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 */
694 * With new tasks being created, their initial util_avgs are extrapolated
695 * based on the cfs_rq's current util_avg:
697 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
699 * However, in many cases, the above util_avg does not give a desired
700 * value. Moreover, the sum of the util_avgs may be divergent, such
701 * as when the series is a harmonic series.
703 * To solve this problem, we also cap the util_avg of successive tasks to
704 * only 1/2 of the left utilization budget:
706 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
708 * where n denotes the nth task.
710 * For example, a simplest series from the beginning would be like:
712 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
713 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
715 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
716 * if util_avg > util_avg_cap.
718 void post_init_entity_util_avg(struct sched_entity *se)
720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
721 struct sched_avg *sa = &se->avg;
722 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
725 if (cfs_rq->avg.util_avg != 0) {
726 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
727 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
729 if (sa->util_avg > cap)
734 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
739 void init_entity_runnable_average(struct sched_entity *se)
742 void post_init_entity_util_avg(struct sched_entity *se)
748 * Update the current task's runtime statistics.
750 static void update_curr(struct cfs_rq *cfs_rq)
752 struct sched_entity *curr = cfs_rq->curr;
753 u64 now = rq_clock_task(rq_of(cfs_rq));
759 delta_exec = now - curr->exec_start;
760 if (unlikely((s64)delta_exec <= 0))
763 curr->exec_start = now;
765 schedstat_set(curr->statistics.exec_max,
766 max(delta_exec, curr->statistics.exec_max));
768 curr->sum_exec_runtime += delta_exec;
769 schedstat_add(cfs_rq, exec_clock, delta_exec);
771 curr->vruntime += calc_delta_fair(delta_exec, curr);
772 update_min_vruntime(cfs_rq);
774 if (entity_is_task(curr)) {
775 struct task_struct *curtask = task_of(curr);
777 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
778 cpuacct_charge(curtask, delta_exec);
779 account_group_exec_runtime(curtask, delta_exec);
782 account_cfs_rq_runtime(cfs_rq, delta_exec);
785 static void update_curr_fair(struct rq *rq)
787 update_curr(cfs_rq_of(&rq->curr->se));
790 #ifdef CONFIG_SCHEDSTATS
792 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
794 u64 wait_start = rq_clock(rq_of(cfs_rq));
796 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
797 likely(wait_start > se->statistics.wait_start))
798 wait_start -= se->statistics.wait_start;
800 se->statistics.wait_start = wait_start;
804 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
806 struct task_struct *p;
809 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
811 if (entity_is_task(se)) {
813 if (task_on_rq_migrating(p)) {
815 * Preserve migrating task's wait time so wait_start
816 * time stamp can be adjusted to accumulate wait time
817 * prior to migration.
819 se->statistics.wait_start = delta;
822 trace_sched_stat_wait(p, delta);
825 se->statistics.wait_max = max(se->statistics.wait_max, delta);
826 se->statistics.wait_count++;
827 se->statistics.wait_sum += delta;
828 se->statistics.wait_start = 0;
832 * Task is being enqueued - update stats:
835 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
838 * Are we enqueueing a waiting task? (for current tasks
839 * a dequeue/enqueue event is a NOP)
841 if (se != cfs_rq->curr)
842 update_stats_wait_start(cfs_rq, se);
846 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
849 * Mark the end of the wait period if dequeueing a
852 if (se != cfs_rq->curr)
853 update_stats_wait_end(cfs_rq, se);
855 if (flags & DEQUEUE_SLEEP) {
856 if (entity_is_task(se)) {
857 struct task_struct *tsk = task_of(se);
859 if (tsk->state & TASK_INTERRUPTIBLE)
860 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
861 if (tsk->state & TASK_UNINTERRUPTIBLE)
862 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
869 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
874 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
879 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
890 * We are picking a new current task - update its stats:
893 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
896 * We are starting a new run period:
898 se->exec_start = rq_clock_task(rq_of(cfs_rq));
901 /**************************************************
902 * Scheduling class queueing methods:
905 #ifdef CONFIG_NUMA_BALANCING
907 * Approximate time to scan a full NUMA task in ms. The task scan period is
908 * calculated based on the tasks virtual memory size and
909 * numa_balancing_scan_size.
911 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
912 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
914 /* Portion of address space to scan in MB */
915 unsigned int sysctl_numa_balancing_scan_size = 256;
917 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
918 unsigned int sysctl_numa_balancing_scan_delay = 1000;
920 static unsigned int task_nr_scan_windows(struct task_struct *p)
922 unsigned long rss = 0;
923 unsigned long nr_scan_pages;
926 * Calculations based on RSS as non-present and empty pages are skipped
927 * by the PTE scanner and NUMA hinting faults should be trapped based
930 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
931 rss = get_mm_rss(p->mm);
935 rss = round_up(rss, nr_scan_pages);
936 return rss / nr_scan_pages;
939 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
940 #define MAX_SCAN_WINDOW 2560
942 static unsigned int task_scan_min(struct task_struct *p)
944 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
945 unsigned int scan, floor;
946 unsigned int windows = 1;
948 if (scan_size < MAX_SCAN_WINDOW)
949 windows = MAX_SCAN_WINDOW / scan_size;
950 floor = 1000 / windows;
952 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
953 return max_t(unsigned int, floor, scan);
956 static unsigned int task_scan_max(struct task_struct *p)
958 unsigned int smin = task_scan_min(p);
961 /* Watch for min being lower than max due to floor calculations */
962 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
963 return max(smin, smax);
966 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
968 rq->nr_numa_running += (p->numa_preferred_nid != -1);
969 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
972 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
974 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
975 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
981 spinlock_t lock; /* nr_tasks, tasks */
987 unsigned long total_faults;
988 unsigned long max_faults_cpu;
990 * Faults_cpu is used to decide whether memory should move
991 * towards the CPU. As a consequence, these stats are weighted
992 * more by CPU use than by memory faults.
994 unsigned long *faults_cpu;
995 unsigned long faults[0];
998 /* Shared or private faults. */
999 #define NR_NUMA_HINT_FAULT_TYPES 2
1001 /* Memory and CPU locality */
1002 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1004 /* Averaged statistics, and temporary buffers. */
1005 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1007 pid_t task_numa_group_id(struct task_struct *p)
1009 return p->numa_group ? p->numa_group->gid : 0;
1013 * The averaged statistics, shared & private, memory & cpu,
1014 * occupy the first half of the array. The second half of the
1015 * array is for current counters, which are averaged into the
1016 * first set by task_numa_placement.
1018 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1020 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1023 static inline unsigned long task_faults(struct task_struct *p, int nid)
1025 if (!p->numa_faults)
1028 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1029 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1032 static inline unsigned long group_faults(struct task_struct *p, int nid)
1037 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1038 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1041 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1043 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1044 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1048 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1049 * considered part of a numa group's pseudo-interleaving set. Migrations
1050 * between these nodes are slowed down, to allow things to settle down.
1052 #define ACTIVE_NODE_FRACTION 3
1054 static bool numa_is_active_node(int nid, struct numa_group *ng)
1056 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1059 /* Handle placement on systems where not all nodes are directly connected. */
1060 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1061 int maxdist, bool task)
1063 unsigned long score = 0;
1067 * All nodes are directly connected, and the same distance
1068 * from each other. No need for fancy placement algorithms.
1070 if (sched_numa_topology_type == NUMA_DIRECT)
1074 * This code is called for each node, introducing N^2 complexity,
1075 * which should be ok given the number of nodes rarely exceeds 8.
1077 for_each_online_node(node) {
1078 unsigned long faults;
1079 int dist = node_distance(nid, node);
1082 * The furthest away nodes in the system are not interesting
1083 * for placement; nid was already counted.
1085 if (dist == sched_max_numa_distance || node == nid)
1089 * On systems with a backplane NUMA topology, compare groups
1090 * of nodes, and move tasks towards the group with the most
1091 * memory accesses. When comparing two nodes at distance
1092 * "hoplimit", only nodes closer by than "hoplimit" are part
1093 * of each group. Skip other nodes.
1095 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1099 /* Add up the faults from nearby nodes. */
1101 faults = task_faults(p, node);
1103 faults = group_faults(p, node);
1106 * On systems with a glueless mesh NUMA topology, there are
1107 * no fixed "groups of nodes". Instead, nodes that are not
1108 * directly connected bounce traffic through intermediate
1109 * nodes; a numa_group can occupy any set of nodes.
1110 * The further away a node is, the less the faults count.
1111 * This seems to result in good task placement.
1113 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1114 faults *= (sched_max_numa_distance - dist);
1115 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1125 * These return the fraction of accesses done by a particular task, or
1126 * task group, on a particular numa node. The group weight is given a
1127 * larger multiplier, in order to group tasks together that are almost
1128 * evenly spread out between numa nodes.
1130 static inline unsigned long task_weight(struct task_struct *p, int nid,
1133 unsigned long faults, total_faults;
1135 if (!p->numa_faults)
1138 total_faults = p->total_numa_faults;
1143 faults = task_faults(p, nid);
1144 faults += score_nearby_nodes(p, nid, dist, true);
1146 return 1000 * faults / total_faults;
1149 static inline unsigned long group_weight(struct task_struct *p, int nid,
1152 unsigned long faults, total_faults;
1157 total_faults = p->numa_group->total_faults;
1162 faults = group_faults(p, nid);
1163 faults += score_nearby_nodes(p, nid, dist, false);
1165 return 1000 * faults / total_faults;
1168 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1169 int src_nid, int dst_cpu)
1171 struct numa_group *ng = p->numa_group;
1172 int dst_nid = cpu_to_node(dst_cpu);
1173 int last_cpupid, this_cpupid;
1175 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1178 * Multi-stage node selection is used in conjunction with a periodic
1179 * migration fault to build a temporal task<->page relation. By using
1180 * a two-stage filter we remove short/unlikely relations.
1182 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1183 * a task's usage of a particular page (n_p) per total usage of this
1184 * page (n_t) (in a given time-span) to a probability.
1186 * Our periodic faults will sample this probability and getting the
1187 * same result twice in a row, given these samples are fully
1188 * independent, is then given by P(n)^2, provided our sample period
1189 * is sufficiently short compared to the usage pattern.
1191 * This quadric squishes small probabilities, making it less likely we
1192 * act on an unlikely task<->page relation.
1194 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1195 if (!cpupid_pid_unset(last_cpupid) &&
1196 cpupid_to_nid(last_cpupid) != dst_nid)
1199 /* Always allow migrate on private faults */
1200 if (cpupid_match_pid(p, last_cpupid))
1203 /* A shared fault, but p->numa_group has not been set up yet. */
1208 * Destination node is much more heavily used than the source
1209 * node? Allow migration.
1211 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1212 ACTIVE_NODE_FRACTION)
1216 * Distribute memory according to CPU & memory use on each node,
1217 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1219 * faults_cpu(dst) 3 faults_cpu(src)
1220 * --------------- * - > ---------------
1221 * faults_mem(dst) 4 faults_mem(src)
1223 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1224 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1227 static unsigned long weighted_cpuload(const int cpu);
1228 static unsigned long source_load(int cpu, int type);
1229 static unsigned long target_load(int cpu, int type);
1230 static unsigned long capacity_of(int cpu);
1231 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1233 /* Cached statistics for all CPUs within a node */
1235 unsigned long nr_running;
1238 /* Total compute capacity of CPUs on a node */
1239 unsigned long compute_capacity;
1241 /* Approximate capacity in terms of runnable tasks on a node */
1242 unsigned long task_capacity;
1243 int has_free_capacity;
1247 * XXX borrowed from update_sg_lb_stats
1249 static void update_numa_stats(struct numa_stats *ns, int nid)
1251 int smt, cpu, cpus = 0;
1252 unsigned long capacity;
1254 memset(ns, 0, sizeof(*ns));
1255 for_each_cpu(cpu, cpumask_of_node(nid)) {
1256 struct rq *rq = cpu_rq(cpu);
1258 ns->nr_running += rq->nr_running;
1259 ns->load += weighted_cpuload(cpu);
1260 ns->compute_capacity += capacity_of(cpu);
1266 * If we raced with hotplug and there are no CPUs left in our mask
1267 * the @ns structure is NULL'ed and task_numa_compare() will
1268 * not find this node attractive.
1270 * We'll either bail at !has_free_capacity, or we'll detect a huge
1271 * imbalance and bail there.
1276 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1277 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1278 capacity = cpus / smt; /* cores */
1280 ns->task_capacity = min_t(unsigned, capacity,
1281 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1282 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1285 struct task_numa_env {
1286 struct task_struct *p;
1288 int src_cpu, src_nid;
1289 int dst_cpu, dst_nid;
1291 struct numa_stats src_stats, dst_stats;
1296 struct task_struct *best_task;
1301 static void task_numa_assign(struct task_numa_env *env,
1302 struct task_struct *p, long imp)
1305 put_task_struct(env->best_task);
1308 env->best_imp = imp;
1309 env->best_cpu = env->dst_cpu;
1312 static bool load_too_imbalanced(long src_load, long dst_load,
1313 struct task_numa_env *env)
1316 long orig_src_load, orig_dst_load;
1317 long src_capacity, dst_capacity;
1320 * The load is corrected for the CPU capacity available on each node.
1323 * ------------ vs ---------
1324 * src_capacity dst_capacity
1326 src_capacity = env->src_stats.compute_capacity;
1327 dst_capacity = env->dst_stats.compute_capacity;
1329 /* We care about the slope of the imbalance, not the direction. */
1330 if (dst_load < src_load)
1331 swap(dst_load, src_load);
1333 /* Is the difference below the threshold? */
1334 imb = dst_load * src_capacity * 100 -
1335 src_load * dst_capacity * env->imbalance_pct;
1340 * The imbalance is above the allowed threshold.
1341 * Compare it with the old imbalance.
1343 orig_src_load = env->src_stats.load;
1344 orig_dst_load = env->dst_stats.load;
1346 if (orig_dst_load < orig_src_load)
1347 swap(orig_dst_load, orig_src_load);
1349 old_imb = orig_dst_load * src_capacity * 100 -
1350 orig_src_load * dst_capacity * env->imbalance_pct;
1352 /* Would this change make things worse? */
1353 return (imb > old_imb);
1357 * This checks if the overall compute and NUMA accesses of the system would
1358 * be improved if the source tasks was migrated to the target dst_cpu taking
1359 * into account that it might be best if task running on the dst_cpu should
1360 * be exchanged with the source task
1362 static void task_numa_compare(struct task_numa_env *env,
1363 long taskimp, long groupimp)
1365 struct rq *src_rq = cpu_rq(env->src_cpu);
1366 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1367 struct task_struct *cur;
1368 long src_load, dst_load;
1370 long imp = env->p->numa_group ? groupimp : taskimp;
1372 int dist = env->dist;
1373 bool assigned = false;
1377 raw_spin_lock_irq(&dst_rq->lock);
1380 * No need to move the exiting task or idle task.
1382 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1386 * The task_struct must be protected here to protect the
1387 * p->numa_faults access in the task_weight since the
1388 * numa_faults could already be freed in the following path:
1389 * finish_task_switch()
1390 * --> put_task_struct()
1391 * --> __put_task_struct()
1392 * --> task_numa_free()
1394 get_task_struct(cur);
1397 raw_spin_unlock_irq(&dst_rq->lock);
1400 * Because we have preemption enabled we can get migrated around and
1401 * end try selecting ourselves (current == env->p) as a swap candidate.
1407 * "imp" is the fault differential for the source task between the
1408 * source and destination node. Calculate the total differential for
1409 * the source task and potential destination task. The more negative
1410 * the value is, the more rmeote accesses that would be expected to
1411 * be incurred if the tasks were swapped.
1414 /* Skip this swap candidate if cannot move to the source cpu */
1415 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1419 * If dst and source tasks are in the same NUMA group, or not
1420 * in any group then look only at task weights.
1422 if (cur->numa_group == env->p->numa_group) {
1423 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1424 task_weight(cur, env->dst_nid, dist);
1426 * Add some hysteresis to prevent swapping the
1427 * tasks within a group over tiny differences.
1429 if (cur->numa_group)
1433 * Compare the group weights. If a task is all by
1434 * itself (not part of a group), use the task weight
1437 if (cur->numa_group)
1438 imp += group_weight(cur, env->src_nid, dist) -
1439 group_weight(cur, env->dst_nid, dist);
1441 imp += task_weight(cur, env->src_nid, dist) -
1442 task_weight(cur, env->dst_nid, dist);
1446 if (imp <= env->best_imp && moveimp <= env->best_imp)
1450 /* Is there capacity at our destination? */
1451 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1452 !env->dst_stats.has_free_capacity)
1458 /* Balance doesn't matter much if we're running a task per cpu */
1459 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1460 dst_rq->nr_running == 1)
1464 * In the overloaded case, try and keep the load balanced.
1467 load = task_h_load(env->p);
1468 dst_load = env->dst_stats.load + load;
1469 src_load = env->src_stats.load - load;
1471 if (moveimp > imp && moveimp > env->best_imp) {
1473 * If the improvement from just moving env->p direction is
1474 * better than swapping tasks around, check if a move is
1475 * possible. Store a slightly smaller score than moveimp,
1476 * so an actually idle CPU will win.
1478 if (!load_too_imbalanced(src_load, dst_load, env)) {
1480 put_task_struct(cur);
1486 if (imp <= env->best_imp)
1490 load = task_h_load(cur);
1495 if (load_too_imbalanced(src_load, dst_load, env))
1499 * One idle CPU per node is evaluated for a task numa move.
1500 * Call select_idle_sibling to maybe find a better one.
1503 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1507 task_numa_assign(env, cur, imp);
1511 * The dst_rq->curr isn't assigned. The protection for task_struct is
1514 if (cur && !assigned)
1515 put_task_struct(cur);
1518 static void task_numa_find_cpu(struct task_numa_env *env,
1519 long taskimp, long groupimp)
1523 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1524 /* Skip this CPU if the source task cannot migrate */
1525 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1529 task_numa_compare(env, taskimp, groupimp);
1533 /* Only move tasks to a NUMA node less busy than the current node. */
1534 static bool numa_has_capacity(struct task_numa_env *env)
1536 struct numa_stats *src = &env->src_stats;
1537 struct numa_stats *dst = &env->dst_stats;
1539 if (src->has_free_capacity && !dst->has_free_capacity)
1543 * Only consider a task move if the source has a higher load
1544 * than the destination, corrected for CPU capacity on each node.
1546 * src->load dst->load
1547 * --------------------- vs ---------------------
1548 * src->compute_capacity dst->compute_capacity
1550 if (src->load * dst->compute_capacity * env->imbalance_pct >
1552 dst->load * src->compute_capacity * 100)
1558 static int task_numa_migrate(struct task_struct *p)
1560 struct task_numa_env env = {
1563 .src_cpu = task_cpu(p),
1564 .src_nid = task_node(p),
1566 .imbalance_pct = 112,
1572 struct sched_domain *sd;
1573 unsigned long taskweight, groupweight;
1575 long taskimp, groupimp;
1578 * Pick the lowest SD_NUMA domain, as that would have the smallest
1579 * imbalance and would be the first to start moving tasks about.
1581 * And we want to avoid any moving of tasks about, as that would create
1582 * random movement of tasks -- counter the numa conditions we're trying
1586 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1588 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1592 * Cpusets can break the scheduler domain tree into smaller
1593 * balance domains, some of which do not cross NUMA boundaries.
1594 * Tasks that are "trapped" in such domains cannot be migrated
1595 * elsewhere, so there is no point in (re)trying.
1597 if (unlikely(!sd)) {
1598 p->numa_preferred_nid = task_node(p);
1602 env.dst_nid = p->numa_preferred_nid;
1603 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1604 taskweight = task_weight(p, env.src_nid, dist);
1605 groupweight = group_weight(p, env.src_nid, dist);
1606 update_numa_stats(&env.src_stats, env.src_nid);
1607 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1608 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1609 update_numa_stats(&env.dst_stats, env.dst_nid);
1611 /* Try to find a spot on the preferred nid. */
1612 if (numa_has_capacity(&env))
1613 task_numa_find_cpu(&env, taskimp, groupimp);
1616 * Look at other nodes in these cases:
1617 * - there is no space available on the preferred_nid
1618 * - the task is part of a numa_group that is interleaved across
1619 * multiple NUMA nodes; in order to better consolidate the group,
1620 * we need to check other locations.
1622 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1623 for_each_online_node(nid) {
1624 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1627 dist = node_distance(env.src_nid, env.dst_nid);
1628 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1630 taskweight = task_weight(p, env.src_nid, dist);
1631 groupweight = group_weight(p, env.src_nid, dist);
1634 /* Only consider nodes where both task and groups benefit */
1635 taskimp = task_weight(p, nid, dist) - taskweight;
1636 groupimp = group_weight(p, nid, dist) - groupweight;
1637 if (taskimp < 0 && groupimp < 0)
1642 update_numa_stats(&env.dst_stats, env.dst_nid);
1643 if (numa_has_capacity(&env))
1644 task_numa_find_cpu(&env, taskimp, groupimp);
1649 * If the task is part of a workload that spans multiple NUMA nodes,
1650 * and is migrating into one of the workload's active nodes, remember
1651 * this node as the task's preferred numa node, so the workload can
1653 * A task that migrated to a second choice node will be better off
1654 * trying for a better one later. Do not set the preferred node here.
1656 if (p->numa_group) {
1657 struct numa_group *ng = p->numa_group;
1659 if (env.best_cpu == -1)
1664 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1665 sched_setnuma(p, env.dst_nid);
1668 /* No better CPU than the current one was found. */
1669 if (env.best_cpu == -1)
1673 * Reset the scan period if the task is being rescheduled on an
1674 * alternative node to recheck if the tasks is now properly placed.
1676 p->numa_scan_period = task_scan_min(p);
1678 if (env.best_task == NULL) {
1679 ret = migrate_task_to(p, env.best_cpu);
1681 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1685 ret = migrate_swap(p, env.best_task);
1687 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1688 put_task_struct(env.best_task);
1692 /* Attempt to migrate a task to a CPU on the preferred node. */
1693 static void numa_migrate_preferred(struct task_struct *p)
1695 unsigned long interval = HZ;
1697 /* This task has no NUMA fault statistics yet */
1698 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1701 /* Periodically retry migrating the task to the preferred node */
1702 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1703 p->numa_migrate_retry = jiffies + interval;
1705 /* Success if task is already running on preferred CPU */
1706 if (task_node(p) == p->numa_preferred_nid)
1709 /* Otherwise, try migrate to a CPU on the preferred node */
1710 task_numa_migrate(p);
1714 * Find out how many nodes on the workload is actively running on. Do this by
1715 * tracking the nodes from which NUMA hinting faults are triggered. This can
1716 * be different from the set of nodes where the workload's memory is currently
1719 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1721 unsigned long faults, max_faults = 0;
1722 int nid, active_nodes = 0;
1724 for_each_online_node(nid) {
1725 faults = group_faults_cpu(numa_group, nid);
1726 if (faults > max_faults)
1727 max_faults = faults;
1730 for_each_online_node(nid) {
1731 faults = group_faults_cpu(numa_group, nid);
1732 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1736 numa_group->max_faults_cpu = max_faults;
1737 numa_group->active_nodes = active_nodes;
1741 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1742 * increments. The more local the fault statistics are, the higher the scan
1743 * period will be for the next scan window. If local/(local+remote) ratio is
1744 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1745 * the scan period will decrease. Aim for 70% local accesses.
1747 #define NUMA_PERIOD_SLOTS 10
1748 #define NUMA_PERIOD_THRESHOLD 7
1751 * Increase the scan period (slow down scanning) if the majority of
1752 * our memory is already on our local node, or if the majority of
1753 * the page accesses are shared with other processes.
1754 * Otherwise, decrease the scan period.
1756 static void update_task_scan_period(struct task_struct *p,
1757 unsigned long shared, unsigned long private)
1759 unsigned int period_slot;
1763 unsigned long remote = p->numa_faults_locality[0];
1764 unsigned long local = p->numa_faults_locality[1];
1767 * If there were no record hinting faults then either the task is
1768 * completely idle or all activity is areas that are not of interest
1769 * to automatic numa balancing. Related to that, if there were failed
1770 * migration then it implies we are migrating too quickly or the local
1771 * node is overloaded. In either case, scan slower
1773 if (local + shared == 0 || p->numa_faults_locality[2]) {
1774 p->numa_scan_period = min(p->numa_scan_period_max,
1775 p->numa_scan_period << 1);
1777 p->mm->numa_next_scan = jiffies +
1778 msecs_to_jiffies(p->numa_scan_period);
1784 * Prepare to scale scan period relative to the current period.
1785 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1786 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1787 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1789 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1790 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1791 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1792 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1795 diff = slot * period_slot;
1797 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1800 * Scale scan rate increases based on sharing. There is an
1801 * inverse relationship between the degree of sharing and
1802 * the adjustment made to the scanning period. Broadly
1803 * speaking the intent is that there is little point
1804 * scanning faster if shared accesses dominate as it may
1805 * simply bounce migrations uselessly
1807 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1808 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1811 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1812 task_scan_min(p), task_scan_max(p));
1813 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1817 * Get the fraction of time the task has been running since the last
1818 * NUMA placement cycle. The scheduler keeps similar statistics, but
1819 * decays those on a 32ms period, which is orders of magnitude off
1820 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1821 * stats only if the task is so new there are no NUMA statistics yet.
1823 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1825 u64 runtime, delta, now;
1826 /* Use the start of this time slice to avoid calculations. */
1827 now = p->se.exec_start;
1828 runtime = p->se.sum_exec_runtime;
1830 if (p->last_task_numa_placement) {
1831 delta = runtime - p->last_sum_exec_runtime;
1832 *period = now - p->last_task_numa_placement;
1834 delta = p->se.avg.load_sum / p->se.load.weight;
1835 *period = LOAD_AVG_MAX;
1838 p->last_sum_exec_runtime = runtime;
1839 p->last_task_numa_placement = now;
1845 * Determine the preferred nid for a task in a numa_group. This needs to
1846 * be done in a way that produces consistent results with group_weight,
1847 * otherwise workloads might not converge.
1849 static int preferred_group_nid(struct task_struct *p, int nid)
1854 /* Direct connections between all NUMA nodes. */
1855 if (sched_numa_topology_type == NUMA_DIRECT)
1859 * On a system with glueless mesh NUMA topology, group_weight
1860 * scores nodes according to the number of NUMA hinting faults on
1861 * both the node itself, and on nearby nodes.
1863 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1864 unsigned long score, max_score = 0;
1865 int node, max_node = nid;
1867 dist = sched_max_numa_distance;
1869 for_each_online_node(node) {
1870 score = group_weight(p, node, dist);
1871 if (score > max_score) {
1880 * Finding the preferred nid in a system with NUMA backplane
1881 * interconnect topology is more involved. The goal is to locate
1882 * tasks from numa_groups near each other in the system, and
1883 * untangle workloads from different sides of the system. This requires
1884 * searching down the hierarchy of node groups, recursively searching
1885 * inside the highest scoring group of nodes. The nodemask tricks
1886 * keep the complexity of the search down.
1888 nodes = node_online_map;
1889 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1890 unsigned long max_faults = 0;
1891 nodemask_t max_group = NODE_MASK_NONE;
1894 /* Are there nodes at this distance from each other? */
1895 if (!find_numa_distance(dist))
1898 for_each_node_mask(a, nodes) {
1899 unsigned long faults = 0;
1900 nodemask_t this_group;
1901 nodes_clear(this_group);
1903 /* Sum group's NUMA faults; includes a==b case. */
1904 for_each_node_mask(b, nodes) {
1905 if (node_distance(a, b) < dist) {
1906 faults += group_faults(p, b);
1907 node_set(b, this_group);
1908 node_clear(b, nodes);
1912 /* Remember the top group. */
1913 if (faults > max_faults) {
1914 max_faults = faults;
1915 max_group = this_group;
1917 * subtle: at the smallest distance there is
1918 * just one node left in each "group", the
1919 * winner is the preferred nid.
1924 /* Next round, evaluate the nodes within max_group. */
1932 static void task_numa_placement(struct task_struct *p)
1934 int seq, nid, max_nid = -1, max_group_nid = -1;
1935 unsigned long max_faults = 0, max_group_faults = 0;
1936 unsigned long fault_types[2] = { 0, 0 };
1937 unsigned long total_faults;
1938 u64 runtime, period;
1939 spinlock_t *group_lock = NULL;
1942 * The p->mm->numa_scan_seq field gets updated without
1943 * exclusive access. Use READ_ONCE() here to ensure
1944 * that the field is read in a single access:
1946 seq = READ_ONCE(p->mm->numa_scan_seq);
1947 if (p->numa_scan_seq == seq)
1949 p->numa_scan_seq = seq;
1950 p->numa_scan_period_max = task_scan_max(p);
1952 total_faults = p->numa_faults_locality[0] +
1953 p->numa_faults_locality[1];
1954 runtime = numa_get_avg_runtime(p, &period);
1956 /* If the task is part of a group prevent parallel updates to group stats */
1957 if (p->numa_group) {
1958 group_lock = &p->numa_group->lock;
1959 spin_lock_irq(group_lock);
1962 /* Find the node with the highest number of faults */
1963 for_each_online_node(nid) {
1964 /* Keep track of the offsets in numa_faults array */
1965 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1966 unsigned long faults = 0, group_faults = 0;
1969 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1970 long diff, f_diff, f_weight;
1972 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1973 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1974 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1975 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1977 /* Decay existing window, copy faults since last scan */
1978 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1979 fault_types[priv] += p->numa_faults[membuf_idx];
1980 p->numa_faults[membuf_idx] = 0;
1983 * Normalize the faults_from, so all tasks in a group
1984 * count according to CPU use, instead of by the raw
1985 * number of faults. Tasks with little runtime have
1986 * little over-all impact on throughput, and thus their
1987 * faults are less important.
1989 f_weight = div64_u64(runtime << 16, period + 1);
1990 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1992 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1993 p->numa_faults[cpubuf_idx] = 0;
1995 p->numa_faults[mem_idx] += diff;
1996 p->numa_faults[cpu_idx] += f_diff;
1997 faults += p->numa_faults[mem_idx];
1998 p->total_numa_faults += diff;
1999 if (p->numa_group) {
2001 * safe because we can only change our own group
2003 * mem_idx represents the offset for a given
2004 * nid and priv in a specific region because it
2005 * is at the beginning of the numa_faults array.
2007 p->numa_group->faults[mem_idx] += diff;
2008 p->numa_group->faults_cpu[mem_idx] += f_diff;
2009 p->numa_group->total_faults += diff;
2010 group_faults += p->numa_group->faults[mem_idx];
2014 if (faults > max_faults) {
2015 max_faults = faults;
2019 if (group_faults > max_group_faults) {
2020 max_group_faults = group_faults;
2021 max_group_nid = nid;
2025 update_task_scan_period(p, fault_types[0], fault_types[1]);
2027 if (p->numa_group) {
2028 numa_group_count_active_nodes(p->numa_group);
2029 spin_unlock_irq(group_lock);
2030 max_nid = preferred_group_nid(p, max_group_nid);
2034 /* Set the new preferred node */
2035 if (max_nid != p->numa_preferred_nid)
2036 sched_setnuma(p, max_nid);
2038 if (task_node(p) != p->numa_preferred_nid)
2039 numa_migrate_preferred(p);
2043 static inline int get_numa_group(struct numa_group *grp)
2045 return atomic_inc_not_zero(&grp->refcount);
2048 static inline void put_numa_group(struct numa_group *grp)
2050 if (atomic_dec_and_test(&grp->refcount))
2051 kfree_rcu(grp, rcu);
2054 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2057 struct numa_group *grp, *my_grp;
2058 struct task_struct *tsk;
2060 int cpu = cpupid_to_cpu(cpupid);
2063 if (unlikely(!p->numa_group)) {
2064 unsigned int size = sizeof(struct numa_group) +
2065 4*nr_node_ids*sizeof(unsigned long);
2067 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2071 atomic_set(&grp->refcount, 1);
2072 grp->active_nodes = 1;
2073 grp->max_faults_cpu = 0;
2074 spin_lock_init(&grp->lock);
2076 /* Second half of the array tracks nids where faults happen */
2077 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2080 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2081 grp->faults[i] = p->numa_faults[i];
2083 grp->total_faults = p->total_numa_faults;
2086 rcu_assign_pointer(p->numa_group, grp);
2090 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2092 if (!cpupid_match_pid(tsk, cpupid))
2095 grp = rcu_dereference(tsk->numa_group);
2099 my_grp = p->numa_group;
2104 * Only join the other group if its bigger; if we're the bigger group,
2105 * the other task will join us.
2107 if (my_grp->nr_tasks > grp->nr_tasks)
2111 * Tie-break on the grp address.
2113 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2116 /* Always join threads in the same process. */
2117 if (tsk->mm == current->mm)
2120 /* Simple filter to avoid false positives due to PID collisions */
2121 if (flags & TNF_SHARED)
2124 /* Update priv based on whether false sharing was detected */
2127 if (join && !get_numa_group(grp))
2135 BUG_ON(irqs_disabled());
2136 double_lock_irq(&my_grp->lock, &grp->lock);
2138 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2139 my_grp->faults[i] -= p->numa_faults[i];
2140 grp->faults[i] += p->numa_faults[i];
2142 my_grp->total_faults -= p->total_numa_faults;
2143 grp->total_faults += p->total_numa_faults;
2148 spin_unlock(&my_grp->lock);
2149 spin_unlock_irq(&grp->lock);
2151 rcu_assign_pointer(p->numa_group, grp);
2153 put_numa_group(my_grp);
2161 void task_numa_free(struct task_struct *p)
2163 struct numa_group *grp = p->numa_group;
2164 void *numa_faults = p->numa_faults;
2165 unsigned long flags;
2169 spin_lock_irqsave(&grp->lock, flags);
2170 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2171 grp->faults[i] -= p->numa_faults[i];
2172 grp->total_faults -= p->total_numa_faults;
2175 spin_unlock_irqrestore(&grp->lock, flags);
2176 RCU_INIT_POINTER(p->numa_group, NULL);
2177 put_numa_group(grp);
2180 p->numa_faults = NULL;
2185 * Got a PROT_NONE fault for a page on @node.
2187 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2189 struct task_struct *p = current;
2190 bool migrated = flags & TNF_MIGRATED;
2191 int cpu_node = task_node(current);
2192 int local = !!(flags & TNF_FAULT_LOCAL);
2193 struct numa_group *ng;
2196 if (!static_branch_likely(&sched_numa_balancing))
2199 /* for example, ksmd faulting in a user's mm */
2203 /* Allocate buffer to track faults on a per-node basis */
2204 if (unlikely(!p->numa_faults)) {
2205 int size = sizeof(*p->numa_faults) *
2206 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2208 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2209 if (!p->numa_faults)
2212 p->total_numa_faults = 0;
2213 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2217 * First accesses are treated as private, otherwise consider accesses
2218 * to be private if the accessing pid has not changed
2220 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2223 priv = cpupid_match_pid(p, last_cpupid);
2224 if (!priv && !(flags & TNF_NO_GROUP))
2225 task_numa_group(p, last_cpupid, flags, &priv);
2229 * If a workload spans multiple NUMA nodes, a shared fault that
2230 * occurs wholly within the set of nodes that the workload is
2231 * actively using should be counted as local. This allows the
2232 * scan rate to slow down when a workload has settled down.
2235 if (!priv && !local && ng && ng->active_nodes > 1 &&
2236 numa_is_active_node(cpu_node, ng) &&
2237 numa_is_active_node(mem_node, ng))
2240 task_numa_placement(p);
2243 * Retry task to preferred node migration periodically, in case it
2244 * case it previously failed, or the scheduler moved us.
2246 if (time_after(jiffies, p->numa_migrate_retry))
2247 numa_migrate_preferred(p);
2250 p->numa_pages_migrated += pages;
2251 if (flags & TNF_MIGRATE_FAIL)
2252 p->numa_faults_locality[2] += pages;
2254 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2255 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2256 p->numa_faults_locality[local] += pages;
2259 static void reset_ptenuma_scan(struct task_struct *p)
2262 * We only did a read acquisition of the mmap sem, so
2263 * p->mm->numa_scan_seq is written to without exclusive access
2264 * and the update is not guaranteed to be atomic. That's not
2265 * much of an issue though, since this is just used for
2266 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2267 * expensive, to avoid any form of compiler optimizations:
2269 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2270 p->mm->numa_scan_offset = 0;
2274 * The expensive part of numa migration is done from task_work context.
2275 * Triggered from task_tick_numa().
2277 void task_numa_work(struct callback_head *work)
2279 unsigned long migrate, next_scan, now = jiffies;
2280 struct task_struct *p = current;
2281 struct mm_struct *mm = p->mm;
2282 u64 runtime = p->se.sum_exec_runtime;
2283 struct vm_area_struct *vma;
2284 unsigned long start, end;
2285 unsigned long nr_pte_updates = 0;
2286 long pages, virtpages;
2288 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2290 work->next = work; /* protect against double add */
2292 * Who cares about NUMA placement when they're dying.
2294 * NOTE: make sure not to dereference p->mm before this check,
2295 * exit_task_work() happens _after_ exit_mm() so we could be called
2296 * without p->mm even though we still had it when we enqueued this
2299 if (p->flags & PF_EXITING)
2302 if (!mm->numa_next_scan) {
2303 mm->numa_next_scan = now +
2304 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2308 * Enforce maximal scan/migration frequency..
2310 migrate = mm->numa_next_scan;
2311 if (time_before(now, migrate))
2314 if (p->numa_scan_period == 0) {
2315 p->numa_scan_period_max = task_scan_max(p);
2316 p->numa_scan_period = task_scan_min(p);
2319 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2320 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2324 * Delay this task enough that another task of this mm will likely win
2325 * the next time around.
2327 p->node_stamp += 2 * TICK_NSEC;
2329 start = mm->numa_scan_offset;
2330 pages = sysctl_numa_balancing_scan_size;
2331 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2332 virtpages = pages * 8; /* Scan up to this much virtual space */
2337 down_read(&mm->mmap_sem);
2338 vma = find_vma(mm, start);
2340 reset_ptenuma_scan(p);
2344 for (; vma; vma = vma->vm_next) {
2345 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2346 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2351 * Shared library pages mapped by multiple processes are not
2352 * migrated as it is expected they are cache replicated. Avoid
2353 * hinting faults in read-only file-backed mappings or the vdso
2354 * as migrating the pages will be of marginal benefit.
2357 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2361 * Skip inaccessible VMAs to avoid any confusion between
2362 * PROT_NONE and NUMA hinting ptes
2364 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2368 start = max(start, vma->vm_start);
2369 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2370 end = min(end, vma->vm_end);
2371 nr_pte_updates = change_prot_numa(vma, start, end);
2374 * Try to scan sysctl_numa_balancing_size worth of
2375 * hpages that have at least one present PTE that
2376 * is not already pte-numa. If the VMA contains
2377 * areas that are unused or already full of prot_numa
2378 * PTEs, scan up to virtpages, to skip through those
2382 pages -= (end - start) >> PAGE_SHIFT;
2383 virtpages -= (end - start) >> PAGE_SHIFT;
2386 if (pages <= 0 || virtpages <= 0)
2390 } while (end != vma->vm_end);
2395 * It is possible to reach the end of the VMA list but the last few
2396 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2397 * would find the !migratable VMA on the next scan but not reset the
2398 * scanner to the start so check it now.
2401 mm->numa_scan_offset = start;
2403 reset_ptenuma_scan(p);
2404 up_read(&mm->mmap_sem);
2407 * Make sure tasks use at least 32x as much time to run other code
2408 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2409 * Usually update_task_scan_period slows down scanning enough; on an
2410 * overloaded system we need to limit overhead on a per task basis.
2412 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2413 u64 diff = p->se.sum_exec_runtime - runtime;
2414 p->node_stamp += 32 * diff;
2419 * Drive the periodic memory faults..
2421 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2423 struct callback_head *work = &curr->numa_work;
2427 * We don't care about NUMA placement if we don't have memory.
2429 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2433 * Using runtime rather than walltime has the dual advantage that
2434 * we (mostly) drive the selection from busy threads and that the
2435 * task needs to have done some actual work before we bother with
2438 now = curr->se.sum_exec_runtime;
2439 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2441 if (now > curr->node_stamp + period) {
2442 if (!curr->node_stamp)
2443 curr->numa_scan_period = task_scan_min(curr);
2444 curr->node_stamp += period;
2446 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2447 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2448 task_work_add(curr, work, true);
2453 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2457 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2461 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2464 #endif /* CONFIG_NUMA_BALANCING */
2467 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2469 update_load_add(&cfs_rq->load, se->load.weight);
2470 if (!parent_entity(se))
2471 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2473 if (entity_is_task(se)) {
2474 struct rq *rq = rq_of(cfs_rq);
2476 account_numa_enqueue(rq, task_of(se));
2477 list_add(&se->group_node, &rq->cfs_tasks);
2480 cfs_rq->nr_running++;
2484 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2486 update_load_sub(&cfs_rq->load, se->load.weight);
2487 if (!parent_entity(se))
2488 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2490 if (entity_is_task(se)) {
2491 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2492 list_del_init(&se->group_node);
2495 cfs_rq->nr_running--;
2498 #ifdef CONFIG_FAIR_GROUP_SCHED
2500 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2502 long tg_weight, load, shares;
2505 * This really should be: cfs_rq->avg.load_avg, but instead we use
2506 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2507 * the shares for small weight interactive tasks.
2509 load = scale_load_down(cfs_rq->load.weight);
2511 tg_weight = atomic_long_read(&tg->load_avg);
2513 /* Ensure tg_weight >= load */
2514 tg_weight -= cfs_rq->tg_load_avg_contrib;
2517 shares = (tg->shares * load);
2519 shares /= tg_weight;
2521 if (shares < MIN_SHARES)
2522 shares = MIN_SHARES;
2523 if (shares > tg->shares)
2524 shares = tg->shares;
2528 # else /* CONFIG_SMP */
2529 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2533 # endif /* CONFIG_SMP */
2535 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2536 unsigned long weight)
2539 /* commit outstanding execution time */
2540 if (cfs_rq->curr == se)
2541 update_curr(cfs_rq);
2542 account_entity_dequeue(cfs_rq, se);
2545 update_load_set(&se->load, weight);
2548 account_entity_enqueue(cfs_rq, se);
2551 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2553 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2555 struct task_group *tg;
2556 struct sched_entity *se;
2560 se = tg->se[cpu_of(rq_of(cfs_rq))];
2561 if (!se || throttled_hierarchy(cfs_rq))
2564 if (likely(se->load.weight == tg->shares))
2567 shares = calc_cfs_shares(cfs_rq, tg);
2569 reweight_entity(cfs_rq_of(se), se, shares);
2571 #else /* CONFIG_FAIR_GROUP_SCHED */
2572 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2575 #endif /* CONFIG_FAIR_GROUP_SCHED */
2578 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2579 static const u32 runnable_avg_yN_inv[] = {
2580 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2581 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2582 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2583 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2584 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2585 0x85aac367, 0x82cd8698,
2589 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2590 * over-estimates when re-combining.
2592 static const u32 runnable_avg_yN_sum[] = {
2593 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2594 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2595 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2599 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2600 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2603 static const u32 __accumulated_sum_N32[] = {
2604 0, 23371, 35056, 40899, 43820, 45281,
2605 46011, 46376, 46559, 46650, 46696, 46719,
2610 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2612 static __always_inline u64 decay_load(u64 val, u64 n)
2614 unsigned int local_n;
2618 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2621 /* after bounds checking we can collapse to 32-bit */
2625 * As y^PERIOD = 1/2, we can combine
2626 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2627 * With a look-up table which covers y^n (n<PERIOD)
2629 * To achieve constant time decay_load.
2631 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2632 val >>= local_n / LOAD_AVG_PERIOD;
2633 local_n %= LOAD_AVG_PERIOD;
2636 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2641 * For updates fully spanning n periods, the contribution to runnable
2642 * average will be: \Sum 1024*y^n
2644 * We can compute this reasonably efficiently by combining:
2645 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2647 static u32 __compute_runnable_contrib(u64 n)
2651 if (likely(n <= LOAD_AVG_PERIOD))
2652 return runnable_avg_yN_sum[n];
2653 else if (unlikely(n >= LOAD_AVG_MAX_N))
2654 return LOAD_AVG_MAX;
2656 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2657 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2658 n %= LOAD_AVG_PERIOD;
2659 contrib = decay_load(contrib, n);
2660 return contrib + runnable_avg_yN_sum[n];
2663 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2666 * We can represent the historical contribution to runnable average as the
2667 * coefficients of a geometric series. To do this we sub-divide our runnable
2668 * history into segments of approximately 1ms (1024us); label the segment that
2669 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2671 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2673 * (now) (~1ms ago) (~2ms ago)
2675 * Let u_i denote the fraction of p_i that the entity was runnable.
2677 * We then designate the fractions u_i as our co-efficients, yielding the
2678 * following representation of historical load:
2679 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2681 * We choose y based on the with of a reasonably scheduling period, fixing:
2684 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2685 * approximately half as much as the contribution to load within the last ms
2688 * When a period "rolls over" and we have new u_0`, multiplying the previous
2689 * sum again by y is sufficient to update:
2690 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2691 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2693 static __always_inline int
2694 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2695 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2697 u64 delta, scaled_delta, periods;
2699 unsigned int delta_w, scaled_delta_w, decayed = 0;
2700 unsigned long scale_freq, scale_cpu;
2702 delta = now - sa->last_update_time;
2704 * This should only happen when time goes backwards, which it
2705 * unfortunately does during sched clock init when we swap over to TSC.
2707 if ((s64)delta < 0) {
2708 sa->last_update_time = now;
2713 * Use 1024ns as the unit of measurement since it's a reasonable
2714 * approximation of 1us and fast to compute.
2719 sa->last_update_time = now;
2721 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2722 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2724 /* delta_w is the amount already accumulated against our next period */
2725 delta_w = sa->period_contrib;
2726 if (delta + delta_w >= 1024) {
2729 /* how much left for next period will start over, we don't know yet */
2730 sa->period_contrib = 0;
2733 * Now that we know we're crossing a period boundary, figure
2734 * out how much from delta we need to complete the current
2735 * period and accrue it.
2737 delta_w = 1024 - delta_w;
2738 scaled_delta_w = cap_scale(delta_w, scale_freq);
2740 sa->load_sum += weight * scaled_delta_w;
2742 cfs_rq->runnable_load_sum +=
2743 weight * scaled_delta_w;
2747 sa->util_sum += scaled_delta_w * scale_cpu;
2751 /* Figure out how many additional periods this update spans */
2752 periods = delta / 1024;
2755 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2757 cfs_rq->runnable_load_sum =
2758 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2760 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2762 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2763 contrib = __compute_runnable_contrib(periods);
2764 contrib = cap_scale(contrib, scale_freq);
2766 sa->load_sum += weight * contrib;
2768 cfs_rq->runnable_load_sum += weight * contrib;
2771 sa->util_sum += contrib * scale_cpu;
2774 /* Remainder of delta accrued against u_0` */
2775 scaled_delta = cap_scale(delta, scale_freq);
2777 sa->load_sum += weight * scaled_delta;
2779 cfs_rq->runnable_load_sum += weight * scaled_delta;
2782 sa->util_sum += scaled_delta * scale_cpu;
2784 sa->period_contrib += delta;
2787 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2789 cfs_rq->runnable_load_avg =
2790 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2792 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2798 #ifdef CONFIG_FAIR_GROUP_SCHED
2800 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2801 * and effective_load (which is not done because it is too costly).
2803 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2805 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2808 * No need to update load_avg for root_task_group as it is not used.
2810 if (cfs_rq->tg == &root_task_group)
2813 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2814 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2815 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2820 * Called within set_task_rq() right before setting a task's cpu. The
2821 * caller only guarantees p->pi_lock is held; no other assumptions,
2822 * including the state of rq->lock, should be made.
2824 void set_task_rq_fair(struct sched_entity *se,
2825 struct cfs_rq *prev, struct cfs_rq *next)
2827 if (!sched_feat(ATTACH_AGE_LOAD))
2831 * We are supposed to update the task to "current" time, then its up to
2832 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2833 * getting what current time is, so simply throw away the out-of-date
2834 * time. This will result in the wakee task is less decayed, but giving
2835 * the wakee more load sounds not bad.
2837 if (se->avg.last_update_time && prev) {
2838 u64 p_last_update_time;
2839 u64 n_last_update_time;
2841 #ifndef CONFIG_64BIT
2842 u64 p_last_update_time_copy;
2843 u64 n_last_update_time_copy;
2846 p_last_update_time_copy = prev->load_last_update_time_copy;
2847 n_last_update_time_copy = next->load_last_update_time_copy;
2851 p_last_update_time = prev->avg.last_update_time;
2852 n_last_update_time = next->avg.last_update_time;
2854 } while (p_last_update_time != p_last_update_time_copy ||
2855 n_last_update_time != n_last_update_time_copy);
2857 p_last_update_time = prev->avg.last_update_time;
2858 n_last_update_time = next->avg.last_update_time;
2860 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2861 &se->avg, 0, 0, NULL);
2862 se->avg.last_update_time = n_last_update_time;
2865 #else /* CONFIG_FAIR_GROUP_SCHED */
2866 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2867 #endif /* CONFIG_FAIR_GROUP_SCHED */
2869 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2871 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2873 struct rq *rq = rq_of(cfs_rq);
2874 int cpu = cpu_of(rq);
2876 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2877 unsigned long max = rq->cpu_capacity_orig;
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_clock(rq),
2896 min(cfs_rq->avg.util_avg, max), max);
2901 * Unsigned subtract and clamp on underflow.
2903 * Explicitly do a load-store to ensure the intermediate value never hits
2904 * memory. This allows lockless observations without ever seeing the negative
2907 #define sub_positive(_ptr, _val) do { \
2908 typeof(_ptr) ptr = (_ptr); \
2909 typeof(*ptr) val = (_val); \
2910 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2914 WRITE_ONCE(*ptr, res); \
2917 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2919 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2921 struct sched_avg *sa = &cfs_rq->avg;
2922 int decayed, removed_load = 0, removed_util = 0;
2924 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2925 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2926 sub_positive(&sa->load_avg, r);
2927 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2931 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2932 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2933 sub_positive(&sa->util_avg, r);
2934 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2938 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2939 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2941 #ifndef CONFIG_64BIT
2943 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2946 if (update_freq && (decayed || removed_util))
2947 cfs_rq_util_change(cfs_rq);
2949 return decayed || removed_load;
2952 /* Update task and its cfs_rq load average */
2953 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2955 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2956 u64 now = cfs_rq_clock_task(cfs_rq);
2957 struct rq *rq = rq_of(cfs_rq);
2958 int cpu = cpu_of(rq);
2961 * Track task load average for carrying it to new CPU after migrated, and
2962 * track group sched_entity load average for task_h_load calc in migration
2964 __update_load_avg(now, cpu, &se->avg,
2965 se->on_rq * scale_load_down(se->load.weight),
2966 cfs_rq->curr == se, NULL);
2968 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2969 update_tg_load_avg(cfs_rq, 0);
2972 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2974 if (!sched_feat(ATTACH_AGE_LOAD))
2978 * If we got migrated (either between CPUs or between cgroups) we'll
2979 * have aged the average right before clearing @last_update_time.
2981 if (se->avg.last_update_time) {
2982 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2983 &se->avg, 0, 0, NULL);
2986 * XXX: we could have just aged the entire load away if we've been
2987 * absent from the fair class for too long.
2992 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2993 cfs_rq->avg.load_avg += se->avg.load_avg;
2994 cfs_rq->avg.load_sum += se->avg.load_sum;
2995 cfs_rq->avg.util_avg += se->avg.util_avg;
2996 cfs_rq->avg.util_sum += se->avg.util_sum;
2998 cfs_rq_util_change(cfs_rq);
3001 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3004 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3005 cfs_rq->curr == se, NULL);
3007 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3008 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3009 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3010 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3012 cfs_rq_util_change(cfs_rq);
3015 /* Add the load generated by se into cfs_rq's load average */
3017 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3019 struct sched_avg *sa = &se->avg;
3020 u64 now = cfs_rq_clock_task(cfs_rq);
3021 int migrated, decayed;
3023 migrated = !sa->last_update_time;
3025 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3026 se->on_rq * scale_load_down(se->load.weight),
3027 cfs_rq->curr == se, NULL);
3030 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3032 cfs_rq->runnable_load_avg += sa->load_avg;
3033 cfs_rq->runnable_load_sum += sa->load_sum;
3036 attach_entity_load_avg(cfs_rq, se);
3038 if (decayed || migrated)
3039 update_tg_load_avg(cfs_rq, 0);
3042 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3044 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3046 update_load_avg(se, 1);
3048 cfs_rq->runnable_load_avg =
3049 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3050 cfs_rq->runnable_load_sum =
3051 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3054 #ifndef CONFIG_64BIT
3055 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3057 u64 last_update_time_copy;
3058 u64 last_update_time;
3061 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3063 last_update_time = cfs_rq->avg.last_update_time;
3064 } while (last_update_time != last_update_time_copy);
3066 return last_update_time;
3069 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3071 return cfs_rq->avg.last_update_time;
3076 * Task first catches up with cfs_rq, and then subtract
3077 * itself from the cfs_rq (task must be off the queue now).
3079 void remove_entity_load_avg(struct sched_entity *se)
3081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3082 u64 last_update_time;
3085 * Newly created task or never used group entity should not be removed
3086 * from its (source) cfs_rq
3088 if (se->avg.last_update_time == 0)
3091 last_update_time = cfs_rq_last_update_time(cfs_rq);
3093 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3094 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3095 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3098 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3100 return cfs_rq->runnable_load_avg;
3103 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3105 return cfs_rq->avg.load_avg;
3108 static int idle_balance(struct rq *this_rq);
3110 #else /* CONFIG_SMP */
3112 static inline void update_load_avg(struct sched_entity *se, int not_used)
3114 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3115 struct rq *rq = rq_of(cfs_rq);
3117 cpufreq_trigger_update(rq_clock(rq));
3121 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3123 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3124 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3127 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3129 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3131 static inline int idle_balance(struct rq *rq)
3136 #endif /* CONFIG_SMP */
3138 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3140 #ifdef CONFIG_SCHEDSTATS
3141 struct task_struct *tsk = NULL;
3143 if (entity_is_task(se))
3146 if (se->statistics.sleep_start) {
3147 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3152 if (unlikely(delta > se->statistics.sleep_max))
3153 se->statistics.sleep_max = delta;
3155 se->statistics.sleep_start = 0;
3156 se->statistics.sum_sleep_runtime += delta;
3159 account_scheduler_latency(tsk, delta >> 10, 1);
3160 trace_sched_stat_sleep(tsk, delta);
3163 if (se->statistics.block_start) {
3164 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3169 if (unlikely(delta > se->statistics.block_max))
3170 se->statistics.block_max = delta;
3172 se->statistics.block_start = 0;
3173 se->statistics.sum_sleep_runtime += delta;
3176 if (tsk->in_iowait) {
3177 se->statistics.iowait_sum += delta;
3178 se->statistics.iowait_count++;
3179 trace_sched_stat_iowait(tsk, delta);
3182 trace_sched_stat_blocked(tsk, delta);
3185 * Blocking time is in units of nanosecs, so shift by
3186 * 20 to get a milliseconds-range estimation of the
3187 * amount of time that the task spent sleeping:
3189 if (unlikely(prof_on == SLEEP_PROFILING)) {
3190 profile_hits(SLEEP_PROFILING,
3191 (void *)get_wchan(tsk),
3194 account_scheduler_latency(tsk, delta >> 10, 0);
3200 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3202 #ifdef CONFIG_SCHED_DEBUG
3203 s64 d = se->vruntime - cfs_rq->min_vruntime;
3208 if (d > 3*sysctl_sched_latency)
3209 schedstat_inc(cfs_rq, nr_spread_over);
3214 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3216 u64 vruntime = cfs_rq->min_vruntime;
3219 * The 'current' period is already promised to the current tasks,
3220 * however the extra weight of the new task will slow them down a
3221 * little, place the new task so that it fits in the slot that
3222 * stays open at the end.
3224 if (initial && sched_feat(START_DEBIT))
3225 vruntime += sched_vslice(cfs_rq, se);
3227 /* sleeps up to a single latency don't count. */
3229 unsigned long thresh = sysctl_sched_latency;
3232 * Halve their sleep time's effect, to allow
3233 * for a gentler effect of sleepers:
3235 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3241 /* ensure we never gain time by being placed backwards. */
3242 se->vruntime = max_vruntime(se->vruntime, vruntime);
3245 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3247 static inline void check_schedstat_required(void)
3249 #ifdef CONFIG_SCHEDSTATS
3250 if (schedstat_enabled())
3253 /* Force schedstat enabled if a dependent tracepoint is active */
3254 if (trace_sched_stat_wait_enabled() ||
3255 trace_sched_stat_sleep_enabled() ||
3256 trace_sched_stat_iowait_enabled() ||
3257 trace_sched_stat_blocked_enabled() ||
3258 trace_sched_stat_runtime_enabled()) {
3259 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3260 "stat_blocked and stat_runtime require the "
3261 "kernel parameter schedstats=enabled or "
3262 "kernel.sched_schedstats=1\n");
3273 * update_min_vruntime()
3274 * vruntime -= min_vruntime
3278 * update_min_vruntime()
3279 * vruntime += min_vruntime
3281 * this way the vruntime transition between RQs is done when both
3282 * min_vruntime are up-to-date.
3286 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3287 * vruntime -= min_vruntime
3291 * update_min_vruntime()
3292 * vruntime += min_vruntime
3294 * this way we don't have the most up-to-date min_vruntime on the originating
3295 * CPU and an up-to-date min_vruntime on the destination CPU.
3299 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3301 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3302 bool curr = cfs_rq->curr == se;
3305 * If we're the current task, we must renormalise before calling
3309 se->vruntime += cfs_rq->min_vruntime;
3311 update_curr(cfs_rq);
3314 * Otherwise, renormalise after, such that we're placed at the current
3315 * moment in time, instead of some random moment in the past. Being
3316 * placed in the past could significantly boost this task to the
3317 * fairness detriment of existing tasks.
3319 if (renorm && !curr)
3320 se->vruntime += cfs_rq->min_vruntime;
3322 enqueue_entity_load_avg(cfs_rq, se);
3323 account_entity_enqueue(cfs_rq, se);
3324 update_cfs_shares(cfs_rq);
3326 if (flags & ENQUEUE_WAKEUP) {
3327 place_entity(cfs_rq, se, 0);
3328 if (schedstat_enabled())
3329 enqueue_sleeper(cfs_rq, se);
3332 check_schedstat_required();
3333 if (schedstat_enabled()) {
3334 update_stats_enqueue(cfs_rq, se);
3335 check_spread(cfs_rq, se);
3338 __enqueue_entity(cfs_rq, se);
3341 if (cfs_rq->nr_running == 1) {
3342 list_add_leaf_cfs_rq(cfs_rq);
3343 check_enqueue_throttle(cfs_rq);
3347 static void __clear_buddies_last(struct sched_entity *se)
3349 for_each_sched_entity(se) {
3350 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3351 if (cfs_rq->last != se)
3354 cfs_rq->last = NULL;
3358 static void __clear_buddies_next(struct sched_entity *se)
3360 for_each_sched_entity(se) {
3361 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3362 if (cfs_rq->next != se)
3365 cfs_rq->next = NULL;
3369 static void __clear_buddies_skip(struct sched_entity *se)
3371 for_each_sched_entity(se) {
3372 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3373 if (cfs_rq->skip != se)
3376 cfs_rq->skip = NULL;
3380 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3382 if (cfs_rq->last == se)
3383 __clear_buddies_last(se);
3385 if (cfs_rq->next == se)
3386 __clear_buddies_next(se);
3388 if (cfs_rq->skip == se)
3389 __clear_buddies_skip(se);
3392 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3395 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3398 * Update run-time statistics of the 'current'.
3400 update_curr(cfs_rq);
3401 dequeue_entity_load_avg(cfs_rq, se);
3403 if (schedstat_enabled())
3404 update_stats_dequeue(cfs_rq, se, flags);
3406 clear_buddies(cfs_rq, se);
3408 if (se != cfs_rq->curr)
3409 __dequeue_entity(cfs_rq, se);
3411 account_entity_dequeue(cfs_rq, se);
3414 * Normalize the entity after updating the min_vruntime because the
3415 * update can refer to the ->curr item and we need to reflect this
3416 * movement in our normalized position.
3418 if (!(flags & DEQUEUE_SLEEP))
3419 se->vruntime -= cfs_rq->min_vruntime;
3421 /* return excess runtime on last dequeue */
3422 return_cfs_rq_runtime(cfs_rq);
3424 update_min_vruntime(cfs_rq);
3425 update_cfs_shares(cfs_rq);
3429 * Preempt the current task with a newly woken task if needed:
3432 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3434 unsigned long ideal_runtime, delta_exec;
3435 struct sched_entity *se;
3438 ideal_runtime = sched_slice(cfs_rq, curr);
3439 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3440 if (delta_exec > ideal_runtime) {
3441 resched_curr(rq_of(cfs_rq));
3443 * The current task ran long enough, ensure it doesn't get
3444 * re-elected due to buddy favours.
3446 clear_buddies(cfs_rq, curr);
3451 * Ensure that a task that missed wakeup preemption by a
3452 * narrow margin doesn't have to wait for a full slice.
3453 * This also mitigates buddy induced latencies under load.
3455 if (delta_exec < sysctl_sched_min_granularity)
3458 se = __pick_first_entity(cfs_rq);
3459 delta = curr->vruntime - se->vruntime;
3464 if (delta > ideal_runtime)
3465 resched_curr(rq_of(cfs_rq));
3469 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3471 /* 'current' is not kept within the tree. */
3474 * Any task has to be enqueued before it get to execute on
3475 * a CPU. So account for the time it spent waiting on the
3478 if (schedstat_enabled())
3479 update_stats_wait_end(cfs_rq, se);
3480 __dequeue_entity(cfs_rq, se);
3481 update_load_avg(se, 1);
3484 update_stats_curr_start(cfs_rq, se);
3486 #ifdef CONFIG_SCHEDSTATS
3488 * Track our maximum slice length, if the CPU's load is at
3489 * least twice that of our own weight (i.e. dont track it
3490 * when there are only lesser-weight tasks around):
3492 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3493 se->statistics.slice_max = max(se->statistics.slice_max,
3494 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3497 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3501 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3504 * Pick the next process, keeping these things in mind, in this order:
3505 * 1) keep things fair between processes/task groups
3506 * 2) pick the "next" process, since someone really wants that to run
3507 * 3) pick the "last" process, for cache locality
3508 * 4) do not run the "skip" process, if something else is available
3510 static struct sched_entity *
3511 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3513 struct sched_entity *left = __pick_first_entity(cfs_rq);
3514 struct sched_entity *se;
3517 * If curr is set we have to see if its left of the leftmost entity
3518 * still in the tree, provided there was anything in the tree at all.
3520 if (!left || (curr && entity_before(curr, left)))
3523 se = left; /* ideally we run the leftmost entity */
3526 * Avoid running the skip buddy, if running something else can
3527 * be done without getting too unfair.
3529 if (cfs_rq->skip == se) {
3530 struct sched_entity *second;
3533 second = __pick_first_entity(cfs_rq);
3535 second = __pick_next_entity(se);
3536 if (!second || (curr && entity_before(curr, second)))
3540 if (second && wakeup_preempt_entity(second, left) < 1)
3545 * Prefer last buddy, try to return the CPU to a preempted task.
3547 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3551 * Someone really wants this to run. If it's not unfair, run it.
3553 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3556 clear_buddies(cfs_rq, se);
3561 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3563 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3566 * If still on the runqueue then deactivate_task()
3567 * was not called and update_curr() has to be done:
3570 update_curr(cfs_rq);
3572 /* throttle cfs_rqs exceeding runtime */
3573 check_cfs_rq_runtime(cfs_rq);
3575 if (schedstat_enabled()) {
3576 check_spread(cfs_rq, prev);
3578 update_stats_wait_start(cfs_rq, prev);
3582 /* Put 'current' back into the tree. */
3583 __enqueue_entity(cfs_rq, prev);
3584 /* in !on_rq case, update occurred at dequeue */
3585 update_load_avg(prev, 0);
3587 cfs_rq->curr = NULL;
3591 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3594 * Update run-time statistics of the 'current'.
3596 update_curr(cfs_rq);
3599 * Ensure that runnable average is periodically updated.
3601 update_load_avg(curr, 1);
3602 update_cfs_shares(cfs_rq);
3604 #ifdef CONFIG_SCHED_HRTICK
3606 * queued ticks are scheduled to match the slice, so don't bother
3607 * validating it and just reschedule.
3610 resched_curr(rq_of(cfs_rq));
3614 * don't let the period tick interfere with the hrtick preemption
3616 if (!sched_feat(DOUBLE_TICK) &&
3617 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3621 if (cfs_rq->nr_running > 1)
3622 check_preempt_tick(cfs_rq, curr);
3626 /**************************************************
3627 * CFS bandwidth control machinery
3630 #ifdef CONFIG_CFS_BANDWIDTH
3632 #ifdef HAVE_JUMP_LABEL
3633 static struct static_key __cfs_bandwidth_used;
3635 static inline bool cfs_bandwidth_used(void)
3637 return static_key_false(&__cfs_bandwidth_used);
3640 void cfs_bandwidth_usage_inc(void)
3642 static_key_slow_inc(&__cfs_bandwidth_used);
3645 void cfs_bandwidth_usage_dec(void)
3647 static_key_slow_dec(&__cfs_bandwidth_used);
3649 #else /* HAVE_JUMP_LABEL */
3650 static bool cfs_bandwidth_used(void)
3655 void cfs_bandwidth_usage_inc(void) {}
3656 void cfs_bandwidth_usage_dec(void) {}
3657 #endif /* HAVE_JUMP_LABEL */
3660 * default period for cfs group bandwidth.
3661 * default: 0.1s, units: nanoseconds
3663 static inline u64 default_cfs_period(void)
3665 return 100000000ULL;
3668 static inline u64 sched_cfs_bandwidth_slice(void)
3670 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3674 * Replenish runtime according to assigned quota and update expiration time.
3675 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3676 * additional synchronization around rq->lock.
3678 * requires cfs_b->lock
3680 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3684 if (cfs_b->quota == RUNTIME_INF)
3687 now = sched_clock_cpu(smp_processor_id());
3688 cfs_b->runtime = cfs_b->quota;
3689 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3692 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3694 return &tg->cfs_bandwidth;
3697 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3698 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3700 if (unlikely(cfs_rq->throttle_count))
3701 return cfs_rq->throttled_clock_task;
3703 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3706 /* returns 0 on failure to allocate runtime */
3707 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3709 struct task_group *tg = cfs_rq->tg;
3710 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3711 u64 amount = 0, min_amount, expires;
3713 /* note: this is a positive sum as runtime_remaining <= 0 */
3714 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3716 raw_spin_lock(&cfs_b->lock);
3717 if (cfs_b->quota == RUNTIME_INF)
3718 amount = min_amount;
3720 start_cfs_bandwidth(cfs_b);
3722 if (cfs_b->runtime > 0) {
3723 amount = min(cfs_b->runtime, min_amount);
3724 cfs_b->runtime -= amount;
3728 expires = cfs_b->runtime_expires;
3729 raw_spin_unlock(&cfs_b->lock);
3731 cfs_rq->runtime_remaining += amount;
3733 * we may have advanced our local expiration to account for allowed
3734 * spread between our sched_clock and the one on which runtime was
3737 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3738 cfs_rq->runtime_expires = expires;
3740 return cfs_rq->runtime_remaining > 0;
3744 * Note: This depends on the synchronization provided by sched_clock and the
3745 * fact that rq->clock snapshots this value.
3747 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3749 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3751 /* if the deadline is ahead of our clock, nothing to do */
3752 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3755 if (cfs_rq->runtime_remaining < 0)
3759 * If the local deadline has passed we have to consider the
3760 * possibility that our sched_clock is 'fast' and the global deadline
3761 * has not truly expired.
3763 * Fortunately we can check determine whether this the case by checking
3764 * whether the global deadline has advanced. It is valid to compare
3765 * cfs_b->runtime_expires without any locks since we only care about
3766 * exact equality, so a partial write will still work.
3769 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3770 /* extend local deadline, drift is bounded above by 2 ticks */
3771 cfs_rq->runtime_expires += TICK_NSEC;
3773 /* global deadline is ahead, expiration has passed */
3774 cfs_rq->runtime_remaining = 0;
3778 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3780 /* dock delta_exec before expiring quota (as it could span periods) */
3781 cfs_rq->runtime_remaining -= delta_exec;
3782 expire_cfs_rq_runtime(cfs_rq);
3784 if (likely(cfs_rq->runtime_remaining > 0))
3788 * if we're unable to extend our runtime we resched so that the active
3789 * hierarchy can be throttled
3791 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3792 resched_curr(rq_of(cfs_rq));
3795 static __always_inline
3796 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3798 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3801 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3804 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3806 return cfs_bandwidth_used() && cfs_rq->throttled;
3809 /* check whether cfs_rq, or any parent, is throttled */
3810 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3812 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3816 * Ensure that neither of the group entities corresponding to src_cpu or
3817 * dest_cpu are members of a throttled hierarchy when performing group
3818 * load-balance operations.
3820 static inline int throttled_lb_pair(struct task_group *tg,
3821 int src_cpu, int dest_cpu)
3823 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3825 src_cfs_rq = tg->cfs_rq[src_cpu];
3826 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3828 return throttled_hierarchy(src_cfs_rq) ||
3829 throttled_hierarchy(dest_cfs_rq);
3832 /* updated child weight may affect parent so we have to do this bottom up */
3833 static int tg_unthrottle_up(struct task_group *tg, void *data)
3835 struct rq *rq = data;
3836 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3838 cfs_rq->throttle_count--;
3840 if (!cfs_rq->throttle_count) {
3841 /* adjust cfs_rq_clock_task() */
3842 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3843 cfs_rq->throttled_clock_task;
3850 static int tg_throttle_down(struct task_group *tg, void *data)
3852 struct rq *rq = data;
3853 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3855 /* group is entering throttled state, stop time */
3856 if (!cfs_rq->throttle_count)
3857 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3858 cfs_rq->throttle_count++;
3863 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3865 struct rq *rq = rq_of(cfs_rq);
3866 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3867 struct sched_entity *se;
3868 long task_delta, dequeue = 1;
3871 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3873 /* freeze hierarchy runnable averages while throttled */
3875 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3878 task_delta = cfs_rq->h_nr_running;
3879 for_each_sched_entity(se) {
3880 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3881 /* throttled entity or throttle-on-deactivate */
3886 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3887 qcfs_rq->h_nr_running -= task_delta;
3889 if (qcfs_rq->load.weight)
3894 sub_nr_running(rq, task_delta);
3896 cfs_rq->throttled = 1;
3897 cfs_rq->throttled_clock = rq_clock(rq);
3898 raw_spin_lock(&cfs_b->lock);
3899 empty = list_empty(&cfs_b->throttled_cfs_rq);
3902 * Add to the _head_ of the list, so that an already-started
3903 * distribute_cfs_runtime will not see us
3905 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3908 * If we're the first throttled task, make sure the bandwidth
3912 start_cfs_bandwidth(cfs_b);
3914 raw_spin_unlock(&cfs_b->lock);
3917 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3919 struct rq *rq = rq_of(cfs_rq);
3920 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3921 struct sched_entity *se;
3925 se = cfs_rq->tg->se[cpu_of(rq)];
3927 cfs_rq->throttled = 0;
3929 update_rq_clock(rq);
3931 raw_spin_lock(&cfs_b->lock);
3932 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3933 list_del_rcu(&cfs_rq->throttled_list);
3934 raw_spin_unlock(&cfs_b->lock);
3936 /* update hierarchical throttle state */
3937 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3939 if (!cfs_rq->load.weight)
3942 task_delta = cfs_rq->h_nr_running;
3943 for_each_sched_entity(se) {
3947 cfs_rq = cfs_rq_of(se);
3949 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3950 cfs_rq->h_nr_running += task_delta;
3952 if (cfs_rq_throttled(cfs_rq))
3957 add_nr_running(rq, task_delta);
3959 /* determine whether we need to wake up potentially idle cpu */
3960 if (rq->curr == rq->idle && rq->cfs.nr_running)
3964 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3965 u64 remaining, u64 expires)
3967 struct cfs_rq *cfs_rq;
3969 u64 starting_runtime = remaining;
3972 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3974 struct rq *rq = rq_of(cfs_rq);
3976 raw_spin_lock(&rq->lock);
3977 if (!cfs_rq_throttled(cfs_rq))
3980 runtime = -cfs_rq->runtime_remaining + 1;
3981 if (runtime > remaining)
3982 runtime = remaining;
3983 remaining -= runtime;
3985 cfs_rq->runtime_remaining += runtime;
3986 cfs_rq->runtime_expires = expires;
3988 /* we check whether we're throttled above */
3989 if (cfs_rq->runtime_remaining > 0)
3990 unthrottle_cfs_rq(cfs_rq);
3993 raw_spin_unlock(&rq->lock);
4000 return starting_runtime - remaining;
4004 * Responsible for refilling a task_group's bandwidth and unthrottling its
4005 * cfs_rqs as appropriate. If there has been no activity within the last
4006 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4007 * used to track this state.
4009 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4011 u64 runtime, runtime_expires;
4014 /* no need to continue the timer with no bandwidth constraint */
4015 if (cfs_b->quota == RUNTIME_INF)
4016 goto out_deactivate;
4018 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4019 cfs_b->nr_periods += overrun;
4022 * idle depends on !throttled (for the case of a large deficit), and if
4023 * we're going inactive then everything else can be deferred
4025 if (cfs_b->idle && !throttled)
4026 goto out_deactivate;
4028 __refill_cfs_bandwidth_runtime(cfs_b);
4031 /* mark as potentially idle for the upcoming period */
4036 /* account preceding periods in which throttling occurred */
4037 cfs_b->nr_throttled += overrun;
4039 runtime_expires = cfs_b->runtime_expires;
4042 * This check is repeated as we are holding onto the new bandwidth while
4043 * we unthrottle. This can potentially race with an unthrottled group
4044 * trying to acquire new bandwidth from the global pool. This can result
4045 * in us over-using our runtime if it is all used during this loop, but
4046 * only by limited amounts in that extreme case.
4048 while (throttled && cfs_b->runtime > 0) {
4049 runtime = cfs_b->runtime;
4050 raw_spin_unlock(&cfs_b->lock);
4051 /* we can't nest cfs_b->lock while distributing bandwidth */
4052 runtime = distribute_cfs_runtime(cfs_b, runtime,
4054 raw_spin_lock(&cfs_b->lock);
4056 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4058 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4062 * While we are ensured activity in the period following an
4063 * unthrottle, this also covers the case in which the new bandwidth is
4064 * insufficient to cover the existing bandwidth deficit. (Forcing the
4065 * timer to remain active while there are any throttled entities.)
4075 /* a cfs_rq won't donate quota below this amount */
4076 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4077 /* minimum remaining period time to redistribute slack quota */
4078 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4079 /* how long we wait to gather additional slack before distributing */
4080 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4083 * Are we near the end of the current quota period?
4085 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4086 * hrtimer base being cleared by hrtimer_start. In the case of
4087 * migrate_hrtimers, base is never cleared, so we are fine.
4089 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4091 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4094 /* if the call-back is running a quota refresh is already occurring */
4095 if (hrtimer_callback_running(refresh_timer))
4098 /* is a quota refresh about to occur? */
4099 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4100 if (remaining < min_expire)
4106 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4108 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4110 /* if there's a quota refresh soon don't bother with slack */
4111 if (runtime_refresh_within(cfs_b, min_left))
4114 hrtimer_start(&cfs_b->slack_timer,
4115 ns_to_ktime(cfs_bandwidth_slack_period),
4119 /* we know any runtime found here is valid as update_curr() precedes return */
4120 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4122 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4123 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4125 if (slack_runtime <= 0)
4128 raw_spin_lock(&cfs_b->lock);
4129 if (cfs_b->quota != RUNTIME_INF &&
4130 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4131 cfs_b->runtime += slack_runtime;
4133 /* we are under rq->lock, defer unthrottling using a timer */
4134 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4135 !list_empty(&cfs_b->throttled_cfs_rq))
4136 start_cfs_slack_bandwidth(cfs_b);
4138 raw_spin_unlock(&cfs_b->lock);
4140 /* even if it's not valid for return we don't want to try again */
4141 cfs_rq->runtime_remaining -= slack_runtime;
4144 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4146 if (!cfs_bandwidth_used())
4149 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4152 __return_cfs_rq_runtime(cfs_rq);
4156 * This is done with a timer (instead of inline with bandwidth return) since
4157 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4159 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4161 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4164 /* confirm we're still not at a refresh boundary */
4165 raw_spin_lock(&cfs_b->lock);
4166 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4167 raw_spin_unlock(&cfs_b->lock);
4171 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4172 runtime = cfs_b->runtime;
4174 expires = cfs_b->runtime_expires;
4175 raw_spin_unlock(&cfs_b->lock);
4180 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4182 raw_spin_lock(&cfs_b->lock);
4183 if (expires == cfs_b->runtime_expires)
4184 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4185 raw_spin_unlock(&cfs_b->lock);
4189 * When a group wakes up we want to make sure that its quota is not already
4190 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4191 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4193 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4195 if (!cfs_bandwidth_used())
4198 /* Synchronize hierarchical throttle counter: */
4199 if (unlikely(!cfs_rq->throttle_uptodate)) {
4200 struct rq *rq = rq_of(cfs_rq);
4201 struct cfs_rq *pcfs_rq;
4202 struct task_group *tg;
4204 cfs_rq->throttle_uptodate = 1;
4206 /* Get closest up-to-date node, because leaves go first: */
4207 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4208 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4209 if (pcfs_rq->throttle_uptodate)
4213 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4214 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4218 /* an active group must be handled by the update_curr()->put() path */
4219 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4222 /* ensure the group is not already throttled */
4223 if (cfs_rq_throttled(cfs_rq))
4226 /* update runtime allocation */
4227 account_cfs_rq_runtime(cfs_rq, 0);
4228 if (cfs_rq->runtime_remaining <= 0)
4229 throttle_cfs_rq(cfs_rq);
4232 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4233 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4235 if (!cfs_bandwidth_used())
4238 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4242 * it's possible for a throttled entity to be forced into a running
4243 * state (e.g. set_curr_task), in this case we're finished.
4245 if (cfs_rq_throttled(cfs_rq))
4248 throttle_cfs_rq(cfs_rq);
4252 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4254 struct cfs_bandwidth *cfs_b =
4255 container_of(timer, struct cfs_bandwidth, slack_timer);
4257 do_sched_cfs_slack_timer(cfs_b);
4259 return HRTIMER_NORESTART;
4262 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4264 struct cfs_bandwidth *cfs_b =
4265 container_of(timer, struct cfs_bandwidth, period_timer);
4269 raw_spin_lock(&cfs_b->lock);
4271 overrun = hrtimer_forward_now(timer, cfs_b->period);
4275 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4278 cfs_b->period_active = 0;
4279 raw_spin_unlock(&cfs_b->lock);
4281 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4284 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4286 raw_spin_lock_init(&cfs_b->lock);
4288 cfs_b->quota = RUNTIME_INF;
4289 cfs_b->period = ns_to_ktime(default_cfs_period());
4291 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4292 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4293 cfs_b->period_timer.function = sched_cfs_period_timer;
4294 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4295 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4298 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4300 cfs_rq->runtime_enabled = 0;
4301 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4304 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4306 lockdep_assert_held(&cfs_b->lock);
4308 if (!cfs_b->period_active) {
4309 cfs_b->period_active = 1;
4310 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4311 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4315 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4317 /* init_cfs_bandwidth() was not called */
4318 if (!cfs_b->throttled_cfs_rq.next)
4321 hrtimer_cancel(&cfs_b->period_timer);
4322 hrtimer_cancel(&cfs_b->slack_timer);
4325 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4327 struct cfs_rq *cfs_rq;
4329 for_each_leaf_cfs_rq(rq, cfs_rq) {
4330 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4332 raw_spin_lock(&cfs_b->lock);
4333 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4334 raw_spin_unlock(&cfs_b->lock);
4338 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4340 struct cfs_rq *cfs_rq;
4342 for_each_leaf_cfs_rq(rq, cfs_rq) {
4343 if (!cfs_rq->runtime_enabled)
4347 * clock_task is not advancing so we just need to make sure
4348 * there's some valid quota amount
4350 cfs_rq->runtime_remaining = 1;
4352 * Offline rq is schedulable till cpu is completely disabled
4353 * in take_cpu_down(), so we prevent new cfs throttling here.
4355 cfs_rq->runtime_enabled = 0;
4357 if (cfs_rq_throttled(cfs_rq))
4358 unthrottle_cfs_rq(cfs_rq);
4362 #else /* CONFIG_CFS_BANDWIDTH */
4363 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4365 return rq_clock_task(rq_of(cfs_rq));
4368 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4369 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4370 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4371 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4373 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4378 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4383 static inline int throttled_lb_pair(struct task_group *tg,
4384 int src_cpu, int dest_cpu)
4389 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4391 #ifdef CONFIG_FAIR_GROUP_SCHED
4392 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4395 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4399 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4400 static inline void update_runtime_enabled(struct rq *rq) {}
4401 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4403 #endif /* CONFIG_CFS_BANDWIDTH */
4405 /**************************************************
4406 * CFS operations on tasks:
4409 #ifdef CONFIG_SCHED_HRTICK
4410 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4412 struct sched_entity *se = &p->se;
4413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4415 WARN_ON(task_rq(p) != rq);
4417 if (cfs_rq->nr_running > 1) {
4418 u64 slice = sched_slice(cfs_rq, se);
4419 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4420 s64 delta = slice - ran;
4427 hrtick_start(rq, delta);
4432 * called from enqueue/dequeue and updates the hrtick when the
4433 * current task is from our class and nr_running is low enough
4436 static void hrtick_update(struct rq *rq)
4438 struct task_struct *curr = rq->curr;
4440 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4443 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4444 hrtick_start_fair(rq, curr);
4446 #else /* !CONFIG_SCHED_HRTICK */
4448 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4452 static inline void hrtick_update(struct rq *rq)
4458 * The enqueue_task method is called before nr_running is
4459 * increased. Here we update the fair scheduling stats and
4460 * then put the task into the rbtree:
4463 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4465 struct cfs_rq *cfs_rq;
4466 struct sched_entity *se = &p->se;
4468 for_each_sched_entity(se) {
4471 cfs_rq = cfs_rq_of(se);
4472 enqueue_entity(cfs_rq, se, flags);
4475 * end evaluation on encountering a throttled cfs_rq
4477 * note: in the case of encountering a throttled cfs_rq we will
4478 * post the final h_nr_running increment below.
4480 if (cfs_rq_throttled(cfs_rq))
4482 cfs_rq->h_nr_running++;
4484 flags = ENQUEUE_WAKEUP;
4487 for_each_sched_entity(se) {
4488 cfs_rq = cfs_rq_of(se);
4489 cfs_rq->h_nr_running++;
4491 if (cfs_rq_throttled(cfs_rq))
4494 update_load_avg(se, 1);
4495 update_cfs_shares(cfs_rq);
4499 add_nr_running(rq, 1);
4504 static void set_next_buddy(struct sched_entity *se);
4507 * The dequeue_task method is called before nr_running is
4508 * decreased. We remove the task from the rbtree and
4509 * update the fair scheduling stats:
4511 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4513 struct cfs_rq *cfs_rq;
4514 struct sched_entity *se = &p->se;
4515 int task_sleep = flags & DEQUEUE_SLEEP;
4517 for_each_sched_entity(se) {
4518 cfs_rq = cfs_rq_of(se);
4519 dequeue_entity(cfs_rq, se, flags);
4522 * end evaluation on encountering a throttled cfs_rq
4524 * note: in the case of encountering a throttled cfs_rq we will
4525 * post the final h_nr_running decrement below.
4527 if (cfs_rq_throttled(cfs_rq))
4529 cfs_rq->h_nr_running--;
4531 /* Don't dequeue parent if it has other entities besides us */
4532 if (cfs_rq->load.weight) {
4533 /* Avoid re-evaluating load for this entity: */
4534 se = parent_entity(se);
4536 * Bias pick_next to pick a task from this cfs_rq, as
4537 * p is sleeping when it is within its sched_slice.
4539 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4543 flags |= DEQUEUE_SLEEP;
4546 for_each_sched_entity(se) {
4547 cfs_rq = cfs_rq_of(se);
4548 cfs_rq->h_nr_running--;
4550 if (cfs_rq_throttled(cfs_rq))
4553 update_load_avg(se, 1);
4554 update_cfs_shares(cfs_rq);
4558 sub_nr_running(rq, 1);
4564 #ifdef CONFIG_NO_HZ_COMMON
4566 * per rq 'load' arrray crap; XXX kill this.
4570 * The exact cpuload calculated at every tick would be:
4572 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4574 * If a cpu misses updates for n ticks (as it was idle) and update gets
4575 * called on the n+1-th tick when cpu may be busy, then we have:
4577 * load_n = (1 - 1/2^i)^n * load_0
4578 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4580 * decay_load_missed() below does efficient calculation of
4582 * load' = (1 - 1/2^i)^n * load
4584 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4585 * This allows us to precompute the above in said factors, thereby allowing the
4586 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4587 * fixed_power_int())
4589 * The calculation is approximated on a 128 point scale.
4591 #define DEGRADE_SHIFT 7
4593 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4594 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4595 { 0, 0, 0, 0, 0, 0, 0, 0 },
4596 { 64, 32, 8, 0, 0, 0, 0, 0 },
4597 { 96, 72, 40, 12, 1, 0, 0, 0 },
4598 { 112, 98, 75, 43, 15, 1, 0, 0 },
4599 { 120, 112, 98, 76, 45, 16, 2, 0 }
4603 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4604 * would be when CPU is idle and so we just decay the old load without
4605 * adding any new load.
4607 static unsigned long
4608 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4612 if (!missed_updates)
4615 if (missed_updates >= degrade_zero_ticks[idx])
4619 return load >> missed_updates;
4621 while (missed_updates) {
4622 if (missed_updates % 2)
4623 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4625 missed_updates >>= 1;
4630 #endif /* CONFIG_NO_HZ_COMMON */
4633 * __cpu_load_update - update the rq->cpu_load[] statistics
4634 * @this_rq: The rq to update statistics for
4635 * @this_load: The current load
4636 * @pending_updates: The number of missed updates
4638 * Update rq->cpu_load[] statistics. This function is usually called every
4639 * scheduler tick (TICK_NSEC).
4641 * This function computes a decaying average:
4643 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4645 * Because of NOHZ it might not get called on every tick which gives need for
4646 * the @pending_updates argument.
4648 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4649 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4650 * = A * (A * load[i]_n-2 + B) + B
4651 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4652 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4653 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4654 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4655 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4657 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4658 * any change in load would have resulted in the tick being turned back on.
4660 * For regular NOHZ, this reduces to:
4662 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4664 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4667 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4668 unsigned long pending_updates)
4670 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4673 this_rq->nr_load_updates++;
4675 /* Update our load: */
4676 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4677 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4678 unsigned long old_load, new_load;
4680 /* scale is effectively 1 << i now, and >> i divides by scale */
4682 old_load = this_rq->cpu_load[i];
4683 #ifdef CONFIG_NO_HZ_COMMON
4684 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4685 if (tickless_load) {
4686 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4688 * old_load can never be a negative value because a
4689 * decayed tickless_load cannot be greater than the
4690 * original tickless_load.
4692 old_load += tickless_load;
4695 new_load = this_load;
4697 * Round up the averaging division if load is increasing. This
4698 * prevents us from getting stuck on 9 if the load is 10, for
4701 if (new_load > old_load)
4702 new_load += scale - 1;
4704 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4707 sched_avg_update(this_rq);
4710 /* Used instead of source_load when we know the type == 0 */
4711 static unsigned long weighted_cpuload(const int cpu)
4713 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4716 #ifdef CONFIG_NO_HZ_COMMON
4718 * There is no sane way to deal with nohz on smp when using jiffies because the
4719 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4720 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4722 * Therefore we need to avoid the delta approach from the regular tick when
4723 * possible since that would seriously skew the load calculation. This is why we
4724 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4725 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4726 * loop exit, nohz_idle_balance, nohz full exit...)
4728 * This means we might still be one tick off for nohz periods.
4731 static void cpu_load_update_nohz(struct rq *this_rq,
4732 unsigned long curr_jiffies,
4735 unsigned long pending_updates;
4737 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4738 if (pending_updates) {
4739 this_rq->last_load_update_tick = curr_jiffies;
4741 * In the regular NOHZ case, we were idle, this means load 0.
4742 * In the NOHZ_FULL case, we were non-idle, we should consider
4743 * its weighted load.
4745 cpu_load_update(this_rq, load, pending_updates);
4750 * Called from nohz_idle_balance() to update the load ratings before doing the
4753 static void cpu_load_update_idle(struct rq *this_rq)
4756 * bail if there's load or we're actually up-to-date.
4758 if (weighted_cpuload(cpu_of(this_rq)))
4761 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4765 * Record CPU load on nohz entry so we know the tickless load to account
4766 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4767 * than other cpu_load[idx] but it should be fine as cpu_load readers
4768 * shouldn't rely into synchronized cpu_load[*] updates.
4770 void cpu_load_update_nohz_start(void)
4772 struct rq *this_rq = this_rq();
4775 * This is all lockless but should be fine. If weighted_cpuload changes
4776 * concurrently we'll exit nohz. And cpu_load write can race with
4777 * cpu_load_update_idle() but both updater would be writing the same.
4779 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4783 * Account the tickless load in the end of a nohz frame.
4785 void cpu_load_update_nohz_stop(void)
4787 unsigned long curr_jiffies = READ_ONCE(jiffies);
4788 struct rq *this_rq = this_rq();
4791 if (curr_jiffies == this_rq->last_load_update_tick)
4794 load = weighted_cpuload(cpu_of(this_rq));
4795 raw_spin_lock(&this_rq->lock);
4796 update_rq_clock(this_rq);
4797 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4798 raw_spin_unlock(&this_rq->lock);
4800 #else /* !CONFIG_NO_HZ_COMMON */
4801 static inline void cpu_load_update_nohz(struct rq *this_rq,
4802 unsigned long curr_jiffies,
4803 unsigned long load) { }
4804 #endif /* CONFIG_NO_HZ_COMMON */
4806 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4808 #ifdef CONFIG_NO_HZ_COMMON
4809 /* See the mess around cpu_load_update_nohz(). */
4810 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4812 cpu_load_update(this_rq, load, 1);
4816 * Called from scheduler_tick()
4818 void cpu_load_update_active(struct rq *this_rq)
4820 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4822 if (tick_nohz_tick_stopped())
4823 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4825 cpu_load_update_periodic(this_rq, load);
4829 * Return a low guess at the load of a migration-source cpu weighted
4830 * according to the scheduling class and "nice" value.
4832 * We want to under-estimate the load of migration sources, to
4833 * balance conservatively.
4835 static unsigned long source_load(int cpu, int type)
4837 struct rq *rq = cpu_rq(cpu);
4838 unsigned long total = weighted_cpuload(cpu);
4840 if (type == 0 || !sched_feat(LB_BIAS))
4843 return min(rq->cpu_load[type-1], total);
4847 * Return a high guess at the load of a migration-target cpu weighted
4848 * according to the scheduling class and "nice" value.
4850 static unsigned long target_load(int cpu, int type)
4852 struct rq *rq = cpu_rq(cpu);
4853 unsigned long total = weighted_cpuload(cpu);
4855 if (type == 0 || !sched_feat(LB_BIAS))
4858 return max(rq->cpu_load[type-1], total);
4861 static unsigned long capacity_of(int cpu)
4863 return cpu_rq(cpu)->cpu_capacity;
4866 static unsigned long capacity_orig_of(int cpu)
4868 return cpu_rq(cpu)->cpu_capacity_orig;
4871 static unsigned long cpu_avg_load_per_task(int cpu)
4873 struct rq *rq = cpu_rq(cpu);
4874 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4875 unsigned long load_avg = weighted_cpuload(cpu);
4878 return load_avg / nr_running;
4883 #ifdef CONFIG_FAIR_GROUP_SCHED
4885 * effective_load() calculates the load change as seen from the root_task_group
4887 * Adding load to a group doesn't make a group heavier, but can cause movement
4888 * of group shares between cpus. Assuming the shares were perfectly aligned one
4889 * can calculate the shift in shares.
4891 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4892 * on this @cpu and results in a total addition (subtraction) of @wg to the
4893 * total group weight.
4895 * Given a runqueue weight distribution (rw_i) we can compute a shares
4896 * distribution (s_i) using:
4898 * s_i = rw_i / \Sum rw_j (1)
4900 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4901 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4902 * shares distribution (s_i):
4904 * rw_i = { 2, 4, 1, 0 }
4905 * s_i = { 2/7, 4/7, 1/7, 0 }
4907 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4908 * task used to run on and the CPU the waker is running on), we need to
4909 * compute the effect of waking a task on either CPU and, in case of a sync
4910 * wakeup, compute the effect of the current task going to sleep.
4912 * So for a change of @wl to the local @cpu with an overall group weight change
4913 * of @wl we can compute the new shares distribution (s'_i) using:
4915 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4917 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4918 * differences in waking a task to CPU 0. The additional task changes the
4919 * weight and shares distributions like:
4921 * rw'_i = { 3, 4, 1, 0 }
4922 * s'_i = { 3/8, 4/8, 1/8, 0 }
4924 * We can then compute the difference in effective weight by using:
4926 * dw_i = S * (s'_i - s_i) (3)
4928 * Where 'S' is the group weight as seen by its parent.
4930 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4931 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4932 * 4/7) times the weight of the group.
4934 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4936 struct sched_entity *se = tg->se[cpu];
4938 if (!tg->parent) /* the trivial, non-cgroup case */
4941 for_each_sched_entity(se) {
4942 struct cfs_rq *cfs_rq = se->my_q;
4943 long W, w = cfs_rq_load_avg(cfs_rq);
4948 * W = @wg + \Sum rw_j
4950 W = wg + atomic_long_read(&tg->load_avg);
4952 /* Ensure \Sum rw_j >= rw_i */
4953 W -= cfs_rq->tg_load_avg_contrib;
4962 * wl = S * s'_i; see (2)
4965 wl = (w * (long)tg->shares) / W;
4970 * Per the above, wl is the new se->load.weight value; since
4971 * those are clipped to [MIN_SHARES, ...) do so now. See
4972 * calc_cfs_shares().
4974 if (wl < MIN_SHARES)
4978 * wl = dw_i = S * (s'_i - s_i); see (3)
4980 wl -= se->avg.load_avg;
4983 * Recursively apply this logic to all parent groups to compute
4984 * the final effective load change on the root group. Since
4985 * only the @tg group gets extra weight, all parent groups can
4986 * only redistribute existing shares. @wl is the shift in shares
4987 * resulting from this level per the above.
4996 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5003 static void record_wakee(struct task_struct *p)
5006 * Only decay a single time; tasks that have less then 1 wakeup per
5007 * jiffy will not have built up many flips.
5009 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5010 current->wakee_flips >>= 1;
5011 current->wakee_flip_decay_ts = jiffies;
5014 if (current->last_wakee != p) {
5015 current->last_wakee = p;
5016 current->wakee_flips++;
5021 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5023 * A waker of many should wake a different task than the one last awakened
5024 * at a frequency roughly N times higher than one of its wakees.
5026 * In order to determine whether we should let the load spread vs consolidating
5027 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5028 * partner, and a factor of lls_size higher frequency in the other.
5030 * With both conditions met, we can be relatively sure that the relationship is
5031 * non-monogamous, with partner count exceeding socket size.
5033 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5034 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5037 static int wake_wide(struct task_struct *p)
5039 unsigned int master = current->wakee_flips;
5040 unsigned int slave = p->wakee_flips;
5041 int factor = this_cpu_read(sd_llc_size);
5044 swap(master, slave);
5045 if (slave < factor || master < slave * factor)
5050 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5052 s64 this_load, load;
5053 s64 this_eff_load, prev_eff_load;
5054 int idx, this_cpu, prev_cpu;
5055 struct task_group *tg;
5056 unsigned long weight;
5060 this_cpu = smp_processor_id();
5061 prev_cpu = task_cpu(p);
5062 load = source_load(prev_cpu, idx);
5063 this_load = target_load(this_cpu, idx);
5066 * If sync wakeup then subtract the (maximum possible)
5067 * effect of the currently running task from the load
5068 * of the current CPU:
5071 tg = task_group(current);
5072 weight = current->se.avg.load_avg;
5074 this_load += effective_load(tg, this_cpu, -weight, -weight);
5075 load += effective_load(tg, prev_cpu, 0, -weight);
5079 weight = p->se.avg.load_avg;
5082 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5083 * due to the sync cause above having dropped this_load to 0, we'll
5084 * always have an imbalance, but there's really nothing you can do
5085 * about that, so that's good too.
5087 * Otherwise check if either cpus are near enough in load to allow this
5088 * task to be woken on this_cpu.
5090 this_eff_load = 100;
5091 this_eff_load *= capacity_of(prev_cpu);
5093 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5094 prev_eff_load *= capacity_of(this_cpu);
5096 if (this_load > 0) {
5097 this_eff_load *= this_load +
5098 effective_load(tg, this_cpu, weight, weight);
5100 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5103 balanced = this_eff_load <= prev_eff_load;
5105 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5110 schedstat_inc(sd, ttwu_move_affine);
5111 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5117 * find_idlest_group finds and returns the least busy CPU group within the
5120 static struct sched_group *
5121 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5122 int this_cpu, int sd_flag)
5124 struct sched_group *idlest = NULL, *group = sd->groups;
5125 unsigned long min_load = ULONG_MAX, this_load = 0;
5126 int load_idx = sd->forkexec_idx;
5127 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5129 if (sd_flag & SD_BALANCE_WAKE)
5130 load_idx = sd->wake_idx;
5133 unsigned long load, avg_load;
5137 /* Skip over this group if it has no CPUs allowed */
5138 if (!cpumask_intersects(sched_group_cpus(group),
5139 tsk_cpus_allowed(p)))
5142 local_group = cpumask_test_cpu(this_cpu,
5143 sched_group_cpus(group));
5145 /* Tally up the load of all CPUs in the group */
5148 for_each_cpu(i, sched_group_cpus(group)) {
5149 /* Bias balancing toward cpus of our domain */
5151 load = source_load(i, load_idx);
5153 load = target_load(i, load_idx);
5158 /* Adjust by relative CPU capacity of the group */
5159 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5162 this_load = avg_load;
5163 } else if (avg_load < min_load) {
5164 min_load = avg_load;
5167 } while (group = group->next, group != sd->groups);
5169 if (!idlest || 100*this_load < imbalance*min_load)
5175 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5178 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5180 unsigned long load, min_load = ULONG_MAX;
5181 unsigned int min_exit_latency = UINT_MAX;
5182 u64 latest_idle_timestamp = 0;
5183 int least_loaded_cpu = this_cpu;
5184 int shallowest_idle_cpu = -1;
5187 /* Traverse only the allowed CPUs */
5188 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5190 struct rq *rq = cpu_rq(i);
5191 struct cpuidle_state *idle = idle_get_state(rq);
5192 if (idle && idle->exit_latency < min_exit_latency) {
5194 * We give priority to a CPU whose idle state
5195 * has the smallest exit latency irrespective
5196 * of any idle timestamp.
5198 min_exit_latency = idle->exit_latency;
5199 latest_idle_timestamp = rq->idle_stamp;
5200 shallowest_idle_cpu = i;
5201 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5202 rq->idle_stamp > latest_idle_timestamp) {
5204 * If equal or no active idle state, then
5205 * the most recently idled CPU might have
5208 latest_idle_timestamp = rq->idle_stamp;
5209 shallowest_idle_cpu = i;
5211 } else if (shallowest_idle_cpu == -1) {
5212 load = weighted_cpuload(i);
5213 if (load < min_load || (load == min_load && i == this_cpu)) {
5215 least_loaded_cpu = i;
5220 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5224 * Try and locate an idle CPU in the sched_domain.
5226 static int select_idle_sibling(struct task_struct *p, int target)
5228 struct sched_domain *sd;
5229 struct sched_group *sg;
5230 int i = task_cpu(p);
5232 if (idle_cpu(target))
5236 * If the prevous cpu is cache affine and idle, don't be stupid.
5238 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5242 * Otherwise, iterate the domains and find an eligible idle cpu.
5244 * A completely idle sched group at higher domains is more
5245 * desirable than an idle group at a lower level, because lower
5246 * domains have smaller groups and usually share hardware
5247 * resources which causes tasks to contend on them, e.g. x86
5248 * hyperthread siblings in the lowest domain (SMT) can contend
5249 * on the shared cpu pipeline.
5251 * However, while we prefer idle groups at higher domains
5252 * finding an idle cpu at the lowest domain is still better than
5253 * returning 'target', which we've already established, isn't
5256 sd = rcu_dereference(per_cpu(sd_llc, target));
5257 for_each_lower_domain(sd) {
5260 if (!cpumask_intersects(sched_group_cpus(sg),
5261 tsk_cpus_allowed(p)))
5264 /* Ensure the entire group is idle */
5265 for_each_cpu(i, sched_group_cpus(sg)) {
5266 if (i == target || !idle_cpu(i))
5271 * It doesn't matter which cpu we pick, the
5272 * whole group is idle.
5274 target = cpumask_first_and(sched_group_cpus(sg),
5275 tsk_cpus_allowed(p));
5279 } while (sg != sd->groups);
5286 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5287 * tasks. The unit of the return value must be the one of capacity so we can
5288 * compare the utilization with the capacity of the CPU that is available for
5289 * CFS task (ie cpu_capacity).
5291 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5292 * recent utilization of currently non-runnable tasks on a CPU. It represents
5293 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5294 * capacity_orig is the cpu_capacity available at the highest frequency
5295 * (arch_scale_freq_capacity()).
5296 * The utilization of a CPU converges towards a sum equal to or less than the
5297 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5298 * the running time on this CPU scaled by capacity_curr.
5300 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5301 * higher than capacity_orig because of unfortunate rounding in
5302 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5303 * the average stabilizes with the new running time. We need to check that the
5304 * utilization stays within the range of [0..capacity_orig] and cap it if
5305 * necessary. Without utilization capping, a group could be seen as overloaded
5306 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5307 * available capacity. We allow utilization to overshoot capacity_curr (but not
5308 * capacity_orig) as it useful for predicting the capacity required after task
5309 * migrations (scheduler-driven DVFS).
5311 static int cpu_util(int cpu)
5313 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5314 unsigned long capacity = capacity_orig_of(cpu);
5316 return (util >= capacity) ? capacity : util;
5320 * select_task_rq_fair: Select target runqueue for the waking task in domains
5321 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5322 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5324 * Balances load by selecting the idlest cpu in the idlest group, or under
5325 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5327 * Returns the target cpu number.
5329 * preempt must be disabled.
5332 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5334 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5335 int cpu = smp_processor_id();
5336 int new_cpu = prev_cpu;
5337 int want_affine = 0;
5338 int sync = wake_flags & WF_SYNC;
5340 if (sd_flag & SD_BALANCE_WAKE) {
5342 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5346 for_each_domain(cpu, tmp) {
5347 if (!(tmp->flags & SD_LOAD_BALANCE))
5351 * If both cpu and prev_cpu are part of this domain,
5352 * cpu is a valid SD_WAKE_AFFINE target.
5354 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5355 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5360 if (tmp->flags & sd_flag)
5362 else if (!want_affine)
5367 sd = NULL; /* Prefer wake_affine over balance flags */
5368 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5373 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5374 new_cpu = select_idle_sibling(p, new_cpu);
5377 struct sched_group *group;
5380 if (!(sd->flags & sd_flag)) {
5385 group = find_idlest_group(sd, p, cpu, sd_flag);
5391 new_cpu = find_idlest_cpu(group, p, cpu);
5392 if (new_cpu == -1 || new_cpu == cpu) {
5393 /* Now try balancing at a lower domain level of cpu */
5398 /* Now try balancing at a lower domain level of new_cpu */
5400 weight = sd->span_weight;
5402 for_each_domain(cpu, tmp) {
5403 if (weight <= tmp->span_weight)
5405 if (tmp->flags & sd_flag)
5408 /* while loop will break here if sd == NULL */
5416 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5417 * cfs_rq_of(p) references at time of call are still valid and identify the
5418 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5420 static void migrate_task_rq_fair(struct task_struct *p)
5423 * As blocked tasks retain absolute vruntime the migration needs to
5424 * deal with this by subtracting the old and adding the new
5425 * min_vruntime -- the latter is done by enqueue_entity() when placing
5426 * the task on the new runqueue.
5428 if (p->state == TASK_WAKING) {
5429 struct sched_entity *se = &p->se;
5430 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5433 #ifndef CONFIG_64BIT
5434 u64 min_vruntime_copy;
5437 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5439 min_vruntime = cfs_rq->min_vruntime;
5440 } while (min_vruntime != min_vruntime_copy);
5442 min_vruntime = cfs_rq->min_vruntime;
5445 se->vruntime -= min_vruntime;
5449 * We are supposed to update the task to "current" time, then its up to date
5450 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5451 * what current time is, so simply throw away the out-of-date time. This
5452 * will result in the wakee task is less decayed, but giving the wakee more
5453 * load sounds not bad.
5455 remove_entity_load_avg(&p->se);
5457 /* Tell new CPU we are migrated */
5458 p->se.avg.last_update_time = 0;
5460 /* We have migrated, no longer consider this task hot */
5461 p->se.exec_start = 0;
5464 static void task_dead_fair(struct task_struct *p)
5466 remove_entity_load_avg(&p->se);
5468 #endif /* CONFIG_SMP */
5470 static unsigned long
5471 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5473 unsigned long gran = sysctl_sched_wakeup_granularity;
5476 * Since its curr running now, convert the gran from real-time
5477 * to virtual-time in his units.
5479 * By using 'se' instead of 'curr' we penalize light tasks, so
5480 * they get preempted easier. That is, if 'se' < 'curr' then
5481 * the resulting gran will be larger, therefore penalizing the
5482 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5483 * be smaller, again penalizing the lighter task.
5485 * This is especially important for buddies when the leftmost
5486 * task is higher priority than the buddy.
5488 return calc_delta_fair(gran, se);
5492 * Should 'se' preempt 'curr'.
5506 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5508 s64 gran, vdiff = curr->vruntime - se->vruntime;
5513 gran = wakeup_gran(curr, se);
5520 static void set_last_buddy(struct sched_entity *se)
5522 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5525 for_each_sched_entity(se)
5526 cfs_rq_of(se)->last = se;
5529 static void set_next_buddy(struct sched_entity *se)
5531 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5534 for_each_sched_entity(se)
5535 cfs_rq_of(se)->next = se;
5538 static void set_skip_buddy(struct sched_entity *se)
5540 for_each_sched_entity(se)
5541 cfs_rq_of(se)->skip = se;
5545 * Preempt the current task with a newly woken task if needed:
5547 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5549 struct task_struct *curr = rq->curr;
5550 struct sched_entity *se = &curr->se, *pse = &p->se;
5551 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5552 int scale = cfs_rq->nr_running >= sched_nr_latency;
5553 int next_buddy_marked = 0;
5555 if (unlikely(se == pse))
5559 * This is possible from callers such as attach_tasks(), in which we
5560 * unconditionally check_prempt_curr() after an enqueue (which may have
5561 * lead to a throttle). This both saves work and prevents false
5562 * next-buddy nomination below.
5564 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5567 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5568 set_next_buddy(pse);
5569 next_buddy_marked = 1;
5573 * We can come here with TIF_NEED_RESCHED already set from new task
5576 * Note: this also catches the edge-case of curr being in a throttled
5577 * group (e.g. via set_curr_task), since update_curr() (in the
5578 * enqueue of curr) will have resulted in resched being set. This
5579 * prevents us from potentially nominating it as a false LAST_BUDDY
5582 if (test_tsk_need_resched(curr))
5585 /* Idle tasks are by definition preempted by non-idle tasks. */
5586 if (unlikely(curr->policy == SCHED_IDLE) &&
5587 likely(p->policy != SCHED_IDLE))
5591 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5592 * is driven by the tick):
5594 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5597 find_matching_se(&se, &pse);
5598 update_curr(cfs_rq_of(se));
5600 if (wakeup_preempt_entity(se, pse) == 1) {
5602 * Bias pick_next to pick the sched entity that is
5603 * triggering this preemption.
5605 if (!next_buddy_marked)
5606 set_next_buddy(pse);
5615 * Only set the backward buddy when the current task is still
5616 * on the rq. This can happen when a wakeup gets interleaved
5617 * with schedule on the ->pre_schedule() or idle_balance()
5618 * point, either of which can * drop the rq lock.
5620 * Also, during early boot the idle thread is in the fair class,
5621 * for obvious reasons its a bad idea to schedule back to it.
5623 if (unlikely(!se->on_rq || curr == rq->idle))
5626 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5630 static struct task_struct *
5631 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5633 struct cfs_rq *cfs_rq = &rq->cfs;
5634 struct sched_entity *se;
5635 struct task_struct *p;
5639 #ifdef CONFIG_FAIR_GROUP_SCHED
5640 if (!cfs_rq->nr_running)
5643 if (prev->sched_class != &fair_sched_class)
5647 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5648 * likely that a next task is from the same cgroup as the current.
5650 * Therefore attempt to avoid putting and setting the entire cgroup
5651 * hierarchy, only change the part that actually changes.
5655 struct sched_entity *curr = cfs_rq->curr;
5658 * Since we got here without doing put_prev_entity() we also
5659 * have to consider cfs_rq->curr. If it is still a runnable
5660 * entity, update_curr() will update its vruntime, otherwise
5661 * forget we've ever seen it.
5665 update_curr(cfs_rq);
5670 * This call to check_cfs_rq_runtime() will do the
5671 * throttle and dequeue its entity in the parent(s).
5672 * Therefore the 'simple' nr_running test will indeed
5675 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5679 se = pick_next_entity(cfs_rq, curr);
5680 cfs_rq = group_cfs_rq(se);
5686 * Since we haven't yet done put_prev_entity and if the selected task
5687 * is a different task than we started out with, try and touch the
5688 * least amount of cfs_rqs.
5691 struct sched_entity *pse = &prev->se;
5693 while (!(cfs_rq = is_same_group(se, pse))) {
5694 int se_depth = se->depth;
5695 int pse_depth = pse->depth;
5697 if (se_depth <= pse_depth) {
5698 put_prev_entity(cfs_rq_of(pse), pse);
5699 pse = parent_entity(pse);
5701 if (se_depth >= pse_depth) {
5702 set_next_entity(cfs_rq_of(se), se);
5703 se = parent_entity(se);
5707 put_prev_entity(cfs_rq, pse);
5708 set_next_entity(cfs_rq, se);
5711 if (hrtick_enabled(rq))
5712 hrtick_start_fair(rq, p);
5719 if (!cfs_rq->nr_running)
5722 put_prev_task(rq, prev);
5725 se = pick_next_entity(cfs_rq, NULL);
5726 set_next_entity(cfs_rq, se);
5727 cfs_rq = group_cfs_rq(se);
5732 if (hrtick_enabled(rq))
5733 hrtick_start_fair(rq, p);
5739 * This is OK, because current is on_cpu, which avoids it being picked
5740 * for load-balance and preemption/IRQs are still disabled avoiding
5741 * further scheduler activity on it and we're being very careful to
5742 * re-start the picking loop.
5744 lockdep_unpin_lock(&rq->lock, cookie);
5745 new_tasks = idle_balance(rq);
5746 lockdep_repin_lock(&rq->lock, cookie);
5748 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5749 * possible for any higher priority task to appear. In that case we
5750 * must re-start the pick_next_entity() loop.
5762 * Account for a descheduled task:
5764 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5766 struct sched_entity *se = &prev->se;
5767 struct cfs_rq *cfs_rq;
5769 for_each_sched_entity(se) {
5770 cfs_rq = cfs_rq_of(se);
5771 put_prev_entity(cfs_rq, se);
5776 * sched_yield() is very simple
5778 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5780 static void yield_task_fair(struct rq *rq)
5782 struct task_struct *curr = rq->curr;
5783 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5784 struct sched_entity *se = &curr->se;
5787 * Are we the only task in the tree?
5789 if (unlikely(rq->nr_running == 1))
5792 clear_buddies(cfs_rq, se);
5794 if (curr->policy != SCHED_BATCH) {
5795 update_rq_clock(rq);
5797 * Update run-time statistics of the 'current'.
5799 update_curr(cfs_rq);
5801 * Tell update_rq_clock() that we've just updated,
5802 * so we don't do microscopic update in schedule()
5803 * and double the fastpath cost.
5805 rq_clock_skip_update(rq, true);
5811 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5813 struct sched_entity *se = &p->se;
5815 /* throttled hierarchies are not runnable */
5816 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5819 /* Tell the scheduler that we'd really like pse to run next. */
5822 yield_task_fair(rq);
5828 /**************************************************
5829 * Fair scheduling class load-balancing methods.
5833 * The purpose of load-balancing is to achieve the same basic fairness the
5834 * per-cpu scheduler provides, namely provide a proportional amount of compute
5835 * time to each task. This is expressed in the following equation:
5837 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5839 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5840 * W_i,0 is defined as:
5842 * W_i,0 = \Sum_j w_i,j (2)
5844 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5845 * is derived from the nice value as per sched_prio_to_weight[].
5847 * The weight average is an exponential decay average of the instantaneous
5850 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5852 * C_i is the compute capacity of cpu i, typically it is the
5853 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5854 * can also include other factors [XXX].
5856 * To achieve this balance we define a measure of imbalance which follows
5857 * directly from (1):
5859 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5861 * We them move tasks around to minimize the imbalance. In the continuous
5862 * function space it is obvious this converges, in the discrete case we get
5863 * a few fun cases generally called infeasible weight scenarios.
5866 * - infeasible weights;
5867 * - local vs global optima in the discrete case. ]
5872 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5873 * for all i,j solution, we create a tree of cpus that follows the hardware
5874 * topology where each level pairs two lower groups (or better). This results
5875 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5876 * tree to only the first of the previous level and we decrease the frequency
5877 * of load-balance at each level inv. proportional to the number of cpus in
5883 * \Sum { --- * --- * 2^i } = O(n) (5)
5885 * `- size of each group
5886 * | | `- number of cpus doing load-balance
5888 * `- sum over all levels
5890 * Coupled with a limit on how many tasks we can migrate every balance pass,
5891 * this makes (5) the runtime complexity of the balancer.
5893 * An important property here is that each CPU is still (indirectly) connected
5894 * to every other cpu in at most O(log n) steps:
5896 * The adjacency matrix of the resulting graph is given by:
5899 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5902 * And you'll find that:
5904 * A^(log_2 n)_i,j != 0 for all i,j (7)
5906 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5907 * The task movement gives a factor of O(m), giving a convergence complexity
5910 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5915 * In order to avoid CPUs going idle while there's still work to do, new idle
5916 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5917 * tree itself instead of relying on other CPUs to bring it work.
5919 * This adds some complexity to both (5) and (8) but it reduces the total idle
5927 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5930 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5935 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5937 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5939 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5942 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5943 * rewrite all of this once again.]
5946 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5948 enum fbq_type { regular, remote, all };
5950 #define LBF_ALL_PINNED 0x01
5951 #define LBF_NEED_BREAK 0x02
5952 #define LBF_DST_PINNED 0x04
5953 #define LBF_SOME_PINNED 0x08
5956 struct sched_domain *sd;
5964 struct cpumask *dst_grpmask;
5966 enum cpu_idle_type idle;
5968 /* The set of CPUs under consideration for load-balancing */
5969 struct cpumask *cpus;
5974 unsigned int loop_break;
5975 unsigned int loop_max;
5977 enum fbq_type fbq_type;
5978 struct list_head tasks;
5982 * Is this task likely cache-hot:
5984 static int task_hot(struct task_struct *p, struct lb_env *env)
5988 lockdep_assert_held(&env->src_rq->lock);
5990 if (p->sched_class != &fair_sched_class)
5993 if (unlikely(p->policy == SCHED_IDLE))
5997 * Buddy candidates are cache hot:
5999 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6000 (&p->se == cfs_rq_of(&p->se)->next ||
6001 &p->se == cfs_rq_of(&p->se)->last))
6004 if (sysctl_sched_migration_cost == -1)
6006 if (sysctl_sched_migration_cost == 0)
6009 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6011 return delta < (s64)sysctl_sched_migration_cost;
6014 #ifdef CONFIG_NUMA_BALANCING
6016 * Returns 1, if task migration degrades locality
6017 * Returns 0, if task migration improves locality i.e migration preferred.
6018 * Returns -1, if task migration is not affected by locality.
6020 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6022 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6023 unsigned long src_faults, dst_faults;
6024 int src_nid, dst_nid;
6026 if (!static_branch_likely(&sched_numa_balancing))
6029 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6032 src_nid = cpu_to_node(env->src_cpu);
6033 dst_nid = cpu_to_node(env->dst_cpu);
6035 if (src_nid == dst_nid)
6038 /* Migrating away from the preferred node is always bad. */
6039 if (src_nid == p->numa_preferred_nid) {
6040 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6046 /* Encourage migration to the preferred node. */
6047 if (dst_nid == p->numa_preferred_nid)
6051 src_faults = group_faults(p, src_nid);
6052 dst_faults = group_faults(p, dst_nid);
6054 src_faults = task_faults(p, src_nid);
6055 dst_faults = task_faults(p, dst_nid);
6058 return dst_faults < src_faults;
6062 static inline int migrate_degrades_locality(struct task_struct *p,
6070 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6073 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6077 lockdep_assert_held(&env->src_rq->lock);
6080 * We do not migrate tasks that are:
6081 * 1) throttled_lb_pair, or
6082 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6083 * 3) running (obviously), or
6084 * 4) are cache-hot on their current CPU.
6086 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6089 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6092 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6094 env->flags |= LBF_SOME_PINNED;
6097 * Remember if this task can be migrated to any other cpu in
6098 * our sched_group. We may want to revisit it if we couldn't
6099 * meet load balance goals by pulling other tasks on src_cpu.
6101 * Also avoid computing new_dst_cpu if we have already computed
6102 * one in current iteration.
6104 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6107 /* Prevent to re-select dst_cpu via env's cpus */
6108 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6109 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6110 env->flags |= LBF_DST_PINNED;
6111 env->new_dst_cpu = cpu;
6119 /* Record that we found atleast one task that could run on dst_cpu */
6120 env->flags &= ~LBF_ALL_PINNED;
6122 if (task_running(env->src_rq, p)) {
6123 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6128 * Aggressive migration if:
6129 * 1) destination numa is preferred
6130 * 2) task is cache cold, or
6131 * 3) too many balance attempts have failed.
6133 tsk_cache_hot = migrate_degrades_locality(p, env);
6134 if (tsk_cache_hot == -1)
6135 tsk_cache_hot = task_hot(p, env);
6137 if (tsk_cache_hot <= 0 ||
6138 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6139 if (tsk_cache_hot == 1) {
6140 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6141 schedstat_inc(p, se.statistics.nr_forced_migrations);
6146 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6151 * detach_task() -- detach the task for the migration specified in env
6153 static void detach_task(struct task_struct *p, struct lb_env *env)
6155 lockdep_assert_held(&env->src_rq->lock);
6157 p->on_rq = TASK_ON_RQ_MIGRATING;
6158 deactivate_task(env->src_rq, p, 0);
6159 set_task_cpu(p, env->dst_cpu);
6163 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6164 * part of active balancing operations within "domain".
6166 * Returns a task if successful and NULL otherwise.
6168 static struct task_struct *detach_one_task(struct lb_env *env)
6170 struct task_struct *p, *n;
6172 lockdep_assert_held(&env->src_rq->lock);
6174 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6175 if (!can_migrate_task(p, env))
6178 detach_task(p, env);
6181 * Right now, this is only the second place where
6182 * lb_gained[env->idle] is updated (other is detach_tasks)
6183 * so we can safely collect stats here rather than
6184 * inside detach_tasks().
6186 schedstat_inc(env->sd, lb_gained[env->idle]);
6192 static const unsigned int sched_nr_migrate_break = 32;
6195 * detach_tasks() -- tries to detach up to imbalance weighted load from
6196 * busiest_rq, as part of a balancing operation within domain "sd".
6198 * Returns number of detached tasks if successful and 0 otherwise.
6200 static int detach_tasks(struct lb_env *env)
6202 struct list_head *tasks = &env->src_rq->cfs_tasks;
6203 struct task_struct *p;
6207 lockdep_assert_held(&env->src_rq->lock);
6209 if (env->imbalance <= 0)
6212 while (!list_empty(tasks)) {
6214 * We don't want to steal all, otherwise we may be treated likewise,
6215 * which could at worst lead to a livelock crash.
6217 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6220 p = list_first_entry(tasks, struct task_struct, se.group_node);
6223 /* We've more or less seen every task there is, call it quits */
6224 if (env->loop > env->loop_max)
6227 /* take a breather every nr_migrate tasks */
6228 if (env->loop > env->loop_break) {
6229 env->loop_break += sched_nr_migrate_break;
6230 env->flags |= LBF_NEED_BREAK;
6234 if (!can_migrate_task(p, env))
6237 load = task_h_load(p);
6239 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6242 if ((load / 2) > env->imbalance)
6245 detach_task(p, env);
6246 list_add(&p->se.group_node, &env->tasks);
6249 env->imbalance -= load;
6251 #ifdef CONFIG_PREEMPT
6253 * NEWIDLE balancing is a source of latency, so preemptible
6254 * kernels will stop after the first task is detached to minimize
6255 * the critical section.
6257 if (env->idle == CPU_NEWLY_IDLE)
6262 * We only want to steal up to the prescribed amount of
6265 if (env->imbalance <= 0)
6270 list_move_tail(&p->se.group_node, tasks);
6274 * Right now, this is one of only two places we collect this stat
6275 * so we can safely collect detach_one_task() stats here rather
6276 * than inside detach_one_task().
6278 schedstat_add(env->sd, lb_gained[env->idle], detached);
6284 * attach_task() -- attach the task detached by detach_task() to its new rq.
6286 static void attach_task(struct rq *rq, struct task_struct *p)
6288 lockdep_assert_held(&rq->lock);
6290 BUG_ON(task_rq(p) != rq);
6291 activate_task(rq, p, 0);
6292 p->on_rq = TASK_ON_RQ_QUEUED;
6293 check_preempt_curr(rq, p, 0);
6297 * attach_one_task() -- attaches the task returned from detach_one_task() to
6300 static void attach_one_task(struct rq *rq, struct task_struct *p)
6302 raw_spin_lock(&rq->lock);
6304 raw_spin_unlock(&rq->lock);
6308 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6311 static void attach_tasks(struct lb_env *env)
6313 struct list_head *tasks = &env->tasks;
6314 struct task_struct *p;
6316 raw_spin_lock(&env->dst_rq->lock);
6318 while (!list_empty(tasks)) {
6319 p = list_first_entry(tasks, struct task_struct, se.group_node);
6320 list_del_init(&p->se.group_node);
6322 attach_task(env->dst_rq, p);
6325 raw_spin_unlock(&env->dst_rq->lock);
6328 #ifdef CONFIG_FAIR_GROUP_SCHED
6329 static void update_blocked_averages(int cpu)
6331 struct rq *rq = cpu_rq(cpu);
6332 struct cfs_rq *cfs_rq;
6333 unsigned long flags;
6335 raw_spin_lock_irqsave(&rq->lock, flags);
6336 update_rq_clock(rq);
6339 * Iterates the task_group tree in a bottom up fashion, see
6340 * list_add_leaf_cfs_rq() for details.
6342 for_each_leaf_cfs_rq(rq, cfs_rq) {
6343 /* throttled entities do not contribute to load */
6344 if (throttled_hierarchy(cfs_rq))
6347 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6348 update_tg_load_avg(cfs_rq, 0);
6350 raw_spin_unlock_irqrestore(&rq->lock, flags);
6354 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6355 * This needs to be done in a top-down fashion because the load of a child
6356 * group is a fraction of its parents load.
6358 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6360 struct rq *rq = rq_of(cfs_rq);
6361 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6362 unsigned long now = jiffies;
6365 if (cfs_rq->last_h_load_update == now)
6368 cfs_rq->h_load_next = NULL;
6369 for_each_sched_entity(se) {
6370 cfs_rq = cfs_rq_of(se);
6371 cfs_rq->h_load_next = se;
6372 if (cfs_rq->last_h_load_update == now)
6377 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6378 cfs_rq->last_h_load_update = now;
6381 while ((se = cfs_rq->h_load_next) != NULL) {
6382 load = cfs_rq->h_load;
6383 load = div64_ul(load * se->avg.load_avg,
6384 cfs_rq_load_avg(cfs_rq) + 1);
6385 cfs_rq = group_cfs_rq(se);
6386 cfs_rq->h_load = load;
6387 cfs_rq->last_h_load_update = now;
6391 static unsigned long task_h_load(struct task_struct *p)
6393 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6395 update_cfs_rq_h_load(cfs_rq);
6396 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6397 cfs_rq_load_avg(cfs_rq) + 1);
6400 static inline void update_blocked_averages(int cpu)
6402 struct rq *rq = cpu_rq(cpu);
6403 struct cfs_rq *cfs_rq = &rq->cfs;
6404 unsigned long flags;
6406 raw_spin_lock_irqsave(&rq->lock, flags);
6407 update_rq_clock(rq);
6408 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6409 raw_spin_unlock_irqrestore(&rq->lock, flags);
6412 static unsigned long task_h_load(struct task_struct *p)
6414 return p->se.avg.load_avg;
6418 /********** Helpers for find_busiest_group ************************/
6427 * sg_lb_stats - stats of a sched_group required for load_balancing
6429 struct sg_lb_stats {
6430 unsigned long avg_load; /*Avg load across the CPUs of the group */
6431 unsigned long group_load; /* Total load over the CPUs of the group */
6432 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6433 unsigned long load_per_task;
6434 unsigned long group_capacity;
6435 unsigned long group_util; /* Total utilization of the group */
6436 unsigned int sum_nr_running; /* Nr tasks running in the group */
6437 unsigned int idle_cpus;
6438 unsigned int group_weight;
6439 enum group_type group_type;
6440 int group_no_capacity;
6441 #ifdef CONFIG_NUMA_BALANCING
6442 unsigned int nr_numa_running;
6443 unsigned int nr_preferred_running;
6448 * sd_lb_stats - Structure to store the statistics of a sched_domain
6449 * during load balancing.
6451 struct sd_lb_stats {
6452 struct sched_group *busiest; /* Busiest group in this sd */
6453 struct sched_group *local; /* Local group in this sd */
6454 unsigned long total_load; /* Total load of all groups in sd */
6455 unsigned long total_capacity; /* Total capacity of all groups in sd */
6456 unsigned long avg_load; /* Average load across all groups in sd */
6458 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6459 struct sg_lb_stats local_stat; /* Statistics of the local group */
6462 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6465 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6466 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6467 * We must however clear busiest_stat::avg_load because
6468 * update_sd_pick_busiest() reads this before assignment.
6470 *sds = (struct sd_lb_stats){
6474 .total_capacity = 0UL,
6477 .sum_nr_running = 0,
6478 .group_type = group_other,
6484 * get_sd_load_idx - Obtain the load index for a given sched domain.
6485 * @sd: The sched_domain whose load_idx is to be obtained.
6486 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6488 * Return: The load index.
6490 static inline int get_sd_load_idx(struct sched_domain *sd,
6491 enum cpu_idle_type idle)
6497 load_idx = sd->busy_idx;
6500 case CPU_NEWLY_IDLE:
6501 load_idx = sd->newidle_idx;
6504 load_idx = sd->idle_idx;
6511 static unsigned long scale_rt_capacity(int cpu)
6513 struct rq *rq = cpu_rq(cpu);
6514 u64 total, used, age_stamp, avg;
6518 * Since we're reading these variables without serialization make sure
6519 * we read them once before doing sanity checks on them.
6521 age_stamp = READ_ONCE(rq->age_stamp);
6522 avg = READ_ONCE(rq->rt_avg);
6523 delta = __rq_clock_broken(rq) - age_stamp;
6525 if (unlikely(delta < 0))
6528 total = sched_avg_period() + delta;
6530 used = div_u64(avg, total);
6532 if (likely(used < SCHED_CAPACITY_SCALE))
6533 return SCHED_CAPACITY_SCALE - used;
6538 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6540 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6541 struct sched_group *sdg = sd->groups;
6543 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6545 capacity *= scale_rt_capacity(cpu);
6546 capacity >>= SCHED_CAPACITY_SHIFT;
6551 cpu_rq(cpu)->cpu_capacity = capacity;
6552 sdg->sgc->capacity = capacity;
6555 void update_group_capacity(struct sched_domain *sd, int cpu)
6557 struct sched_domain *child = sd->child;
6558 struct sched_group *group, *sdg = sd->groups;
6559 unsigned long capacity;
6560 unsigned long interval;
6562 interval = msecs_to_jiffies(sd->balance_interval);
6563 interval = clamp(interval, 1UL, max_load_balance_interval);
6564 sdg->sgc->next_update = jiffies + interval;
6567 update_cpu_capacity(sd, cpu);
6573 if (child->flags & SD_OVERLAP) {
6575 * SD_OVERLAP domains cannot assume that child groups
6576 * span the current group.
6579 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6580 struct sched_group_capacity *sgc;
6581 struct rq *rq = cpu_rq(cpu);
6584 * build_sched_domains() -> init_sched_groups_capacity()
6585 * gets here before we've attached the domains to the
6588 * Use capacity_of(), which is set irrespective of domains
6589 * in update_cpu_capacity().
6591 * This avoids capacity from being 0 and
6592 * causing divide-by-zero issues on boot.
6594 if (unlikely(!rq->sd)) {
6595 capacity += capacity_of(cpu);
6599 sgc = rq->sd->groups->sgc;
6600 capacity += sgc->capacity;
6604 * !SD_OVERLAP domains can assume that child groups
6605 * span the current group.
6608 group = child->groups;
6610 capacity += group->sgc->capacity;
6611 group = group->next;
6612 } while (group != child->groups);
6615 sdg->sgc->capacity = capacity;
6619 * Check whether the capacity of the rq has been noticeably reduced by side
6620 * activity. The imbalance_pct is used for the threshold.
6621 * Return true is the capacity is reduced
6624 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6626 return ((rq->cpu_capacity * sd->imbalance_pct) <
6627 (rq->cpu_capacity_orig * 100));
6631 * Group imbalance indicates (and tries to solve) the problem where balancing
6632 * groups is inadequate due to tsk_cpus_allowed() constraints.
6634 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6635 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6638 * { 0 1 2 3 } { 4 5 6 7 }
6641 * If we were to balance group-wise we'd place two tasks in the first group and
6642 * two tasks in the second group. Clearly this is undesired as it will overload
6643 * cpu 3 and leave one of the cpus in the second group unused.
6645 * The current solution to this issue is detecting the skew in the first group
6646 * by noticing the lower domain failed to reach balance and had difficulty
6647 * moving tasks due to affinity constraints.
6649 * When this is so detected; this group becomes a candidate for busiest; see
6650 * update_sd_pick_busiest(). And calculate_imbalance() and
6651 * find_busiest_group() avoid some of the usual balance conditions to allow it
6652 * to create an effective group imbalance.
6654 * This is a somewhat tricky proposition since the next run might not find the
6655 * group imbalance and decide the groups need to be balanced again. A most
6656 * subtle and fragile situation.
6659 static inline int sg_imbalanced(struct sched_group *group)
6661 return group->sgc->imbalance;
6665 * group_has_capacity returns true if the group has spare capacity that could
6666 * be used by some tasks.
6667 * We consider that a group has spare capacity if the * number of task is
6668 * smaller than the number of CPUs or if the utilization is lower than the
6669 * available capacity for CFS tasks.
6670 * For the latter, we use a threshold to stabilize the state, to take into
6671 * account the variance of the tasks' load and to return true if the available
6672 * capacity in meaningful for the load balancer.
6673 * As an example, an available capacity of 1% can appear but it doesn't make
6674 * any benefit for the load balance.
6677 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6679 if (sgs->sum_nr_running < sgs->group_weight)
6682 if ((sgs->group_capacity * 100) >
6683 (sgs->group_util * env->sd->imbalance_pct))
6690 * group_is_overloaded returns true if the group has more tasks than it can
6692 * group_is_overloaded is not equals to !group_has_capacity because a group
6693 * with the exact right number of tasks, has no more spare capacity but is not
6694 * overloaded so both group_has_capacity and group_is_overloaded return
6698 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6700 if (sgs->sum_nr_running <= sgs->group_weight)
6703 if ((sgs->group_capacity * 100) <
6704 (sgs->group_util * env->sd->imbalance_pct))
6711 group_type group_classify(struct sched_group *group,
6712 struct sg_lb_stats *sgs)
6714 if (sgs->group_no_capacity)
6715 return group_overloaded;
6717 if (sg_imbalanced(group))
6718 return group_imbalanced;
6724 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6725 * @env: The load balancing environment.
6726 * @group: sched_group whose statistics are to be updated.
6727 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6728 * @local_group: Does group contain this_cpu.
6729 * @sgs: variable to hold the statistics for this group.
6730 * @overload: Indicate more than one runnable task for any CPU.
6732 static inline void update_sg_lb_stats(struct lb_env *env,
6733 struct sched_group *group, int load_idx,
6734 int local_group, struct sg_lb_stats *sgs,
6740 memset(sgs, 0, sizeof(*sgs));
6742 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6743 struct rq *rq = cpu_rq(i);
6745 /* Bias balancing toward cpus of our domain */
6747 load = target_load(i, load_idx);
6749 load = source_load(i, load_idx);
6751 sgs->group_load += load;
6752 sgs->group_util += cpu_util(i);
6753 sgs->sum_nr_running += rq->cfs.h_nr_running;
6755 nr_running = rq->nr_running;
6759 #ifdef CONFIG_NUMA_BALANCING
6760 sgs->nr_numa_running += rq->nr_numa_running;
6761 sgs->nr_preferred_running += rq->nr_preferred_running;
6763 sgs->sum_weighted_load += weighted_cpuload(i);
6765 * No need to call idle_cpu() if nr_running is not 0
6767 if (!nr_running && idle_cpu(i))
6771 /* Adjust by relative CPU capacity of the group */
6772 sgs->group_capacity = group->sgc->capacity;
6773 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6775 if (sgs->sum_nr_running)
6776 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6778 sgs->group_weight = group->group_weight;
6780 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6781 sgs->group_type = group_classify(group, sgs);
6785 * update_sd_pick_busiest - return 1 on busiest group
6786 * @env: The load balancing environment.
6787 * @sds: sched_domain statistics
6788 * @sg: sched_group candidate to be checked for being the busiest
6789 * @sgs: sched_group statistics
6791 * Determine if @sg is a busier group than the previously selected
6794 * Return: %true if @sg is a busier group than the previously selected
6795 * busiest group. %false otherwise.
6797 static bool update_sd_pick_busiest(struct lb_env *env,
6798 struct sd_lb_stats *sds,
6799 struct sched_group *sg,
6800 struct sg_lb_stats *sgs)
6802 struct sg_lb_stats *busiest = &sds->busiest_stat;
6804 if (sgs->group_type > busiest->group_type)
6807 if (sgs->group_type < busiest->group_type)
6810 if (sgs->avg_load <= busiest->avg_load)
6813 /* This is the busiest node in its class. */
6814 if (!(env->sd->flags & SD_ASYM_PACKING))
6817 /* No ASYM_PACKING if target cpu is already busy */
6818 if (env->idle == CPU_NOT_IDLE)
6821 * ASYM_PACKING needs to move all the work to the lowest
6822 * numbered CPUs in the group, therefore mark all groups
6823 * higher than ourself as busy.
6825 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6829 /* Prefer to move from highest possible cpu's work */
6830 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6837 #ifdef CONFIG_NUMA_BALANCING
6838 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6840 if (sgs->sum_nr_running > sgs->nr_numa_running)
6842 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6847 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6849 if (rq->nr_running > rq->nr_numa_running)
6851 if (rq->nr_running > rq->nr_preferred_running)
6856 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6861 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6865 #endif /* CONFIG_NUMA_BALANCING */
6868 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6869 * @env: The load balancing environment.
6870 * @sds: variable to hold the statistics for this sched_domain.
6872 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6874 struct sched_domain *child = env->sd->child;
6875 struct sched_group *sg = env->sd->groups;
6876 struct sg_lb_stats tmp_sgs;
6877 int load_idx, prefer_sibling = 0;
6878 bool overload = false;
6880 if (child && child->flags & SD_PREFER_SIBLING)
6883 load_idx = get_sd_load_idx(env->sd, env->idle);
6886 struct sg_lb_stats *sgs = &tmp_sgs;
6889 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6892 sgs = &sds->local_stat;
6894 if (env->idle != CPU_NEWLY_IDLE ||
6895 time_after_eq(jiffies, sg->sgc->next_update))
6896 update_group_capacity(env->sd, env->dst_cpu);
6899 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6906 * In case the child domain prefers tasks go to siblings
6907 * first, lower the sg capacity so that we'll try
6908 * and move all the excess tasks away. We lower the capacity
6909 * of a group only if the local group has the capacity to fit
6910 * these excess tasks. The extra check prevents the case where
6911 * you always pull from the heaviest group when it is already
6912 * under-utilized (possible with a large weight task outweighs
6913 * the tasks on the system).
6915 if (prefer_sibling && sds->local &&
6916 group_has_capacity(env, &sds->local_stat) &&
6917 (sgs->sum_nr_running > 1)) {
6918 sgs->group_no_capacity = 1;
6919 sgs->group_type = group_classify(sg, sgs);
6922 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6924 sds->busiest_stat = *sgs;
6928 /* Now, start updating sd_lb_stats */
6929 sds->total_load += sgs->group_load;
6930 sds->total_capacity += sgs->group_capacity;
6933 } while (sg != env->sd->groups);
6935 if (env->sd->flags & SD_NUMA)
6936 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6938 if (!env->sd->parent) {
6939 /* update overload indicator if we are at root domain */
6940 if (env->dst_rq->rd->overload != overload)
6941 env->dst_rq->rd->overload = overload;
6947 * check_asym_packing - Check to see if the group is packed into the
6950 * This is primarily intended to used at the sibling level. Some
6951 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6952 * case of POWER7, it can move to lower SMT modes only when higher
6953 * threads are idle. When in lower SMT modes, the threads will
6954 * perform better since they share less core resources. Hence when we
6955 * have idle threads, we want them to be the higher ones.
6957 * This packing function is run on idle threads. It checks to see if
6958 * the busiest CPU in this domain (core in the P7 case) has a higher
6959 * CPU number than the packing function is being run on. Here we are
6960 * assuming lower CPU number will be equivalent to lower a SMT thread
6963 * Return: 1 when packing is required and a task should be moved to
6964 * this CPU. The amount of the imbalance is returned in *imbalance.
6966 * @env: The load balancing environment.
6967 * @sds: Statistics of the sched_domain which is to be packed
6969 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6973 if (!(env->sd->flags & SD_ASYM_PACKING))
6976 if (env->idle == CPU_NOT_IDLE)
6982 busiest_cpu = group_first_cpu(sds->busiest);
6983 if (env->dst_cpu > busiest_cpu)
6986 env->imbalance = DIV_ROUND_CLOSEST(
6987 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6988 SCHED_CAPACITY_SCALE);
6994 * fix_small_imbalance - Calculate the minor imbalance that exists
6995 * amongst the groups of a sched_domain, during
6997 * @env: The load balancing environment.
6998 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7001 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7003 unsigned long tmp, capa_now = 0, capa_move = 0;
7004 unsigned int imbn = 2;
7005 unsigned long scaled_busy_load_per_task;
7006 struct sg_lb_stats *local, *busiest;
7008 local = &sds->local_stat;
7009 busiest = &sds->busiest_stat;
7011 if (!local->sum_nr_running)
7012 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7013 else if (busiest->load_per_task > local->load_per_task)
7016 scaled_busy_load_per_task =
7017 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7018 busiest->group_capacity;
7020 if (busiest->avg_load + scaled_busy_load_per_task >=
7021 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7022 env->imbalance = busiest->load_per_task;
7027 * OK, we don't have enough imbalance to justify moving tasks,
7028 * however we may be able to increase total CPU capacity used by
7032 capa_now += busiest->group_capacity *
7033 min(busiest->load_per_task, busiest->avg_load);
7034 capa_now += local->group_capacity *
7035 min(local->load_per_task, local->avg_load);
7036 capa_now /= SCHED_CAPACITY_SCALE;
7038 /* Amount of load we'd subtract */
7039 if (busiest->avg_load > scaled_busy_load_per_task) {
7040 capa_move += busiest->group_capacity *
7041 min(busiest->load_per_task,
7042 busiest->avg_load - scaled_busy_load_per_task);
7045 /* Amount of load we'd add */
7046 if (busiest->avg_load * busiest->group_capacity <
7047 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7048 tmp = (busiest->avg_load * busiest->group_capacity) /
7049 local->group_capacity;
7051 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7052 local->group_capacity;
7054 capa_move += local->group_capacity *
7055 min(local->load_per_task, local->avg_load + tmp);
7056 capa_move /= SCHED_CAPACITY_SCALE;
7058 /* Move if we gain throughput */
7059 if (capa_move > capa_now)
7060 env->imbalance = busiest->load_per_task;
7064 * calculate_imbalance - Calculate the amount of imbalance present within the
7065 * groups of a given sched_domain during load balance.
7066 * @env: load balance environment
7067 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7069 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7071 unsigned long max_pull, load_above_capacity = ~0UL;
7072 struct sg_lb_stats *local, *busiest;
7074 local = &sds->local_stat;
7075 busiest = &sds->busiest_stat;
7077 if (busiest->group_type == group_imbalanced) {
7079 * In the group_imb case we cannot rely on group-wide averages
7080 * to ensure cpu-load equilibrium, look at wider averages. XXX
7082 busiest->load_per_task =
7083 min(busiest->load_per_task, sds->avg_load);
7087 * Avg load of busiest sg can be less and avg load of local sg can
7088 * be greater than avg load across all sgs of sd because avg load
7089 * factors in sg capacity and sgs with smaller group_type are
7090 * skipped when updating the busiest sg:
7092 if (busiest->avg_load <= sds->avg_load ||
7093 local->avg_load >= sds->avg_load) {
7095 return fix_small_imbalance(env, sds);
7099 * If there aren't any idle cpus, avoid creating some.
7101 if (busiest->group_type == group_overloaded &&
7102 local->group_type == group_overloaded) {
7103 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7104 if (load_above_capacity > busiest->group_capacity) {
7105 load_above_capacity -= busiest->group_capacity;
7106 load_above_capacity *= NICE_0_LOAD;
7107 load_above_capacity /= busiest->group_capacity;
7109 load_above_capacity = ~0UL;
7113 * We're trying to get all the cpus to the average_load, so we don't
7114 * want to push ourselves above the average load, nor do we wish to
7115 * reduce the max loaded cpu below the average load. At the same time,
7116 * we also don't want to reduce the group load below the group
7117 * capacity. Thus we look for the minimum possible imbalance.
7119 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7121 /* How much load to actually move to equalise the imbalance */
7122 env->imbalance = min(
7123 max_pull * busiest->group_capacity,
7124 (sds->avg_load - local->avg_load) * local->group_capacity
7125 ) / SCHED_CAPACITY_SCALE;
7128 * if *imbalance is less than the average load per runnable task
7129 * there is no guarantee that any tasks will be moved so we'll have
7130 * a think about bumping its value to force at least one task to be
7133 if (env->imbalance < busiest->load_per_task)
7134 return fix_small_imbalance(env, sds);
7137 /******* find_busiest_group() helpers end here *********************/
7140 * find_busiest_group - Returns the busiest group within the sched_domain
7141 * if there is an imbalance.
7143 * Also calculates the amount of weighted load which should be moved
7144 * to restore balance.
7146 * @env: The load balancing environment.
7148 * Return: - The busiest group if imbalance exists.
7150 static struct sched_group *find_busiest_group(struct lb_env *env)
7152 struct sg_lb_stats *local, *busiest;
7153 struct sd_lb_stats sds;
7155 init_sd_lb_stats(&sds);
7158 * Compute the various statistics relavent for load balancing at
7161 update_sd_lb_stats(env, &sds);
7162 local = &sds.local_stat;
7163 busiest = &sds.busiest_stat;
7165 /* ASYM feature bypasses nice load balance check */
7166 if (check_asym_packing(env, &sds))
7169 /* There is no busy sibling group to pull tasks from */
7170 if (!sds.busiest || busiest->sum_nr_running == 0)
7173 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7174 / sds.total_capacity;
7177 * If the busiest group is imbalanced the below checks don't
7178 * work because they assume all things are equal, which typically
7179 * isn't true due to cpus_allowed constraints and the like.
7181 if (busiest->group_type == group_imbalanced)
7184 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7185 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7186 busiest->group_no_capacity)
7190 * If the local group is busier than the selected busiest group
7191 * don't try and pull any tasks.
7193 if (local->avg_load >= busiest->avg_load)
7197 * Don't pull any tasks if this group is already above the domain
7200 if (local->avg_load >= sds.avg_load)
7203 if (env->idle == CPU_IDLE) {
7205 * This cpu is idle. If the busiest group is not overloaded
7206 * and there is no imbalance between this and busiest group
7207 * wrt idle cpus, it is balanced. The imbalance becomes
7208 * significant if the diff is greater than 1 otherwise we
7209 * might end up to just move the imbalance on another group
7211 if ((busiest->group_type != group_overloaded) &&
7212 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7216 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7217 * imbalance_pct to be conservative.
7219 if (100 * busiest->avg_load <=
7220 env->sd->imbalance_pct * local->avg_load)
7225 /* Looks like there is an imbalance. Compute it */
7226 calculate_imbalance(env, &sds);
7235 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7237 static struct rq *find_busiest_queue(struct lb_env *env,
7238 struct sched_group *group)
7240 struct rq *busiest = NULL, *rq;
7241 unsigned long busiest_load = 0, busiest_capacity = 1;
7244 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7245 unsigned long capacity, wl;
7249 rt = fbq_classify_rq(rq);
7252 * We classify groups/runqueues into three groups:
7253 * - regular: there are !numa tasks
7254 * - remote: there are numa tasks that run on the 'wrong' node
7255 * - all: there is no distinction
7257 * In order to avoid migrating ideally placed numa tasks,
7258 * ignore those when there's better options.
7260 * If we ignore the actual busiest queue to migrate another
7261 * task, the next balance pass can still reduce the busiest
7262 * queue by moving tasks around inside the node.
7264 * If we cannot move enough load due to this classification
7265 * the next pass will adjust the group classification and
7266 * allow migration of more tasks.
7268 * Both cases only affect the total convergence complexity.
7270 if (rt > env->fbq_type)
7273 capacity = capacity_of(i);
7275 wl = weighted_cpuload(i);
7278 * When comparing with imbalance, use weighted_cpuload()
7279 * which is not scaled with the cpu capacity.
7282 if (rq->nr_running == 1 && wl > env->imbalance &&
7283 !check_cpu_capacity(rq, env->sd))
7287 * For the load comparisons with the other cpu's, consider
7288 * the weighted_cpuload() scaled with the cpu capacity, so
7289 * that the load can be moved away from the cpu that is
7290 * potentially running at a lower capacity.
7292 * Thus we're looking for max(wl_i / capacity_i), crosswise
7293 * multiplication to rid ourselves of the division works out
7294 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7295 * our previous maximum.
7297 if (wl * busiest_capacity > busiest_load * capacity) {
7299 busiest_capacity = capacity;
7308 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7309 * so long as it is large enough.
7311 #define MAX_PINNED_INTERVAL 512
7313 /* Working cpumask for load_balance and load_balance_newidle. */
7314 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7316 static int need_active_balance(struct lb_env *env)
7318 struct sched_domain *sd = env->sd;
7320 if (env->idle == CPU_NEWLY_IDLE) {
7323 * ASYM_PACKING needs to force migrate tasks from busy but
7324 * higher numbered CPUs in order to pack all tasks in the
7325 * lowest numbered CPUs.
7327 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7332 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7333 * It's worth migrating the task if the src_cpu's capacity is reduced
7334 * because of other sched_class or IRQs if more capacity stays
7335 * available on dst_cpu.
7337 if ((env->idle != CPU_NOT_IDLE) &&
7338 (env->src_rq->cfs.h_nr_running == 1)) {
7339 if ((check_cpu_capacity(env->src_rq, sd)) &&
7340 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7344 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7347 static int active_load_balance_cpu_stop(void *data);
7349 static int should_we_balance(struct lb_env *env)
7351 struct sched_group *sg = env->sd->groups;
7352 struct cpumask *sg_cpus, *sg_mask;
7353 int cpu, balance_cpu = -1;
7356 * In the newly idle case, we will allow all the cpu's
7357 * to do the newly idle load balance.
7359 if (env->idle == CPU_NEWLY_IDLE)
7362 sg_cpus = sched_group_cpus(sg);
7363 sg_mask = sched_group_mask(sg);
7364 /* Try to find first idle cpu */
7365 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7366 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7373 if (balance_cpu == -1)
7374 balance_cpu = group_balance_cpu(sg);
7377 * First idle cpu or the first cpu(busiest) in this sched group
7378 * is eligible for doing load balancing at this and above domains.
7380 return balance_cpu == env->dst_cpu;
7384 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7385 * tasks if there is an imbalance.
7387 static int load_balance(int this_cpu, struct rq *this_rq,
7388 struct sched_domain *sd, enum cpu_idle_type idle,
7389 int *continue_balancing)
7391 int ld_moved, cur_ld_moved, active_balance = 0;
7392 struct sched_domain *sd_parent = sd->parent;
7393 struct sched_group *group;
7395 unsigned long flags;
7396 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7398 struct lb_env env = {
7400 .dst_cpu = this_cpu,
7402 .dst_grpmask = sched_group_cpus(sd->groups),
7404 .loop_break = sched_nr_migrate_break,
7407 .tasks = LIST_HEAD_INIT(env.tasks),
7411 * For NEWLY_IDLE load_balancing, we don't need to consider
7412 * other cpus in our group
7414 if (idle == CPU_NEWLY_IDLE)
7415 env.dst_grpmask = NULL;
7417 cpumask_copy(cpus, cpu_active_mask);
7419 schedstat_inc(sd, lb_count[idle]);
7422 if (!should_we_balance(&env)) {
7423 *continue_balancing = 0;
7427 group = find_busiest_group(&env);
7429 schedstat_inc(sd, lb_nobusyg[idle]);
7433 busiest = find_busiest_queue(&env, group);
7435 schedstat_inc(sd, lb_nobusyq[idle]);
7439 BUG_ON(busiest == env.dst_rq);
7441 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7443 env.src_cpu = busiest->cpu;
7444 env.src_rq = busiest;
7447 if (busiest->nr_running > 1) {
7449 * Attempt to move tasks. If find_busiest_group has found
7450 * an imbalance but busiest->nr_running <= 1, the group is
7451 * still unbalanced. ld_moved simply stays zero, so it is
7452 * correctly treated as an imbalance.
7454 env.flags |= LBF_ALL_PINNED;
7455 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7458 raw_spin_lock_irqsave(&busiest->lock, flags);
7461 * cur_ld_moved - load moved in current iteration
7462 * ld_moved - cumulative load moved across iterations
7464 cur_ld_moved = detach_tasks(&env);
7467 * We've detached some tasks from busiest_rq. Every
7468 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7469 * unlock busiest->lock, and we are able to be sure
7470 * that nobody can manipulate the tasks in parallel.
7471 * See task_rq_lock() family for the details.
7474 raw_spin_unlock(&busiest->lock);
7478 ld_moved += cur_ld_moved;
7481 local_irq_restore(flags);
7483 if (env.flags & LBF_NEED_BREAK) {
7484 env.flags &= ~LBF_NEED_BREAK;
7489 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7490 * us and move them to an alternate dst_cpu in our sched_group
7491 * where they can run. The upper limit on how many times we
7492 * iterate on same src_cpu is dependent on number of cpus in our
7495 * This changes load balance semantics a bit on who can move
7496 * load to a given_cpu. In addition to the given_cpu itself
7497 * (or a ilb_cpu acting on its behalf where given_cpu is
7498 * nohz-idle), we now have balance_cpu in a position to move
7499 * load to given_cpu. In rare situations, this may cause
7500 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7501 * _independently_ and at _same_ time to move some load to
7502 * given_cpu) causing exceess load to be moved to given_cpu.
7503 * This however should not happen so much in practice and
7504 * moreover subsequent load balance cycles should correct the
7505 * excess load moved.
7507 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7509 /* Prevent to re-select dst_cpu via env's cpus */
7510 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7512 env.dst_rq = cpu_rq(env.new_dst_cpu);
7513 env.dst_cpu = env.new_dst_cpu;
7514 env.flags &= ~LBF_DST_PINNED;
7516 env.loop_break = sched_nr_migrate_break;
7519 * Go back to "more_balance" rather than "redo" since we
7520 * need to continue with same src_cpu.
7526 * We failed to reach balance because of affinity.
7529 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7531 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7532 *group_imbalance = 1;
7535 /* All tasks on this runqueue were pinned by CPU affinity */
7536 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7537 cpumask_clear_cpu(cpu_of(busiest), cpus);
7538 if (!cpumask_empty(cpus)) {
7540 env.loop_break = sched_nr_migrate_break;
7543 goto out_all_pinned;
7548 schedstat_inc(sd, lb_failed[idle]);
7550 * Increment the failure counter only on periodic balance.
7551 * We do not want newidle balance, which can be very
7552 * frequent, pollute the failure counter causing
7553 * excessive cache_hot migrations and active balances.
7555 if (idle != CPU_NEWLY_IDLE)
7556 sd->nr_balance_failed++;
7558 if (need_active_balance(&env)) {
7559 raw_spin_lock_irqsave(&busiest->lock, flags);
7561 /* don't kick the active_load_balance_cpu_stop,
7562 * if the curr task on busiest cpu can't be
7565 if (!cpumask_test_cpu(this_cpu,
7566 tsk_cpus_allowed(busiest->curr))) {
7567 raw_spin_unlock_irqrestore(&busiest->lock,
7569 env.flags |= LBF_ALL_PINNED;
7570 goto out_one_pinned;
7574 * ->active_balance synchronizes accesses to
7575 * ->active_balance_work. Once set, it's cleared
7576 * only after active load balance is finished.
7578 if (!busiest->active_balance) {
7579 busiest->active_balance = 1;
7580 busiest->push_cpu = this_cpu;
7583 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7585 if (active_balance) {
7586 stop_one_cpu_nowait(cpu_of(busiest),
7587 active_load_balance_cpu_stop, busiest,
7588 &busiest->active_balance_work);
7591 /* We've kicked active balancing, force task migration. */
7592 sd->nr_balance_failed = sd->cache_nice_tries+1;
7595 sd->nr_balance_failed = 0;
7597 if (likely(!active_balance)) {
7598 /* We were unbalanced, so reset the balancing interval */
7599 sd->balance_interval = sd->min_interval;
7602 * If we've begun active balancing, start to back off. This
7603 * case may not be covered by the all_pinned logic if there
7604 * is only 1 task on the busy runqueue (because we don't call
7607 if (sd->balance_interval < sd->max_interval)
7608 sd->balance_interval *= 2;
7615 * We reach balance although we may have faced some affinity
7616 * constraints. Clear the imbalance flag if it was set.
7619 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7621 if (*group_imbalance)
7622 *group_imbalance = 0;
7627 * We reach balance because all tasks are pinned at this level so
7628 * we can't migrate them. Let the imbalance flag set so parent level
7629 * can try to migrate them.
7631 schedstat_inc(sd, lb_balanced[idle]);
7633 sd->nr_balance_failed = 0;
7636 /* tune up the balancing interval */
7637 if (((env.flags & LBF_ALL_PINNED) &&
7638 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7639 (sd->balance_interval < sd->max_interval))
7640 sd->balance_interval *= 2;
7647 static inline unsigned long
7648 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7650 unsigned long interval = sd->balance_interval;
7653 interval *= sd->busy_factor;
7655 /* scale ms to jiffies */
7656 interval = msecs_to_jiffies(interval);
7657 interval = clamp(interval, 1UL, max_load_balance_interval);
7663 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7665 unsigned long interval, next;
7667 interval = get_sd_balance_interval(sd, cpu_busy);
7668 next = sd->last_balance + interval;
7670 if (time_after(*next_balance, next))
7671 *next_balance = next;
7675 * idle_balance is called by schedule() if this_cpu is about to become
7676 * idle. Attempts to pull tasks from other CPUs.
7678 static int idle_balance(struct rq *this_rq)
7680 unsigned long next_balance = jiffies + HZ;
7681 int this_cpu = this_rq->cpu;
7682 struct sched_domain *sd;
7683 int pulled_task = 0;
7687 * We must set idle_stamp _before_ calling idle_balance(), such that we
7688 * measure the duration of idle_balance() as idle time.
7690 this_rq->idle_stamp = rq_clock(this_rq);
7692 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7693 !this_rq->rd->overload) {
7695 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7697 update_next_balance(sd, 0, &next_balance);
7703 raw_spin_unlock(&this_rq->lock);
7705 update_blocked_averages(this_cpu);
7707 for_each_domain(this_cpu, sd) {
7708 int continue_balancing = 1;
7709 u64 t0, domain_cost;
7711 if (!(sd->flags & SD_LOAD_BALANCE))
7714 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7715 update_next_balance(sd, 0, &next_balance);
7719 if (sd->flags & SD_BALANCE_NEWIDLE) {
7720 t0 = sched_clock_cpu(this_cpu);
7722 pulled_task = load_balance(this_cpu, this_rq,
7724 &continue_balancing);
7726 domain_cost = sched_clock_cpu(this_cpu) - t0;
7727 if (domain_cost > sd->max_newidle_lb_cost)
7728 sd->max_newidle_lb_cost = domain_cost;
7730 curr_cost += domain_cost;
7733 update_next_balance(sd, 0, &next_balance);
7736 * Stop searching for tasks to pull if there are
7737 * now runnable tasks on this rq.
7739 if (pulled_task || this_rq->nr_running > 0)
7744 raw_spin_lock(&this_rq->lock);
7746 if (curr_cost > this_rq->max_idle_balance_cost)
7747 this_rq->max_idle_balance_cost = curr_cost;
7750 * While browsing the domains, we released the rq lock, a task could
7751 * have been enqueued in the meantime. Since we're not going idle,
7752 * pretend we pulled a task.
7754 if (this_rq->cfs.h_nr_running && !pulled_task)
7758 /* Move the next balance forward */
7759 if (time_after(this_rq->next_balance, next_balance))
7760 this_rq->next_balance = next_balance;
7762 /* Is there a task of a high priority class? */
7763 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7767 this_rq->idle_stamp = 0;
7773 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7774 * running tasks off the busiest CPU onto idle CPUs. It requires at
7775 * least 1 task to be running on each physical CPU where possible, and
7776 * avoids physical / logical imbalances.
7778 static int active_load_balance_cpu_stop(void *data)
7780 struct rq *busiest_rq = data;
7781 int busiest_cpu = cpu_of(busiest_rq);
7782 int target_cpu = busiest_rq->push_cpu;
7783 struct rq *target_rq = cpu_rq(target_cpu);
7784 struct sched_domain *sd;
7785 struct task_struct *p = NULL;
7787 raw_spin_lock_irq(&busiest_rq->lock);
7789 /* make sure the requested cpu hasn't gone down in the meantime */
7790 if (unlikely(busiest_cpu != smp_processor_id() ||
7791 !busiest_rq->active_balance))
7794 /* Is there any task to move? */
7795 if (busiest_rq->nr_running <= 1)
7799 * This condition is "impossible", if it occurs
7800 * we need to fix it. Originally reported by
7801 * Bjorn Helgaas on a 128-cpu setup.
7803 BUG_ON(busiest_rq == target_rq);
7805 /* Search for an sd spanning us and the target CPU. */
7807 for_each_domain(target_cpu, sd) {
7808 if ((sd->flags & SD_LOAD_BALANCE) &&
7809 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7814 struct lb_env env = {
7816 .dst_cpu = target_cpu,
7817 .dst_rq = target_rq,
7818 .src_cpu = busiest_rq->cpu,
7819 .src_rq = busiest_rq,
7823 schedstat_inc(sd, alb_count);
7825 p = detach_one_task(&env);
7827 schedstat_inc(sd, alb_pushed);
7828 /* Active balancing done, reset the failure counter. */
7829 sd->nr_balance_failed = 0;
7831 schedstat_inc(sd, alb_failed);
7836 busiest_rq->active_balance = 0;
7837 raw_spin_unlock(&busiest_rq->lock);
7840 attach_one_task(target_rq, p);
7847 static inline int on_null_domain(struct rq *rq)
7849 return unlikely(!rcu_dereference_sched(rq->sd));
7852 #ifdef CONFIG_NO_HZ_COMMON
7854 * idle load balancing details
7855 * - When one of the busy CPUs notice that there may be an idle rebalancing
7856 * needed, they will kick the idle load balancer, which then does idle
7857 * load balancing for all the idle CPUs.
7860 cpumask_var_t idle_cpus_mask;
7862 unsigned long next_balance; /* in jiffy units */
7863 } nohz ____cacheline_aligned;
7865 static inline int find_new_ilb(void)
7867 int ilb = cpumask_first(nohz.idle_cpus_mask);
7869 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7876 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7877 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7878 * CPU (if there is one).
7880 static void nohz_balancer_kick(void)
7884 nohz.next_balance++;
7886 ilb_cpu = find_new_ilb();
7888 if (ilb_cpu >= nr_cpu_ids)
7891 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7894 * Use smp_send_reschedule() instead of resched_cpu().
7895 * This way we generate a sched IPI on the target cpu which
7896 * is idle. And the softirq performing nohz idle load balance
7897 * will be run before returning from the IPI.
7899 smp_send_reschedule(ilb_cpu);
7903 void nohz_balance_exit_idle(unsigned int cpu)
7905 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7907 * Completely isolated CPUs don't ever set, so we must test.
7909 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7910 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7911 atomic_dec(&nohz.nr_cpus);
7913 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7917 static inline void set_cpu_sd_state_busy(void)
7919 struct sched_domain *sd;
7920 int cpu = smp_processor_id();
7923 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7925 if (!sd || !sd->nohz_idle)
7929 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7934 void set_cpu_sd_state_idle(void)
7936 struct sched_domain *sd;
7937 int cpu = smp_processor_id();
7940 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7942 if (!sd || sd->nohz_idle)
7946 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7952 * This routine will record that the cpu is going idle with tick stopped.
7953 * This info will be used in performing idle load balancing in the future.
7955 void nohz_balance_enter_idle(int cpu)
7958 * If this cpu is going down, then nothing needs to be done.
7960 if (!cpu_active(cpu))
7963 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7967 * If we're a completely isolated CPU, we don't play.
7969 if (on_null_domain(cpu_rq(cpu)))
7972 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7973 atomic_inc(&nohz.nr_cpus);
7974 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7978 static DEFINE_SPINLOCK(balancing);
7981 * Scale the max load_balance interval with the number of CPUs in the system.
7982 * This trades load-balance latency on larger machines for less cross talk.
7984 void update_max_interval(void)
7986 max_load_balance_interval = HZ*num_online_cpus()/10;
7990 * It checks each scheduling domain to see if it is due to be balanced,
7991 * and initiates a balancing operation if so.
7993 * Balancing parameters are set up in init_sched_domains.
7995 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7997 int continue_balancing = 1;
7999 unsigned long interval;
8000 struct sched_domain *sd;
8001 /* Earliest time when we have to do rebalance again */
8002 unsigned long next_balance = jiffies + 60*HZ;
8003 int update_next_balance = 0;
8004 int need_serialize, need_decay = 0;
8007 update_blocked_averages(cpu);
8010 for_each_domain(cpu, sd) {
8012 * Decay the newidle max times here because this is a regular
8013 * visit to all the domains. Decay ~1% per second.
8015 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8016 sd->max_newidle_lb_cost =
8017 (sd->max_newidle_lb_cost * 253) / 256;
8018 sd->next_decay_max_lb_cost = jiffies + HZ;
8021 max_cost += sd->max_newidle_lb_cost;
8023 if (!(sd->flags & SD_LOAD_BALANCE))
8027 * Stop the load balance at this level. There is another
8028 * CPU in our sched group which is doing load balancing more
8031 if (!continue_balancing) {
8037 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8039 need_serialize = sd->flags & SD_SERIALIZE;
8040 if (need_serialize) {
8041 if (!spin_trylock(&balancing))
8045 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8046 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8048 * The LBF_DST_PINNED logic could have changed
8049 * env->dst_cpu, so we can't know our idle
8050 * state even if we migrated tasks. Update it.
8052 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8054 sd->last_balance = jiffies;
8055 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8058 spin_unlock(&balancing);
8060 if (time_after(next_balance, sd->last_balance + interval)) {
8061 next_balance = sd->last_balance + interval;
8062 update_next_balance = 1;
8067 * Ensure the rq-wide value also decays but keep it at a
8068 * reasonable floor to avoid funnies with rq->avg_idle.
8070 rq->max_idle_balance_cost =
8071 max((u64)sysctl_sched_migration_cost, max_cost);
8076 * next_balance will be updated only when there is a need.
8077 * When the cpu is attached to null domain for ex, it will not be
8080 if (likely(update_next_balance)) {
8081 rq->next_balance = next_balance;
8083 #ifdef CONFIG_NO_HZ_COMMON
8085 * If this CPU has been elected to perform the nohz idle
8086 * balance. Other idle CPUs have already rebalanced with
8087 * nohz_idle_balance() and nohz.next_balance has been
8088 * updated accordingly. This CPU is now running the idle load
8089 * balance for itself and we need to update the
8090 * nohz.next_balance accordingly.
8092 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8093 nohz.next_balance = rq->next_balance;
8098 #ifdef CONFIG_NO_HZ_COMMON
8100 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8101 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8103 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8105 int this_cpu = this_rq->cpu;
8108 /* Earliest time when we have to do rebalance again */
8109 unsigned long next_balance = jiffies + 60*HZ;
8110 int update_next_balance = 0;
8112 if (idle != CPU_IDLE ||
8113 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8116 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8117 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8121 * If this cpu gets work to do, stop the load balancing
8122 * work being done for other cpus. Next load
8123 * balancing owner will pick it up.
8128 rq = cpu_rq(balance_cpu);
8131 * If time for next balance is due,
8134 if (time_after_eq(jiffies, rq->next_balance)) {
8135 raw_spin_lock_irq(&rq->lock);
8136 update_rq_clock(rq);
8137 cpu_load_update_idle(rq);
8138 raw_spin_unlock_irq(&rq->lock);
8139 rebalance_domains(rq, CPU_IDLE);
8142 if (time_after(next_balance, rq->next_balance)) {
8143 next_balance = rq->next_balance;
8144 update_next_balance = 1;
8149 * next_balance will be updated only when there is a need.
8150 * When the CPU is attached to null domain for ex, it will not be
8153 if (likely(update_next_balance))
8154 nohz.next_balance = next_balance;
8156 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8160 * Current heuristic for kicking the idle load balancer in the presence
8161 * of an idle cpu in the system.
8162 * - This rq has more than one task.
8163 * - This rq has at least one CFS task and the capacity of the CPU is
8164 * significantly reduced because of RT tasks or IRQs.
8165 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8166 * multiple busy cpu.
8167 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8168 * domain span are idle.
8170 static inline bool nohz_kick_needed(struct rq *rq)
8172 unsigned long now = jiffies;
8173 struct sched_domain *sd;
8174 struct sched_group_capacity *sgc;
8175 int nr_busy, cpu = rq->cpu;
8178 if (unlikely(rq->idle_balance))
8182 * We may be recently in ticked or tickless idle mode. At the first
8183 * busy tick after returning from idle, we will update the busy stats.
8185 set_cpu_sd_state_busy();
8186 nohz_balance_exit_idle(cpu);
8189 * None are in tickless mode and hence no need for NOHZ idle load
8192 if (likely(!atomic_read(&nohz.nr_cpus)))
8195 if (time_before(now, nohz.next_balance))
8198 if (rq->nr_running >= 2)
8202 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8204 sgc = sd->groups->sgc;
8205 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8214 sd = rcu_dereference(rq->sd);
8216 if ((rq->cfs.h_nr_running >= 1) &&
8217 check_cpu_capacity(rq, sd)) {
8223 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8224 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8225 sched_domain_span(sd)) < cpu)) {
8235 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8239 * run_rebalance_domains is triggered when needed from the scheduler tick.
8240 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8242 static void run_rebalance_domains(struct softirq_action *h)
8244 struct rq *this_rq = this_rq();
8245 enum cpu_idle_type idle = this_rq->idle_balance ?
8246 CPU_IDLE : CPU_NOT_IDLE;
8249 * If this cpu has a pending nohz_balance_kick, then do the
8250 * balancing on behalf of the other idle cpus whose ticks are
8251 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8252 * give the idle cpus a chance to load balance. Else we may
8253 * load balance only within the local sched_domain hierarchy
8254 * and abort nohz_idle_balance altogether if we pull some load.
8256 nohz_idle_balance(this_rq, idle);
8257 rebalance_domains(this_rq, idle);
8261 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8263 void trigger_load_balance(struct rq *rq)
8265 /* Don't need to rebalance while attached to NULL domain */
8266 if (unlikely(on_null_domain(rq)))
8269 if (time_after_eq(jiffies, rq->next_balance))
8270 raise_softirq(SCHED_SOFTIRQ);
8271 #ifdef CONFIG_NO_HZ_COMMON
8272 if (nohz_kick_needed(rq))
8273 nohz_balancer_kick();
8277 static void rq_online_fair(struct rq *rq)
8281 update_runtime_enabled(rq);
8284 static void rq_offline_fair(struct rq *rq)
8288 /* Ensure any throttled groups are reachable by pick_next_task */
8289 unthrottle_offline_cfs_rqs(rq);
8292 #endif /* CONFIG_SMP */
8295 * scheduler tick hitting a task of our scheduling class:
8297 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8299 struct cfs_rq *cfs_rq;
8300 struct sched_entity *se = &curr->se;
8302 for_each_sched_entity(se) {
8303 cfs_rq = cfs_rq_of(se);
8304 entity_tick(cfs_rq, se, queued);
8307 if (static_branch_unlikely(&sched_numa_balancing))
8308 task_tick_numa(rq, curr);
8312 * called on fork with the child task as argument from the parent's context
8313 * - child not yet on the tasklist
8314 * - preemption disabled
8316 static void task_fork_fair(struct task_struct *p)
8318 struct cfs_rq *cfs_rq;
8319 struct sched_entity *se = &p->se, *curr;
8320 int this_cpu = smp_processor_id();
8321 struct rq *rq = this_rq();
8322 unsigned long flags;
8324 raw_spin_lock_irqsave(&rq->lock, flags);
8326 update_rq_clock(rq);
8328 cfs_rq = task_cfs_rq(current);
8329 curr = cfs_rq->curr;
8332 * Not only the cpu but also the task_group of the parent might have
8333 * been changed after parent->se.parent,cfs_rq were copied to
8334 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8335 * of child point to valid ones.
8338 __set_task_cpu(p, this_cpu);
8341 update_curr(cfs_rq);
8344 se->vruntime = curr->vruntime;
8345 place_entity(cfs_rq, se, 1);
8347 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8349 * Upon rescheduling, sched_class::put_prev_task() will place
8350 * 'current' within the tree based on its new key value.
8352 swap(curr->vruntime, se->vruntime);
8356 se->vruntime -= cfs_rq->min_vruntime;
8358 raw_spin_unlock_irqrestore(&rq->lock, flags);
8362 * Priority of the task has changed. Check to see if we preempt
8366 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8368 if (!task_on_rq_queued(p))
8372 * Reschedule if we are currently running on this runqueue and
8373 * our priority decreased, or if we are not currently running on
8374 * this runqueue and our priority is higher than the current's
8376 if (rq->curr == p) {
8377 if (p->prio > oldprio)
8380 check_preempt_curr(rq, p, 0);
8383 static inline bool vruntime_normalized(struct task_struct *p)
8385 struct sched_entity *se = &p->se;
8388 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8389 * the dequeue_entity(.flags=0) will already have normalized the
8396 * When !on_rq, vruntime of the task has usually NOT been normalized.
8397 * But there are some cases where it has already been normalized:
8399 * - A forked child which is waiting for being woken up by
8400 * wake_up_new_task().
8401 * - A task which has been woken up by try_to_wake_up() and
8402 * waiting for actually being woken up by sched_ttwu_pending().
8404 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8410 static void detach_task_cfs_rq(struct task_struct *p)
8412 struct sched_entity *se = &p->se;
8413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8415 if (!vruntime_normalized(p)) {
8417 * Fix up our vruntime so that the current sleep doesn't
8418 * cause 'unlimited' sleep bonus.
8420 place_entity(cfs_rq, se, 0);
8421 se->vruntime -= cfs_rq->min_vruntime;
8424 /* Catch up with the cfs_rq and remove our load when we leave */
8425 detach_entity_load_avg(cfs_rq, se);
8428 static void attach_task_cfs_rq(struct task_struct *p)
8430 struct sched_entity *se = &p->se;
8431 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8433 #ifdef CONFIG_FAIR_GROUP_SCHED
8435 * Since the real-depth could have been changed (only FAIR
8436 * class maintain depth value), reset depth properly.
8438 se->depth = se->parent ? se->parent->depth + 1 : 0;
8441 /* Synchronize task with its cfs_rq */
8442 attach_entity_load_avg(cfs_rq, se);
8444 if (!vruntime_normalized(p))
8445 se->vruntime += cfs_rq->min_vruntime;
8448 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8450 detach_task_cfs_rq(p);
8453 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8455 attach_task_cfs_rq(p);
8457 if (task_on_rq_queued(p)) {
8459 * We were most likely switched from sched_rt, so
8460 * kick off the schedule if running, otherwise just see
8461 * if we can still preempt the current task.
8466 check_preempt_curr(rq, p, 0);
8470 /* Account for a task changing its policy or group.
8472 * This routine is mostly called to set cfs_rq->curr field when a task
8473 * migrates between groups/classes.
8475 static void set_curr_task_fair(struct rq *rq)
8477 struct sched_entity *se = &rq->curr->se;
8479 for_each_sched_entity(se) {
8480 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8482 set_next_entity(cfs_rq, se);
8483 /* ensure bandwidth has been allocated on our new cfs_rq */
8484 account_cfs_rq_runtime(cfs_rq, 0);
8488 void init_cfs_rq(struct cfs_rq *cfs_rq)
8490 cfs_rq->tasks_timeline = RB_ROOT;
8491 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8492 #ifndef CONFIG_64BIT
8493 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8496 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8497 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8501 #ifdef CONFIG_FAIR_GROUP_SCHED
8502 static void task_move_group_fair(struct task_struct *p)
8504 detach_task_cfs_rq(p);
8505 set_task_rq(p, task_cpu(p));
8508 /* Tell se's cfs_rq has been changed -- migrated */
8509 p->se.avg.last_update_time = 0;
8511 attach_task_cfs_rq(p);
8514 void free_fair_sched_group(struct task_group *tg)
8518 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8520 for_each_possible_cpu(i) {
8522 kfree(tg->cfs_rq[i]);
8531 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8533 struct sched_entity *se;
8534 struct cfs_rq *cfs_rq;
8538 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8541 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8545 tg->shares = NICE_0_LOAD;
8547 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8549 for_each_possible_cpu(i) {
8552 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8553 GFP_KERNEL, cpu_to_node(i));
8557 se = kzalloc_node(sizeof(struct sched_entity),
8558 GFP_KERNEL, cpu_to_node(i));
8562 init_cfs_rq(cfs_rq);
8563 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8564 init_entity_runnable_average(se);
8566 raw_spin_lock_irq(&rq->lock);
8567 post_init_entity_util_avg(se);
8568 raw_spin_unlock_irq(&rq->lock);
8579 void unregister_fair_sched_group(struct task_group *tg)
8581 unsigned long flags;
8585 for_each_possible_cpu(cpu) {
8587 remove_entity_load_avg(tg->se[cpu]);
8590 * Only empty task groups can be destroyed; so we can speculatively
8591 * check on_list without danger of it being re-added.
8593 if (!tg->cfs_rq[cpu]->on_list)
8598 raw_spin_lock_irqsave(&rq->lock, flags);
8599 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8600 raw_spin_unlock_irqrestore(&rq->lock, flags);
8604 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8605 struct sched_entity *se, int cpu,
8606 struct sched_entity *parent)
8608 struct rq *rq = cpu_rq(cpu);
8612 init_cfs_rq_runtime(cfs_rq);
8614 tg->cfs_rq[cpu] = cfs_rq;
8617 /* se could be NULL for root_task_group */
8622 se->cfs_rq = &rq->cfs;
8625 se->cfs_rq = parent->my_q;
8626 se->depth = parent->depth + 1;
8630 /* guarantee group entities always have weight */
8631 update_load_set(&se->load, NICE_0_LOAD);
8632 se->parent = parent;
8635 static DEFINE_MUTEX(shares_mutex);
8637 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8640 unsigned long flags;
8643 * We can't change the weight of the root cgroup.
8648 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8650 mutex_lock(&shares_mutex);
8651 if (tg->shares == shares)
8654 tg->shares = shares;
8655 for_each_possible_cpu(i) {
8656 struct rq *rq = cpu_rq(i);
8657 struct sched_entity *se;
8660 /* Propagate contribution to hierarchy */
8661 raw_spin_lock_irqsave(&rq->lock, flags);
8663 /* Possible calls to update_curr() need rq clock */
8664 update_rq_clock(rq);
8665 for_each_sched_entity(se)
8666 update_cfs_shares(group_cfs_rq(se));
8667 raw_spin_unlock_irqrestore(&rq->lock, flags);
8671 mutex_unlock(&shares_mutex);
8674 #else /* CONFIG_FAIR_GROUP_SCHED */
8676 void free_fair_sched_group(struct task_group *tg) { }
8678 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8683 void unregister_fair_sched_group(struct task_group *tg) { }
8685 #endif /* CONFIG_FAIR_GROUP_SCHED */
8688 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8690 struct sched_entity *se = &task->se;
8691 unsigned int rr_interval = 0;
8694 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8697 if (rq->cfs.load.weight)
8698 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8704 * All the scheduling class methods:
8706 const struct sched_class fair_sched_class = {
8707 .next = &idle_sched_class,
8708 .enqueue_task = enqueue_task_fair,
8709 .dequeue_task = dequeue_task_fair,
8710 .yield_task = yield_task_fair,
8711 .yield_to_task = yield_to_task_fair,
8713 .check_preempt_curr = check_preempt_wakeup,
8715 .pick_next_task = pick_next_task_fair,
8716 .put_prev_task = put_prev_task_fair,
8719 .select_task_rq = select_task_rq_fair,
8720 .migrate_task_rq = migrate_task_rq_fair,
8722 .rq_online = rq_online_fair,
8723 .rq_offline = rq_offline_fair,
8725 .task_dead = task_dead_fair,
8726 .set_cpus_allowed = set_cpus_allowed_common,
8729 .set_curr_task = set_curr_task_fair,
8730 .task_tick = task_tick_fair,
8731 .task_fork = task_fork_fair,
8733 .prio_changed = prio_changed_fair,
8734 .switched_from = switched_from_fair,
8735 .switched_to = switched_to_fair,
8737 .get_rr_interval = get_rr_interval_fair,
8739 .update_curr = update_curr_fair,
8741 #ifdef CONFIG_FAIR_GROUP_SCHED
8742 .task_move_group = task_move_group_fair,
8746 #ifdef CONFIG_SCHED_DEBUG
8747 void print_cfs_stats(struct seq_file *m, int cpu)
8749 struct cfs_rq *cfs_rq;
8752 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8753 print_cfs_rq(m, cpu, cfs_rq);
8757 #ifdef CONFIG_NUMA_BALANCING
8758 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8761 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8763 for_each_online_node(node) {
8764 if (p->numa_faults) {
8765 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8766 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8768 if (p->numa_group) {
8769 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8770 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8772 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8775 #endif /* CONFIG_NUMA_BALANCING */
8776 #endif /* CONFIG_SCHED_DEBUG */
8778 __init void init_sched_fair_class(void)
8781 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8783 #ifdef CONFIG_NO_HZ_COMMON
8784 nohz.next_balance = jiffies;
8785 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);