239 /* ensure the group is not already throttled */
4240 if (cfs_rq_throttled(cfs_rq))
4243 /* update runtime allocation */
4244 account_cfs_rq_runtime(cfs_rq, 0);
4245 if (cfs_rq->runtime_remaining <= 0)
4246 throttle_cfs_rq(cfs_rq);
4249 static void sync_throttle(struct task_group *tg, int cpu)
4251 struct cfs_rq *pcfs_rq, *cfs_rq;
4253 if (!cfs_bandwidth_used())
4259 cfs_rq = tg->cfs_rq[cpu];
4260 pcfs_rq = tg->parent->cfs_rq[cpu];
4262 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4263 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4266 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4267 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4269 if (!cfs_bandwidth_used())
4272 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4276 * it's possible for a throttled entity to be forced into a running
4277 * state (e.g. set_curr_task), in this case we're finished.
4279 if (cfs_rq_throttled(cfs_rq))
4282 throttle_cfs_rq(cfs_rq);
4286 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4288 struct cfs_bandwidth *cfs_b =
4289 container_of(timer, struct cfs_bandwidth, slack_timer);
4291 do_sched_cfs_slack_timer(cfs_b);
4293 return HRTIMER_NORESTART;
4296 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4298 struct cfs_bandwidth *cfs_b =
4299 container_of(timer, struct cfs_bandwidth, period_timer);
4303 raw_spin_lock(&cfs_b->lock);
4305 overrun = hrtimer_forward_now(timer, cfs_b->period);
4309 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4312 cfs_b->period_active = 0;
4313 raw_spin_unlock(&cfs_b->lock);
4315 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4318 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4320 raw_spin_lock_init(&cfs_b->lock);
4322 cfs_b->quota = RUNTIME_INF;
4323 cfs_b->period = ns_to_ktime(default_cfs_period());
4325 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4326 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4327 cfs_b->period_timer.function = sched_cfs_period_timer;
4328 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4329 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4332 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4334 cfs_rq->runtime_enabled = 0;
4335 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4338 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4340 lockdep_assert_held(&cfs_b->lock);
4342 if (!cfs_b->period_active) {
4343 cfs_b->period_active = 1;
4344 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4345 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4349 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4351 /* init_cfs_bandwidth() was not called */
4352 if (!cfs_b->throttled_cfs_rq.next)
4355 hrtimer_cancel(&cfs_b->period_timer);
4356 hrtimer_cancel(&cfs_b->slack_timer);
4359 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4361 struct cfs_rq *cfs_rq;
4363 for_each_leaf_cfs_rq(rq, cfs_rq) {
4364 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4366 raw_spin_lock(&cfs_b->lock);
4367 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4368 raw_spin_unlock(&cfs_b->lock);
4372 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4374 struct cfs_rq *cfs_rq;
4376 for_each_leaf_cfs_rq(rq, cfs_rq) {
4377 if (!cfs_rq->runtime_enabled)
4381 * clock_task is not advancing so we just need to make sure
4382 * there's some valid quota amount
4384 cfs_rq->runtime_remaining = 1;
4386 * Offline rq is schedulable till cpu is completely disabled
4387 * in take_cpu_down(), so we prevent new cfs throttling here.
4389 cfs_rq->runtime_enabled = 0;
4391 if (cfs_rq_throttled(cfs_rq))
4392 unthrottle_cfs_rq(cfs_rq);
4396 #else /* CONFIG_CFS_BANDWIDTH */
4397 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4399 return rq_clock_task(rq_of(cfs_rq));
4402 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4403 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4404 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4405 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4406 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4408 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4413 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4418 static inline int throttled_lb_pair(struct task_group *tg,
4419 int src_cpu, int dest_cpu)
4424 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4426 #ifdef CONFIG_FAIR_GROUP_SCHED
4427 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4430 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4434 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4435 static inline void update_runtime_enabled(struct rq *rq) {}
4436 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4438 #endif /* CONFIG_CFS_BANDWIDTH */
4440 /**************************************************
4441 * CFS operations on tasks:
4444 #ifdef CONFIG_SCHED_HRTICK
4445 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4447 struct sched_entity *se = &p->se;
4448 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4450 WARN_ON(task_rq(p) != rq);
4452 if (cfs_rq->nr_running > 1) {
4453 u64 slice = sched_slice(cfs_rq, se);
4454 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4455 s64 delta = slice - ran;
4462 hrtick_start(rq, delta);
4467 * called from enqueue/dequeue and updates the hrtick when the
4468 * current task is from our class and nr_running is low enough
4471 static void hrtick_update(struct rq *rq)
4473 struct task_struct *curr = rq->curr;
4475 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4478 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4479 hrtick_start_fair(rq, curr);
4481 #else /* !CONFIG_SCHED_HRTICK */
4483 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4487 static inline void hrtick_update(struct rq *rq)
4493 * The enqueue_task method is called before nr_running is
4494 * increased. Here we update the fair scheduling stats and
4495 * then put the task into the rbtree:
4498 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4500 struct cfs_rq *cfs_rq;
4501 struct sched_entity *se = &p->se;
4504 * If in_iowait is set, the code below may not trigger any cpufreq
4505 * utilization updates, so do it here explicitly with the IOWAIT flag
4509 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4511 for_each_sched_entity(se) {
4514 cfs_rq = cfs_rq_of(se);
4515 enqueue_entity(cfs_rq, se, flags);
4518 * end evaluation on encountering a throttled cfs_rq
4520 * note: in the case of encountering a throttled cfs_rq we will
4521 * post the final h_nr_running increment below.
4523 if (cfs_rq_throttled(cfs_rq))
4525 cfs_rq->h_nr_running++;
4527 flags = ENQUEUE_WAKEUP;
4530 for_each_sched_entity(se) {
4531 cfs_rq = cfs_rq_of(se);
4532 cfs_rq->h_nr_running++;
4534 if (cfs_rq_throttled(cfs_rq))
4537 update_load_avg(se, 1);
4538 update_cfs_shares(cfs_rq);
4542 add_nr_running(rq, 1);
4547 static void set_next_buddy(struct sched_entity *se);
4550 * The dequeue_task method is called before nr_running is
4551 * decreased. We remove the task from the rbtree and
4552 * update the fair scheduling stats:
4554 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4556 struct cfs_rq *cfs_rq;
4557 struct sched_entity *se = &p->se;
4558 int task_sleep = flags & DEQUEUE_SLEEP;
4560 for_each_sched_entity(se) {
4561 cfs_rq = cfs_rq_of(se);
4562 dequeue_entity(cfs_rq, se, flags);
4565 * end evaluation on encountering a throttled cfs_rq
4567 * note: in the case of encountering a throttled cfs_rq we will
4568 * post the final h_nr_running decrement below.
4570 if (cfs_rq_throttled(cfs_rq))
4572 cfs_rq->h_nr_running--;
4574 /* Don't dequeue parent if it has other entities besides us */
4575 if (cfs_rq->load.weight) {
4576 /* Avoid re-evaluating load for this entity: */
4577 se = parent_entity(se);
4579 * Bias pick_next to pick a task from this cfs_rq, as
4580 * p is sleeping when it is within its sched_slice.
4582 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4586 flags |= DEQUEUE_SLEEP;
4589 for_each_sched_entity(se) {
4590 cfs_rq = cfs_rq_of(se);
4591 cfs_rq->h_nr_running--;
4593 if (cfs_rq_throttled(cfs_rq))
4596 update_load_avg(se, 1);
4597 update_cfs_shares(cfs_rq);
4601 sub_nr_running(rq, 1);
4607 #ifdef CONFIG_NO_HZ_COMMON
4609 * per rq 'load' arrray crap; XXX kill this.
4613 * The exact cpuload calculated at every tick would be:
4615 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4617 * If a cpu misses updates for n ticks (as it was idle) and update gets
4618 * called on the n+1-th tick when cpu may be busy, then we have:
4620 * load_n = (1 - 1/2^i)^n * load_0
4621 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4623 * decay_load_missed() below does efficient calculation of
4625 * load' = (1 - 1/2^i)^n * load
4627 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4628 * This allows us to precompute the above in said factors, thereby allowing the
4629 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4630 * fixed_power_int())
4632 * The calculation is approximated on a 128 point scale.
4634 #define DEGRADE_SHIFT 7
4636 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4637 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4638 { 0, 0, 0, 0, 0, 0, 0, 0 },
4639 { 64, 32, 8, 0, 0, 0, 0, 0 },
4640 { 96, 72, 40, 12, 1, 0, 0, 0 },
4641 { 112, 98, 75, 43, 15, 1, 0, 0 },
4642 { 120, 112, 98, 76, 45, 16, 2, 0 }
4646 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4647 * would be when CPU is idle and so we just decay the old load without
4648 * adding any new load.
4650 static unsigned long
4651 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4655 if (!missed_updates)
4658 if (missed_updates >= degrade_zero_ticks[idx])
4662 return load >> missed_updates;
4664 while (missed_updates) {
4665 if (missed_updates % 2)
4666 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4668 missed_updates >>= 1;
4673 #endif /* CONFIG_NO_HZ_COMMON */
4676 * __cpu_load_update - update the rq->cpu_load[] statistics
4677 * @this_rq: The rq to update statistics for
4678 * @this_load: The current load
4679 * @pending_updates: The number of missed updates
4681 * Update rq->cpu_load[] statistics. This function is usually called every
4682 * scheduler tick (TICK_NSEC).
4684 * This function computes a decaying average:
4686 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4688 * Because of NOHZ it might not get called on every tick which gives need for
4689 * the @pending_updates argument.
4691 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4692 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4693 * = A * (A * load[i]_n-2 + B) + B
4694 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4695 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4696 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4697 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4698 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4700 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4701 * any change in load would have resulted in the tick being turned back on.
4703 * For regular NOHZ, this reduces to:
4705 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4707 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4710 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4711 unsigned long pending_updates)
4713 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4716 this_rq->nr_load_updates++;
4718 /* Update our load: */
4719 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4720 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4721 unsigned long old_load, new_load;
4723 /* scale is effectively 1 << i now, and >> i divides by scale */
4725 old_load = this_rq->cpu_load[i];
4726 #ifdef CONFIG_NO_HZ_COMMON
4727 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4728 if (tickless_load) {
4729 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4731 * old_load can never be a negative value because a
4732 * decayed tickless_load cannot be greater than the
4733 * original tickless_load.
4735 old_load += tickless_load;
4738 new_load = this_load;
4740 * Round up the averaging division if load is increasing. This
4741 * prevents us from getting stuck on 9 if the load is 10, for
4744 if (new_load > old_load)
4745 new_load += scale - 1;
4747 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4750 sched_avg_update(this_rq);
4753 /* Used instead of source_load when we know the type == 0 */
4754 static unsigned long weighted_cpuload(const int cpu)
4756 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4759 #ifdef CONFIG_NO_HZ_COMMON
4761 * There is no sane way to deal with nohz on smp when using jiffies because the
4762 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4763 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4765 * Therefore we need to avoid the delta approach from the regular tick when
4766 * possible since that would seriously skew the load calculation. This is why we
4767 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4768 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4769 * loop exit, nohz_idle_balance, nohz full exit...)
4771 * This means we might still be one tick off for nohz periods.
4774 static void cpu_load_update_nohz(struct rq *this_rq,
4775 unsigned long curr_jiffies,
4778 unsigned long pending_updates;
4780 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4781 if (pending_updates) {
4782 this_rq->last_load_update_tick = curr_jiffies;
4784 * In the regular NOHZ case, we were idle, this means load 0.
4785 * In the NOHZ_FULL case, we were non-idle, we should consider
4786 * its weighted load.
4788 cpu_load_update(this_rq, load, pending_updates);
4793 * Called from nohz_idle_balance() to update the load ratings before doing the
4796 static void cpu_load_update_idle(struct rq *this_rq)
4799 * bail if there's load or we're actually up-to-date.
4801 if (weighted_cpuload(cpu_of(this_rq)))
4804 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4808 * Record CPU load on nohz entry so we know the tickless load to account
4809 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4810 * than other cpu_load[idx] but it should be fine as cpu_load readers
4811 * shouldn't rely into synchronized cpu_load[*] updates.
4813 void cpu_load_update_nohz_start(void)
4815 struct rq *this_rq = this_rq();
4818 * This is all lockless but should be fine. If weighted_cpuload changes
4819 * concurrently we'll exit nohz. And cpu_load write can race with
4820 * cpu_load_update_idle() but both updater would be writing the same.
4822 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4826 * Account the tickless load in the end of a nohz frame.
4828 void cpu_load_update_nohz_stop(void)
4830 unsigned long curr_jiffies = READ_ONCE(jiffies);
4831 struct rq *this_rq = this_rq();
4834 if (curr_jiffies == this_rq->last_load_update_tick)
4837 load = weighted_cpuload(cpu_of(this_rq));
4838 raw_spin_lock(&this_rq->lock);
4839 update_rq_clock(this_rq);
4840 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4841 raw_spin_unlock(&this_rq->lock);
4843 #else /* !CONFIG_NO_HZ_COMMON */
4844 static inline void cpu_load_update_nohz(struct rq *this_rq,
4845 unsigned long curr_jiffies,
4846 unsigned long load) { }
4847 #endif /* CONFIG_NO_HZ_COMMON */
4849 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4851 #ifdef CONFIG_NO_HZ_COMMON
4852 /* See the mess around cpu_load_update_nohz(). */
4853 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4855 cpu_load_update(this_rq, load, 1);
4859 * Called from scheduler_tick()
4861 void cpu_load_update_active(struct rq *this_rq)
4863 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4865 if (tick_nohz_tick_stopped())
4866 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4868 cpu_load_update_periodic(this_rq, load);
4872 * Return a low guess at the load of a migration-source cpu weighted
4873 * according to the scheduling class and "nice" value.
4875 * We want to under-estimate the load of migration sources, to
4876 * balance conservatively.
4878 static unsigned long source_load(int cpu, int type)
4880 struct rq *rq = cpu_rq(cpu);
4881 unsigned long total = weighted_cpuload(cpu);
4883 if (type == 0 || !sched_feat(LB_BIAS))
4886 return min(rq->cpu_load[type-1], total);
4890 * Return a high guess at the load of a migration-target cpu weighted
4891 * according to the scheduling class and "nice" value.
4893 static unsigned long target_load(int cpu, int type)
4895 struct rq *rq = cpu_rq(cpu);
4896 unsigned long total = weighted_cpuload(cpu);
4898 if (type == 0 || !sched_feat(LB_BIAS))
4901 return max(rq->cpu_load[type-1], total);
4904 static unsigned long capacity_of(int cpu)
4906 return cpu_rq(cpu)->cpu_capacity;
4909 static unsigned long capacity_orig_of(int cpu)
4911 return cpu_rq(cpu)->cpu_capacity_orig;
4914 static unsigned long cpu_avg_load_per_task(int cpu)
4916 struct rq *rq = cpu_rq(cpu);
4917 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4918 unsigned long load_avg = weighted_cpuload(cpu);
4921 return load_avg / nr_running;
4926 #ifdef CONFIG_FAIR_GROUP_SCHED
4928 * effective_load() calculates the load change as seen from the root_task_group
4930 * Adding load to a group doesn't make a group heavier, but can cause movement
4931 * of group shares between cpus. Assuming the shares were perfectly aligned one
4932 * can calculate the shift in shares.
4934 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4935 * on this @cpu and results in a total addition (subtraction) of @wg to the
4936 * total group weight.
4938 * Given a runqueue weight distribution (rw_i) we can compute a shares
4939 * distribution (s_i) using:
4941 * s_i = rw_i / \Sum rw_j (1)
4943 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4944 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4945 * shares distribution (s_i):
4947 * rw_i = { 2, 4, 1, 0 }
4948 * s_i = { 2/7, 4/7, 1/7, 0 }
4950 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4951 * task used to run on and the CPU the waker is running on), we need to
4952 * compute the effect of waking a task on either CPU and, in case of a sync
4953 * wakeup, compute the effect of the current task going to sleep.
4955 * So for a change of @wl to the local @cpu with an overall group weight change
4956 * of @wl we can compute the new shares distribution (s'_i) using:
4958 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4960 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4961 * differences in waking a task to CPU 0. The additional task changes the
4962 * weight and shares distributions like:
4964 * rw'_i = { 3, 4, 1, 0 }
4965 * s'_i = { 3/8, 4/8, 1/8, 0 }
4967 * We can then compute the difference in effective weight by using:
4969 * dw_i = S * (s'_i - s_i) (3)
4971 * Where 'S' is the group weight as seen by its parent.
4973 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4974 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4975 * 4/7) times the weight of the group.
4977 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4979 struct sched_entity *se = tg->se[cpu];
4981 if (!tg->parent) /* the trivial, non-cgroup case */
4984 for_each_sched_entity(se) {
4985 struct cfs_rq *cfs_rq = se->my_q;
4986 long W, w = cfs_rq_load_avg(cfs_rq);
4991 * W = @wg + \Sum rw_j
4993 W = wg + atomic_long_read(&tg->load_avg);
4995 /* Ensure \Sum rw_j >= rw_i */
4996 W -= cfs_rq->tg_load_avg_contrib;
5005 * wl = S * s'_i; see (2)
5008 wl = (w * (long)tg->shares) / W;
5013 * Per the above, wl is the new se->load.weight value; since
5014 * those are clipped to [MIN_SHARES, ...) do so now. See
5015 * calc_cfs_shares().
5017 if (wl < MIN_SHARES)
5021 * wl = dw_i = S * (s'_i - s_i); see (3)
5023 wl -= se->avg.load_avg;
5026 * Recursively apply this logic to all parent groups to compute
5027 * the final effective load change on the root group. Since
5028 * only the @tg group gets extra weight, all parent groups can
5029 * only redistribute existing shares. @wl is the shift in shares
5030 * resulting from this level per the above.
5039 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5046 static void record_wakee(struct task_struct *p)
5049 * Only decay a single time; tasks that have less then 1 wakeup per
5050 * jiffy will not have built up many flips.
5052 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5053 current->wakee_flips >>= 1;
5054 current->wakee_flip_decay_ts = jiffies;
5057 if (current->last_wakee != p) {
5058 current->last_wakee = p;
5059 current->wakee_flips++;
5064 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5066 * A waker of many should wake a different task than the one last awakened
5067 * at a frequency roughly N times higher than one of its wakees.
5069 * In order to determine whether we should let the load spread vs consolidating
5070 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5071 * partner, and a factor of lls_size higher frequency in the other.
5073 * With both conditions met, we can be relatively sure that the relationship is
5074 * non-monogamous, with partner count exceeding socket size.
5076 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5077 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5080 static int wake_wide(struct task_struct *p)
5082 unsigned int master = current->wakee_flips;
5083 unsigned int slave = p->wakee_flips;
5084 int factor = this_cpu_read(sd_llc_size);
5087 swap(master, slave);
5088 if (slave < factor || master < slave * factor)
5093 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5095 s64 this_load, load;
5096 s64 this_eff_load, prev_eff_load;
5097 int idx, this_cpu, prev_cpu;
5098 struct task_group *tg;
5099 unsigned long weight;
5103 this_cpu = smp_processor_id();
5104 prev_cpu = task_cpu(p);
5105 load = source_load(prev_cpu, idx);
5106 this_load = target_load(this_cpu, idx);
5109 * If sync wakeup then subtract the (maximum possible)
5110 * effect of the currently running task from the load
5111 * of the current CPU:
5114 tg = task_group(current);
5115 weight = current->se.avg.load_avg;
5117 this_load += effective_load(tg, this_cpu, -weight, -weight);
5118 load += effective_load(tg, prev_cpu, 0, -weight);
5122 weight = p->se.avg.load_avg;
5125 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5126 * due to the sync cause above having dropped this_load to 0, we'll
5127 * always have an imbalance, but there's really nothing you can do
5128 * about that, so that's good too.
5130 * Otherwise check if either cpus are near enough in load to allow this
5131 * task to be woken on this_cpu.
5133 this_eff_load = 100;
5134 this_eff_load *= capacity_of(prev_cpu);
5136 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5137 prev_eff_load *= capacity_of(this_cpu);
5139 if (this_load > 0) {
5140 this_eff_load *= this_load +
5141 effective_load(tg, this_cpu, weight, weight);
5143 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5146 balanced = this_eff_load <= prev_eff_load;
5148 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5153 schedstat_inc(sd, ttwu_move_affine);
5154 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5160 * find_idlest_group finds and returns the least busy CPU group within the
5163 static struct sched_group *
5164 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5165 int this_cpu, int sd_flag)
5167 struct sched_group *idlest = NULL, *group = sd->groups;
5168 unsigned long min_load = ULONG_MAX, this_load = 0;
5169 int load_idx = sd->forkexec_idx;
5170 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5172 if (sd_flag & SD_BALANCE_WAKE)
5173 load_idx = sd->wake_idx;
5176 unsigned long load, avg_load;
5180 /* Skip over this group if it has no CPUs allowed */
5181 if (!cpumask_intersects(sched_group_cpus(group),
5182 tsk_cpus_allowed(p)))
5185 local_group = cpumask_test_cpu(this_cpu,
5186 sched_group_cpus(group));
5188 /* Tally up the load of all CPUs in the group */
5191 for_each_cpu(i, sched_group_cpus(group)) {
5192 /* Bias balancing toward cpus of our domain */
5194 load = source_load(i, load_idx);
5196 load = target_load(i, load_idx);
5201 /* Adjust by relative CPU capacity of the group */
5202 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5205 this_load = avg_load;
5206 } else if (avg_load < min_load) {
5207 min_load = avg_load;
5210 } while (group = group->next, group != sd->groups);
5212 if (!idlest || 100*this_load < imbalance*min_load)
5218 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5221 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5223 unsigned long load, min_load = ULONG_MAX;
5224 unsigned int min_exit_latency = UINT_MAX;
5225 u64 latest_idle_timestamp = 0;
5226 int least_loaded_cpu = this_cpu;
5227 int shallowest_idle_cpu = -1;
5230 /* Traverse only the allowed CPUs */
5231 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5233 struct rq *rq = cpu_rq(i);
5234 struct cpuidle_state *idle = idle_get_state(rq);
5235 if (idle && idle->exit_latency < min_exit_latency) {
5237 * We give priority to a CPU whose idle state
5238 * has the smallest exit latency irrespective
5239 * of any idle timestamp.
5241 min_exit_latency = idle->exit_latency;
5242 latest_idle_timestamp = rq->idle_stamp;
5243 shallowest_idle_cpu = i;
5244 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5245 rq->idle_stamp > latest_idle_timestamp) {
5247 * If equal or no active idle state, then
5248 * the most recently idled CPU might have
5251 latest_idle_timestamp = rq->idle_stamp;
5252 shallowest_idle_cpu = i;
5254 } else if (shallowest_idle_cpu == -1) {
5255 load = weighted_cpuload(i);
5256 if (load < min_load || (load == min_load && i == this_cpu)) {
5258 least_loaded_cpu = i;
5263 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5267 * Try and locate an idle CPU in the sched_domain.
5269 static int select_idle_sibling(struct task_struct *p, int target)
5271 struct sched_domain *sd;
5272 struct sched_group *sg;
5273 int i = task_cpu(p);
5275 if (idle_cpu(target))
5279 * If the prevous cpu is cache affine and idle, don't be stupid.
5281 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5285 * Otherwise, iterate the domains and find an eligible idle cpu.
5287 * A completely idle sched group at higher domains is more
5288 * desirable than an idle group at a lower level, because lower
5289 * domains have smaller groups and usually share hardware
5290 * resources which causes tasks to contend on them, e.g. x86
5291 * hyperthread siblings in the lowest domain (SMT) can contend
5292 * on the shared cpu pipeline.
5294 * However, while we prefer idle groups at higher domains
5295 * finding an idle cpu at the lowest domain is still better than
5296 * returning 'target', which we've already established, isn't
5299 sd = rcu_dereference(per_cpu(sd_llc, target));
5300 for_each_lower_domain(sd) {
5303 if (!cpumask_intersects(sched_group_cpus(sg),
5304 tsk_cpus_allowed(p)))
5307 /* Ensure the entire group is idle */
5308 for_each_cpu(i, sched_group_cpus(sg)) {
5309 if (i == target || !idle_cpu(i))
5314 * It doesn't matter which cpu we pick, the
5315 * whole group is idle.
5317 target = cpumask_first_and(sched_group_cpus(sg),
5318 tsk_cpus_allowed(p));
5322 } while (sg != sd->groups);
5329 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5330 * tasks. The unit of the return value must be the one of capacity so we can
5331 * compare the utilization with the capacity of the CPU that is available for
5332 * CFS task (ie cpu_capacity).
5334 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5335 * recent utilization of currently non-runnable tasks on a CPU. It represents
5336 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5337 * capacity_orig is the cpu_capacity available at the highest frequency
5338 * (arch_scale_freq_capacity()).
5339 * The utilization of a CPU converges towards a sum equal to or less than the
5340 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5341 * the running time on this CPU scaled by capacity_curr.
5343 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5344 * higher than capacity_orig because of unfortunate rounding in
5345 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5346 * the average stabilizes with the new running time. We need to check that the
5347 * utilization stays within the range of [0..capacity_orig] and cap it if
5348 * necessary. Without utilization capping, a group could be seen as overloaded
5349 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5350 * available capacity. We allow utilization to overshoot capacity_curr (but not
5351 * capacity_orig) as it useful for predicting the capacity required after task
5352 * migrations (scheduler-driven DVFS).
5354 static int cpu_util(int cpu)
5356 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5357 unsigned long capacity = capacity_orig_of(cpu);
5359 return (util >= capacity) ? capacity : util;
5363 * select_task_rq_fair: Select target runqueue for the waking task in domains
5364 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5365 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5367 * Balances load by selecting the idlest cpu in the idlest group, or under
5368 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5370 * Returns the target cpu number.
5372 * preempt must be disabled.
5375 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5377 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5378 int cpu = smp_processor_id();
5379 int new_cpu = prev_cpu;
5380 int want_affine = 0;
5381 int sync = wake_flags & WF_SYNC;
5383 if (sd_flag & SD_BALANCE_WAKE) {
5385 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5389 for_each_domain(cpu, tmp) {
5390 if (!(tmp->flags & SD_LOAD_BALANCE))
5394 * If both cpu and prev_cpu are part of this domain,
5395 * cpu is a valid SD_WAKE_AFFINE target.
5397 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5398 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5403 if (tmp->flags & sd_flag)
5405 else if (!want_affine)
5410 sd = NULL; /* Prefer wake_affine over balance flags */
5411 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5416 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5417 new_cpu = select_idle_sibling(p, new_cpu);
5420 struct sched_group *group;
5423 if (!(sd->flags & sd_flag)) {
5428 group = find_idlest_group(sd, p, cpu, sd_flag);
5434 new_cpu = find_idlest_cpu(group, p, cpu);
5435 if (new_cpu == -1 || new_cpu == cpu) {
5436 /* Now try balancing at a lower domain level of cpu */
5441 /* Now try balancing at a lower domain level of new_cpu */
5443 weight = sd->span_weight;
5445 for_each_domain(cpu, tmp) {
5446 if (weight <= tmp->span_weight)
5448 if (tmp->flags & sd_flag)
5451 /* while loop will break here if sd == NULL */
5459 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5460 * cfs_rq_of(p) references at time of call are still valid and identify the
5461 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5463 static void migrate_task_rq_fair(struct task_struct *p)
5466 * As blocked tasks retain absolute vruntime the migration needs to
5467 * deal with this by subtracting the old and adding the new
5468 * min_vruntime -- the latter is done by enqueue_entity() when placing
5469 * the task on the new runqueue.
5471 if (p->state == TASK_WAKING) {
5472 struct sched_entity *se = &p->se;
5473 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5476 #ifndef CONFIG_64BIT
5477 u64 min_vruntime_copy;
5480 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5482 min_vruntime = cfs_rq->min_vruntime;
5483 } while (min_vruntime != min_vruntime_copy);
5485 min_vruntime = cfs_rq->min_vruntime;
5488 se->vruntime -= min_vruntime;
5492 * We are supposed to update the task to "current" time, then its up to date
5493 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5494 * what current time is, so simply throw away the out-of-date time. This
5495 * will result in the wakee task is less decayed, but giving the wakee more
5496 * load sounds not bad.
5498 remove_entity_load_avg(&p->se);
5500 /* Tell new CPU we are migrated */
5501 p->se.avg.last_update_time = 0;
5503 /* We have migrated, no longer consider this task hot */
5504 p->se.exec_start = 0;
5507 static void task_dead_fair(struct task_struct *p)
5509 remove_entity_load_avg(&p->se);
5511 #endif /* CONFIG_SMP */
5513 static unsigned long
5514 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5516 unsigned long gran = sysctl_sched_wakeup_granularity;
5519 * Since its curr running now, convert the gran from real-time
5520 * to virtual-time in his units.
5522 * By using 'se' instead of 'curr' we penalize light tasks, so
5523 * they get preempted easier. That is, if 'se' < 'curr' then
5524 * the resulting gran will be larger, therefore penalizing the
5525 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5526 * be smaller, again penalizing the lighter task.
5528 * This is especially important for buddies when the leftmost
5529 * task is higher priority than the buddy.
5531 return calc_delta_fair(gran, se);
5535 * Should 'se' preempt 'curr'.
5549 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5551 s64 gran, vdiff = curr->vruntime - se->vruntime;
5556 gran = wakeup_gran(curr, se);
5563 static void set_last_buddy(struct sched_entity *se)
5565 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5568 for_each_sched_entity(se)
5569 cfs_rq_of(se)->last = se;
5572 static void set_next_buddy(struct sched_entity *se)
5574 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5577 for_each_sched_entity(se)
5578 cfs_rq_of(se)->next = se;
5581 static void set_skip_buddy(struct sched_entity *se)
5583 for_each_sched_entity(se)
5584 cfs_rq_of(se)->skip = se;
5588 * Preempt the current task with a newly woken task if needed:
5590 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5592 struct task_struct *curr = rq->curr;
5593 struct sched_entity *se = &curr->se, *pse = &p->se;
5594 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5595 int scale = cfs_rq->nr_running >= sched_nr_latency;
5596 int next_buddy_marked = 0;
5598 if (unlikely(se == pse))
5602 * This is possible from callers such as attach_tasks(), in which we
5603 * unconditionally check_prempt_curr() after an enqueue (which may have
5604 * lead to a throttle). This both saves work and prevents false
5605 * next-buddy nomination below.
5607 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5610 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5611 set_next_buddy(pse);
5612 next_buddy_marked = 1;
5616 * We can come here with TIF_NEED_RESCHED already set from new task
5619 * Note: this also catches the edge-case of curr being in a throttled
5620 * group (e.g. via set_curr_task), since update_curr() (in the
5621 * enqueue of curr) will have resulted in resched being set. This
5622 * prevents us from potentially nominating it as a false LAST_BUDDY
5625 if (test_tsk_need_resched(curr))
5628 /* Idle tasks are by definition preempted by non-idle tasks. */
5629 if (unlikely(curr->policy == SCHED_IDLE) &&
5630 likely(p->policy != SCHED_IDLE))
5634 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5635 * is driven by the tick):
5637 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5640 find_matching_se(&se, &pse);
5641 update_curr(cfs_rq_of(se));
5643 if (wakeup_preempt_entity(se, pse) == 1) {
5645 * Bias pick_next to pick the sched entity that is
5646 * triggering this preemption.
5648 if (!next_buddy_marked)
5649 set_next_buddy(pse);
5658 * Only set the backward buddy when the current task is still
5659 * on the rq. This can happen when a wakeup gets interleaved
5660 * with schedule on the ->pre_schedule() or idle_balance()
5661 * point, either of which can * drop the rq lock.
5663 * Also, during early boot the idle thread is in the fair class,
5664 * for obvious reasons its a bad idea to schedule back to it.
5666 if (unlikely(!se->on_rq || curr == rq->idle))
5669 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5673 static struct task_struct *
5674 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5676 struct cfs_rq *cfs_rq = &rq->cfs;
5677 struct sched_entity *se;
5678 struct task_struct *p;
5682 #ifdef CONFIG_FAIR_GROUP_SCHED
5683 if (!cfs_rq->nr_running)
5686 if (prev->sched_class != &fair_sched_class)
5690 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5691 * likely that a next task is from the same cgroup as the current.
5693 * Therefore attempt to avoid putting and setting the entire cgroup
5694 * hierarchy, only change the part that actually changes.
5698 struct sched_entity *curr = cfs_rq->curr;
5701 * Since we got here without doing put_prev_entity() we also
5702 * have to consider cfs_rq->curr. If it is still a runnable
5703 * entity, update_curr() will update its vruntime, otherwise
5704 * forget we've ever seen it.
5708 update_curr(cfs_rq);
5713 * This call to check_cfs_rq_runtime() will do the
5714 * throttle and dequeue its entity in the parent(s).
5715 * Therefore the 'simple' nr_running test will indeed
5718 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5722 se = pick_next_entity(cfs_rq, curr);
5723 cfs_rq = group_cfs_rq(se);
5729 * Since we haven't yet done put_prev_entity and if the selected task
5730 * is a different task than we started out with, try and touch the
5731 * least amount of cfs_rqs.
5734 struct sched_entity *pse = &prev->se;
5736 while (!(cfs_rq = is_same_group(se, pse))) {
5737 int se_depth = se->depth;
5738 int pse_depth = pse->depth;
5740 if (se_depth <= pse_depth) {
5741 put_prev_entity(cfs_rq_of(pse), pse);
5742 pse = parent_entity(pse);
5744 if (se_depth >= pse_depth) {
5745 set_next_entity(cfs_rq_of(se), se);
5746 se = parent_entity(se);
5750 put_prev_entity(cfs_rq, pse);
5751 set_next_entity(cfs_rq, se);
5754 if (hrtick_enabled(rq))
5755 hrtick_start_fair(rq, p);
5762 if (!cfs_rq->nr_running)
5765 put_prev_task(rq, prev);
5768 se = pick_next_entity(cfs_rq, NULL);
5769 set_next_entity(cfs_rq, se);
5770 cfs_rq = group_cfs_rq(se);
5775 if (hrtick_enabled(rq))
5776 hrtick_start_fair(rq, p);
5782 * This is OK, because current is on_cpu, which avoids it being picked
5783 * for load-balance and preemption/IRQs are still disabled avoiding
5784 * further scheduler activity on it and we're being very careful to
5785 * re-start the picking loop.
5787 lockdep_unpin_lock(&rq->lock, cookie);
5788 new_tasks = idle_balance(rq);
5789 lockdep_repin_lock(&rq->lock, cookie);
5791 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5792 * possible for any higher priority task to appear. In that case we
5793 * must re-start the pick_next_entity() loop.
5805 * Account for a descheduled task:
5807 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5809 struct sched_entity *se = &prev->se;
5810 struct cfs_rq *cfs_rq;
5812 for_each_sched_entity(se) {
5813 cfs_rq = cfs_rq_of(se);
5814 put_prev_entity(cfs_rq, se);
5819 * sched_yield() is very simple
5821 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5823 static void yield_task_fair(struct rq *rq)
5825 struct task_struct *curr = rq->curr;
5826 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5827 struct sched_entity *se = &curr->se;
5830 * Are we the only task in the tree?
5832 if (unlikely(rq->nr_running == 1))
5835 clear_buddies(cfs_rq, se);
5837 if (curr->policy != SCHED_BATCH) {
5838 update_rq_clock(rq);
5840 * Update run-time statistics of the 'current'.
5842 update_curr(cfs_rq);
5844 * Tell update_rq_clock() that we've just updated,
5845 * so we don't do microscopic update in schedule()
5846 * and double the fastpath cost.
5848 rq_clock_skip_update(rq, true);
5854 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5856 struct sched_entity *se = &p->se;
5858 /* throttled hierarchies are not runnable */
5859 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5862 /* Tell the scheduler that we'd really like pse to run next. */
5865 yield_task_fair(rq);
5871 /**************************************************
5872 * Fair scheduling class load-balancing methods.
5876 * The purpose of load-balancing is to achieve the same basic fairness the
5877 * per-cpu scheduler provides, namely provide a proportional amount of compute
5878 * time to each task. This is expressed in the following equation:
5880 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5882 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5883 * W_i,0 is defined as:
5885 * W_i,0 = \Sum_j w_i,j (2)
5887 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5888 * is derived from the nice value as per sched_prio_to_weight[].
5890 * The weight average is an exponential decay average of the instantaneous
5893 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5895 * C_i is the compute capacity of cpu i, typically it is the
5896 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5897 * can also include other factors [XXX].
5899 * To achieve this balance we define a measure of imbalance which follows
5900 * directly from (1):
5902 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5904 * We them move tasks around to minimize the imbalance. In the continuous
5905 * function space it is obvious this converges, in the discrete case we get
5906 * a few fun cases generally called infeasible weight scenarios.
5909 * - infeasible weights;
5910 * - local vs global optima in the discrete case. ]
5915 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5916 * for all i,j solution, we create a tree of cpus that follows the hardware
5917 * topology where each level pairs two lower groups (or better). This results
5918 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5919 * tree to only the first of the previous level and we decrease the frequency
5920 * of load-balance at each level inv. proportional to the number of cpus in
5926 * \Sum { --- * --- * 2^i } = O(n) (5)
5928 * `- size of each group
5929 * | | `- number of cpus doing load-balance
5931 * `- sum over all levels
5933 * Coupled with a limit on how many tasks we can migrate every balance pass,
5934 * this makes (5) the runtime complexity of the balancer.
5936 * An important property here is that each CPU is still (indirectly) connected
5937 * to every other cpu in at most O(log n) steps:
5939 * The adjacency matrix of the resulting graph is given by:
5942 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5945 * And you'll find that:
5947 * A^(log_2 n)_i,j != 0 for all i,j (7)
5949 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5950 * The task movement gives a factor of O(m), giving a convergence complexity
5953 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5958 * In order to avoid CPUs going idle while there's still work to do, new idle
5959 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5960 * tree itself instead of relying on other CPUs to bring it work.
5962 * This adds some complexity to both (5) and (8) but it reduces the total idle
5970 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5973 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5978 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5980 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5982 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5985 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5986 * rewrite all of this once again.]
5989 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5991 enum fbq_type { regular, remote, all };
5993 #define LBF_ALL_PINNED 0x01
5994 #define LBF_NEED_BREAK 0x02
5995 #define LBF_DST_PINNED 0x04
5996 #define LBF_SOME_PINNED 0x08
5999 struct sched_domain *sd;
6007 struct cpumask *dst_grpmask;
6009 enum cpu_idle_type idle;
6011 /* The set of CPUs under consideration for load-balancing */
6012 struct cpumask *cpus;
6017 unsigned int loop_break;
6018 unsigned int loop_max;
6020 enum fbq_type fbq_type;
6021 struct list_head tasks;
6025 * Is this task likely cache-hot:
6027 static int task_hot(struct task_struct *p, struct lb_env *env)
6031 lockdep_assert_held(&env->src_rq->lock);
6033 if (p->sched_class != &fair_sched_class)
6036 if (unlikely(p->policy == SCHED_IDLE))
6040 * Buddy candidates are cache hot:
6042 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6043 (&p->se == cfs_rq_of(&p->se)->next ||
6044 &p->se == cfs_rq_of(&p->se)->last))
6047 if (sysctl_sched_migration_cost == -1)
6049 if (sysctl_sched_migration_cost == 0)
6052 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6054 return delta < (s64)sysctl_sched_migration_cost;
6057 #ifdef CONFIG_NUMA_BALANCING
6059 * Returns 1, if task migration degrades locality
6060 * Returns 0, if task migration improves locality i.e migration preferred.
6061 * Returns -1, if task migration is not affected by locality.
6063 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6065 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6066 unsigned long src_faults, dst_faults;
6067 int src_nid, dst_nid;
6069 if (!static_branch_likely(&sched_numa_balancing))
6072 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6075 src_nid = cpu_to_node(env->src_cpu);
6076 dst_nid = cpu_to_node(env->dst_cpu);
6078 if (src_nid == dst_nid)
6081 /* Migrating away from the preferred node is always bad. */
6082 if (src_nid == p->numa_preferred_nid) {
6083 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6089 /* Encourage migration to the preferred node. */
6090 if (dst_nid == p->numa_preferred_nid)
6094 src_faults = group_faults(p, src_nid);
6095 dst_faults = group_faults(p, dst_nid);
6097 src_faults = task_faults(p, src_nid);
6098 dst_faults = task_faults(p, dst_nid);
6101 return dst_faults < src_faults;
6105 static inline int migrate_degrades_locality(struct task_struct *p,
6113 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6116 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6120 lockdep_assert_held(&env->src_rq->lock);
6123 * We do not migrate tasks that are:
6124 * 1) throttled_lb_pair, or
6125 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6126 * 3) running (obviously), or
6127 * 4) are cache-hot on their current CPU.
6129 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6132 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6135 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6137 env->flags |= LBF_SOME_PINNED;
6140 * Remember if this task can be migrated to any other cpu in
6141 * our sched_group. We may want to revisit it if we couldn't
6142 * meet load balance goals by pulling other tasks on src_cpu.
6144 * Also avoid computing new_dst_cpu if we have already computed
6145 * one in current iteration.
6147 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6150 /* Prevent to re-select dst_cpu via env's cpus */
6151 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6152 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6153 env->flags |= LBF_DST_PINNED;
6154 env->new_dst_cpu = cpu;
6162 /* Record that we found atleast one task that could run on dst_cpu */
6163 env->flags &= ~LBF_ALL_PINNED;
6165 if (task_running(env->src_rq, p)) {
6166 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6171 * Aggressive migration if:
6172 * 1) destination numa is preferred
6173 * 2) task is cache cold, or
6174 * 3) too many balance attempts have failed.
6176 tsk_cache_hot = migrate_degrades_locality(p, env);
6177 if (tsk_cache_hot == -1)
6178 tsk_cache_hot = task_hot(p, env);
6180 if (tsk_cache_hot <= 0 ||
6181 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6182 if (tsk_cache_hot == 1) {
6183 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6184 schedstat_inc(p, se.statistics.nr_forced_migrations);
6189 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6194 * detach_task() -- detach the task for the migration specified in env
6196 static void detach_task(struct task_struct *p, struct lb_env *env)
6198 lockdep_assert_held(&env->src_rq->lock);
6200 p->on_rq = TASK_ON_RQ_MIGRATING;
6201 deactivate_task(env->src_rq, p, 0);
6202 set_task_cpu(p, env->dst_cpu);
6206 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6207 * part of active balancing operations within "domain".
6209 * Returns a task if successful and NULL otherwise.
6211 static struct task_struct *detach_one_task(struct lb_env *env)
6213 struct task_struct *p, *n;
6215 lockdep_assert_held(&env->src_rq->lock);
6217 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6218 if (!can_migrate_task(p, env))
6221 detach_task(p, env);
6224 * Right now, this is only the second place where
6225 * lb_gained[env->idle] is updated (other is detach_tasks)
6226 * so we can safely collect stats here rather than
6227 * inside detach_tasks().
6229 schedstat_inc(env->sd, lb_gained[env->idle]);
6235 static const unsigned int sched_nr_migrate_break = 32;
6238 * detach_tasks() -- tries to detach up to imbalance weighted load from
6239 * busiest_rq, as part of a balancing operation within domain "sd".
6241 * Returns number of detached tasks if successful and 0 otherwise.
6243 static int detach_tasks(struct lb_env *env)
6245 struct list_head *tasks = &env->src_rq->cfs_tasks;
6246 struct task_struct *p;
6250 lockdep_assert_held(&env->src_rq->lock);
6252 if (env->imbalance <= 0)
6255 while (!list_empty(tasks)) {
6257 * We don't want to steal all, otherwise we may be treated likewise,
6258 * which could at worst lead to a livelock crash.
6260 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6263 p = list_first_entry(tasks, struct task_struct, se.group_node);
6266 /* We've more or less seen every task there is, call it quits */
6267 if (env->loop > env->loop_max)
6270 /* take a breather every nr_migrate tasks */
6271 if (env->loop > env->loop_break) {
6272 env->loop_break += sched_nr_migrate_break;
6273 env->flags |= LBF_NEED_BREAK;
6277 if (!can_migrate_task(p, env))
6280 load = task_h_load(p);
6282 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6285 if ((load / 2) > env->imbalance)
6288 detach_task(p, env);
6289 list_add(&p->se.group_node, &env->tasks);
6292 env->imbalance -= load;
6294 #ifdef CONFIG_PREEMPT
6296 * NEWIDLE balancing is a source of latency, so preemptible
6297 * kernels will stop after the first task is detached to minimize
6298 * the critical section.
6300 if (env->idle == CPU_NEWLY_IDLE)
6305 * We only want to steal up to the prescribed amount of
6308 if (env->imbalance <= 0)
6313 list_move_tail(&p->se.group_node, tasks);
6317 * Right now, this is one of only two places we collect this stat
6318 * so we can safely collect detach_one_task() stats here rather
6319 * than inside detach_one_task().
6321 schedstat_add(env->sd, lb_gained[env->idle], detached);
6327 * attach_task() -- attach the task detached by detach_task() to its new rq.
6329 static void attach_task(struct rq *rq, struct task_struct *p)
6331 lockdep_assert_held(&rq->lock);
6333 BUG_ON(task_rq(p) != rq);
6334 activate_task(rq, p, 0);
6335 p->on_rq = TASK_ON_RQ_QUEUED;
6336 check_preempt_curr(rq, p, 0);
6340 * attach_one_task() -- attaches the task returned from detach_one_task() to
6343 static void attach_one_task(struct rq *rq, struct task_struct *p)
6345 raw_spin_lock(&rq->lock);
6347 raw_spin_unlock(&rq->lock);
6351 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6354 static void attach_tasks(struct lb_env *env)
6356 struct list_head *tasks = &env->tasks;
6357 struct task_struct *p;
6359 raw_spin_lock(&env->dst_rq->lock);
6361 while (!list_empty(tasks)) {
6362 p = list_first_entry(tasks, struct task_struct, se.group_node);
6363 list_del_init(&p->se.group_node);
6365 attach_task(env->dst_rq, p);
6368 raw_spin_unlock(&env->dst_rq->lock);
6371 #ifdef CONFIG_FAIR_GROUP_SCHED
6372 static void update_blocked_averages(int cpu)
6374 struct rq *rq = cpu_rq(cpu);
6375 struct cfs_rq *cfs_rq;
6376 unsigned long flags;
6378 raw_spin_lock_irqsave(&rq->lock, flags);
6379 update_rq_clock(rq);
6382 * Iterates the task_group tree in a bottom up fashion, see
6383 * list_add_leaf_cfs_rq() for details.
6385 for_each_leaf_cfs_rq(rq, cfs_rq) {
6386 /* throttled entities do not contribute to load */
6387 if (throttled_hierarchy(cfs_rq))
6390 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6391 update_tg_load_avg(cfs_rq, 0);
6393 raw_spin_unlock_irqrestore(&rq->lock, flags);
6397 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6398 * This needs to be done in a top-down fashion because the load of a child
6399 * group is a fraction of its parents load.
6401 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6403 struct rq *rq = rq_of(cfs_rq);
6404 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6405 unsigned long now = jiffies;
6408 if (cfs_rq->last_h_load_update == now)
6411 cfs_rq->h_load_next = NULL;
6412 for_each_sched_entity(se) {
6413 cfs_rq = cfs_rq_of(se);
6414 cfs_rq->h_load_next = se;
6415 if (cfs_rq->last_h_load_update == now)
6420 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6421 cfs_rq->last_h_load_update = now;
6424 while ((se = cfs_rq->h_load_next) != NULL) {
6425 load = cfs_rq->h_load;
6426 load = div64_ul(load * se->avg.load_avg,
6427 cfs_rq_load_avg(cfs_rq) + 1);
6428 cfs_rq = group_cfs_rq(se);
6429 cfs_rq->h_load = load;
6430 cfs_rq->last_h_load_update = now;
6434 static unsigned long task_h_load(struct task_struct *p)
6436 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6438 update_cfs_rq_h_load(cfs_rq);
6439 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6440 cfs_rq_load_avg(cfs_rq) + 1);
6443 static inline void update_blocked_averages(int cpu)
6445 struct rq *rq = cpu_rq(cpu);
6446 struct cfs_rq *cfs_rq = &rq->cfs;
6447 unsigned long flags;
6449 raw_spin_lock_irqsave(&rq->lock, flags);
6450 update_rq_clock(rq);
6451 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6452 raw_spin_unlock_irqrestore(&rq->lock, flags);
6455 static unsigned long task_h_load(struct task_struct *p)
6457 return p->se.avg.load_avg;
6461 /********** Helpers for find_busiest_group ************************/
6470 * sg_lb_stats - stats of a sched_group required for load_balancing
6472 struct sg_lb_stats {
6473 unsigned long avg_load; /*Avg load across the CPUs of the group */
6474 unsigned long group_load; /* Total load over the CPUs of the group */
6475 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6476 unsigned long load_per_task;
6477 unsigned long group_capacity;
6478 unsigned long group_util; /* Total utilization of the group */
6479 unsigned int sum_nr_running; /* Nr tasks running in the group */
6480 unsigned int idle_cpus;
6481 unsigned int group_weight;
6482 enum group_type group_type;
6483 int group_no_capacity;
6484 #ifdef CONFIG_NUMA_BALANCING
6485 unsigned int nr_numa_running;
6486 unsigned int nr_preferred_running;
6491 * sd_lb_stats - Structure to store the statistics of a sched_domain
6492 * during load balancing.
6494 struct sd_lb_stats {
6495 struct sched_group *busiest; /* Busiest group in this sd */
6496 struct sched_group *local; /* Local group in this sd */
6497 unsigned long total_load; /* Total load of all groups in sd */
6498 unsigned long total_capacity; /* Total capacity of all groups in sd */
6499 unsigned long avg_load; /* Average load across all groups in sd */
6501 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6502 struct sg_lb_stats local_stat; /* Statistics of the local group */
6505 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6508 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6509 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6510 * We must however clear busiest_stat::avg_load because
6511 * update_sd_pick_busiest() reads this before assignment.
6513 *sds = (struct sd_lb_stats){
6517 .total_capacity = 0UL,
6520 .sum_nr_running = 0,
6521 .group_type = group_other,
6527 * get_sd_load_idx - Obtain the load index for a given sched domain.
6528 * @sd: The sched_domain whose load_idx is to be obtained.
6529 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6531 * Return: The load index.
6533 static inline int get_sd_load_idx(struct sched_domain *sd,
6534 enum cpu_idle_type idle)
6540 load_idx = sd->busy_idx;
6543 case CPU_NEWLY_IDLE:
6544 load_idx = sd->newidle_idx;
6547 load_idx = sd->idle_idx;
6554 static unsigned long scale_rt_capacity(int cpu)
6556 struct rq *rq = cpu_rq(cpu);
6557 u64 total, used, age_stamp, avg;
6561 * Since we're reading these variables without serialization make sure
6562 * we read them once before doing sanity checks on them.
6564 age_stamp = READ_ONCE(rq->age_stamp);
6565 avg = READ_ONCE(rq->rt_avg);
6566 delta = __rq_clock_broken(rq) - age_stamp;
6568 if (unlikely(delta < 0))
6571 total = sched_avg_period() + delta;
6573 used = div_u64(avg, total);
6575 if (likely(used < SCHED_CAPACITY_SCALE))
6576 return SCHED_CAPACITY_SCALE - used;
6581 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6583 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6584 struct sched_group *sdg = sd->groups;
6586 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6588 capacity *= scale_rt_capacity(cpu);
6589 capacity >>= SCHED_CAPACITY_SHIFT;
6594 cpu_rq(cpu)->cpu_capacity = capacity;
6595 sdg->sgc->capacity = capacity;
6598 void update_group_capacity(struct sched_domain *sd, int cpu)
6600 struct sched_domain *child = sd->child;
6601 struct sched_group *group, *sdg = sd->groups;
6602 unsigned long capacity;
6603 unsigned long interval;
6605 interval = msecs_to_jiffies(sd->balance_interval);
6606 interval = clamp(interval, 1UL, max_load_balance_interval);
6607 sdg->sgc->next_update = jiffies + interval;
6610 update_cpu_capacity(sd, cpu);
6616 if (child->flags & SD_OVERLAP) {
6618 * SD_OVERLAP domains cannot assume that child groups
6619 * span the current group.
6622 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6623 struct sched_group_capacity *sgc;
6624 struct rq *rq = cpu_rq(cpu);
6627 * build_sched_domains() -> init_sched_groups_capacity()
6628 * gets here before we've attached the domains to the
6631 * Use capacity_of(), which is set irrespective of domains
6632 * in update_cpu_capacity().
6634 * This avoids capacity from being 0 and
6635 * causing divide-by-zero issues on boot.
6637 if (unlikely(!rq->sd)) {
6638 capacity += capacity_of(cpu);
6642 sgc = rq->sd->groups->sgc;
6643 capacity += sgc->capacity;
6647 * !SD_OVERLAP domains can assume that child groups
6648 * span the current group.
6651 group = child->groups;
6653 capacity += group->sgc->capacity;
6654 group = group->next;
6655 } while (group != child->groups);
6658 sdg->sgc->capacity = capacity;
6662 * Check whether the capacity of the rq has been noticeably reduced by side
6663 * activity. The imbalance_pct is used for the threshold.
6664 * Return true is the capacity is reduced
6667 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6669 return ((rq->cpu_capacity * sd->imbalance_pct) <
6670 (rq->cpu_capacity_orig * 100));
6674 * Group imbalance indicates (and tries to solve) the problem where balancing
6675 * groups is inadequate due to tsk_cpus_allowed() constraints.
6677 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6678 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6681 * { 0 1 2 3 } { 4 5 6 7 }
6684 * If we were to balance group-wise we'd place two tasks in the first group and
6685 * two tasks in the second group. Clearly this is undesired as it will overload
6686 * cpu 3 and leave one of the cpus in the second group unused.
6688 * The current solution to this issue is detecting the skew in the first group
6689 * by noticing the lower domain failed to reach balance and had difficulty
6690 * moving tasks due to affinity constraints.
6692 * When this is so detected; this group becomes a candidate for busiest; see
6693 * update_sd_pick_busiest(). And calculate_imbalance() and
6694 * find_busiest_group() avoid some of the usual balance conditions to allow it
6695 * to create an effective group imbalance.
6697 * This is a somewhat tricky proposition since the next run might not find the
6698 * group imbalance and decide the groups need to be balanced again. A most
6699 * subtle and fragile situation.
6702 static inline int sg_imbalanced(struct sched_group *group)
6704 return group->sgc->imbalance;
6708 * group_has_capacity returns true if the group has spare capacity that could
6709 * be used by some tasks.
6710 * We consider that a group has spare capacity if the * number of task is
6711 * smaller than the number of CPUs or if the utilization is lower than the
6712 * available capacity for CFS tasks.
6713 * For the latter, we use a threshold to stabilize the state, to take into
6714 * account the variance of the tasks' load and to return true if the available
6715 * capacity in meaningful for the load balancer.
6716 * As an example, an available capacity of 1% can appear but it doesn't make
6717 * any benefit for the load balance.
6720 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6722 if (sgs->sum_nr_running < sgs->group_weight)
6725 if ((sgs->group_capacity * 100) >
6726 (sgs->group_util * env->sd->imbalance_pct))
6733 * group_is_overloaded returns true if the group has more tasks than it can
6735 * group_is_overloaded is not equals to !group_has_capacity because a group
6736 * with the exact right number of tasks, has no more spare capacity but is not
6737 * overloaded so both group_has_capacity and group_is_overloaded return
6741 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6743 if (sgs->sum_nr_running <= sgs->group_weight)
6746 if ((sgs->group_capacity * 100) <
6747 (sgs->group_util * env->sd->imbalance_pct))
6754 group_type group_classify(struct sched_group *group,
6755 struct sg_lb_stats *sgs)
6757 if (sgs->group_no_capacity)
6758 return group_overloaded;
6760 if (sg_imbalanced(group))
6761 return group_imbalanced;
6767 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6768 * @env: The load balancing environment.
6769 * @group: sched_group whose statistics are to be updated.
6770 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6771 * @local_group: Does group contain this_cpu.
6772 * @sgs: variable to hold the statistics for this group.
6773 * @overload: Indicate more than one runnable task for any CPU.
6775 static inline void update_sg_lb_stats(struct lb_env *env,
6776 struct sched_group *group, int load_idx,
6777 int local_group, struct sg_lb_stats *sgs,
6783 memset(sgs, 0, sizeof(*sgs));
6785 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6786 struct rq *rq = cpu_rq(i);
6788 /* Bias balancing toward cpus of our domain */
6790 load = target_load(i, load_idx);
6792 load = source_load(i, load_idx);
6794 sgs->group_load += load;
6795 sgs->group_util += cpu_util(i);
6796 sgs->sum_nr_running += rq->cfs.h_nr_running;
6798 nr_running = rq->nr_running;
6802 #ifdef CONFIG_NUMA_BALANCING
6803 sgs->nr_numa_running += rq->nr_numa_running;
6804 sgs->nr_preferred_running += rq->nr_preferred_running;
6806 sgs->sum_weighted_load += weighted_cpuload(i);
6808 * No need to call idle_cpu() if nr_running is not 0
6810 if (!nr_running && idle_cpu(i))
6814 /* Adjust by relative CPU capacity of the group */
6815 sgs->group_capacity = group->sgc->capacity;
6816 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6818 if (sgs->sum_nr_running)
6819 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6821 sgs->group_weight = group->group_weight;
6823 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6824 sgs->group_type = group_classify(group, sgs);
6828 * update_sd_pick_busiest - return 1 on busiest group
6829 * @env: The load balancing environment.
6830 * @sds: sched_domain statistics
6831 * @sg: sched_group candidate to be checked for being the busiest
6832 * @sgs: sched_group statistics
6834 * Determine if @sg is a busier group than the previously selected
6837 * Return: %true if @sg is a busier group than the previously selected
6838 * busiest group. %false otherwise.
6840 static bool update_sd_pick_busiest(struct lb_env *env,
6841 struct sd_lb_stats *sds,
6842 struct sched_group *sg,
6843 struct sg_lb_stats *sgs)
6845 struct sg_lb_stats *busiest = &sds->busiest_stat;
6847 if (sgs->group_type > busiest->group_type)
6850 if (sgs->group_type < busiest->group_type)
6853 if (sgs->avg_load <= busiest->avg_load)
6856 /* This is the busiest node in its class. */
6857 if (!(env->sd->flags & SD_ASYM_PACKING))
6860 /* No ASYM_PACKING if target cpu is already busy */
6861 if (env->idle == CPU_NOT_IDLE)
6864 * ASYM_PACKING needs to move all the work to the lowest
6865 * numbered CPUs in the group, therefore mark all groups
6866 * higher than ourself as busy.
6868 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6872 /* Prefer to move from highest possible cpu's work */
6873 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6880 #ifdef CONFIG_NUMA_BALANCING
6881 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6883 if (sgs->sum_nr_running > sgs->nr_numa_running)
6885 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6890 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6892 if (rq->nr_running > rq->nr_numa_running)
6894 if (rq->nr_running > rq->nr_preferred_running)
6899 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6904 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6908 #endif /* CONFIG_NUMA_BALANCING */
6911 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6912 * @env: The load balancing environment.
6913 * @sds: variable to hold the statistics for this sched_domain.
6915 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6917 struct sched_domain *child = env->sd->child;
6918 struct sched_group *sg = env->sd->groups;
6919 struct sg_lb_stats tmp_sgs;
6920 int load_idx, prefer_sibling = 0;
6921 bool overload = false;
6923 if (child && child->flags & SD_PREFER_SIBLING)
6926 load_idx = get_sd_load_idx(env->sd, env->idle);
6929 struct sg_lb_stats *sgs = &tmp_sgs;
6932 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6935 sgs = &sds->local_stat;
6937 if (env->idle != CPU_NEWLY_IDLE ||
6938 time_after_eq(jiffies, sg->sgc->next_update))
6939 update_group_capacity(env->sd, env->dst_cpu);
6942 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6949 * In case the child domain prefers tasks go to siblings
6950 * first, lower the sg capacity so that we'll try
6951 * and move all the excess tasks away. We lower the capacity
6952 * of a group only if the local group has the capacity to fit
6953 * these excess tasks. The extra check prevents the case where
6954 * you always pull from the heaviest group when it is already
6955 * under-utilized (possible with a large weight task outweighs
6956 * the tasks on the system).
6958 if (prefer_sibling && sds->local &&
6959 group_has_capacity(env, &sds->local_stat) &&
6960 (sgs->sum_nr_running > 1)) {
6961 sgs->group_no_capacity = 1;
6962 sgs->group_type = group_classify(sg, sgs);
6965 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6967 sds->busiest_stat = *sgs;
6971 /* Now, start updating sd_lb_stats */
6972 sds->total_load += sgs->group_load;
6973 sds->total_capacity += sgs->group_capacity;
6976 } while (sg != env->sd->groups);
6978 if (env->sd->flags & SD_NUMA)
6979 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6981 if (!env->sd->parent) {
6982 /* update overload indicator if we are at root domain */
6983 if (env->dst_rq->rd->overload != overload)
6984 env->dst_rq->rd->overload = overload;
6990 * check_asym_packing - Check to see if the group is packed into the
6993 * This is primarily intended to used at the sibling level. Some
6994 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6995 * case of POWER7, it can move to lower SMT modes only when higher
6996 * threads are idle. When in lower SMT modes, the threads will
6997 * perform better since they share less core resources. Hence when we
6998 * have idle threads, we want them to be the higher ones.
7000 * This packing function is run on idle threads. It checks to see if
7001 * the busiest CPU in this domain (core in the P7 case) has a higher
7002 * CPU number than the packing function is being run on. Here we are
7003 * assuming lower CPU number will be equivalent to lower a SMT thread
7006 * Return: 1 when packing is required and a task should be moved to
7007 * this CPU. The amount of the imbalance is returned in *imbalance.
7009 * @env: The load balancing environment.
7010 * @sds: Statistics of the sched_domain which is to be packed
7012 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7016 if (!(env->sd->flags & SD_ASYM_PACKING))
7019 if (env->idle == CPU_NOT_IDLE)
7025 busiest_cpu = group_first_cpu(sds->busiest);
7026 if (env->dst_cpu > busiest_cpu)
7029 env->imbalance = DIV_ROUND_CLOSEST(
7030 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7031 SCHED_CAPACITY_SCALE);
7037 * fix_small_imbalance - Calculate the minor imbalance that exists
7038 * amongst the groups of a sched_domain, during
7040 * @env: The load balancing environment.
7041 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7044 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7046 unsigned long tmp, capa_now = 0, capa_move = 0;
7047 unsigned int imbn = 2;
7048 unsigned long scaled_busy_load_per_task;
7049 struct sg_lb_stats *local, *busiest;
7051 local = &sds->local_stat;
7052 busiest = &sds->busiest_stat;
7054 if (!local->sum_nr_running)
7055 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7056 else if (busiest->load_per_task > local->load_per_task)
7059 scaled_busy_load_per_task =
7060 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7061 busiest->group_capacity;
7063 if (busiest->avg_load + scaled_busy_load_per_task >=
7064 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7065 env->imbalance = busiest->load_per_task;
7070 * OK, we don't have enough imbalance to justify moving tasks,
7071 * however we may be able to increase total CPU capacity used by
7075 capa_now += busiest->group_capacity *
7076 min(busiest->load_per_task, busiest->avg_load);
7077 capa_now += local->group_capacity *
7078 min(local->load_per_task, local->avg_load);
7079 capa_now /= SCHED_CAPACITY_SCALE;
7081 /* Amount of load we'd subtract */
7082 if (busiest->avg_load > scaled_busy_load_per_task) {
7083 capa_move += busiest->group_capacity *
7084 min(busiest->load_per_task,
7085 busiest->avg_load - scaled_busy_load_per_task);
7088 /* Amount of load we'd add */
7089 if (busiest->avg_load * busiest->group_capacity <
7090 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7091 tmp = (busiest->avg_load * busiest->group_capacity) /
7092 local->group_capacity;
7094 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7095 local->group_capacity;
7097 capa_move += local->group_capacity *
7098 min(local->load_per_task, local->avg_load + tmp);
7099 capa_move /= SCHED_CAPACITY_SCALE;
7101 /* Move if we gain throughput */
7102 if (capa_move > capa_now)
7103 env->imbalance = busiest->load_per_task;
7107 * calculate_imbalance - Calculate the amount of imbalance present within the
7108 * groups of a given sched_domain during load balance.
7109 * @env: load balance environment
7110 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7112 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7114 unsigned long max_pull, load_above_capacity = ~0UL;
7115 struct sg_lb_stats *local, *busiest;
7117 local = &sds->local_stat;
7118 busiest = &sds->busiest_stat;
7120 if (busiest->group_type == group_imbalanced) {
7122 * In the group_imb case we cannot rely on group-wide averages
7123 * to ensure cpu-load equilibrium, look at wider averages. XXX
7125 busiest->load_per_task =
7126 min(busiest->load_per_task, sds->avg_load);
7130 * Avg load of busiest sg can be less and avg load of local sg can
7131 * be greater than avg load across all sgs of sd because avg load
7132 * factors in sg capacity and sgs with smaller group_type are
7133 * skipped when updating the busiest sg:
7135 if (busiest->avg_load <= sds->avg_load ||
7136 local->avg_load >= sds->avg_load) {
7138 return fix_small_imbalance(env, sds);
7142 * If there aren't any idle cpus, avoid creating some.
7144 if (busiest->group_type == group_overloaded &&
7145 local->group_type == group_overloaded) {
7146 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7147 if (load_above_capacity > busiest->group_capacity) {
7148 load_above_capacity -= busiest->group_capacity;
7149 load_above_capacity *= NICE_0_LOAD;
7150 load_above_capacity /= busiest->group_capacity;
7152 load_above_capacity = ~0UL;
7156 * We're trying to get all the cpus to the average_load, so we don't
7157 * want to push ourselves above the average load, nor do we wish to
7158 * reduce the max loaded cpu below the average load. At the same time,
7159 * we also don't want to reduce the group load below the group
7160 * capacity. Thus we look for the minimum possible imbalance.
7162 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7164 /* How much load to actually move to equalise the imbalance */
7165 env->imbalance = min(
7166 max_pull * busiest->group_capacity,
7167 (sds->avg_load - local->avg_load) * local->group_capacity
7168 ) / SCHED_CAPACITY_SCALE;
7171 * if *imbalance is less than the average load per runnable task
7172 * there is no guarantee that any tasks will be moved so we'll have
7173 * a think about bumping its value to force at least one task to be
7176 if (env->imbalance < busiest->load_per_task)
7177 return fix_small_imbalance(env, sds);
7180 /******* find_busiest_group() helpers end here *********************/
7183 * find_busiest_group - Returns the busiest group within the sched_domain
7184 * if there is an imbalance.
7186 * Also calculates the amount of weighted load which should be moved
7187 * to restore balance.
7189 * @env: The load balancing environment.
7191 * Return: - The busiest group if imbalance exists.
7193 static struct sched_group *find_busiest_group(struct lb_env *env)
7195 struct sg_lb_stats *local, *busiest;
7196 struct sd_lb_stats sds;
7198 init_sd_lb_stats(&sds);
7201 * Compute the various statistics relavent for load balancing at
7204 update_sd_lb_stats(env, &sds);
7205 local = &sds.local_stat;
7206 busiest = &sds.busiest_stat;
7208 /* ASYM feature bypasses nice load balance check */
7209 if (check_asym_packing(env, &sds))
7212 /* There is no busy sibling group to pull tasks from */
7213 if (!sds.busiest || busiest->sum_nr_running == 0)
7216 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7217 / sds.total_capacity;
7220 * If the busiest group is imbalanced the below checks don't
7221 * work because they assume all things are equal, which typically
7222 * isn't true due to cpus_allowed constraints and the like.
7224 if (busiest->group_type == group_imbalanced)
7227 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7228 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7229 busiest->group_no_capacity)
7233 * If the local group is busier than the selected busiest group
7234 * don't try and pull any tasks.
7236 if (local->avg_load >= busiest->avg_load)
7240 * Don't pull any tasks if this group is already above the domain
7243 if (local->avg_load >= sds.avg_load)
7246 if (env->idle == CPU_IDLE) {
7248 * This cpu is idle. If the busiest group is not overloaded
7249 * and there is no imbalance between this and busiest group
7250 * wrt idle cpus, it is balanced. The imbalance becomes
7251 * significant if the diff is greater than 1 otherwise we
7252 * might end up to just move the imbalance on another group
7254 if ((busiest->group_type != group_overloaded) &&
7255 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7259 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7260 * imbalance_pct to be conservative.
7262 if (100 * busiest->avg_load <=
7263 env->sd->imbalance_pct * local->avg_load)
7268 /* Looks like there is an imbalance. Compute it */
7269 calculate_imbalance(env, &sds);
7278 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7280 static struct rq *find_busiest_queue(struct lb_env *env,
7281 struct sched_group *group)
7283 struct rq *busiest = NULL, *rq;
7284 unsigned long busiest_load = 0, busiest_capacity = 1;
7287 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7288 unsigned long capacity, wl;
7292 rt = fbq_classify_rq(rq);
7295 * We classify groups/runqueues into three groups:
7296 * - regular: there are !numa tasks
7297 * - remote: there are numa tasks that run on the 'wrong' node
7298 * - all: there is no distinction
7300 * In order to avoid migrating ideally placed numa tasks,
7301 * ignore those when there's better options.
7303 * If we ignore the actual busiest queue to migrate another
7304 * task, the next balance pass can still reduce the busiest
7305 * queue by moving tasks around inside the node.
7307 * If we cannot move enough load due to this classification
7308 * the next pass will adjust the group classification and
7309 * allow migration of more tasks.
7311 * Both cases only affect the total convergence complexity.
7313 if (rt > env->fbq_type)
7316 capacity = capacity_of(i);
7318 wl = weighted_cpuload(i);
7321 * When comparing with imbalance, use weighted_cpuload()
7322 * which is not scaled with the cpu capacity.
7325 if (rq->nr_running == 1 && wl > env->imbalance &&
7326 !check_cpu_capacity(rq, env->sd))
7330 * For the load comparisons with the other cpu's, consider
7331 * the weighted_cpuload() scaled with the cpu capacity, so
7332 * that the load can be moved away from the cpu that is
7333 * potentially running at a lower capacity.
7335 * Thus we're looking for max(wl_i / capacity_i), crosswise
7336 * multiplication to rid ourselves of the division works out
7337 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7338 * our previous maximum.
7340 if (wl * busiest_capacity > busiest_load * capacity) {
7342 busiest_capacity = capacity;
7351 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7352 * so long as it is large enough.
7354 #define MAX_PINNED_INTERVAL 512
7356 /* Working cpumask for load_balance and load_balance_newidle. */
7357 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7359 static int need_active_balance(struct lb_env *env)
7361 struct sched_domain *sd = env->sd;
7363 if (env->idle == CPU_NEWLY_IDLE) {
7366 * ASYM_PACKING needs to force migrate tasks from busy but
7367 * higher numbered CPUs in order to pack all tasks in the
7368 * lowest numbered CPUs.
7370 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7375 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7376 * It's worth migrating the task if the src_cpu's capacity is reduced
7377 * because of other sched_class or IRQs if more capacity stays
7378 * available on dst_cpu.
7380 if ((env->idle != CPU_NOT_IDLE) &&
7381 (env->src_rq->cfs.h_nr_running == 1)) {
7382 if ((check_cpu_capacity(env->src_rq, sd)) &&
7383 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7387 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7390 static int active_load_balance_cpu_stop(void *data);
7392 static int should_we_balance(struct lb_env *env)
7394 struct sched_group *sg = env->sd->groups;
7395 struct cpumask *sg_cpus, *sg_mask;
7396 int cpu, balance_cpu = -1;
7399 * In the newly idle case, we will allow all the cpu's
7400 * to do the newly idle load balance.
7402 if (env->idle == CPU_NEWLY_IDLE)
7405 sg_cpus = sched_group_cpus(sg);
7406 sg_mask = sched_group_mask(sg);
7407 /* Try to find first idle cpu */
7408 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7409 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7416 if (balance_cpu == -1)
7417 balance_cpu = group_balance_cpu(sg);
7420 * First idle cpu or the first cpu(busiest) in this sched group
7421 * is eligible for doing load balancing at this and above domains.
7423 return balance_cpu == env->dst_cpu;
7427 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7428 * tasks if there is an imbalance.
7430 static int load_balance(int this_cpu, struct rq *this_rq,
7431 struct sched_domain *sd, enum cpu_idle_type idle,
7432 int *continue_balancing)
7434 int ld_moved, cur_ld_moved, active_balance = 0;
7435 struct sched_domain *sd_parent = sd->parent;
7436 struct sched_group *group;
7438 unsigned long flags;
7439 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7441 struct lb_env env = {
7443 .dst_cpu = this_cpu,
7445 .dst_grpmask = sched_group_cpus(sd->groups),
7447 .loop_break = sched_nr_migrate_break,
7450 .tasks = LIST_HEAD_INIT(env.tasks),
7454 * For NEWLY_IDLE load_balancing, we don't need to consider
7455 * other cpus in our group
7457 if (idle == CPU_NEWLY_IDLE)
7458 env.dst_grpmask = NULL;
7460 cpumask_copy(cpus, cpu_active_mask);
7462 schedstat_inc(sd, lb_count[idle]);
7465 if (!should_we_balance(&env)) {
7466 *continue_balancing = 0;
7470 group = find_busiest_group(&env);
7472 schedstat_inc(sd, lb_nobusyg[idle]);
7476 busiest = find_busiest_queue(&env, group);
7478 schedstat_inc(sd, lb_nobusyq[idle]);
7482 BUG_ON(busiest == env.dst_rq);
7484 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7486 env.src_cpu = busiest->cpu;
7487 env.src_rq = busiest;
7490 if (busiest->nr_running > 1) {
7492 * Attempt to move tasks. If find_busiest_group has found
7493 * an imbalance but busiest->nr_running <= 1, the group is
7494 * still unbalanced. ld_moved simply stays zero, so it is
7495 * correctly treated as an imbalance.
7497 env.flags |= LBF_ALL_PINNED;
7498 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7501 raw_spin_lock_irqsave(&busiest->lock, flags);
7504 * cur_ld_moved - load moved in current iteration
7505 * ld_moved - cumulative load moved across iterations
7507 cur_ld_moved = detach_tasks(&env);
7510 * We've detached some tasks from busiest_rq. Every
7511 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7512 * unlock busiest->lock, and we are able to be sure
7513 * that nobody can manipulate the tasks in parallel.
7514 * See task_rq_lock() family for the details.
7517 raw_spin_unlock(&busiest->lock);
7521 ld_moved += cur_ld_moved;
7524 local_irq_restore(flags);
7526 if (env.flags & LBF_NEED_BREAK) {
7527 env.flags &= ~LBF_NEED_BREAK;
7532 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7533 * us and move them to an alternate dst_cpu in our sched_group
7534 * where they can run. The upper limit on how many times we
7535 * iterate on same src_cpu is dependent on number of cpus in our
7538 * This changes load balance semantics a bit on who can move
7539 * load to a given_cpu. In addition to the given_cpu itself
7540 * (or a ilb_cpu acting on its behalf where given_cpu is
7541 * nohz-idle), we now have balance_cpu in a position to move
7542 * load to given_cpu. In rare situations, this may cause
7543 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7544 * _independently_ and at _same_ time to move some load to
7545 * given_cpu) causing exceess load to be moved to given_cpu.
7546 * This however should not happen so much in practice and
7547 * moreover subsequent load balance cycles should correct the
7548 * excess load moved.
7550 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7552 /* Prevent to re-select dst_cpu via env's cpus */
7553 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7555 env.dst_rq = cpu_rq(env.new_dst_cpu);
7556 env.dst_cpu = env.new_dst_cpu;
7557 env.flags &= ~LBF_DST_PINNED;
7559 env.loop_break = sched_nr_migrate_break;
7562 * Go back to "more_balance" rather than "redo" since we
7563 * need to continue with same src_cpu.
7569 * We failed to reach balance because of affinity.
7572 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7574 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7575 *group_imbalance = 1;
7578 /* All tasks on this runqueue were pinned by CPU affinity */
7579 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7580 cpumask_clear_cpu(cpu_of(busiest), cpus);
7581 if (!cpumask_empty(cpus)) {
7583 env.loop_break = sched_nr_migrate_break;
7586 goto out_all_pinned;
7591 schedstat_inc(sd, lb_failed[idle]);
7593 * Increment the failure counter only on periodic balance.
7594 * We do not want newidle balance, which can be very
7595 * frequent, pollute the failure counter causing
7596 * excessive cache_hot migrations and active balances.
7598 if (idle != CPU_NEWLY_IDLE)
7599 sd->nr_balance_failed++;
7601 if (need_active_balance(&env)) {
7602 raw_spin_lock_irqsave(&busiest->lock, flags);
7604 /* don't kick the active_load_balance_cpu_stop,
7605 * if the curr task on busiest cpu can't be
7608 if (!cpumask_test_cpu(this_cpu,
7609 tsk_cpus_allowed(busiest->curr))) {
7610 raw_spin_unlock_irqrestore(&busiest->lock,
7612 env.flags |= LBF_ALL_PINNED;
7613 goto out_one_pinned;
7617 * ->active_balance synchronizes accesses to
7618 * ->active_balance_work. Once set, it's cleared
7619 * only after active load balance is finished.
7621 if (!busiest->active_balance) {
7622 busiest->active_balance = 1;
7623 busiest->push_cpu = this_cpu;
7626 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7628 if (active_balance) {
7629 stop_one_cpu_nowait(cpu_of(busiest),
7630 active_load_balance_cpu_stop, busiest,
7631 &busiest->active_balance_work);
7634 /* We've kicked active balancing, force task migration. */
7635 sd->nr_balance_failed = sd->cache_nice_tries+1;
7638 sd->nr_balance_failed = 0;
7640 if (likely(!active_balance)) {
7641 /* We were unbalanced, so reset the balancing interval */
7642 sd->balance_interval = sd->min_interval;
7645 * If we've begun active balancing, start to back off. This
7646 * case may not be covered by the all_pinned logic if there
7647 * is only 1 task on the busy runqueue (because we don't call
7650 if (sd->balance_interval < sd->max_interval)
7651 sd->balance_interval *= 2;
7658 * We reach balance although we may have faced some affinity
7659 * constraints. Clear the imbalance flag if it was set.
7662 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7664 if (*group_imbalance)
7665 *group_imbalance = 0;
7670 * We reach balance because all tasks are pinned at this level so
7671 * we can't migrate them. Let the imbalance flag set so parent level
7672 * can try to migrate them.
7674 schedstat_inc(sd, lb_balanced[idle]);
7676 sd->nr_balance_failed = 0;
7679 /* tune up the balancing interval */
7680 if (((env.flags & LBF_ALL_PINNED) &&
7681 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7682 (sd->balance_interval < sd->max_interval))
7683 sd->balance_interval *= 2;
7690 static inline unsigned long
7691 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7693 unsigned long interval = sd->balance_interval;
7696 interval *= sd->busy_factor;
7698 /* scale ms to jiffies */
7699 interval = msecs_to_jiffies(interval);
7700 interval = clamp(interval, 1UL, max_load_balance_interval);
7706 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7708 unsigned long interval, next;
7710 interval = get_sd_balance_interval(sd, cpu_busy);
7711 next = sd->last_balance + interval;
7713 if (time_after(*next_balance, next))
7714 *next_balance = next;
7718 * idle_balance is called by schedule() if this_cpu is about to become
7719 * idle. Attempts to pull tasks from other CPUs.
7721 static int idle_balance(struct rq *this_rq)
7723 unsigned long next_balance = jiffies + HZ;
7724 int this_cpu = this_rq->cpu;
7725 struct sched_domain *sd;
7726 int pulled_task = 0;
7730 * We must set idle_stamp _before_ calling idle_balance(), such that we
7731 * measure the duration of idle_balance() as idle time.
7733 this_rq->idle_stamp = rq_clock(this_rq);
7735 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7736 !this_rq->rd->overload) {
7738 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7740 update_next_balance(sd, 0, &next_balance);
7746 raw_spin_unlock(&this_rq->lock);
7748 update_blocked_averages(this_cpu);
7750 for_each_domain(this_cpu, sd) {
7751 int continue_balancing = 1;
7752 u64 t0, domain_cost;
7754 if (!(sd->flags & SD_LOAD_BALANCE))
7757 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7758 update_next_balance(sd, 0, &next_balance);
7762 if (sd->flags & SD_BALANCE_NEWIDLE) {
7763 t0 = sched_clock_cpu(this_cpu);
7765 pulled_task = load_balance(this_cpu, this_rq,
7767 &continue_balancing);
7769 domain_cost = sched_clock_cpu(this_cpu) - t0;
7770 if (domain_cost > sd->max_newidle_lb_cost)
7771 sd->max_newidle_lb_cost = domain_cost;
7773 curr_cost += domain_cost;
7776 update_next_balance(sd, 0, &next_balance);
7779 * Stop searching for tasks to pull if there are
7780 * now runnable tasks on this rq.
7782 if (pulled_task || this_rq->nr_running > 0)
7787 raw_spin_lock(&this_rq->lock);
7789 if (curr_cost > this_rq->max_idle_balance_cost)
7790 this_rq->max_idle_balance_cost = curr_cost;
7793 * While browsing the domains, we released the rq lock, a task could
7794 * have been enqueued in the meantime. Since we're not going idle,
7795 * pretend we pulled a task.
7797 if (this_rq->cfs.h_nr_running && !pulled_task)
7801 /* Move the next balance forward */
7802 if (time_after(this_rq->next_balance, next_balance))
7803 this_rq->next_balance = next_balance;
7805 /* Is there a task of a high priority class? */
7806 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7810 this_rq->idle_stamp = 0;
7816 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7817 * running tasks off the busiest CPU onto idle CPUs. It requires at
7818 * least 1 task to be running on each physical CPU where possible, and
7819 * avoids physical / logical imbalances.
7821 static int active_load_balance_cpu_stop(void *data)
7823 struct rq *busiest_rq = data;
7824 int busiest_cpu = cpu_of(busiest_rq);
7825 int target_cpu = busiest_rq->push_cpu;
7826 struct rq *target_rq = cpu_rq(target_cpu);
7827 struct sched_domain *sd;
7828 struct task_struct *p = NULL;
7830 raw_spin_lock_irq(&busiest_rq->lock);
7832 /* make sure the requested cpu hasn't gone down in the meantime */
7833 if (unlikely(busiest_cpu != smp_processor_id() ||
7834 !busiest_rq->active_balance))
7837 /* Is there any task to move? */
7838 if (busiest_rq->nr_running <= 1)
7842 * This condition is "impossible", if it occurs
7843 * we need to fix it. Originally reported by
7844 * Bjorn Helgaas on a 128-cpu setup.
7846 BUG_ON(busiest_rq == target_rq);
7848 /* Search for an sd spanning us and the target CPU. */
7850 for_each_domain(target_cpu, sd) {
7851 if ((sd->flags & SD_LOAD_BALANCE) &&
7852 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7857 struct lb_env env = {
7859 .dst_cpu = target_cpu,
7860 .dst_rq = target_rq,
7861 .src_cpu = busiest_rq->cpu,
7862 .src_rq = busiest_rq,
7866 schedstat_inc(sd, alb_count);
7868 p = detach_one_task(&env);
7870 schedstat_inc(sd, alb_pushed);
7871 /* Active balancing done, reset the failure counter. */
7872 sd->nr_balance_failed = 0;
7874 schedstat_inc(sd, alb_failed);
7879 busiest_rq->active_balance = 0;
7880 raw_spin_unlock(&busiest_rq->lock);
7883 attach_one_task(target_rq, p);
7890 static inline int on_null_domain(struct rq *rq)
7892 return unlikely(!rcu_dereference_sched(rq->sd));
7895 #ifdef CONFIG_NO_HZ_COMMON
7897 * idle load balancing details
7898 * - When one of the busy CPUs notice that there may be an idle rebalancing
7899 * needed, they will kick the idle load balancer, which then does idle
7900 * load balancing for all the idle CPUs.
7903 cpumask_var_t idle_cpus_mask;
7905 unsigned long next_balance; /* in jiffy units */
7906 } nohz ____cacheline_aligned;
7908 static inline int find_new_ilb(void)
7910 int ilb = cpumask_first(nohz.idle_cpus_mask);
7912 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7919 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7920 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7921 * CPU (if there is one).
7923 static void nohz_balancer_kick(void)
7927 nohz.next_balance++;
7929 ilb_cpu = find_new_ilb();
7931 if (ilb_cpu >= nr_cpu_ids)
7934 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7937 * Use smp_send_reschedule() instead of resched_cpu().
7938 * This way we generate a sched IPI on the target cpu which
7939 * is idle. And the softirq performing nohz idle load balance
7940 * will be run before returning from the IPI.
7942 smp_send_reschedule(ilb_cpu);
7946 void nohz_balance_exit_idle(unsigned int cpu)
7948 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7950 * Completely isolated CPUs don't ever set, so we must test.
7952 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7953 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7954 atomic_dec(&nohz.nr_cpus);
7956 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7960 static inline void set_cpu_sd_state_busy(void)
7962 struct sched_domain *sd;
7963 int cpu = smp_processor_id();
7966 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7968 if (!sd || !sd->nohz_idle)
7972 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7977 void set_cpu_sd_state_idle(void)
7979 struct sched_domain *sd;
7980 int cpu = smp_processor_id();
7983 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7985 if (!sd || sd->nohz_idle)
7989 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7995 * This routine will record that the cpu is going idle with tick stopped.
7996 * This info will be used in performing idle load balancing in the future.
7998 void nohz_balance_enter_idle(int cpu)
8001 * If this cpu is going down, then nothing needs to be done.
8003 if (!cpu_active(cpu))
8006 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8010 * If we're a completely isolated CPU, we don't play.
8012 if (on_null_domain(cpu_rq(cpu)))
8015 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8016 atomic_inc(&nohz.nr_cpus);
8017 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8021 static DEFINE_SPINLOCK(balancing);
8024 * Scale the max load_balance interval with the number of CPUs in the system.
8025 * This trades load-balance latency on larger machines for less cross talk.
8027 void update_max_interval(void)
8029 max_load_balance_interval = HZ*num_online_cpus()/10;
8033 * It checks each scheduling domain to see if it is due to be balanced,
8034 * and initiates a balancing operation if so.
8036 * Balancing parameters are set up in init_sched_domains.
8038 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8040 int continue_balancing = 1;
8042 unsigned long interval;
8043 struct sched_domain *sd;
8044 /* Earliest time when we have to do rebalance again */
8045 unsigned long next_balance = jiffies + 60*HZ;
8046 int update_next_balance = 0;
8047 int need_serialize, need_decay = 0;
8050 update_blocked_averages(cpu);
8053 for_each_domain(cpu, sd) {
8055 * Decay the newidle max times here because this is a regular
8056 * visit to all the domains. Decay ~1% per second.
8058 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8059 sd->max_newidle_lb_cost =
8060 (sd->max_newidle_lb_cost * 253) / 256;
8061 sd->next_decay_max_lb_cost = jiffies + HZ;
8064 max_cost += sd->max_newidle_lb_cost;
8066 if (!(sd->flags & SD_LOAD_BALANCE))
8070 * Stop the load balance at this level. There is another
8071 * CPU in our sched group which is doing load balancing more
8074 if (!continue_balancing) {
8080 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8082 need_serialize = sd->flags & SD_SERIALIZE;
8083 if (need_serialize) {
8084 if (!spin_trylock(&balancing))
8088 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8089 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8091 * The LBF_DST_PINNED logic could have changed
8092 * env->dst_cpu, so we can't know our idle
8093 * state even if we migrated tasks. Update it.
8095 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8097 sd->last_balance = jiffies;
8098 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8101 spin_unlock(&balancing);
8103 if (time_after(next_balance, sd->last_balance + interval)) {
8104 next_balance = sd->last_balance + interval;
8105 update_next_balance = 1;
8110 * Ensure the rq-wide value also decays but keep it at a
8111 * reasonable floor to avoid funnies with rq->avg_idle.
8113 rq->max_idle_balance_cost =
8114 max((u64)sysctl_sched_migration_cost, max_cost);
8119 * next_balance will be updated only when there is a need.
8120 * When the cpu is attached to null domain for ex, it will not be
8123 if (likely(update_next_balance)) {
8124 rq->next_balance = next_balance;
8126 #ifdef CONFIG_NO_HZ_COMMON
8128 * If this CPU has been elected to perform the nohz idle
8129 * balance. Other idle CPUs have already rebalanced with
8130 * nohz_idle_balance() and nohz.next_balance has been
8131 * updated accordingly. This CPU is now running the idle load
8132 * balance for itself and we need to update the
8133 * nohz.next_balance accordingly.
8135 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8136 nohz.next_balance = rq->next_balance;
8141 #ifdef CONFIG_NO_HZ_COMMON
8143 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8144 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8146 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8148 int this_cpu = this_rq->cpu;
8151 /* Earliest time when we have to do rebalance again */
8152 unsigned long next_balance = jiffies + 60*HZ;
8153 int update_next_balance = 0;
8155 if (idle != CPU_IDLE ||
8156 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8159 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8160 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8164 * If this cpu gets work to do, stop the load balancing
8165 * work being done for other cpus. Next load
8166 * balancing owner will pick it up.
8171 rq = cpu_rq(balance_cpu);
8174 * If time for next balance is due,
8177 if (time_after_eq(jiffies, rq->next_balance)) {
8178 raw_spin_lock_irq(&rq->lock);
8179 update_rq_clock(rq);
8180 cpu_load_update_idle(rq);
8181 raw_spin_unlock_irq(&rq->lock);
8182 rebalance_domains(rq, CPU_IDLE);
8185 if (time_after(next_balance, rq->next_balance)) {
8186 next_balance = rq->next_balance;
8187 update_next_balance = 1;
8192 * next_balance will be updated only when there is a need.
8193 * When the CPU is attached to null domain for ex, it will not be
8196 if (likely(update_next_balance))
8197 nohz.next_balance = next_balance;
8199 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8203 * Current heuristic for kicking the idle load balancer in the presence
8204 * of an idle cpu in the system.
8205 * - This rq has more than one task.
8206 * - This rq has at least one CFS task and the capacity of the CPU is
8207 * significantly reduced because of RT tasks or IRQs.
8208 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8209 * multiple busy cpu.
8210 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8211 * domain span are idle.
8213 static inline bool nohz_kick_needed(struct rq *rq)
8215 unsigned long now = jiffies;
8216 struct sched_domain *sd;
8217 struct sched_group_capacity *sgc;
8218 int nr_busy, cpu = rq->cpu;
8221 if (unlikely(rq->idle_balance))
8225 * We may be recently in ticked or tickless idle mode. At the first
8226 * busy tick after returning from idle, we will update the busy stats.
8228 set_cpu_sd_state_busy();
8229 nohz_balance_exit_idle(cpu);
8232 * None are in tickless mode and hence no need for NOHZ idle load
8235 if (likely(!atomic_read(&nohz.nr_cpus)))
8238 if (time_before(now, nohz.next_balance))
8241 if (rq->nr_running >= 2)
8245 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8247 sgc = sd->groups->sgc;
8248 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8257 sd = rcu_dereference(rq->sd);
8259 if ((rq->cfs.h_nr_running >= 1) &&
8260 check_cpu_capacity(rq, sd)) {
8266 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8267 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8268 sched_domain_span(sd)) < cpu)) {
8278 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8282 * run_rebalance_domains is triggered when needed from the scheduler tick.
8283 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8285 static void run_rebalance_domains(struct softirq_action *h)
8287 struct rq *this_rq = this_rq();
8288 enum cpu_idle_type idle = this_rq->idle_balance ?
8289 CPU_IDLE : CPU_NOT_IDLE;
8292 * If this cpu has a pending nohz_balance_kick, then do the
8293 * balancing on behalf of the other idle cpus whose ticks are
8294 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8295 * give the idle cpus a chance to load balance. Else we may
8296 * load balance only within the local sched_domain hierarchy
8297 * and abort nohz_idle_balance altogether if we pull some load.
8299 nohz_idle_balance(this_rq, idle);
8300 rebalance_domains(this_rq, idle);
8304 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8306 void trigger_load_balance(struct rq *rq)
8308 /* Don't need to rebalance while attached to NULL domain */
8309 if (unlikely(on_null_domain(rq)))
8312 if (time_after_eq(jiffies, rq->next_balance))
8313 raise_softirq(SCHED_SOFTIRQ);
8314 #ifdef CONFIG_NO_HZ_COMMON
8315 if (nohz_kick_needed(rq))
8316 nohz_balancer_kick();
8320 static void rq_online_fair(struct rq *rq)
8324 update_runtime_enabled(rq);
8327 static void rq_offline_fair(struct rq *rq)
8331 /* Ensure any throttled groups are reachable by pick_next_task */
8332 unthrottle_offline_cfs_rqs(rq);
8335 #endif /* CONFIG_SMP */
8338 * scheduler tick hitting a task of our scheduling class:
8340 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8342 struct cfs_rq *cfs_rq;
8343 struct sched_entity *se = &curr->se;
8345 for_each_sched_entity(se) {
8346 cfs_rq = cfs_rq_of(se);
8347 entity_tick(cfs_rq, se, queued);
8350 if (static_branch_unlikely(&sched_numa_balancing))
8351 task_tick_numa(rq, curr);
8355 * called on fork with the child task as argument from the parent's context
8356 * - child not yet on the tasklist
8357 * - preemption disabled
8359 static void task_fork_fair(struct task_struct *p)
8361 struct cfs_rq *cfs_rq;
8362 struct sched_entity *se = &p->se, *curr;
8363 struct rq *rq = this_rq();
8365 raw_spin_lock(&rq->lock);
8366 update_rq_clock(rq);
8368 cfs_rq = task_cfs_rq(current);
8369 curr = cfs_rq->curr;
8371 update_curr(cfs_rq);
8372 se->vruntime = curr->vruntime;
8374 place_entity(cfs_rq, se, 1);
8376 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8378 * Upon rescheduling, sched_class::put_prev_task() will place
8379 * 'current' within the tree based on its new key value.
8381 swap(curr->vruntime, se->vruntime);
8385 se->vruntime -= cfs_rq->min_vruntime;
8386 raw_spin_unlock(&rq->lock);
8390 * Priority of the task has changed. Check to see if we preempt
8394 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8396 if (!task_on_rq_queued(p))
8400 * Reschedule if we are currently running on this runqueue and
8401 * our priority decreased, or if we are not currently running on
8402 * this runqueue and our priority is higher than the current's
8404 if (rq->curr == p) {
8405 if (p->prio > oldprio)
8408 check_preempt_curr(rq, p, 0);
8411 static inline bool vruntime_normalized(struct task_struct *p)
8413 struct sched_entity *se = &p->se;
8416 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8417 * the dequeue_entity(.flags=0) will already have normalized the
8424 * When !on_rq, vruntime of the task has usually NOT been normalized.
8425 * But there are some cases where it has already been normalized:
8427 * - A forked child which is waiting for being woken up by
8428 * wake_up_new_task().
8429 * - A task which has been woken up by try_to_wake_up() and
8430 * waiting for actually being woken up by sched_ttwu_pending().
8432 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8438 static void detach_task_cfs_rq(struct task_struct *p)
8440 struct sched_entity *se = &p->se;
8441 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8442 u64 now = cfs_rq_clock_task(cfs_rq);
8445 if (!vruntime_normalized(p)) {
8447 * Fix up our vruntime so that the current sleep doesn't
8448 * cause 'unlimited' sleep bonus.
8450 place_entity(cfs_rq, se, 0);
8451 se->vruntime -= cfs_rq->min_vruntime;
8454 /* Catch up with the cfs_rq and remove our load when we leave */
8455 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
8456 detach_entity_load_avg(cfs_rq, se);
8458 update_tg_load_avg(cfs_rq, false);
8461 static void attach_task_cfs_rq(struct task_struct *p)
8463 struct sched_entity *se = &p->se;
8464 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8465 u64 now = cfs_rq_clock_task(cfs_rq);
8468 #ifdef CONFIG_FAIR_GROUP_SCHED
8470 * Since the real-depth could have been changed (only FAIR
8471 * class maintain depth value), reset depth properly.
8473 se->depth = se->parent ? se->parent->depth + 1 : 0;
8476 /* Synchronize task with its cfs_rq */
8477 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
8478 attach_entity_load_avg(cfs_rq, se);
8480 update_tg_load_avg(cfs_rq, false);
8482 if (!vruntime_normalized(p))
8483 se->vruntime += cfs_rq->min_vruntime;
8486 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8488 detach_task_cfs_rq(p);
8491 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8493 attach_task_cfs_rq(p);
8495 if (task_on_rq_queued(p)) {
8497 * We were most likely switched from sched_rt, so
8498 * kick off the schedule if running, otherwise just see
8499 * if we can still preempt the current task.
8504 check_preempt_curr(rq, p, 0);
8508 /* Account for a task changing its policy or group.
8510 * This routine is mostly called to set cfs_rq->curr field when a task
8511 * migrates between groups/classes.
8513 static void set_curr_task_fair(struct rq *rq)
8515 struct sched_entity *se = &rq->curr->se;
8517 for_each_sched_entity(se) {
8518 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8520 set_next_entity(cfs_rq, se);
8521 /* ensure bandwidth has been allocated on our new cfs_rq */
8522 account_cfs_rq_runtime(cfs_rq, 0);
8526 void init_cfs_rq(struct cfs_rq *cfs_rq)
8528 cfs_rq->tasks_timeline = RB_ROOT;
8529 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8530 #ifndef CONFIG_64BIT
8531 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8534 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8535 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8539 #ifdef CONFIG_FAIR_GROUP_SCHED
8540 static void task_set_group_fair(struct task_struct *p)
8542 struct sched_entity *se = &p->se;
8544 set_task_rq(p, task_cpu(p));
8545 se->depth = se->parent ? se->parent->depth + 1 : 0;
8548 static void task_move_group_fair(struct task_struct *p)
8550 detach_task_cfs_rq(p);
8551 set_task_rq(p, task_cpu(p));
8554 /* Tell se's cfs_rq has been changed -- migrated */
8555 p->se.avg.last_update_time = 0;
8557 attach_task_cfs_rq(p);
8560 static void task_change_group_fair(struct task_struct *p, int type)
8563 case TASK_SET_GROUP:
8564 task_set_group_fair(p);
8567 case TASK_MOVE_GROUP:
8568 task_move_group_fair(p);
8573 void free_fair_sched_group(struct task_group *tg)
8577 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8579 for_each_possible_cpu(i) {
8581 kfree(tg->cfs_rq[i]);
8590 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8592 struct sched_entity *se;
8593 struct cfs_rq *cfs_rq;
8597 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8600 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8604 tg->shares = NICE_0_LOAD;
8606 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8608 for_each_possible_cpu(i) {
8611 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8612 GFP_KERNEL, cpu_to_node(i));
8616 se = kzalloc_node(sizeof(struct sched_entity),
8617 GFP_KERNEL, cpu_to_node(i));
8621 init_cfs_rq(cfs_rq);
8622 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8623 init_entity_runnable_average(se);
8634 void online_fair_sched_group(struct task_group *tg)
8636 struct sched_entity *se;
8640 for_each_possible_cpu(i) {
8644 raw_spin_lock_irq(&rq->lock);
8645 post_init_entity_util_avg(se);
8646 sync_throttle(tg, i);
8647 raw_spin_unlock_irq(&rq->lock);
8651 void unregister_fair_sched_group(struct task_group *tg)
8653 unsigned long flags;
8657 for_each_possible_cpu(cpu) {
8659 remove_entity_load_avg(tg->se[cpu]);
8662 * Only empty task groups can be destroyed; so we can speculatively
8663 * check on_list without danger of it being re-added.
8665 if (!tg->cfs_rq[cpu]->on_list)
8670 raw_spin_lock_irqsave(&rq->lock, flags);
8671 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8672 raw_spin_unlock_irqrestore(&rq->lock, flags);
8676 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8677 struct sched_entity *se, int cpu,
8678 struct sched_entity *parent)
8680 struct rq *rq = cpu_rq(cpu);
8684 init_cfs_rq_runtime(cfs_rq);
8686 tg->cfs_rq[cpu] = cfs_rq;
8689 /* se could be NULL for root_task_group */
8694 se->cfs_rq = &rq->cfs;
8697 se->cfs_rq = parent->my_q;
8698 se->depth = parent->depth + 1;
8702 /* guarantee group entities always have weight */
8703 update_load_set(&se->load, NICE_0_LOAD);
8704 se->parent = parent;
8707 static DEFINE_MUTEX(shares_mutex);
8709 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8712 unsigned long flags;
8715 * We can't change the weight of the root cgroup.
8720 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8722 mutex_lock(&shares_mutex);
8723 if (tg->shares == shares)
8726 tg->shares = shares;
8727 for_each_possible_cpu(i) {
8728 struct rq *rq = cpu_rq(i);
8729 struct sched_entity *se;
8732 /* Propagate contribution to hierarchy */
8733 raw_spin_lock_irqsave(&rq->lock, flags);
8735 /* Possible calls to update_curr() need rq clock */
8736 update_rq_clock(rq);
8737 for_each_sched_entity(se)
8738 update_cfs_shares(group_cfs_rq(se));
8739 raw_spin_unlock_irqrestore(&rq->lock, flags);
8743 mutex_unlock(&shares_mutex);
8746 #else /* CONFIG_FAIR_GROUP_SCHED */
8748 void free_fair_sched_group(struct task_group *tg) { }
8750 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8755 void online_fair_sched_group(struct task_group *tg) { }
8757 void unregister_fair_sched_group(struct task_group *tg) { }
8759 #endif /* CONFIG_FAIR_GROUP_SCHED */
8762 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8764 struct sched_entity *se = &task->se;
8765 unsigned int rr_interval = 0;
8768 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8771 if (rq->cfs.load.weight)
8772 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8778 * All the scheduling class methods:
8780 const struct sched_class fair_sched_class = {
8781 .next = &idle_sched_class,
8782 .enqueue_task = enqueue_task_fair,
8783 .dequeue_task = dequeue_task_fair,
8784 .yield_task = yield_task_fair,
8785 .yield_to_task = yield_to_task_fair,
8787 .check_preempt_curr = check_preempt_wakeup,
8789 .pick_next_task = pick_next_task_fair,
8790 .put_prev_task = put_prev_task_fair,
8793 .select_task_rq = select_task_rq_fair,
8794 .migrate_task_rq = migrate_task_rq_fair,
8796 .rq_online = rq_online_fair,
8797 .rq_offline = rq_offline_fair,
8799 .task_dead = task_dead_fair,
8800 .set_cpus_allowed = set_cpus_allowed_common,
8803 .set_curr_task = set_curr_task_fair,
8804 .task_tick = task_tick_fair,
8805 .task_fork = task_fork_fair,
8807 .prio_changed = prio_changed_fair,
8808 .switched_from = switched_from_fair,
8809 .switched_to = switched_to_fair,
8811 .get_rr_interval = get_rr_interval_fair,
8813 .update_curr = update_curr_fair,
8815 #ifdef CONFIG_FAIR_GROUP_SCHED
8816 .task_change_group = task_change_group_fair,
8820 #ifdef CONFIG_SCHED_DEBUG
8821 void print_cfs_stats(struct seq_file *m, int cpu)
8823 struct cfs_rq *cfs_rq;
8826 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8827 print_cfs_rq(m, cpu, cfs_rq);
8831 #ifdef CONFIG_NUMA_BALANCING
8832 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8835 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8837 for_each_online_node(node) {
8838 if (p->numa_faults) {
8839 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8840 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8842 if (p->numa_group) {
8843 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8844 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8846 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8849 #endif /* CONFIG_NUMA_BALANCING */
8850 #endif /* CONFIG_SCHED_DEBUG */
8852 __init void init_sched_fair_class(void)
8855 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8857 #ifdef CONFIG_NO_HZ_COMMON
8858 nohz.next_balance = jiffies;
8859 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);