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);
1310 env->best_imp = imp;
1311 env->best_cpu = env->dst_cpu;
1314 static bool load_too_imbalanced(long src_load, long dst_load,
1315 struct task_numa_env *env)
1318 long orig_src_load, orig_dst_load;
1319 long src_capacity, dst_capacity;
1322 * The load is corrected for the CPU capacity available on each node.
1325 * ------------ vs ---------
1326 * src_capacity dst_capacity
1328 src_capacity = env->src_stats.compute_capacity;
1329 dst_capacity = env->dst_stats.compute_capacity;
1331 /* We care about the slope of the imbalance, not the direction. */
1332 if (dst_load < src_load)
1333 swap(dst_load, src_load);
1335 /* Is the difference below the threshold? */
1336 imb = dst_load * src_capacity * 100 -
1337 src_load * dst_capacity * env->imbalance_pct;
1342 * The imbalance is above the allowed threshold.
1343 * Compare it with the old imbalance.
1345 orig_src_load = env->src_stats.load;
1346 orig_dst_load = env->dst_stats.load;
1348 if (orig_dst_load < orig_src_load)
1349 swap(orig_dst_load, orig_src_load);
1351 old_imb = orig_dst_load * src_capacity * 100 -
1352 orig_src_load * dst_capacity * env->imbalance_pct;
1354 /* Would this change make things worse? */
1355 return (imb > old_imb);
1359 * This checks if the overall compute and NUMA accesses of the system would
1360 * be improved if the source tasks was migrated to the target dst_cpu taking
1361 * into account that it might be best if task running on the dst_cpu should
1362 * be exchanged with the source task
1364 static void task_numa_compare(struct task_numa_env *env,
1365 long taskimp, long groupimp)
1367 struct rq *src_rq = cpu_rq(env->src_cpu);
1368 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1369 struct task_struct *cur;
1370 long src_load, dst_load;
1372 long imp = env->p->numa_group ? groupimp : taskimp;
1374 int dist = env->dist;
1377 cur = task_rcu_dereference(&dst_rq->curr);
1378 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1382 * Because we have preemption enabled we can get migrated around and
1383 * end try selecting ourselves (current == env->p) as a swap candidate.
1389 * "imp" is the fault differential for the source task between the
1390 * source and destination node. Calculate the total differential for
1391 * the source task and potential destination task. The more negative
1392 * the value is, the more rmeote accesses that would be expected to
1393 * be incurred if the tasks were swapped.
1396 /* Skip this swap candidate if cannot move to the source cpu */
1397 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1401 * If dst and source tasks are in the same NUMA group, or not
1402 * in any group then look only at task weights.
1404 if (cur->numa_group == env->p->numa_group) {
1405 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1406 task_weight(cur, env->dst_nid, dist);
1408 * Add some hysteresis to prevent swapping the
1409 * tasks within a group over tiny differences.
1411 if (cur->numa_group)
1415 * Compare the group weights. If a task is all by
1416 * itself (not part of a group), use the task weight
1419 if (cur->numa_group)
1420 imp += group_weight(cur, env->src_nid, dist) -
1421 group_weight(cur, env->dst_nid, dist);
1423 imp += task_weight(cur, env->src_nid, dist) -
1424 task_weight(cur, env->dst_nid, dist);
1428 if (imp <= env->best_imp && moveimp <= env->best_imp)
1432 /* Is there capacity at our destination? */
1433 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1434 !env->dst_stats.has_free_capacity)
1440 /* Balance doesn't matter much if we're running a task per cpu */
1441 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1442 dst_rq->nr_running == 1)
1446 * In the overloaded case, try and keep the load balanced.
1449 load = task_h_load(env->p);
1450 dst_load = env->dst_stats.load + load;
1451 src_load = env->src_stats.load - load;
1453 if (moveimp > imp && moveimp > env->best_imp) {
1455 * If the improvement from just moving env->p direction is
1456 * better than swapping tasks around, check if a move is
1457 * possible. Store a slightly smaller score than moveimp,
1458 * so an actually idle CPU will win.
1460 if (!load_too_imbalanced(src_load, dst_load, env)) {
1467 if (imp <= env->best_imp)
1471 load = task_h_load(cur);
1476 if (load_too_imbalanced(src_load, dst_load, env))
1480 * One idle CPU per node is evaluated for a task numa move.
1481 * Call select_idle_sibling to maybe find a better one.
1484 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1487 task_numa_assign(env, cur, imp);
1492 static void task_numa_find_cpu(struct task_numa_env *env,
1493 long taskimp, long groupimp)
1497 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1498 /* Skip this CPU if the source task cannot migrate */
1499 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1503 task_numa_compare(env, taskimp, groupimp);
1507 /* Only move tasks to a NUMA node less busy than the current node. */
1508 static bool numa_has_capacity(struct task_numa_env *env)
1510 struct numa_stats *src = &env->src_stats;
1511 struct numa_stats *dst = &env->dst_stats;
1513 if (src->has_free_capacity && !dst->has_free_capacity)
1517 * Only consider a task move if the source has a higher load
1518 * than the destination, corrected for CPU capacity on each node.
1520 * src->load dst->load
1521 * --------------------- vs ---------------------
1522 * src->compute_capacity dst->compute_capacity
1524 if (src->load * dst->compute_capacity * env->imbalance_pct >
1526 dst->load * src->compute_capacity * 100)
1532 static int task_numa_migrate(struct task_struct *p)
1534 struct task_numa_env env = {
1537 .src_cpu = task_cpu(p),
1538 .src_nid = task_node(p),
1540 .imbalance_pct = 112,
1546 struct sched_domain *sd;
1547 unsigned long taskweight, groupweight;
1549 long taskimp, groupimp;
1552 * Pick the lowest SD_NUMA domain, as that would have the smallest
1553 * imbalance and would be the first to start moving tasks about.
1555 * And we want to avoid any moving of tasks about, as that would create
1556 * random movement of tasks -- counter the numa conditions we're trying
1560 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1562 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1566 * Cpusets can break the scheduler domain tree into smaller
1567 * balance domains, some of which do not cross NUMA boundaries.
1568 * Tasks that are "trapped" in such domains cannot be migrated
1569 * elsewhere, so there is no point in (re)trying.
1571 if (unlikely(!sd)) {
1572 p->numa_preferred_nid = task_node(p);
1576 env.dst_nid = p->numa_preferred_nid;
1577 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1578 taskweight = task_weight(p, env.src_nid, dist);
1579 groupweight = group_weight(p, env.src_nid, dist);
1580 update_numa_stats(&env.src_stats, env.src_nid);
1581 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1582 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1583 update_numa_stats(&env.dst_stats, env.dst_nid);
1585 /* Try to find a spot on the preferred nid. */
1586 if (numa_has_capacity(&env))
1587 task_numa_find_cpu(&env, taskimp, groupimp);
1590 * Look at other nodes in these cases:
1591 * - there is no space available on the preferred_nid
1592 * - the task is part of a numa_group that is interleaved across
1593 * multiple NUMA nodes; in order to better consolidate the group,
1594 * we need to check other locations.
1596 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1597 for_each_online_node(nid) {
1598 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1601 dist = node_distance(env.src_nid, env.dst_nid);
1602 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1604 taskweight = task_weight(p, env.src_nid, dist);
1605 groupweight = group_weight(p, env.src_nid, dist);
1608 /* Only consider nodes where both task and groups benefit */
1609 taskimp = task_weight(p, nid, dist) - taskweight;
1610 groupimp = group_weight(p, nid, dist) - groupweight;
1611 if (taskimp < 0 && groupimp < 0)
1616 update_numa_stats(&env.dst_stats, env.dst_nid);
1617 if (numa_has_capacity(&env))
1618 task_numa_find_cpu(&env, taskimp, groupimp);
1623 * If the task is part of a workload that spans multiple NUMA nodes,
1624 * and is migrating into one of the workload's active nodes, remember
1625 * this node as the task's preferred numa node, so the workload can
1627 * A task that migrated to a second choice node will be better off
1628 * trying for a better one later. Do not set the preferred node here.
1630 if (p->numa_group) {
1631 struct numa_group *ng = p->numa_group;
1633 if (env.best_cpu == -1)
1638 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1639 sched_setnuma(p, env.dst_nid);
1642 /* No better CPU than the current one was found. */
1643 if (env.best_cpu == -1)
1647 * Reset the scan period if the task is being rescheduled on an
1648 * alternative node to recheck if the tasks is now properly placed.
1650 p->numa_scan_period = task_scan_min(p);
1652 if (env.best_task == NULL) {
1653 ret = migrate_task_to(p, env.best_cpu);
1655 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1659 ret = migrate_swap(p, env.best_task);
1661 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1662 put_task_struct(env.best_task);
1666 /* Attempt to migrate a task to a CPU on the preferred node. */
1667 static void numa_migrate_preferred(struct task_struct *p)
1669 unsigned long interval = HZ;
1671 /* This task has no NUMA fault statistics yet */
1672 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1675 /* Periodically retry migrating the task to the preferred node */
1676 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1677 p->numa_migrate_retry = jiffies + interval;
1679 /* Success if task is already running on preferred CPU */
1680 if (task_node(p) == p->numa_preferred_nid)
1683 /* Otherwise, try migrate to a CPU on the preferred node */
1684 task_numa_migrate(p);
1688 * Find out how many nodes on the workload is actively running on. Do this by
1689 * tracking the nodes from which NUMA hinting faults are triggered. This can
1690 * be different from the set of nodes where the workload's memory is currently
1693 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1695 unsigned long faults, max_faults = 0;
1696 int nid, active_nodes = 0;
1698 for_each_online_node(nid) {
1699 faults = group_faults_cpu(numa_group, nid);
1700 if (faults > max_faults)
1701 max_faults = faults;
1704 for_each_online_node(nid) {
1705 faults = group_faults_cpu(numa_group, nid);
1706 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1710 numa_group->max_faults_cpu = max_faults;
1711 numa_group->active_nodes = active_nodes;
1715 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1716 * increments. The more local the fault statistics are, the higher the scan
1717 * period will be for the next scan window. If local/(local+remote) ratio is
1718 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1719 * the scan period will decrease. Aim for 70% local accesses.
1721 #define NUMA_PERIOD_SLOTS 10
1722 #define NUMA_PERIOD_THRESHOLD 7
1725 * Increase the scan period (slow down scanning) if the majority of
1726 * our memory is already on our local node, or if the majority of
1727 * the page accesses are shared with other processes.
1728 * Otherwise, decrease the scan period.
1730 static void update_task_scan_period(struct task_struct *p,
1731 unsigned long shared, unsigned long private)
1733 unsigned int period_slot;
1737 unsigned long remote = p->numa_faults_locality[0];
1738 unsigned long local = p->numa_faults_locality[1];
1741 * If there were no record hinting faults then either the task is
1742 * completely idle or all activity is areas that are not of interest
1743 * to automatic numa balancing. Related to that, if there were failed
1744 * migration then it implies we are migrating too quickly or the local
1745 * node is overloaded. In either case, scan slower
1747 if (local + shared == 0 || p->numa_faults_locality[2]) {
1748 p->numa_scan_period = min(p->numa_scan_period_max,
1749 p->numa_scan_period << 1);
1751 p->mm->numa_next_scan = jiffies +
1752 msecs_to_jiffies(p->numa_scan_period);
1758 * Prepare to scale scan period relative to the current period.
1759 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1760 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1761 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1763 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1764 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1765 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1766 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1769 diff = slot * period_slot;
1771 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1774 * Scale scan rate increases based on sharing. There is an
1775 * inverse relationship between the degree of sharing and
1776 * the adjustment made to the scanning period. Broadly
1777 * speaking the intent is that there is little point
1778 * scanning faster if shared accesses dominate as it may
1779 * simply bounce migrations uselessly
1781 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1782 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1785 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1786 task_scan_min(p), task_scan_max(p));
1787 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1791 * Get the fraction of time the task has been running since the last
1792 * NUMA placement cycle. The scheduler keeps similar statistics, but
1793 * decays those on a 32ms period, which is orders of magnitude off
1794 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1795 * stats only if the task is so new there are no NUMA statistics yet.
1797 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1799 u64 runtime, delta, now;
1800 /* Use the start of this time slice to avoid calculations. */
1801 now = p->se.exec_start;
1802 runtime = p->se.sum_exec_runtime;
1804 if (p->last_task_numa_placement) {
1805 delta = runtime - p->last_sum_exec_runtime;
1806 *period = now - p->last_task_numa_placement;
1808 delta = p->se.avg.load_sum / p->se.load.weight;
1809 *period = LOAD_AVG_MAX;
1812 p->last_sum_exec_runtime = runtime;
1813 p->last_task_numa_placement = now;
1819 * Determine the preferred nid for a task in a numa_group. This needs to
1820 * be done in a way that produces consistent results with group_weight,
1821 * otherwise workloads might not converge.
1823 static int preferred_group_nid(struct task_struct *p, int nid)
1828 /* Direct connections between all NUMA nodes. */
1829 if (sched_numa_topology_type == NUMA_DIRECT)
1833 * On a system with glueless mesh NUMA topology, group_weight
1834 * scores nodes according to the number of NUMA hinting faults on
1835 * both the node itself, and on nearby nodes.
1837 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1838 unsigned long score, max_score = 0;
1839 int node, max_node = nid;
1841 dist = sched_max_numa_distance;
1843 for_each_online_node(node) {
1844 score = group_weight(p, node, dist);
1845 if (score > max_score) {
1854 * Finding the preferred nid in a system with NUMA backplane
1855 * interconnect topology is more involved. The goal is to locate
1856 * tasks from numa_groups near each other in the system, and
1857 * untangle workloads from different sides of the system. This requires
1858 * searching down the hierarchy of node groups, recursively searching
1859 * inside the highest scoring group of nodes. The nodemask tricks
1860 * keep the complexity of the search down.
1862 nodes = node_online_map;
1863 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1864 unsigned long max_faults = 0;
1865 nodemask_t max_group = NODE_MASK_NONE;
1868 /* Are there nodes at this distance from each other? */
1869 if (!find_numa_distance(dist))
1872 for_each_node_mask(a, nodes) {
1873 unsigned long faults = 0;
1874 nodemask_t this_group;
1875 nodes_clear(this_group);
1877 /* Sum group's NUMA faults; includes a==b case. */
1878 for_each_node_mask(b, nodes) {
1879 if (node_distance(a, b) < dist) {
1880 faults += group_faults(p, b);
1881 node_set(b, this_group);
1882 node_clear(b, nodes);
1886 /* Remember the top group. */
1887 if (faults > max_faults) {
1888 max_faults = faults;
1889 max_group = this_group;
1891 * subtle: at the smallest distance there is
1892 * just one node left in each "group", the
1893 * winner is the preferred nid.
1898 /* Next round, evaluate the nodes within max_group. */
1906 static void task_numa_placement(struct task_struct *p)
1908 int seq, nid, max_nid = -1, max_group_nid = -1;
1909 unsigned long max_faults = 0, max_group_faults = 0;
1910 unsigned long fault_types[2] = { 0, 0 };
1911 unsigned long total_faults;
1912 u64 runtime, period;
1913 spinlock_t *group_lock = NULL;
1916 * The p->mm->numa_scan_seq field gets updated without
1917 * exclusive access. Use READ_ONCE() here to ensure
1918 * that the field is read in a single access:
1920 seq = READ_ONCE(p->mm->numa_scan_seq);
1921 if (p->numa_scan_seq == seq)
1923 p->numa_scan_seq = seq;
1924 p->numa_scan_period_max = task_scan_max(p);
1926 total_faults = p->numa_faults_locality[0] +
1927 p->numa_faults_locality[1];
1928 runtime = numa_get_avg_runtime(p, &period);
1930 /* If the task is part of a group prevent parallel updates to group stats */
1931 if (p->numa_group) {
1932 group_lock = &p->numa_group->lock;
1933 spin_lock_irq(group_lock);
1936 /* Find the node with the highest number of faults */
1937 for_each_online_node(nid) {
1938 /* Keep track of the offsets in numa_faults array */
1939 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1940 unsigned long faults = 0, group_faults = 0;
1943 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1944 long diff, f_diff, f_weight;
1946 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1947 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1948 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1949 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1951 /* Decay existing window, copy faults since last scan */
1952 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1953 fault_types[priv] += p->numa_faults[membuf_idx];
1954 p->numa_faults[membuf_idx] = 0;
1957 * Normalize the faults_from, so all tasks in a group
1958 * count according to CPU use, instead of by the raw
1959 * number of faults. Tasks with little runtime have
1960 * little over-all impact on throughput, and thus their
1961 * faults are less important.
1963 f_weight = div64_u64(runtime << 16, period + 1);
1964 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1966 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1967 p->numa_faults[cpubuf_idx] = 0;
1969 p->numa_faults[mem_idx] += diff;
1970 p->numa_faults[cpu_idx] += f_diff;
1971 faults += p->numa_faults[mem_idx];
1972 p->total_numa_faults += diff;
1973 if (p->numa_group) {
1975 * safe because we can only change our own group
1977 * mem_idx represents the offset for a given
1978 * nid and priv in a specific region because it
1979 * is at the beginning of the numa_faults array.
1981 p->numa_group->faults[mem_idx] += diff;
1982 p->numa_group->faults_cpu[mem_idx] += f_diff;
1983 p->numa_group->total_faults += diff;
1984 group_faults += p->numa_group->faults[mem_idx];
1988 if (faults > max_faults) {
1989 max_faults = faults;
1993 if (group_faults > max_group_faults) {
1994 max_group_faults = group_faults;
1995 max_group_nid = nid;
1999 update_task_scan_period(p, fault_types[0], fault_types[1]);
2001 if (p->numa_group) {
2002 numa_group_count_active_nodes(p->numa_group);
2003 spin_unlock_irq(group_lock);
2004 max_nid = preferred_group_nid(p, max_group_nid);
2008 /* Set the new preferred node */
2009 if (max_nid != p->numa_preferred_nid)
2010 sched_setnuma(p, max_nid);
2012 if (task_node(p) != p->numa_preferred_nid)
2013 numa_migrate_preferred(p);
2017 static inline int get_numa_group(struct numa_group *grp)
2019 return atomic_inc_not_zero(&grp->refcount);
2022 static inline void put_numa_group(struct numa_group *grp)
2024 if (atomic_dec_and_test(&grp->refcount))
2025 kfree_rcu(grp, rcu);
2028 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2031 struct numa_group *grp, *my_grp;
2032 struct task_struct *tsk;
2034 int cpu = cpupid_to_cpu(cpupid);
2037 if (unlikely(!p->numa_group)) {
2038 unsigned int size = sizeof(struct numa_group) +
2039 4*nr_node_ids*sizeof(unsigned long);
2041 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2045 atomic_set(&grp->refcount, 1);
2046 grp->active_nodes = 1;
2047 grp->max_faults_cpu = 0;
2048 spin_lock_init(&grp->lock);
2050 /* Second half of the array tracks nids where faults happen */
2051 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2054 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2055 grp->faults[i] = p->numa_faults[i];
2057 grp->total_faults = p->total_numa_faults;
2060 rcu_assign_pointer(p->numa_group, grp);
2064 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2066 if (!cpupid_match_pid(tsk, cpupid))
2069 grp = rcu_dereference(tsk->numa_group);
2073 my_grp = p->numa_group;
2078 * Only join the other group if its bigger; if we're the bigger group,
2079 * the other task will join us.
2081 if (my_grp->nr_tasks > grp->nr_tasks)
2085 * Tie-break on the grp address.
2087 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2090 /* Always join threads in the same process. */
2091 if (tsk->mm == current->mm)
2094 /* Simple filter to avoid false positives due to PID collisions */
2095 if (flags & TNF_SHARED)
2098 /* Update priv based on whether false sharing was detected */
2101 if (join && !get_numa_group(grp))
2109 BUG_ON(irqs_disabled());
2110 double_lock_irq(&my_grp->lock, &grp->lock);
2112 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2113 my_grp->faults[i] -= p->numa_faults[i];
2114 grp->faults[i] += p->numa_faults[i];
2116 my_grp->total_faults -= p->total_numa_faults;
2117 grp->total_faults += p->total_numa_faults;
2122 spin_unlock(&my_grp->lock);
2123 spin_unlock_irq(&grp->lock);
2125 rcu_assign_pointer(p->numa_group, grp);
2127 put_numa_group(my_grp);
2135 void task_numa_free(struct task_struct *p)
2137 struct numa_group *grp = p->numa_group;
2138 void *numa_faults = p->numa_faults;
2139 unsigned long flags;
2143 spin_lock_irqsave(&grp->lock, flags);
2144 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2145 grp->faults[i] -= p->numa_faults[i];
2146 grp->total_faults -= p->total_numa_faults;
2149 spin_unlock_irqrestore(&grp->lock, flags);
2150 RCU_INIT_POINTER(p->numa_group, NULL);
2151 put_numa_group(grp);
2154 p->numa_faults = NULL;
2159 * Got a PROT_NONE fault for a page on @node.
2161 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2163 struct task_struct *p = current;
2164 bool migrated = flags & TNF_MIGRATED;
2165 int cpu_node = task_node(current);
2166 int local = !!(flags & TNF_FAULT_LOCAL);
2167 struct numa_group *ng;
2170 if (!static_branch_likely(&sched_numa_balancing))
2173 /* for example, ksmd faulting in a user's mm */
2177 /* Allocate buffer to track faults on a per-node basis */
2178 if (unlikely(!p->numa_faults)) {
2179 int size = sizeof(*p->numa_faults) *
2180 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2182 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2183 if (!p->numa_faults)
2186 p->total_numa_faults = 0;
2187 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2191 * First accesses are treated as private, otherwise consider accesses
2192 * to be private if the accessing pid has not changed
2194 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2197 priv = cpupid_match_pid(p, last_cpupid);
2198 if (!priv && !(flags & TNF_NO_GROUP))
2199 task_numa_group(p, last_cpupid, flags, &priv);
2203 * If a workload spans multiple NUMA nodes, a shared fault that
2204 * occurs wholly within the set of nodes that the workload is
2205 * actively using should be counted as local. This allows the
2206 * scan rate to slow down when a workload has settled down.
2209 if (!priv && !local && ng && ng->active_nodes > 1 &&
2210 numa_is_active_node(cpu_node, ng) &&
2211 numa_is_active_node(mem_node, ng))
2214 task_numa_placement(p);
2217 * Retry task to preferred node migration periodically, in case it
2218 * case it previously failed, or the scheduler moved us.
2220 if (time_after(jiffies, p->numa_migrate_retry))
2221 numa_migrate_preferred(p);
2224 p->numa_pages_migrated += pages;
2225 if (flags & TNF_MIGRATE_FAIL)
2226 p->numa_faults_locality[2] += pages;
2228 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2229 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2230 p->numa_faults_locality[local] += pages;
2233 static void reset_ptenuma_scan(struct task_struct *p)
2236 * We only did a read acquisition of the mmap sem, so
2237 * p->mm->numa_scan_seq is written to without exclusive access
2238 * and the update is not guaranteed to be atomic. That's not
2239 * much of an issue though, since this is just used for
2240 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2241 * expensive, to avoid any form of compiler optimizations:
2243 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2244 p->mm->numa_scan_offset = 0;
2248 * The expensive part of numa migration is done from task_work context.
2249 * Triggered from task_tick_numa().
2251 void task_numa_work(struct callback_head *work)
2253 unsigned long migrate, next_scan, now = jiffies;
2254 struct task_struct *p = current;
2255 struct mm_struct *mm = p->mm;
2256 u64 runtime = p->se.sum_exec_runtime;
2257 struct vm_area_struct *vma;
2258 unsigned long start, end;
2259 unsigned long nr_pte_updates = 0;
2260 long pages, virtpages;
2262 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2264 work->next = work; /* protect against double add */
2266 * Who cares about NUMA placement when they're dying.
2268 * NOTE: make sure not to dereference p->mm before this check,
2269 * exit_task_work() happens _after_ exit_mm() so we could be called
2270 * without p->mm even though we still had it when we enqueued this
2273 if (p->flags & PF_EXITING)
2276 if (!mm->numa_next_scan) {
2277 mm->numa_next_scan = now +
2278 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2282 * Enforce maximal scan/migration frequency..
2284 migrate = mm->numa_next_scan;
2285 if (time_before(now, migrate))
2288 if (p->numa_scan_period == 0) {
2289 p->numa_scan_period_max = task_scan_max(p);
2290 p->numa_scan_period = task_scan_min(p);
2293 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2294 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2298 * Delay this task enough that another task of this mm will likely win
2299 * the next time around.
2301 p->node_stamp += 2 * TICK_NSEC;
2303 start = mm->numa_scan_offset;
2304 pages = sysctl_numa_balancing_scan_size;
2305 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2306 virtpages = pages * 8; /* Scan up to this much virtual space */
2311 down_read(&mm->mmap_sem);
2312 vma = find_vma(mm, start);
2314 reset_ptenuma_scan(p);
2318 for (; vma; vma = vma->vm_next) {
2319 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2320 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2325 * Shared library pages mapped by multiple processes are not
2326 * migrated as it is expected they are cache replicated. Avoid
2327 * hinting faults in read-only file-backed mappings or the vdso
2328 * as migrating the pages will be of marginal benefit.
2331 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2335 * Skip inaccessible VMAs to avoid any confusion between
2336 * PROT_NONE and NUMA hinting ptes
2338 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2342 start = max(start, vma->vm_start);
2343 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2344 end = min(end, vma->vm_end);
2345 nr_pte_updates = change_prot_numa(vma, start, end);
2348 * Try to scan sysctl_numa_balancing_size worth of
2349 * hpages that have at least one present PTE that
2350 * is not already pte-numa. If the VMA contains
2351 * areas that are unused or already full of prot_numa
2352 * PTEs, scan up to virtpages, to skip through those
2356 pages -= (end - start) >> PAGE_SHIFT;
2357 virtpages -= (end - start) >> PAGE_SHIFT;
2360 if (pages <= 0 || virtpages <= 0)
2364 } while (end != vma->vm_end);
2369 * It is possible to reach the end of the VMA list but the last few
2370 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2371 * would find the !migratable VMA on the next scan but not reset the
2372 * scanner to the start so check it now.
2375 mm->numa_scan_offset = start;
2377 reset_ptenuma_scan(p);
2378 up_read(&mm->mmap_sem);
2381 * Make sure tasks use at least 32x as much time to run other code
2382 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2383 * Usually update_task_scan_period slows down scanning enough; on an
2384 * overloaded system we need to limit overhead on a per task basis.
2386 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2387 u64 diff = p->se.sum_exec_runtime - runtime;
2388 p->node_stamp += 32 * diff;
2393 * Drive the periodic memory faults..
2395 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2397 struct callback_head *work = &curr->numa_work;
2401 * We don't care about NUMA placement if we don't have memory.
2403 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2407 * Using runtime rather than walltime has the dual advantage that
2408 * we (mostly) drive the selection from busy threads and that the
2409 * task needs to have done some actual work before we bother with
2412 now = curr->se.sum_exec_runtime;
2413 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2415 if (now > curr->node_stamp + period) {
2416 if (!curr->node_stamp)
2417 curr->numa_scan_period = task_scan_min(curr);
2418 curr->node_stamp += period;
2420 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2421 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2422 task_work_add(curr, work, true);
2427 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2431 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2435 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2438 #endif /* CONFIG_NUMA_BALANCING */
2441 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2443 update_load_add(&cfs_rq->load, se->load.weight);
2444 if (!parent_entity(se))
2445 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2447 if (entity_is_task(se)) {
2448 struct rq *rq = rq_of(cfs_rq);
2450 account_numa_enqueue(rq, task_of(se));
2451 list_add(&se->group_node, &rq->cfs_tasks);
2454 cfs_rq->nr_running++;
2458 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2460 update_load_sub(&cfs_rq->load, se->load.weight);
2461 if (!parent_entity(se))
2462 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2464 if (entity_is_task(se)) {
2465 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2466 list_del_init(&se->group_node);
2469 cfs_rq->nr_running--;
2472 #ifdef CONFIG_FAIR_GROUP_SCHED
2474 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2476 long tg_weight, load, shares;
2479 * This really should be: cfs_rq->avg.load_avg, but instead we use
2480 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2481 * the shares for small weight interactive tasks.
2483 load = scale_load_down(cfs_rq->load.weight);
2485 tg_weight = atomic_long_read(&tg->load_avg);
2487 /* Ensure tg_weight >= load */
2488 tg_weight -= cfs_rq->tg_load_avg_contrib;
2491 shares = (tg->shares * load);
2493 shares /= tg_weight;
2495 if (shares < MIN_SHARES)
2496 shares = MIN_SHARES;
2497 if (shares > tg->shares)
2498 shares = tg->shares;
2502 # else /* CONFIG_SMP */
2503 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2507 # endif /* CONFIG_SMP */
2509 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2510 unsigned long weight)
2513 /* commit outstanding execution time */
2514 if (cfs_rq->curr == se)
2515 update_curr(cfs_rq);
2516 account_entity_dequeue(cfs_rq, se);
2519 update_load_set(&se->load, weight);
2522 account_entity_enqueue(cfs_rq, se);
2525 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2527 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2529 struct task_group *tg;
2530 struct sched_entity *se;
2534 se = tg->se[cpu_of(rq_of(cfs_rq))];
2535 if (!se || throttled_hierarchy(cfs_rq))
2538 if (likely(se->load.weight == tg->shares))
2541 shares = calc_cfs_shares(cfs_rq, tg);
2543 reweight_entity(cfs_rq_of(se), se, shares);
2545 #else /* CONFIG_FAIR_GROUP_SCHED */
2546 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2549 #endif /* CONFIG_FAIR_GROUP_SCHED */
2552 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2553 static const u32 runnable_avg_yN_inv[] = {
2554 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2555 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2556 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2557 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2558 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2559 0x85aac367, 0x82cd8698,
2563 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2564 * over-estimates when re-combining.
2566 static const u32 runnable_avg_yN_sum[] = {
2567 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2568 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2569 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2573 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2574 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2577 static const u32 __accumulated_sum_N32[] = {
2578 0, 23371, 35056, 40899, 43820, 45281,
2579 46011, 46376, 46559, 46650, 46696, 46719,
2584 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2586 static __always_inline u64 decay_load(u64 val, u64 n)
2588 unsigned int local_n;
2592 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2595 /* after bounds checking we can collapse to 32-bit */
2599 * As y^PERIOD = 1/2, we can combine
2600 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2601 * With a look-up table which covers y^n (n<PERIOD)
2603 * To achieve constant time decay_load.
2605 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2606 val >>= local_n / LOAD_AVG_PERIOD;
2607 local_n %= LOAD_AVG_PERIOD;
2610 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2615 * For updates fully spanning n periods, the contribution to runnable
2616 * average will be: \Sum 1024*y^n
2618 * We can compute this reasonably efficiently by combining:
2619 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2621 static u32 __compute_runnable_contrib(u64 n)
2625 if (likely(n <= LOAD_AVG_PERIOD))
2626 return runnable_avg_yN_sum[n];
2627 else if (unlikely(n >= LOAD_AVG_MAX_N))
2628 return LOAD_AVG_MAX;
2630 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2631 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2632 n %= LOAD_AVG_PERIOD;
2633 contrib = decay_load(contrib, n);
2634 return contrib + runnable_avg_yN_sum[n];
2637 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2640 * We can represent the historical contribution to runnable average as the
2641 * coefficients of a geometric series. To do this we sub-divide our runnable
2642 * history into segments of approximately 1ms (1024us); label the segment that
2643 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2645 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2647 * (now) (~1ms ago) (~2ms ago)
2649 * Let u_i denote the fraction of p_i that the entity was runnable.
2651 * We then designate the fractions u_i as our co-efficients, yielding the
2652 * following representation of historical load:
2653 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2655 * We choose y based on the with of a reasonably scheduling period, fixing:
2658 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2659 * approximately half as much as the contribution to load within the last ms
2662 * When a period "rolls over" and we have new u_0`, multiplying the previous
2663 * sum again by y is sufficient to update:
2664 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2665 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2667 static __always_inline int
2668 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2669 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2671 u64 delta, scaled_delta, periods;
2673 unsigned int delta_w, scaled_delta_w, decayed = 0;
2674 unsigned long scale_freq, scale_cpu;
2676 delta = now - sa->last_update_time;
2678 * This should only happen when time goes backwards, which it
2679 * unfortunately does during sched clock init when we swap over to TSC.
2681 if ((s64)delta < 0) {
2682 sa->last_update_time = now;
2687 * Use 1024ns as the unit of measurement since it's a reasonable
2688 * approximation of 1us and fast to compute.
2693 sa->last_update_time = now;
2695 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2696 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2698 /* delta_w is the amount already accumulated against our next period */
2699 delta_w = sa->period_contrib;
2700 if (delta + delta_w >= 1024) {
2703 /* how much left for next period will start over, we don't know yet */
2704 sa->period_contrib = 0;
2707 * Now that we know we're crossing a period boundary, figure
2708 * out how much from delta we need to complete the current
2709 * period and accrue it.
2711 delta_w = 1024 - delta_w;
2712 scaled_delta_w = cap_scale(delta_w, scale_freq);
2714 sa->load_sum += weight * scaled_delta_w;
2716 cfs_rq->runnable_load_sum +=
2717 weight * scaled_delta_w;
2721 sa->util_sum += scaled_delta_w * scale_cpu;
2725 /* Figure out how many additional periods this update spans */
2726 periods = delta / 1024;
2729 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2731 cfs_rq->runnable_load_sum =
2732 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2734 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2736 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2737 contrib = __compute_runnable_contrib(periods);
2738 contrib = cap_scale(contrib, scale_freq);
2740 sa->load_sum += weight * contrib;
2742 cfs_rq->runnable_load_sum += weight * contrib;
2745 sa->util_sum += contrib * scale_cpu;
2748 /* Remainder of delta accrued against u_0` */
2749 scaled_delta = cap_scale(delta, scale_freq);
2751 sa->load_sum += weight * scaled_delta;
2753 cfs_rq->runnable_load_sum += weight * scaled_delta;
2756 sa->util_sum += scaled_delta * scale_cpu;
2758 sa->period_contrib += delta;
2761 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2763 cfs_rq->runnable_load_avg =
2764 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2766 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2772 #ifdef CONFIG_FAIR_GROUP_SCHED
2774 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2775 * and effective_load (which is not done because it is too costly).
2777 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2779 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2782 * No need to update load_avg for root_task_group as it is not used.
2784 if (cfs_rq->tg == &root_task_group)
2787 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2788 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2789 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2794 * Called within set_task_rq() right before setting a task's cpu. The
2795 * caller only guarantees p->pi_lock is held; no other assumptions,
2796 * including the state of rq->lock, should be made.
2798 void set_task_rq_fair(struct sched_entity *se,
2799 struct cfs_rq *prev, struct cfs_rq *next)
2801 if (!sched_feat(ATTACH_AGE_LOAD))
2805 * We are supposed to update the task to "current" time, then its up to
2806 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2807 * getting what current time is, so simply throw away the out-of-date
2808 * time. This will result in the wakee task is less decayed, but giving
2809 * the wakee more load sounds not bad.
2811 if (se->avg.last_update_time && prev) {
2812 u64 p_last_update_time;
2813 u64 n_last_update_time;
2815 #ifndef CONFIG_64BIT
2816 u64 p_last_update_time_copy;
2817 u64 n_last_update_time_copy;
2820 p_last_update_time_copy = prev->load_last_update_time_copy;
2821 n_last_update_time_copy = next->load_last_update_time_copy;
2825 p_last_update_time = prev->avg.last_update_time;
2826 n_last_update_time = next->avg.last_update_time;
2828 } while (p_last_update_time != p_last_update_time_copy ||
2829 n_last_update_time != n_last_update_time_copy);
2831 p_last_update_time = prev->avg.last_update_time;
2832 n_last_update_time = next->avg.last_update_time;
2834 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2835 &se->avg, 0, 0, NULL);
2836 se->avg.last_update_time = n_last_update_time;
2839 #else /* CONFIG_FAIR_GROUP_SCHED */
2840 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2841 #endif /* CONFIG_FAIR_GROUP_SCHED */
2843 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2845 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2847 struct rq *rq = rq_of(cfs_rq);
2848 int cpu = cpu_of(rq);
2850 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2851 unsigned long max = rq->cpu_capacity_orig;
2854 * There are a few boundary cases this might miss but it should
2855 * get called often enough that that should (hopefully) not be
2856 * a real problem -- added to that it only calls on the local
2857 * CPU, so if we enqueue remotely we'll miss an update, but
2858 * the next tick/schedule should update.
2860 * It will not get called when we go idle, because the idle
2861 * thread is a different class (!fair), nor will the utilization
2862 * number include things like RT tasks.
2864 * As is, the util number is not freq-invariant (we'd have to
2865 * implement arch_scale_freq_capacity() for that).
2869 cpufreq_update_util(rq_clock(rq),
2870 min(cfs_rq->avg.util_avg, max), max);
2875 * Unsigned subtract and clamp on underflow.
2877 * Explicitly do a load-store to ensure the intermediate value never hits
2878 * memory. This allows lockless observations without ever seeing the negative
2881 #define sub_positive(_ptr, _val) do { \
2882 typeof(_ptr) ptr = (_ptr); \
2883 typeof(*ptr) val = (_val); \
2884 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2888 WRITE_ONCE(*ptr, res); \
2891 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2893 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2895 struct sched_avg *sa = &cfs_rq->avg;
2896 int decayed, removed_load = 0, removed_util = 0;
2898 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2899 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2900 sub_positive(&sa->load_avg, r);
2901 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2905 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2906 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2907 sub_positive(&sa->util_avg, r);
2908 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2912 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2913 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2915 #ifndef CONFIG_64BIT
2917 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2920 if (update_freq && (decayed || removed_util))
2921 cfs_rq_util_change(cfs_rq);
2923 return decayed || removed_load;
2926 /* Update task and its cfs_rq load average */
2927 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2929 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2930 u64 now = cfs_rq_clock_task(cfs_rq);
2931 struct rq *rq = rq_of(cfs_rq);
2932 int cpu = cpu_of(rq);
2935 * Track task load average for carrying it to new CPU after migrated, and
2936 * track group sched_entity load average for task_h_load calc in migration
2938 __update_load_avg(now, cpu, &se->avg,
2939 se->on_rq * scale_load_down(se->load.weight),
2940 cfs_rq->curr == se, NULL);
2942 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2943 update_tg_load_avg(cfs_rq, 0);
2946 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2948 if (!sched_feat(ATTACH_AGE_LOAD))
2952 * If we got migrated (either between CPUs or between cgroups) we'll
2953 * have aged the average right before clearing @last_update_time.
2955 if (se->avg.last_update_time) {
2956 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2957 &se->avg, 0, 0, NULL);
2960 * XXX: we could have just aged the entire load away if we've been
2961 * absent from the fair class for too long.
2966 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2967 cfs_rq->avg.load_avg += se->avg.load_avg;
2968 cfs_rq->avg.load_sum += se->avg.load_sum;
2969 cfs_rq->avg.util_avg += se->avg.util_avg;
2970 cfs_rq->avg.util_sum += se->avg.util_sum;
2972 cfs_rq_util_change(cfs_rq);
2975 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2977 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2978 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2979 cfs_rq->curr == se, NULL);
2981 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2982 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2983 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2984 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2986 cfs_rq_util_change(cfs_rq);
2989 /* Add the load generated by se into cfs_rq's load average */
2991 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2993 struct sched_avg *sa = &se->avg;
2994 u64 now = cfs_rq_clock_task(cfs_rq);
2995 int migrated, decayed;
2997 migrated = !sa->last_update_time;
2999 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3000 se->on_rq * scale_load_down(se->load.weight),
3001 cfs_rq->curr == se, NULL);
3004 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3006 cfs_rq->runnable_load_avg += sa->load_avg;
3007 cfs_rq->runnable_load_sum += sa->load_sum;
3010 attach_entity_load_avg(cfs_rq, se);
3012 if (decayed || migrated)
3013 update_tg_load_avg(cfs_rq, 0);
3016 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3018 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3020 update_load_avg(se, 1);
3022 cfs_rq->runnable_load_avg =
3023 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3024 cfs_rq->runnable_load_sum =
3025 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3028 #ifndef CONFIG_64BIT
3029 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3031 u64 last_update_time_copy;
3032 u64 last_update_time;
3035 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3037 last_update_time = cfs_rq->avg.last_update_time;
3038 } while (last_update_time != last_update_time_copy);
3040 return last_update_time;
3043 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3045 return cfs_rq->avg.last_update_time;
3050 * Task first catches up with cfs_rq, and then subtract
3051 * itself from the cfs_rq (task must be off the queue now).
3053 void remove_entity_load_avg(struct sched_entity *se)
3055 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3056 u64 last_update_time;
3059 * Newly created task or never used group entity should not be removed
3060 * from its (source) cfs_rq
3062 if (se->avg.last_update_time == 0)
3065 last_update_time = cfs_rq_last_update_time(cfs_rq);
3067 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3068 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3069 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3072 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3074 return cfs_rq->runnable_load_avg;
3077 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3079 return cfs_rq->avg.load_avg;
3082 static int idle_balance(struct rq *this_rq);
3084 #else /* CONFIG_SMP */
3086 static inline void update_load_avg(struct sched_entity *se, int not_used)
3088 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3089 struct rq *rq = rq_of(cfs_rq);
3091 cpufreq_trigger_update(rq_clock(rq));
3095 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3097 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3098 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3101 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3103 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3105 static inline int idle_balance(struct rq *rq)
3110 #endif /* CONFIG_SMP */
3112 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3114 #ifdef CONFIG_SCHEDSTATS
3115 struct task_struct *tsk = NULL;
3117 if (entity_is_task(se))
3120 if (se->statistics.sleep_start) {
3121 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3126 if (unlikely(delta > se->statistics.sleep_max))
3127 se->statistics.sleep_max = delta;
3129 se->statistics.sleep_start = 0;
3130 se->statistics.sum_sleep_runtime += delta;
3133 account_scheduler_latency(tsk, delta >> 10, 1);
3134 trace_sched_stat_sleep(tsk, delta);
3137 if (se->statistics.block_start) {
3138 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3143 if (unlikely(delta > se->statistics.block_max))
3144 se->statistics.block_max = delta;
3146 se->statistics.block_start = 0;
3147 se->statistics.sum_sleep_runtime += delta;
3150 if (tsk->in_iowait) {
3151 se->statistics.iowait_sum += delta;
3152 se->statistics.iowait_count++;
3153 trace_sched_stat_iowait(tsk, delta);
3156 trace_sched_stat_blocked(tsk, delta);
3159 * Blocking time is in units of nanosecs, so shift by
3160 * 20 to get a milliseconds-range estimation of the
3161 * amount of time that the task spent sleeping:
3163 if (unlikely(prof_on == SLEEP_PROFILING)) {
3164 profile_hits(SLEEP_PROFILING,
3165 (void *)get_wchan(tsk),
3168 account_scheduler_latency(tsk, delta >> 10, 0);
3174 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3176 #ifdef CONFIG_SCHED_DEBUG
3177 s64 d = se->vruntime - cfs_rq->min_vruntime;
3182 if (d > 3*sysctl_sched_latency)
3183 schedstat_inc(cfs_rq, nr_spread_over);
3188 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3190 u64 vruntime = cfs_rq->min_vruntime;
3193 * The 'current' period is already promised to the current tasks,
3194 * however the extra weight of the new task will slow them down a
3195 * little, place the new task so that it fits in the slot that
3196 * stays open at the end.
3198 if (initial && sched_feat(START_DEBIT))
3199 vruntime += sched_vslice(cfs_rq, se);
3201 /* sleeps up to a single latency don't count. */
3203 unsigned long thresh = sysctl_sched_latency;
3206 * Halve their sleep time's effect, to allow
3207 * for a gentler effect of sleepers:
3209 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3215 /* ensure we never gain time by being placed backwards. */
3216 se->vruntime = max_vruntime(se->vruntime, vruntime);
3219 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3221 static inline void check_schedstat_required(void)
3223 #ifdef CONFIG_SCHEDSTATS
3224 if (schedstat_enabled())
3227 /* Force schedstat enabled if a dependent tracepoint is active */
3228 if (trace_sched_stat_wait_enabled() ||
3229 trace_sched_stat_sleep_enabled() ||
3230 trace_sched_stat_iowait_enabled() ||
3231 trace_sched_stat_blocked_enabled() ||
3232 trace_sched_stat_runtime_enabled()) {
3233 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3234 "stat_blocked and stat_runtime require the "
3235 "kernel parameter schedstats=enabled or "
3236 "kernel.sched_schedstats=1\n");
3247 * update_min_vruntime()
3248 * vruntime -= min_vruntime
3252 * update_min_vruntime()
3253 * vruntime += min_vruntime
3255 * this way the vruntime transition between RQs is done when both
3256 * min_vruntime are up-to-date.
3260 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3261 * vruntime -= min_vruntime
3265 * update_min_vruntime()
3266 * vruntime += min_vruntime
3268 * this way we don't have the most up-to-date min_vruntime on the originating
3269 * CPU and an up-to-date min_vruntime on the destination CPU.
3273 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3275 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3276 bool curr = cfs_rq->curr == se;
3279 * If we're the current task, we must renormalise before calling
3283 se->vruntime += cfs_rq->min_vruntime;
3285 update_curr(cfs_rq);
3288 * Otherwise, renormalise after, such that we're placed at the current
3289 * moment in time, instead of some random moment in the past. Being
3290 * placed in the past could significantly boost this task to the
3291 * fairness detriment of existing tasks.
3293 if (renorm && !curr)
3294 se->vruntime += cfs_rq->min_vruntime;
3296 enqueue_entity_load_avg(cfs_rq, se);
3297 account_entity_enqueue(cfs_rq, se);
3298 update_cfs_shares(cfs_rq);
3300 if (flags & ENQUEUE_WAKEUP) {
3301 place_entity(cfs_rq, se, 0);
3302 if (schedstat_enabled())
3303 enqueue_sleeper(cfs_rq, se);
3306 check_schedstat_required();
3307 if (schedstat_enabled()) {
3308 update_stats_enqueue(cfs_rq, se);
3309 check_spread(cfs_rq, se);
3312 __enqueue_entity(cfs_rq, se);
3315 if (cfs_rq->nr_running == 1) {
3316 list_add_leaf_cfs_rq(cfs_rq);
3317 check_enqueue_throttle(cfs_rq);
3321 static void __clear_buddies_last(struct sched_entity *se)
3323 for_each_sched_entity(se) {
3324 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3325 if (cfs_rq->last != se)
3328 cfs_rq->last = NULL;
3332 static void __clear_buddies_next(struct sched_entity *se)
3334 for_each_sched_entity(se) {
3335 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3336 if (cfs_rq->next != se)
3339 cfs_rq->next = NULL;
3343 static void __clear_buddies_skip(struct sched_entity *se)
3345 for_each_sched_entity(se) {
3346 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3347 if (cfs_rq->skip != se)
3350 cfs_rq->skip = NULL;
3354 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3356 if (cfs_rq->last == se)
3357 __clear_buddies_last(se);
3359 if (cfs_rq->next == se)
3360 __clear_buddies_next(se);
3362 if (cfs_rq->skip == se)
3363 __clear_buddies_skip(se);
3366 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3369 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3372 * Update run-time statistics of the 'current'.
3374 update_curr(cfs_rq);
3375 dequeue_entity_load_avg(cfs_rq, se);
3377 if (schedstat_enabled())
3378 update_stats_dequeue(cfs_rq, se, flags);
3380 clear_buddies(cfs_rq, se);
3382 if (se != cfs_rq->curr)
3383 __dequeue_entity(cfs_rq, se);
3385 account_entity_dequeue(cfs_rq, se);
3388 * Normalize the entity after updating the min_vruntime because the
3389 * update can refer to the ->curr item and we need to reflect this
3390 * movement in our normalized position.
3392 if (!(flags & DEQUEUE_SLEEP))
3393 se->vruntime -= cfs_rq->min_vruntime;
3395 /* return excess runtime on last dequeue */
3396 return_cfs_rq_runtime(cfs_rq);
3398 update_min_vruntime(cfs_rq);
3399 update_cfs_shares(cfs_rq);
3403 * Preempt the current task with a newly woken task if needed:
3406 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3408 unsigned long ideal_runtime, delta_exec;
3409 struct sched_entity *se;
3412 ideal_runtime = sched_slice(cfs_rq, curr);
3413 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3414 if (delta_exec > ideal_runtime) {
3415 resched_curr(rq_of(cfs_rq));
3417 * The current task ran long enough, ensure it doesn't get
3418 * re-elected due to buddy favours.
3420 clear_buddies(cfs_rq, curr);
3425 * Ensure that a task that missed wakeup preemption by a
3426 * narrow margin doesn't have to wait for a full slice.
3427 * This also mitigates buddy induced latencies under load.
3429 if (delta_exec < sysctl_sched_min_granularity)
3432 se = __pick_first_entity(cfs_rq);
3433 delta = curr->vruntime - se->vruntime;
3438 if (delta > ideal_runtime)
3439 resched_curr(rq_of(cfs_rq));
3443 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3445 /* 'current' is not kept within the tree. */
3448 * Any task has to be enqueued before it get to execute on
3449 * a CPU. So account for the time it spent waiting on the
3452 if (schedstat_enabled())
3453 update_stats_wait_end(cfs_rq, se);
3454 __dequeue_entity(cfs_rq, se);
3455 update_load_avg(se, 1);
3458 update_stats_curr_start(cfs_rq, se);
3460 #ifdef CONFIG_SCHEDSTATS
3462 * Track our maximum slice length, if the CPU's load is at
3463 * least twice that of our own weight (i.e. dont track it
3464 * when there are only lesser-weight tasks around):
3466 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3467 se->statistics.slice_max = max(se->statistics.slice_max,
3468 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3471 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3475 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3478 * Pick the next process, keeping these things in mind, in this order:
3479 * 1) keep things fair between processes/task groups
3480 * 2) pick the "next" process, since someone really wants that to run
3481 * 3) pick the "last" process, for cache locality
3482 * 4) do not run the "skip" process, if something else is available
3484 static struct sched_entity *
3485 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3487 struct sched_entity *left = __pick_first_entity(cfs_rq);
3488 struct sched_entity *se;
3491 * If curr is set we have to see if its left of the leftmost entity
3492 * still in the tree, provided there was anything in the tree at all.
3494 if (!left || (curr && entity_before(curr, left)))
3497 se = left; /* ideally we run the leftmost entity */
3500 * Avoid running the skip buddy, if running something else can
3501 * be done without getting too unfair.
3503 if (cfs_rq->skip == se) {
3504 struct sched_entity *second;
3507 second = __pick_first_entity(cfs_rq);
3509 second = __pick_next_entity(se);
3510 if (!second || (curr && entity_before(curr, second)))
3514 if (second && wakeup_preempt_entity(second, left) < 1)
3519 * Prefer last buddy, try to return the CPU to a preempted task.
3521 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3525 * Someone really wants this to run. If it's not unfair, run it.
3527 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3530 clear_buddies(cfs_rq, se);
3535 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3537 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3540 * If still on the runqueue then deactivate_task()
3541 * was not called and update_curr() has to be done:
3544 update_curr(cfs_rq);
3546 /* throttle cfs_rqs exceeding runtime */
3547 check_cfs_rq_runtime(cfs_rq);
3549 if (schedstat_enabled()) {
3550 check_spread(cfs_rq, prev);
3552 update_stats_wait_start(cfs_rq, prev);
3556 /* Put 'current' back into the tree. */
3557 __enqueue_entity(cfs_rq, prev);
3558 /* in !on_rq case, update occurred at dequeue */
3559 update_load_avg(prev, 0);
3561 cfs_rq->curr = NULL;
3565 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3568 * Update run-time statistics of the 'current'.
3570 update_curr(cfs_rq);
3573 * Ensure that runnable average is periodically updated.
3575 update_load_avg(curr, 1);
3576 update_cfs_shares(cfs_rq);
3578 #ifdef CONFIG_SCHED_HRTICK
3580 * queued ticks are scheduled to match the slice, so don't bother
3581 * validating it and just reschedule.
3584 resched_curr(rq_of(cfs_rq));
3588 * don't let the period tick interfere with the hrtick preemption
3590 if (!sched_feat(DOUBLE_TICK) &&
3591 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3595 if (cfs_rq->nr_running > 1)
3596 check_preempt_tick(cfs_rq, curr);
3600 /**************************************************
3601 * CFS bandwidth control machinery
3604 #ifdef CONFIG_CFS_BANDWIDTH
3606 #ifdef HAVE_JUMP_LABEL
3607 static struct static_key __cfs_bandwidth_used;
3609 static inline bool cfs_bandwidth_used(void)
3611 return static_key_false(&__cfs_bandwidth_used);
3614 void cfs_bandwidth_usage_inc(void)
3616 static_key_slow_inc(&__cfs_bandwidth_used);
3619 void cfs_bandwidth_usage_dec(void)
3621 static_key_slow_dec(&__cfs_bandwidth_used);
3623 #else /* HAVE_JUMP_LABEL */
3624 static bool cfs_bandwidth_used(void)
3629 void cfs_bandwidth_usage_inc(void) {}
3630 void cfs_bandwidth_usage_dec(void) {}
3631 #endif /* HAVE_JUMP_LABEL */
3634 * default period for cfs group bandwidth.
3635 * default: 0.1s, units: nanoseconds
3637 static inline u64 default_cfs_period(void)
3639 return 100000000ULL;
3642 static inline u64 sched_cfs_bandwidth_slice(void)
3644 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3648 * Replenish runtime according to assigned quota and update expiration time.
3649 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3650 * additional synchronization around rq->lock.
3652 * requires cfs_b->lock
3654 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3658 if (cfs_b->quota == RUNTIME_INF)
3661 now = sched_clock_cpu(smp_processor_id());
3662 cfs_b->runtime = cfs_b->quota;
3663 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3666 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3668 return &tg->cfs_bandwidth;
3671 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3672 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3674 if (unlikely(cfs_rq->throttle_count))
3675 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3677 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3680 /* returns 0 on failure to allocate runtime */
3681 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3683 struct task_group *tg = cfs_rq->tg;
3684 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3685 u64 amount = 0, min_amount, expires;
3687 /* note: this is a positive sum as runtime_remaining <= 0 */
3688 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3690 raw_spin_lock(&cfs_b->lock);
3691 if (cfs_b->quota == RUNTIME_INF)
3692 amount = min_amount;
3694 start_cfs_bandwidth(cfs_b);
3696 if (cfs_b->runtime > 0) {
3697 amount = min(cfs_b->runtime, min_amount);
3698 cfs_b->runtime -= amount;
3702 expires = cfs_b->runtime_expires;
3703 raw_spin_unlock(&cfs_b->lock);
3705 cfs_rq->runtime_remaining += amount;
3707 * we may have advanced our local expiration to account for allowed
3708 * spread between our sched_clock and the one on which runtime was
3711 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3712 cfs_rq->runtime_expires = expires;
3714 return cfs_rq->runtime_remaining > 0;
3718 * Note: This depends on the synchronization provided by sched_clock and the
3719 * fact that rq->clock snapshots this value.
3721 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3723 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3725 /* if the deadline is ahead of our clock, nothing to do */
3726 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3729 if (cfs_rq->runtime_remaining < 0)
3733 * If the local deadline has passed we have to consider the
3734 * possibility that our sched_clock is 'fast' and the global deadline
3735 * has not truly expired.
3737 * Fortunately we can check determine whether this the case by checking
3738 * whether the global deadline has advanced. It is valid to compare
3739 * cfs_b->runtime_expires without any locks since we only care about
3740 * exact equality, so a partial write will still work.
3743 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3744 /* extend local deadline, drift is bounded above by 2 ticks */
3745 cfs_rq->runtime_expires += TICK_NSEC;
3747 /* global deadline is ahead, expiration has passed */
3748 cfs_rq->runtime_remaining = 0;
3752 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3754 /* dock delta_exec before expiring quota (as it could span periods) */
3755 cfs_rq->runtime_remaining -= delta_exec;
3756 expire_cfs_rq_runtime(cfs_rq);
3758 if (likely(cfs_rq->runtime_remaining > 0))
3762 * if we're unable to extend our runtime we resched so that the active
3763 * hierarchy can be throttled
3765 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3766 resched_curr(rq_of(cfs_rq));
3769 static __always_inline
3770 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3772 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3775 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3778 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3780 return cfs_bandwidth_used() && cfs_rq->throttled;
3783 /* check whether cfs_rq, or any parent, is throttled */
3784 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3786 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3790 * Ensure that neither of the group entities corresponding to src_cpu or
3791 * dest_cpu are members of a throttled hierarchy when performing group
3792 * load-balance operations.
3794 static inline int throttled_lb_pair(struct task_group *tg,
3795 int src_cpu, int dest_cpu)
3797 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3799 src_cfs_rq = tg->cfs_rq[src_cpu];
3800 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3802 return throttled_hierarchy(src_cfs_rq) ||
3803 throttled_hierarchy(dest_cfs_rq);
3806 /* updated child weight may affect parent so we have to do this bottom up */
3807 static int tg_unthrottle_up(struct task_group *tg, void *data)
3809 struct rq *rq = data;
3810 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3812 cfs_rq->throttle_count--;
3813 if (!cfs_rq->throttle_count) {
3814 /* adjust cfs_rq_clock_task() */
3815 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3816 cfs_rq->throttled_clock_task;
3822 static int tg_throttle_down(struct task_group *tg, void *data)
3824 struct rq *rq = data;
3825 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3827 /* group is entering throttled state, stop time */
3828 if (!cfs_rq->throttle_count)
3829 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3830 cfs_rq->throttle_count++;
3835 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3837 struct rq *rq = rq_of(cfs_rq);
3838 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3839 struct sched_entity *se;
3840 long task_delta, dequeue = 1;
3843 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3845 /* freeze hierarchy runnable averages while throttled */
3847 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3850 task_delta = cfs_rq->h_nr_running;
3851 for_each_sched_entity(se) {
3852 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3853 /* throttled entity or throttle-on-deactivate */
3858 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3859 qcfs_rq->h_nr_running -= task_delta;
3861 if (qcfs_rq->load.weight)
3866 sub_nr_running(rq, task_delta);
3868 cfs_rq->throttled = 1;
3869 cfs_rq->throttled_clock = rq_clock(rq);
3870 raw_spin_lock(&cfs_b->lock);
3871 empty = list_empty(&cfs_b->throttled_cfs_rq);
3874 * Add to the _head_ of the list, so that an already-started
3875 * distribute_cfs_runtime will not see us
3877 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3880 * If we're the first throttled task, make sure the bandwidth
3884 start_cfs_bandwidth(cfs_b);
3886 raw_spin_unlock(&cfs_b->lock);
3889 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3891 struct rq *rq = rq_of(cfs_rq);
3892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3893 struct sched_entity *se;
3897 se = cfs_rq->tg->se[cpu_of(rq)];
3899 cfs_rq->throttled = 0;
3901 update_rq_clock(rq);
3903 raw_spin_lock(&cfs_b->lock);
3904 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3905 list_del_rcu(&cfs_rq->throttled_list);
3906 raw_spin_unlock(&cfs_b->lock);
3908 /* update hierarchical throttle state */
3909 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3911 if (!cfs_rq->load.weight)
3914 task_delta = cfs_rq->h_nr_running;
3915 for_each_sched_entity(se) {
3919 cfs_rq = cfs_rq_of(se);
3921 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3922 cfs_rq->h_nr_running += task_delta;
3924 if (cfs_rq_throttled(cfs_rq))
3929 add_nr_running(rq, task_delta);
3931 /* determine whether we need to wake up potentially idle cpu */
3932 if (rq->curr == rq->idle && rq->cfs.nr_running)
3936 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3937 u64 remaining, u64 expires)
3939 struct cfs_rq *cfs_rq;
3941 u64 starting_runtime = remaining;
3944 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3946 struct rq *rq = rq_of(cfs_rq);
3948 raw_spin_lock(&rq->lock);
3949 if (!cfs_rq_throttled(cfs_rq))
3952 runtime = -cfs_rq->runtime_remaining + 1;
3953 if (runtime > remaining)
3954 runtime = remaining;
3955 remaining -= runtime;
3957 cfs_rq->runtime_remaining += runtime;
3958 cfs_rq->runtime_expires = expires;
3960 /* we check whether we're throttled above */
3961 if (cfs_rq->runtime_remaining > 0)
3962 unthrottle_cfs_rq(cfs_rq);
3965 raw_spin_unlock(&rq->lock);
3972 return starting_runtime - remaining;
3976 * Responsible for refilling a task_group's bandwidth and unthrottling its
3977 * cfs_rqs as appropriate. If there has been no activity within the last
3978 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3979 * used to track this state.
3981 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3983 u64 runtime, runtime_expires;
3986 /* no need to continue the timer with no bandwidth constraint */
3987 if (cfs_b->quota == RUNTIME_INF)
3988 goto out_deactivate;
3990 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3991 cfs_b->nr_periods += overrun;
3994 * idle depends on !throttled (for the case of a large deficit), and if
3995 * we're going inactive then everything else can be deferred
3997 if (cfs_b->idle && !throttled)
3998 goto out_deactivate;
4000 __refill_cfs_bandwidth_runtime(cfs_b);
4003 /* mark as potentially idle for the upcoming period */
4008 /* account preceding periods in which throttling occurred */
4009 cfs_b->nr_throttled += overrun;
4011 runtime_expires = cfs_b->runtime_expires;
4014 * This check is repeated as we are holding onto the new bandwidth while
4015 * we unthrottle. This can potentially race with an unthrottled group
4016 * trying to acquire new bandwidth from the global pool. This can result
4017 * in us over-using our runtime if it is all used during this loop, but
4018 * only by limited amounts in that extreme case.
4020 while (throttled && cfs_b->runtime > 0) {
4021 runtime = cfs_b->runtime;
4022 raw_spin_unlock(&cfs_b->lock);
4023 /* we can't nest cfs_b->lock while distributing bandwidth */
4024 runtime = distribute_cfs_runtime(cfs_b, runtime,
4026 raw_spin_lock(&cfs_b->lock);
4028 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4030 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4034 * While we are ensured activity in the period following an
4035 * unthrottle, this also covers the case in which the new bandwidth is
4036 * insufficient to cover the existing bandwidth deficit. (Forcing the
4037 * timer to remain active while there are any throttled entities.)
4047 /* a cfs_rq won't donate quota below this amount */
4048 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4049 /* minimum remaining period time to redistribute slack quota */
4050 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4051 /* how long we wait to gather additional slack before distributing */
4052 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4055 * Are we near the end of the current quota period?
4057 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4058 * hrtimer base being cleared by hrtimer_start. In the case of
4059 * migrate_hrtimers, base is never cleared, so we are fine.
4061 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4063 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4066 /* if the call-back is running a quota refresh is already occurring */
4067 if (hrtimer_callback_running(refresh_timer))
4070 /* is a quota refresh about to occur? */
4071 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4072 if (remaining < min_expire)
4078 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4080 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4082 /* if there's a quota refresh soon don't bother with slack */
4083 if (runtime_refresh_within(cfs_b, min_left))
4086 hrtimer_start(&cfs_b->slack_timer,
4087 ns_to_ktime(cfs_bandwidth_slack_period),
4091 /* we know any runtime found here is valid as update_curr() precedes return */
4092 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4094 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4095 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4097 if (slack_runtime <= 0)
4100 raw_spin_lock(&cfs_b->lock);
4101 if (cfs_b->quota != RUNTIME_INF &&
4102 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4103 cfs_b->runtime += slack_runtime;
4105 /* we are under rq->lock, defer unthrottling using a timer */
4106 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4107 !list_empty(&cfs_b->throttled_cfs_rq))
4108 start_cfs_slack_bandwidth(cfs_b);
4110 raw_spin_unlock(&cfs_b->lock);
4112 /* even if it's not valid for return we don't want to try again */
4113 cfs_rq->runtime_remaining -= slack_runtime;
4116 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4118 if (!cfs_bandwidth_used())
4121 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4124 __return_cfs_rq_runtime(cfs_rq);
4128 * This is done with a timer (instead of inline with bandwidth return) since
4129 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4131 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4133 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4136 /* confirm we're still not at a refresh boundary */
4137 raw_spin_lock(&cfs_b->lock);
4138 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4139 raw_spin_unlock(&cfs_b->lock);
4143 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4144 runtime = cfs_b->runtime;
4146 expires = cfs_b->runtime_expires;
4147 raw_spin_unlock(&cfs_b->lock);
4152 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4154 raw_spin_lock(&cfs_b->lock);
4155 if (expires == cfs_b->runtime_expires)
4156 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4157 raw_spin_unlock(&cfs_b->lock);
4161 * When a group wakes up we want to make sure that its quota is not already
4162 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4163 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4165 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4167 if (!cfs_bandwidth_used())
4170 /* Synchronize hierarchical throttle counter: */
4171 if (unlikely(!cfs_rq->throttle_uptodate)) {
4172 struct rq *rq = rq_of(cfs_rq);
4173 struct cfs_rq *pcfs_rq;
4174 struct task_group *tg;
4176 cfs_rq->throttle_uptodate = 1;
4178 /* Get closest up-to-date node, because leaves go first: */
4179 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4180 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4181 if (pcfs_rq->throttle_uptodate)
4185 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4186 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4190 /* an active group must be handled by the update_curr()->put() path */
4191 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4194 /* ensure the group is not already throttled */
4195 if (cfs_rq_throttled(cfs_rq))
4198 /* update runtime allocation */
4199 account_cfs_rq_runtime(cfs_rq, 0);
4200 if (cfs_rq->runtime_remaining <= 0)
4201 throttle_cfs_rq(cfs_rq);
4204 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4205 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4207 if (!cfs_bandwidth_used())
4210 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4214 * it's possible for a throttled entity to be forced into a running
4215 * state (e.g. set_curr_task), in this case we're finished.
4217 if (cfs_rq_throttled(cfs_rq))
4220 throttle_cfs_rq(cfs_rq);
4224 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4226 struct cfs_bandwidth *cfs_b =
4227 container_of(timer, struct cfs_bandwidth, slack_timer);
4229 do_sched_cfs_slack_timer(cfs_b);
4231 return HRTIMER_NORESTART;
4234 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4236 struct cfs_bandwidth *cfs_b =
4237 container_of(timer, struct cfs_bandwidth, period_timer);
4241 raw_spin_lock(&cfs_b->lock);
4243 overrun = hrtimer_forward_now(timer, cfs_b->period);
4247 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4250 cfs_b->period_active = 0;
4251 raw_spin_unlock(&cfs_b->lock);
4253 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4256 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4258 raw_spin_lock_init(&cfs_b->lock);
4260 cfs_b->quota = RUNTIME_INF;
4261 cfs_b->period = ns_to_ktime(default_cfs_period());
4263 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4264 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4265 cfs_b->period_timer.function = sched_cfs_period_timer;
4266 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4267 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4270 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4272 cfs_rq->runtime_enabled = 0;
4273 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4276 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4278 lockdep_assert_held(&cfs_b->lock);
4280 if (!cfs_b->period_active) {
4281 cfs_b->period_active = 1;
4282 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4283 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4287 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4289 /* init_cfs_bandwidth() was not called */
4290 if (!cfs_b->throttled_cfs_rq.next)
4293 hrtimer_cancel(&cfs_b->period_timer);
4294 hrtimer_cancel(&cfs_b->slack_timer);
4297 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4299 struct cfs_rq *cfs_rq;
4301 for_each_leaf_cfs_rq(rq, cfs_rq) {
4302 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4304 raw_spin_lock(&cfs_b->lock);
4305 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4306 raw_spin_unlock(&cfs_b->lock);
4310 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4312 struct cfs_rq *cfs_rq;
4314 for_each_leaf_cfs_rq(rq, cfs_rq) {
4315 if (!cfs_rq->runtime_enabled)
4319 * clock_task is not advancing so we just need to make sure
4320 * there's some valid quota amount
4322 cfs_rq->runtime_remaining = 1;
4324 * Offline rq is schedulable till cpu is completely disabled
4325 * in take_cpu_down(), so we prevent new cfs throttling here.
4327 cfs_rq->runtime_enabled = 0;
4329 if (cfs_rq_throttled(cfs_rq))
4330 unthrottle_cfs_rq(cfs_rq);
4334 #else /* CONFIG_CFS_BANDWIDTH */
4335 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4337 return rq_clock_task(rq_of(cfs_rq));
4340 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4341 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4342 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4343 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4345 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4350 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4355 static inline int throttled_lb_pair(struct task_group *tg,
4356 int src_cpu, int dest_cpu)
4361 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4363 #ifdef CONFIG_FAIR_GROUP_SCHED
4364 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4367 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4371 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4372 static inline void update_runtime_enabled(struct rq *rq) {}
4373 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4375 #endif /* CONFIG_CFS_BANDWIDTH */
4377 /**************************************************
4378 * CFS operations on tasks:
4381 #ifdef CONFIG_SCHED_HRTICK
4382 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4384 struct sched_entity *se = &p->se;
4385 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4387 WARN_ON(task_rq(p) != rq);
4389 if (cfs_rq->nr_running > 1) {
4390 u64 slice = sched_slice(cfs_rq, se);
4391 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4392 s64 delta = slice - ran;
4399 hrtick_start(rq, delta);
4404 * called from enqueue/dequeue and updates the hrtick when the
4405 * current task is from our class and nr_running is low enough
4408 static void hrtick_update(struct rq *rq)
4410 struct task_struct *curr = rq->curr;
4412 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4415 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4416 hrtick_start_fair(rq, curr);
4418 #else /* !CONFIG_SCHED_HRTICK */
4420 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4424 static inline void hrtick_update(struct rq *rq)
4430 * The enqueue_task method is called before nr_running is
4431 * increased. Here we update the fair scheduling stats and
4432 * then put the task into the rbtree:
4435 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4437 struct cfs_rq *cfs_rq;
4438 struct sched_entity *se = &p->se;
4440 for_each_sched_entity(se) {
4443 cfs_rq = cfs_rq_of(se);
4444 enqueue_entity(cfs_rq, se, flags);
4447 * end evaluation on encountering a throttled cfs_rq
4449 * note: in the case of encountering a throttled cfs_rq we will
4450 * post the final h_nr_running increment below.
4452 if (cfs_rq_throttled(cfs_rq))
4454 cfs_rq->h_nr_running++;
4456 flags = ENQUEUE_WAKEUP;
4459 for_each_sched_entity(se) {
4460 cfs_rq = cfs_rq_of(se);
4461 cfs_rq->h_nr_running++;
4463 if (cfs_rq_throttled(cfs_rq))
4466 update_load_avg(se, 1);
4467 update_cfs_shares(cfs_rq);
4471 add_nr_running(rq, 1);
4476 static void set_next_buddy(struct sched_entity *se);
4479 * The dequeue_task method is called before nr_running is
4480 * decreased. We remove the task from the rbtree and
4481 * update the fair scheduling stats:
4483 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4485 struct cfs_rq *cfs_rq;
4486 struct sched_entity *se = &p->se;
4487 int task_sleep = flags & DEQUEUE_SLEEP;
4489 for_each_sched_entity(se) {
4490 cfs_rq = cfs_rq_of(se);
4491 dequeue_entity(cfs_rq, se, flags);
4494 * end evaluation on encountering a throttled cfs_rq
4496 * note: in the case of encountering a throttled cfs_rq we will
4497 * post the final h_nr_running decrement below.
4499 if (cfs_rq_throttled(cfs_rq))
4501 cfs_rq->h_nr_running--;
4503 /* Don't dequeue parent if it has other entities besides us */
4504 if (cfs_rq->load.weight) {
4505 /* Avoid re-evaluating load for this entity: */
4506 se = parent_entity(se);
4508 * Bias pick_next to pick a task from this cfs_rq, as
4509 * p is sleeping when it is within its sched_slice.
4511 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4515 flags |= DEQUEUE_SLEEP;
4518 for_each_sched_entity(se) {
4519 cfs_rq = cfs_rq_of(se);
4520 cfs_rq->h_nr_running--;
4522 if (cfs_rq_throttled(cfs_rq))
4525 update_load_avg(se, 1);
4526 update_cfs_shares(cfs_rq);
4530 sub_nr_running(rq, 1);
4536 #ifdef CONFIG_NO_HZ_COMMON
4538 * per rq 'load' arrray crap; XXX kill this.
4542 * The exact cpuload calculated at every tick would be:
4544 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4546 * If a cpu misses updates for n ticks (as it was idle) and update gets
4547 * called on the n+1-th tick when cpu may be busy, then we have:
4549 * load_n = (1 - 1/2^i)^n * load_0
4550 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4552 * decay_load_missed() below does efficient calculation of
4554 * load' = (1 - 1/2^i)^n * load
4556 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4557 * This allows us to precompute the above in said factors, thereby allowing the
4558 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4559 * fixed_power_int())
4561 * The calculation is approximated on a 128 point scale.
4563 #define DEGRADE_SHIFT 7
4565 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4566 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4567 { 0, 0, 0, 0, 0, 0, 0, 0 },
4568 { 64, 32, 8, 0, 0, 0, 0, 0 },
4569 { 96, 72, 40, 12, 1, 0, 0, 0 },
4570 { 112, 98, 75, 43, 15, 1, 0, 0 },
4571 { 120, 112, 98, 76, 45, 16, 2, 0 }
4575 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4576 * would be when CPU is idle and so we just decay the old load without
4577 * adding any new load.
4579 static unsigned long
4580 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4584 if (!missed_updates)
4587 if (missed_updates >= degrade_zero_ticks[idx])
4591 return load >> missed_updates;
4593 while (missed_updates) {
4594 if (missed_updates % 2)
4595 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4597 missed_updates >>= 1;
4602 #endif /* CONFIG_NO_HZ_COMMON */
4605 * __cpu_load_update - update the rq->cpu_load[] statistics
4606 * @this_rq: The rq to update statistics for
4607 * @this_load: The current load
4608 * @pending_updates: The number of missed updates
4610 * Update rq->cpu_load[] statistics. This function is usually called every
4611 * scheduler tick (TICK_NSEC).
4613 * This function computes a decaying average:
4615 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4617 * Because of NOHZ it might not get called on every tick which gives need for
4618 * the @pending_updates argument.
4620 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4621 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4622 * = A * (A * load[i]_n-2 + B) + B
4623 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4624 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4625 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4626 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4627 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4629 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4630 * any change in load would have resulted in the tick being turned back on.
4632 * For regular NOHZ, this reduces to:
4634 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4636 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4639 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4640 unsigned long pending_updates)
4642 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4645 this_rq->nr_load_updates++;
4647 /* Update our load: */
4648 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4649 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4650 unsigned long old_load, new_load;
4652 /* scale is effectively 1 << i now, and >> i divides by scale */
4654 old_load = this_rq->cpu_load[i];
4655 #ifdef CONFIG_NO_HZ_COMMON
4656 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4657 if (tickless_load) {
4658 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4660 * old_load can never be a negative value because a
4661 * decayed tickless_load cannot be greater than the
4662 * original tickless_load.
4664 old_load += tickless_load;
4667 new_load = this_load;
4669 * Round up the averaging division if load is increasing. This
4670 * prevents us from getting stuck on 9 if the load is 10, for
4673 if (new_load > old_load)
4674 new_load += scale - 1;
4676 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4679 sched_avg_update(this_rq);
4682 /* Used instead of source_load when we know the type == 0 */
4683 static unsigned long weighted_cpuload(const int cpu)
4685 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4688 #ifdef CONFIG_NO_HZ_COMMON
4690 * There is no sane way to deal with nohz on smp when using jiffies because the
4691 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4692 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4694 * Therefore we need to avoid the delta approach from the regular tick when
4695 * possible since that would seriously skew the load calculation. This is why we
4696 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4697 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4698 * loop exit, nohz_idle_balance, nohz full exit...)
4700 * This means we might still be one tick off for nohz periods.
4703 static void cpu_load_update_nohz(struct rq *this_rq,
4704 unsigned long curr_jiffies,
4707 unsigned long pending_updates;
4709 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4710 if (pending_updates) {
4711 this_rq->last_load_update_tick = curr_jiffies;
4713 * In the regular NOHZ case, we were idle, this means load 0.
4714 * In the NOHZ_FULL case, we were non-idle, we should consider
4715 * its weighted load.
4717 cpu_load_update(this_rq, load, pending_updates);
4722 * Called from nohz_idle_balance() to update the load ratings before doing the
4725 static void cpu_load_update_idle(struct rq *this_rq)
4728 * bail if there's load or we're actually up-to-date.
4730 if (weighted_cpuload(cpu_of(this_rq)))
4733 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4737 * Record CPU load on nohz entry so we know the tickless load to account
4738 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4739 * than other cpu_load[idx] but it should be fine as cpu_load readers
4740 * shouldn't rely into synchronized cpu_load[*] updates.
4742 void cpu_load_update_nohz_start(void)
4744 struct rq *this_rq = this_rq();
4747 * This is all lockless but should be fine. If weighted_cpuload changes
4748 * concurrently we'll exit nohz. And cpu_load write can race with
4749 * cpu_load_update_idle() but both updater would be writing the same.
4751 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4755 * Account the tickless load in the end of a nohz frame.
4757 void cpu_load_update_nohz_stop(void)
4759 unsigned long curr_jiffies = READ_ONCE(jiffies);
4760 struct rq *this_rq = this_rq();
4763 if (curr_jiffies == this_rq->last_load_update_tick)
4766 load = weighted_cpuload(cpu_of(this_rq));
4767 raw_spin_lock(&this_rq->lock);
4768 update_rq_clock(this_rq);
4769 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4770 raw_spin_unlock(&this_rq->lock);
4772 #else /* !CONFIG_NO_HZ_COMMON */
4773 static inline void cpu_load_update_nohz(struct rq *this_rq,
4774 unsigned long curr_jiffies,
4775 unsigned long load) { }
4776 #endif /* CONFIG_NO_HZ_COMMON */
4778 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4780 #ifdef CONFIG_NO_HZ_COMMON
4781 /* See the mess around cpu_load_update_nohz(). */
4782 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4784 cpu_load_update(this_rq, load, 1);
4788 * Called from scheduler_tick()
4790 void cpu_load_update_active(struct rq *this_rq)
4792 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4794 if (tick_nohz_tick_stopped())
4795 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4797 cpu_load_update_periodic(this_rq, load);
4801 * Return a low guess at the load of a migration-source cpu weighted
4802 * according to the scheduling class and "nice" value.
4804 * We want to under-estimate the load of migration sources, to
4805 * balance conservatively.
4807 static unsigned long source_load(int cpu, int type)
4809 struct rq *rq = cpu_rq(cpu);
4810 unsigned long total = weighted_cpuload(cpu);
4812 if (type == 0 || !sched_feat(LB_BIAS))
4815 return min(rq->cpu_load[type-1], total);
4819 * Return a high guess at the load of a migration-target cpu weighted
4820 * according to the scheduling class and "nice" value.
4822 static unsigned long target_load(int cpu, int type)
4824 struct rq *rq = cpu_rq(cpu);
4825 unsigned long total = weighted_cpuload(cpu);
4827 if (type == 0 || !sched_feat(LB_BIAS))
4830 return max(rq->cpu_load[type-1], total);
4833 static unsigned long capacity_of(int cpu)
4835 return cpu_rq(cpu)->cpu_capacity;
4838 static unsigned long capacity_orig_of(int cpu)
4840 return cpu_rq(cpu)->cpu_capacity_orig;
4843 static unsigned long cpu_avg_load_per_task(int cpu)
4845 struct rq *rq = cpu_rq(cpu);
4846 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4847 unsigned long load_avg = weighted_cpuload(cpu);
4850 return load_avg / nr_running;
4855 #ifdef CONFIG_FAIR_GROUP_SCHED
4857 * effective_load() calculates the load change as seen from the root_task_group
4859 * Adding load to a group doesn't make a group heavier, but can cause movement
4860 * of group shares between cpus. Assuming the shares were perfectly aligned one
4861 * can calculate the shift in shares.
4863 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4864 * on this @cpu and results in a total addition (subtraction) of @wg to the
4865 * total group weight.
4867 * Given a runqueue weight distribution (rw_i) we can compute a shares
4868 * distribution (s_i) using:
4870 * s_i = rw_i / \Sum rw_j (1)
4872 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4873 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4874 * shares distribution (s_i):
4876 * rw_i = { 2, 4, 1, 0 }
4877 * s_i = { 2/7, 4/7, 1/7, 0 }
4879 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4880 * task used to run on and the CPU the waker is running on), we need to
4881 * compute the effect of waking a task on either CPU and, in case of a sync
4882 * wakeup, compute the effect of the current task going to sleep.
4884 * So for a change of @wl to the local @cpu with an overall group weight change
4885 * of @wl we can compute the new shares distribution (s'_i) using:
4887 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4889 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4890 * differences in waking a task to CPU 0. The additional task changes the
4891 * weight and shares distributions like:
4893 * rw'_i = { 3, 4, 1, 0 }
4894 * s'_i = { 3/8, 4/8, 1/8, 0 }
4896 * We can then compute the difference in effective weight by using:
4898 * dw_i = S * (s'_i - s_i) (3)
4900 * Where 'S' is the group weight as seen by its parent.
4902 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4903 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4904 * 4/7) times the weight of the group.
4906 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4908 struct sched_entity *se = tg->se[cpu];
4910 if (!tg->parent) /* the trivial, non-cgroup case */
4913 for_each_sched_entity(se) {
4914 struct cfs_rq *cfs_rq = se->my_q;
4915 long W, w = cfs_rq_load_avg(cfs_rq);
4920 * W = @wg + \Sum rw_j
4922 W = wg + atomic_long_read(&tg->load_avg);
4924 /* Ensure \Sum rw_j >= rw_i */
4925 W -= cfs_rq->tg_load_avg_contrib;
4934 * wl = S * s'_i; see (2)
4937 wl = (w * (long)tg->shares) / W;
4942 * Per the above, wl is the new se->load.weight value; since
4943 * those are clipped to [MIN_SHARES, ...) do so now. See
4944 * calc_cfs_shares().
4946 if (wl < MIN_SHARES)
4950 * wl = dw_i = S * (s'_i - s_i); see (3)
4952 wl -= se->avg.load_avg;
4955 * Recursively apply this logic to all parent groups to compute
4956 * the final effective load change on the root group. Since
4957 * only the @tg group gets extra weight, all parent groups can
4958 * only redistribute existing shares. @wl is the shift in shares
4959 * resulting from this level per the above.
4968 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4975 static void record_wakee(struct task_struct *p)
4978 * Only decay a single time; tasks that have less then 1 wakeup per
4979 * jiffy will not have built up many flips.
4981 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4982 current->wakee_flips >>= 1;
4983 current->wakee_flip_decay_ts = jiffies;
4986 if (current->last_wakee != p) {
4987 current->last_wakee = p;
4988 current->wakee_flips++;
4993 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4995 * A waker of many should wake a different task than the one last awakened
4996 * at a frequency roughly N times higher than one of its wakees.
4998 * In order to determine whether we should let the load spread vs consolidating
4999 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5000 * partner, and a factor of lls_size higher frequency in the other.
5002 * With both conditions met, we can be relatively sure that the relationship is
5003 * non-monogamous, with partner count exceeding socket size.
5005 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5006 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5009 static int wake_wide(struct task_struct *p)
5011 unsigned int master = current->wakee_flips;
5012 unsigned int slave = p->wakee_flips;
5013 int factor = this_cpu_read(sd_llc_size);
5016 swap(master, slave);
5017 if (slave < factor || master < slave * factor)
5022 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5024 s64 this_load, load;
5025 s64 this_eff_load, prev_eff_load;
5026 int idx, this_cpu, prev_cpu;
5027 struct task_group *tg;
5028 unsigned long weight;
5032 this_cpu = smp_processor_id();
5033 prev_cpu = task_cpu(p);
5034 load = source_load(prev_cpu, idx);
5035 this_load = target_load(this_cpu, idx);
5038 * If sync wakeup then subtract the (maximum possible)
5039 * effect of the currently running task from the load
5040 * of the current CPU:
5043 tg = task_group(current);
5044 weight = current->se.avg.load_avg;
5046 this_load += effective_load(tg, this_cpu, -weight, -weight);
5047 load += effective_load(tg, prev_cpu, 0, -weight);
5051 weight = p->se.avg.load_avg;
5054 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5055 * due to the sync cause above having dropped this_load to 0, we'll
5056 * always have an imbalance, but there's really nothing you can do
5057 * about that, so that's good too.
5059 * Otherwise check if either cpus are near enough in load to allow this
5060 * task to be woken on this_cpu.
5062 this_eff_load = 100;
5063 this_eff_load *= capacity_of(prev_cpu);
5065 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5066 prev_eff_load *= capacity_of(this_cpu);
5068 if (this_load > 0) {
5069 this_eff_load *= this_load +
5070 effective_load(tg, this_cpu, weight, weight);
5072 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5075 balanced = this_eff_load <= prev_eff_load;
5077 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5082 schedstat_inc(sd, ttwu_move_affine);
5083 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5089 * find_idlest_group finds and returns the least busy CPU group within the
5092 static struct sched_group *
5093 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5094 int this_cpu, int sd_flag)
5096 struct sched_group *idlest = NULL, *group = sd->groups;
5097 unsigned long min_load = ULONG_MAX, this_load = 0;
5098 int load_idx = sd->forkexec_idx;
5099 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5101 if (sd_flag & SD_BALANCE_WAKE)
5102 load_idx = sd->wake_idx;
5105 unsigned long load, avg_load;
5109 /* Skip over this group if it has no CPUs allowed */
5110 if (!cpumask_intersects(sched_group_cpus(group),
5111 tsk_cpus_allowed(p)))
5114 local_group = cpumask_test_cpu(this_cpu,
5115 sched_group_cpus(group));
5117 /* Tally up the load of all CPUs in the group */
5120 for_each_cpu(i, sched_group_cpus(group)) {
5121 /* Bias balancing toward cpus of our domain */
5123 load = source_load(i, load_idx);
5125 load = target_load(i, load_idx);
5130 /* Adjust by relative CPU capacity of the group */
5131 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5134 this_load = avg_load;
5135 } else if (avg_load < min_load) {
5136 min_load = avg_load;
5139 } while (group = group->next, group != sd->groups);
5141 if (!idlest || 100*this_load < imbalance*min_load)
5147 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5150 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5152 unsigned long load, min_load = ULONG_MAX;
5153 unsigned int min_exit_latency = UINT_MAX;
5154 u64 latest_idle_timestamp = 0;
5155 int least_loaded_cpu = this_cpu;
5156 int shallowest_idle_cpu = -1;
5159 /* Traverse only the allowed CPUs */
5160 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5162 struct rq *rq = cpu_rq(i);
5163 struct cpuidle_state *idle = idle_get_state(rq);
5164 if (idle && idle->exit_latency < min_exit_latency) {
5166 * We give priority to a CPU whose idle state
5167 * has the smallest exit latency irrespective
5168 * of any idle timestamp.
5170 min_exit_latency = idle->exit_latency;
5171 latest_idle_timestamp = rq->idle_stamp;
5172 shallowest_idle_cpu = i;
5173 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5174 rq->idle_stamp > latest_idle_timestamp) {
5176 * If equal or no active idle state, then
5177 * the most recently idled CPU might have
5180 latest_idle_timestamp = rq->idle_stamp;
5181 shallowest_idle_cpu = i;
5183 } else if (shallowest_idle_cpu == -1) {
5184 load = weighted_cpuload(i);
5185 if (load < min_load || (load == min_load && i == this_cpu)) {
5187 least_loaded_cpu = i;
5192 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5196 * Try and locate an idle CPU in the sched_domain.
5198 static int select_idle_sibling(struct task_struct *p, int target)
5200 struct sched_domain *sd;
5201 struct sched_group *sg;
5202 int i = task_cpu(p);
5204 if (idle_cpu(target))
5208 * If the prevous cpu is cache affine and idle, don't be stupid.
5210 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5214 * Otherwise, iterate the domains and find an eligible idle cpu.
5216 * A completely idle sched group at higher domains is more
5217 * desirable than an idle group at a lower level, because lower
5218 * domains have smaller groups and usually share hardware
5219 * resources which causes tasks to contend on them, e.g. x86
5220 * hyperthread siblings in the lowest domain (SMT) can contend
5221 * on the shared cpu pipeline.
5223 * However, while we prefer idle groups at higher domains
5224 * finding an idle cpu at the lowest domain is still better than
5225 * returning 'target', which we've already established, isn't
5228 sd = rcu_dereference(per_cpu(sd_llc, target));
5229 for_each_lower_domain(sd) {
5232 if (!cpumask_intersects(sched_group_cpus(sg),
5233 tsk_cpus_allowed(p)))
5236 /* Ensure the entire group is idle */
5237 for_each_cpu(i, sched_group_cpus(sg)) {
5238 if (i == target || !idle_cpu(i))
5243 * It doesn't matter which cpu we pick, the
5244 * whole group is idle.
5246 target = cpumask_first_and(sched_group_cpus(sg),
5247 tsk_cpus_allowed(p));
5251 } while (sg != sd->groups);
5258 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5259 * tasks. The unit of the return value must be the one of capacity so we can
5260 * compare the utilization with the capacity of the CPU that is available for
5261 * CFS task (ie cpu_capacity).
5263 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5264 * recent utilization of currently non-runnable tasks on a CPU. It represents
5265 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5266 * capacity_orig is the cpu_capacity available at the highest frequency
5267 * (arch_scale_freq_capacity()).
5268 * The utilization of a CPU converges towards a sum equal to or less than the
5269 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5270 * the running time on this CPU scaled by capacity_curr.
5272 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5273 * higher than capacity_orig because of unfortunate rounding in
5274 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5275 * the average stabilizes with the new running time. We need to check that the
5276 * utilization stays within the range of [0..capacity_orig] and cap it if
5277 * necessary. Without utilization capping, a group could be seen as overloaded
5278 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5279 * available capacity. We allow utilization to overshoot capacity_curr (but not
5280 * capacity_orig) as it useful for predicting the capacity required after task
5281 * migrations (scheduler-driven DVFS).
5283 static int cpu_util(int cpu)
5285 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5286 unsigned long capacity = capacity_orig_of(cpu);
5288 return (util >= capacity) ? capacity : util;
5292 * select_task_rq_fair: Select target runqueue for the waking task in domains
5293 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5294 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5296 * Balances load by selecting the idlest cpu in the idlest group, or under
5297 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5299 * Returns the target cpu number.
5301 * preempt must be disabled.
5304 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5306 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5307 int cpu = smp_processor_id();
5308 int new_cpu = prev_cpu;
5309 int want_affine = 0;
5310 int sync = wake_flags & WF_SYNC;
5312 if (sd_flag & SD_BALANCE_WAKE) {
5314 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5318 for_each_domain(cpu, tmp) {
5319 if (!(tmp->flags & SD_LOAD_BALANCE))
5323 * If both cpu and prev_cpu are part of this domain,
5324 * cpu is a valid SD_WAKE_AFFINE target.
5326 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5327 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5332 if (tmp->flags & sd_flag)
5334 else if (!want_affine)
5339 sd = NULL; /* Prefer wake_affine over balance flags */
5340 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5345 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5346 new_cpu = select_idle_sibling(p, new_cpu);
5349 struct sched_group *group;
5352 if (!(sd->flags & sd_flag)) {
5357 group = find_idlest_group(sd, p, cpu, sd_flag);
5363 new_cpu = find_idlest_cpu(group, p, cpu);
5364 if (new_cpu == -1 || new_cpu == cpu) {
5365 /* Now try balancing at a lower domain level of cpu */
5370 /* Now try balancing at a lower domain level of new_cpu */
5372 weight = sd->span_weight;
5374 for_each_domain(cpu, tmp) {
5375 if (weight <= tmp->span_weight)
5377 if (tmp->flags & sd_flag)
5380 /* while loop will break here if sd == NULL */
5388 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5389 * cfs_rq_of(p) references at time of call are still valid and identify the
5390 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5392 static void migrate_task_rq_fair(struct task_struct *p)
5395 * As blocked tasks retain absolute vruntime the migration needs to
5396 * deal with this by subtracting the old and adding the new
5397 * min_vruntime -- the latter is done by enqueue_entity() when placing
5398 * the task on the new runqueue.
5400 if (p->state == TASK_WAKING) {
5401 struct sched_entity *se = &p->se;
5402 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5405 #ifndef CONFIG_64BIT
5406 u64 min_vruntime_copy;
5409 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5411 min_vruntime = cfs_rq->min_vruntime;
5412 } while (min_vruntime != min_vruntime_copy);
5414 min_vruntime = cfs_rq->min_vruntime;
5417 se->vruntime -= min_vruntime;
5421 * We are supposed to update the task to "current" time, then its up to date
5422 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5423 * what current time is, so simply throw away the out-of-date time. This
5424 * will result in the wakee task is less decayed, but giving the wakee more
5425 * load sounds not bad.
5427 remove_entity_load_avg(&p->se);
5429 /* Tell new CPU we are migrated */
5430 p->se.avg.last_update_time = 0;
5432 /* We have migrated, no longer consider this task hot */
5433 p->se.exec_start = 0;
5436 static void task_dead_fair(struct task_struct *p)
5438 remove_entity_load_avg(&p->se);
5440 #endif /* CONFIG_SMP */
5442 static unsigned long
5443 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5445 unsigned long gran = sysctl_sched_wakeup_granularity;
5448 * Since its curr running now, convert the gran from real-time
5449 * to virtual-time in his units.
5451 * By using 'se' instead of 'curr' we penalize light tasks, so
5452 * they get preempted easier. That is, if 'se' < 'curr' then
5453 * the resulting gran will be larger, therefore penalizing the
5454 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5455 * be smaller, again penalizing the lighter task.
5457 * This is especially important for buddies when the leftmost
5458 * task is higher priority than the buddy.
5460 return calc_delta_fair(gran, se);
5464 * Should 'se' preempt 'curr'.
5478 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5480 s64 gran, vdiff = curr->vruntime - se->vruntime;
5485 gran = wakeup_gran(curr, se);
5492 static void set_last_buddy(struct sched_entity *se)
5494 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5497 for_each_sched_entity(se)
5498 cfs_rq_of(se)->last = se;
5501 static void set_next_buddy(struct sched_entity *se)
5503 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5506 for_each_sched_entity(se)
5507 cfs_rq_of(se)->next = se;
5510 static void set_skip_buddy(struct sched_entity *se)
5512 for_each_sched_entity(se)
5513 cfs_rq_of(se)->skip = se;
5517 * Preempt the current task with a newly woken task if needed:
5519 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5521 struct task_struct *curr = rq->curr;
5522 struct sched_entity *se = &curr->se, *pse = &p->se;
5523 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5524 int scale = cfs_rq->nr_running >= sched_nr_latency;
5525 int next_buddy_marked = 0;
5527 if (unlikely(se == pse))
5531 * This is possible from callers such as attach_tasks(), in which we
5532 * unconditionally check_prempt_curr() after an enqueue (which may have
5533 * lead to a throttle). This both saves work and prevents false
5534 * next-buddy nomination below.
5536 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5539 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5540 set_next_buddy(pse);
5541 next_buddy_marked = 1;
5545 * We can come here with TIF_NEED_RESCHED already set from new task
5548 * Note: this also catches the edge-case of curr being in a throttled
5549 * group (e.g. via set_curr_task), since update_curr() (in the
5550 * enqueue of curr) will have resulted in resched being set. This
5551 * prevents us from potentially nominating it as a false LAST_BUDDY
5554 if (test_tsk_need_resched(curr))
5557 /* Idle tasks are by definition preempted by non-idle tasks. */
5558 if (unlikely(curr->policy == SCHED_IDLE) &&
5559 likely(p->policy != SCHED_IDLE))
5563 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5564 * is driven by the tick):
5566 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5569 find_matching_se(&se, &pse);
5570 update_curr(cfs_rq_of(se));
5572 if (wakeup_preempt_entity(se, pse) == 1) {
5574 * Bias pick_next to pick the sched entity that is
5575 * triggering this preemption.
5577 if (!next_buddy_marked)
5578 set_next_buddy(pse);
5587 * Only set the backward buddy when the current task is still
5588 * on the rq. This can happen when a wakeup gets interleaved
5589 * with schedule on the ->pre_schedule() or idle_balance()
5590 * point, either of which can * drop the rq lock.
5592 * Also, during early boot the idle thread is in the fair class,
5593 * for obvious reasons its a bad idea to schedule back to it.
5595 if (unlikely(!se->on_rq || curr == rq->idle))
5598 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5602 static struct task_struct *
5603 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5605 struct cfs_rq *cfs_rq = &rq->cfs;
5606 struct sched_entity *se;
5607 struct task_struct *p;
5611 #ifdef CONFIG_FAIR_GROUP_SCHED
5612 if (!cfs_rq->nr_running)
5615 if (prev->sched_class != &fair_sched_class)
5619 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5620 * likely that a next task is from the same cgroup as the current.
5622 * Therefore attempt to avoid putting and setting the entire cgroup
5623 * hierarchy, only change the part that actually changes.
5627 struct sched_entity *curr = cfs_rq->curr;
5630 * Since we got here without doing put_prev_entity() we also
5631 * have to consider cfs_rq->curr. If it is still a runnable
5632 * entity, update_curr() will update its vruntime, otherwise
5633 * forget we've ever seen it.
5637 update_curr(cfs_rq);
5642 * This call to check_cfs_rq_runtime() will do the
5643 * throttle and dequeue its entity in the parent(s).
5644 * Therefore the 'simple' nr_running test will indeed
5647 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5651 se = pick_next_entity(cfs_rq, curr);
5652 cfs_rq = group_cfs_rq(se);
5658 * Since we haven't yet done put_prev_entity and if the selected task
5659 * is a different task than we started out with, try and touch the
5660 * least amount of cfs_rqs.
5663 struct sched_entity *pse = &prev->se;
5665 while (!(cfs_rq = is_same_group(se, pse))) {
5666 int se_depth = se->depth;
5667 int pse_depth = pse->depth;
5669 if (se_depth <= pse_depth) {
5670 put_prev_entity(cfs_rq_of(pse), pse);
5671 pse = parent_entity(pse);
5673 if (se_depth >= pse_depth) {
5674 set_next_entity(cfs_rq_of(se), se);
5675 se = parent_entity(se);
5679 put_prev_entity(cfs_rq, pse);
5680 set_next_entity(cfs_rq, se);
5683 if (hrtick_enabled(rq))
5684 hrtick_start_fair(rq, p);
5691 if (!cfs_rq->nr_running)
5694 put_prev_task(rq, prev);
5697 se = pick_next_entity(cfs_rq, NULL);
5698 set_next_entity(cfs_rq, se);
5699 cfs_rq = group_cfs_rq(se);
5704 if (hrtick_enabled(rq))
5705 hrtick_start_fair(rq, p);
5711 * This is OK, because current is on_cpu, which avoids it being picked
5712 * for load-balance and preemption/IRQs are still disabled avoiding
5713 * further scheduler activity on it and we're being very careful to
5714 * re-start the picking loop.
5716 lockdep_unpin_lock(&rq->lock, cookie);
5717 new_tasks = idle_balance(rq);
5718 lockdep_repin_lock(&rq->lock, cookie);
5720 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5721 * possible for any higher priority task to appear. In that case we
5722 * must re-start the pick_next_entity() loop.
5734 * Account for a descheduled task:
5736 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5738 struct sched_entity *se = &prev->se;
5739 struct cfs_rq *cfs_rq;
5741 for_each_sched_entity(se) {
5742 cfs_rq = cfs_rq_of(se);
5743 put_prev_entity(cfs_rq, se);
5748 * sched_yield() is very simple
5750 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5752 static void yield_task_fair(struct rq *rq)
5754 struct task_struct *curr = rq->curr;
5755 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5756 struct sched_entity *se = &curr->se;
5759 * Are we the only task in the tree?
5761 if (unlikely(rq->nr_running == 1))
5764 clear_buddies(cfs_rq, se);
5766 if (curr->policy != SCHED_BATCH) {
5767 update_rq_clock(rq);
5769 * Update run-time statistics of the 'current'.
5771 update_curr(cfs_rq);
5773 * Tell update_rq_clock() that we've just updated,
5774 * so we don't do microscopic update in schedule()
5775 * and double the fastpath cost.
5777 rq_clock_skip_update(rq, true);
5783 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5785 struct sched_entity *se = &p->se;
5787 /* throttled hierarchies are not runnable */
5788 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5791 /* Tell the scheduler that we'd really like pse to run next. */
5794 yield_task_fair(rq);
5800 /**************************************************
5801 * Fair scheduling class load-balancing methods.
5805 * The purpose of load-balancing is to achieve the same basic fairness the
5806 * per-cpu scheduler provides, namely provide a proportional amount of compute
5807 * time to each task. This is expressed in the following equation:
5809 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5811 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5812 * W_i,0 is defined as:
5814 * W_i,0 = \Sum_j w_i,j (2)
5816 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5817 * is derived from the nice value as per sched_prio_to_weight[].
5819 * The weight average is an exponential decay average of the instantaneous
5822 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5824 * C_i is the compute capacity of cpu i, typically it is the
5825 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5826 * can also include other factors [XXX].
5828 * To achieve this balance we define a measure of imbalance which follows
5829 * directly from (1):
5831 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5833 * We them move tasks around to minimize the imbalance. In the continuous
5834 * function space it is obvious this converges, in the discrete case we get
5835 * a few fun cases generally called infeasible weight scenarios.
5838 * - infeasible weights;
5839 * - local vs global optima in the discrete case. ]
5844 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5845 * for all i,j solution, we create a tree of cpus that follows the hardware
5846 * topology where each level pairs two lower groups (or better). This results
5847 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5848 * tree to only the first of the previous level and we decrease the frequency
5849 * of load-balance at each level inv. proportional to the number of cpus in
5855 * \Sum { --- * --- * 2^i } = O(n) (5)
5857 * `- size of each group
5858 * | | `- number of cpus doing load-balance
5860 * `- sum over all levels
5862 * Coupled with a limit on how many tasks we can migrate every balance pass,
5863 * this makes (5) the runtime complexity of the balancer.
5865 * An important property here is that each CPU is still (indirectly) connected
5866 * to every other cpu in at most O(log n) steps:
5868 * The adjacency matrix of the resulting graph is given by:
5871 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5874 * And you'll find that:
5876 * A^(log_2 n)_i,j != 0 for all i,j (7)
5878 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5879 * The task movement gives a factor of O(m), giving a convergence complexity
5882 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5887 * In order to avoid CPUs going idle while there's still work to do, new idle
5888 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5889 * tree itself instead of relying on other CPUs to bring it work.
5891 * This adds some complexity to both (5) and (8) but it reduces the total idle
5899 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5902 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5907 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5909 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5911 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5914 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5915 * rewrite all of this once again.]
5918 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5920 enum fbq_type { regular, remote, all };
5922 #define LBF_ALL_PINNED 0x01
5923 #define LBF_NEED_BREAK 0x02
5924 #define LBF_DST_PINNED 0x04
5925 #define LBF_SOME_PINNED 0x08
5928 struct sched_domain *sd;
5936 struct cpumask *dst_grpmask;
5938 enum cpu_idle_type idle;
5940 /* The set of CPUs under consideration for load-balancing */
5941 struct cpumask *cpus;
5946 unsigned int loop_break;
5947 unsigned int loop_max;
5949 enum fbq_type fbq_type;
5950 struct list_head tasks;
5954 * Is this task likely cache-hot:
5956 static int task_hot(struct task_struct *p, struct lb_env *env)
5960 lockdep_assert_held(&env->src_rq->lock);
5962 if (p->sched_class != &fair_sched_class)
5965 if (unlikely(p->policy == SCHED_IDLE))
5969 * Buddy candidates are cache hot:
5971 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5972 (&p->se == cfs_rq_of(&p->se)->next ||
5973 &p->se == cfs_rq_of(&p->se)->last))
5976 if (sysctl_sched_migration_cost == -1)
5978 if (sysctl_sched_migration_cost == 0)
5981 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5983 return delta < (s64)sysctl_sched_migration_cost;
5986 #ifdef CONFIG_NUMA_BALANCING
5988 * Returns 1, if task migration degrades locality
5989 * Returns 0, if task migration improves locality i.e migration preferred.
5990 * Returns -1, if task migration is not affected by locality.
5992 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5994 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5995 unsigned long src_faults, dst_faults;
5996 int src_nid, dst_nid;
5998 if (!static_branch_likely(&sched_numa_balancing))
6001 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6004 src_nid = cpu_to_node(env->src_cpu);
6005 dst_nid = cpu_to_node(env->dst_cpu);
6007 if (src_nid == dst_nid)
6010 /* Migrating away from the preferred node is always bad. */
6011 if (src_nid == p->numa_preferred_nid) {
6012 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6018 /* Encourage migration to the preferred node. */
6019 if (dst_nid == p->numa_preferred_nid)
6023 src_faults = group_faults(p, src_nid);
6024 dst_faults = group_faults(p, dst_nid);
6026 src_faults = task_faults(p, src_nid);
6027 dst_faults = task_faults(p, dst_nid);
6030 return dst_faults < src_faults;
6034 static inline int migrate_degrades_locality(struct task_struct *p,
6042 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6045 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6049 lockdep_assert_held(&env->src_rq->lock);
6052 * We do not migrate tasks that are:
6053 * 1) throttled_lb_pair, or
6054 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6055 * 3) running (obviously), or
6056 * 4) are cache-hot on their current CPU.
6058 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6061 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6064 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6066 env->flags |= LBF_SOME_PINNED;
6069 * Remember if this task can be migrated to any other cpu in
6070 * our sched_group. We may want to revisit it if we couldn't
6071 * meet load balance goals by pulling other tasks on src_cpu.
6073 * Also avoid computing new_dst_cpu if we have already computed
6074 * one in current iteration.
6076 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6079 /* Prevent to re-select dst_cpu via env's cpus */
6080 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6081 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6082 env->flags |= LBF_DST_PINNED;
6083 env->new_dst_cpu = cpu;
6091 /* Record that we found atleast one task that could run on dst_cpu */
6092 env->flags &= ~LBF_ALL_PINNED;
6094 if (task_running(env->src_rq, p)) {
6095 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6100 * Aggressive migration if:
6101 * 1) destination numa is preferred
6102 * 2) task is cache cold, or
6103 * 3) too many balance attempts have failed.
6105 tsk_cache_hot = migrate_degrades_locality(p, env);
6106 if (tsk_cache_hot == -1)
6107 tsk_cache_hot = task_hot(p, env);
6109 if (tsk_cache_hot <= 0 ||
6110 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6111 if (tsk_cache_hot == 1) {
6112 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6113 schedstat_inc(p, se.statistics.nr_forced_migrations);
6118 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6123 * detach_task() -- detach the task for the migration specified in env
6125 static void detach_task(struct task_struct *p, struct lb_env *env)
6127 lockdep_assert_held(&env->src_rq->lock);
6129 p->on_rq = TASK_ON_RQ_MIGRATING;
6130 deactivate_task(env->src_rq, p, 0);
6131 set_task_cpu(p, env->dst_cpu);
6135 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6136 * part of active balancing operations within "domain".
6138 * Returns a task if successful and NULL otherwise.
6140 static struct task_struct *detach_one_task(struct lb_env *env)
6142 struct task_struct *p, *n;
6144 lockdep_assert_held(&env->src_rq->lock);
6146 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6147 if (!can_migrate_task(p, env))
6150 detach_task(p, env);
6153 * Right now, this is only the second place where
6154 * lb_gained[env->idle] is updated (other is detach_tasks)
6155 * so we can safely collect stats here rather than
6156 * inside detach_tasks().
6158 schedstat_inc(env->sd, lb_gained[env->idle]);
6164 static const unsigned int sched_nr_migrate_break = 32;
6167 * detach_tasks() -- tries to detach up to imbalance weighted load from
6168 * busiest_rq, as part of a balancing operation within domain "sd".
6170 * Returns number of detached tasks if successful and 0 otherwise.
6172 static int detach_tasks(struct lb_env *env)
6174 struct list_head *tasks = &env->src_rq->cfs_tasks;
6175 struct task_struct *p;
6179 lockdep_assert_held(&env->src_rq->lock);
6181 if (env->imbalance <= 0)
6184 while (!list_empty(tasks)) {
6186 * We don't want to steal all, otherwise we may be treated likewise,
6187 * which could at worst lead to a livelock crash.
6189 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6192 p = list_first_entry(tasks, struct task_struct, se.group_node);
6195 /* We've more or less seen every task there is, call it quits */
6196 if (env->loop > env->loop_max)
6199 /* take a breather every nr_migrate tasks */
6200 if (env->loop > env->loop_break) {
6201 env->loop_break += sched_nr_migrate_break;
6202 env->flags |= LBF_NEED_BREAK;
6206 if (!can_migrate_task(p, env))
6209 load = task_h_load(p);
6211 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6214 if ((load / 2) > env->imbalance)
6217 detach_task(p, env);
6218 list_add(&p->se.group_node, &env->tasks);
6221 env->imbalance -= load;
6223 #ifdef CONFIG_PREEMPT
6225 * NEWIDLE balancing is a source of latency, so preemptible
6226 * kernels will stop after the first task is detached to minimize
6227 * the critical section.
6229 if (env->idle == CPU_NEWLY_IDLE)
6234 * We only want to steal up to the prescribed amount of
6237 if (env->imbalance <= 0)
6242 list_move_tail(&p->se.group_node, tasks);
6246 * Right now, this is one of only two places we collect this stat
6247 * so we can safely collect detach_one_task() stats here rather
6248 * than inside detach_one_task().
6250 schedstat_add(env->sd, lb_gained[env->idle], detached);
6256 * attach_task() -- attach the task detached by detach_task() to its new rq.
6258 static void attach_task(struct rq *rq, struct task_struct *p)
6260 lockdep_assert_held(&rq->lock);
6262 BUG_ON(task_rq(p) != rq);
6263 activate_task(rq, p, 0);
6264 p->on_rq = TASK_ON_RQ_QUEUED;
6265 check_preempt_curr(rq, p, 0);
6269 * attach_one_task() -- attaches the task returned from detach_one_task() to
6272 static void attach_one_task(struct rq *rq, struct task_struct *p)
6274 raw_spin_lock(&rq->lock);
6276 raw_spin_unlock(&rq->lock);
6280 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6283 static void attach_tasks(struct lb_env *env)
6285 struct list_head *tasks = &env->tasks;
6286 struct task_struct *p;
6288 raw_spin_lock(&env->dst_rq->lock);
6290 while (!list_empty(tasks)) {
6291 p = list_first_entry(tasks, struct task_struct, se.group_node);
6292 list_del_init(&p->se.group_node);
6294 attach_task(env->dst_rq, p);
6297 raw_spin_unlock(&env->dst_rq->lock);
6300 #ifdef CONFIG_FAIR_GROUP_SCHED
6301 static void update_blocked_averages(int cpu)
6303 struct rq *rq = cpu_rq(cpu);
6304 struct cfs_rq *cfs_rq;
6305 unsigned long flags;
6307 raw_spin_lock_irqsave(&rq->lock, flags);
6308 update_rq_clock(rq);
6311 * Iterates the task_group tree in a bottom up fashion, see
6312 * list_add_leaf_cfs_rq() for details.
6314 for_each_leaf_cfs_rq(rq, cfs_rq) {
6315 /* throttled entities do not contribute to load */
6316 if (throttled_hierarchy(cfs_rq))
6319 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6320 update_tg_load_avg(cfs_rq, 0);
6322 raw_spin_unlock_irqrestore(&rq->lock, flags);
6326 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6327 * This needs to be done in a top-down fashion because the load of a child
6328 * group is a fraction of its parents load.
6330 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6332 struct rq *rq = rq_of(cfs_rq);
6333 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6334 unsigned long now = jiffies;
6337 if (cfs_rq->last_h_load_update == now)
6340 cfs_rq->h_load_next = NULL;
6341 for_each_sched_entity(se) {
6342 cfs_rq = cfs_rq_of(se);
6343 cfs_rq->h_load_next = se;
6344 if (cfs_rq->last_h_load_update == now)
6349 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6350 cfs_rq->last_h_load_update = now;
6353 while ((se = cfs_rq->h_load_next) != NULL) {
6354 load = cfs_rq->h_load;
6355 load = div64_ul(load * se->avg.load_avg,
6356 cfs_rq_load_avg(cfs_rq) + 1);
6357 cfs_rq = group_cfs_rq(se);
6358 cfs_rq->h_load = load;
6359 cfs_rq->last_h_load_update = now;
6363 static unsigned long task_h_load(struct task_struct *p)
6365 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6367 update_cfs_rq_h_load(cfs_rq);
6368 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6369 cfs_rq_load_avg(cfs_rq) + 1);
6372 static inline void update_blocked_averages(int cpu)
6374 struct rq *rq = cpu_rq(cpu);
6375 struct cfs_rq *cfs_rq = &rq->cfs;
6376 unsigned long flags;
6378 raw_spin_lock_irqsave(&rq->lock, flags);
6379 update_rq_clock(rq);
6380 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6381 raw_spin_unlock_irqrestore(&rq->lock, flags);
6384 static unsigned long task_h_load(struct task_struct *p)
6386 return p->se.avg.load_avg;
6390 /********** Helpers for find_busiest_group ************************/
6399 * sg_lb_stats - stats of a sched_group required for load_balancing
6401 struct sg_lb_stats {
6402 unsigned long avg_load; /*Avg load across the CPUs of the group */
6403 unsigned long group_load; /* Total load over the CPUs of the group */
6404 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6405 unsigned long load_per_task;
6406 unsigned long group_capacity;
6407 unsigned long group_util; /* Total utilization of the group */
6408 unsigned int sum_nr_running; /* Nr tasks running in the group */
6409 unsigned int idle_cpus;
6410 unsigned int group_weight;
6411 enum group_type group_type;
6412 int group_no_capacity;
6413 #ifdef CONFIG_NUMA_BALANCING
6414 unsigned int nr_numa_running;
6415 unsigned int nr_preferred_running;
6420 * sd_lb_stats - Structure to store the statistics of a sched_domain
6421 * during load balancing.
6423 struct sd_lb_stats {
6424 struct sched_group *busiest; /* Busiest group in this sd */
6425 struct sched_group *local; /* Local group in this sd */
6426 unsigned long total_load; /* Total load of all groups in sd */
6427 unsigned long total_capacity; /* Total capacity of all groups in sd */
6428 unsigned long avg_load; /* Average load across all groups in sd */
6430 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6431 struct sg_lb_stats local_stat; /* Statistics of the local group */
6434 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6437 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6438 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6439 * We must however clear busiest_stat::avg_load because
6440 * update_sd_pick_busiest() reads this before assignment.
6442 *sds = (struct sd_lb_stats){
6446 .total_capacity = 0UL,
6449 .sum_nr_running = 0,
6450 .group_type = group_other,
6456 * get_sd_load_idx - Obtain the load index for a given sched domain.
6457 * @sd: The sched_domain whose load_idx is to be obtained.
6458 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6460 * Return: The load index.
6462 static inline int get_sd_load_idx(struct sched_domain *sd,
6463 enum cpu_idle_type idle)
6469 load_idx = sd->busy_idx;
6472 case CPU_NEWLY_IDLE:
6473 load_idx = sd->newidle_idx;
6476 load_idx = sd->idle_idx;
6483 static unsigned long scale_rt_capacity(int cpu)
6485 struct rq *rq = cpu_rq(cpu);
6486 u64 total, used, age_stamp, avg;
6490 * Since we're reading these variables without serialization make sure
6491 * we read them once before doing sanity checks on them.
6493 age_stamp = READ_ONCE(rq->age_stamp);
6494 avg = READ_ONCE(rq->rt_avg);
6495 delta = __rq_clock_broken(rq) - age_stamp;
6497 if (unlikely(delta < 0))
6500 total = sched_avg_period() + delta;
6502 used = div_u64(avg, total);
6504 if (likely(used < SCHED_CAPACITY_SCALE))
6505 return SCHED_CAPACITY_SCALE - used;
6510 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6512 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6513 struct sched_group *sdg = sd->groups;
6515 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6517 capacity *= scale_rt_capacity(cpu);
6518 capacity >>= SCHED_CAPACITY_SHIFT;
6523 cpu_rq(cpu)->cpu_capacity = capacity;
6524 sdg->sgc->capacity = capacity;
6527 void update_group_capacity(struct sched_domain *sd, int cpu)
6529 struct sched_domain *child = sd->child;
6530 struct sched_group *group, *sdg = sd->groups;
6531 unsigned long capacity;
6532 unsigned long interval;
6534 interval = msecs_to_jiffies(sd->balance_interval);
6535 interval = clamp(interval, 1UL, max_load_balance_interval);
6536 sdg->sgc->next_update = jiffies + interval;
6539 update_cpu_capacity(sd, cpu);
6545 if (child->flags & SD_OVERLAP) {
6547 * SD_OVERLAP domains cannot assume that child groups
6548 * span the current group.
6551 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6552 struct sched_group_capacity *sgc;
6553 struct rq *rq = cpu_rq(cpu);
6556 * build_sched_domains() -> init_sched_groups_capacity()
6557 * gets here before we've attached the domains to the
6560 * Use capacity_of(), which is set irrespective of domains
6561 * in update_cpu_capacity().
6563 * This avoids capacity from being 0 and
6564 * causing divide-by-zero issues on boot.
6566 if (unlikely(!rq->sd)) {
6567 capacity += capacity_of(cpu);
6571 sgc = rq->sd->groups->sgc;
6572 capacity += sgc->capacity;
6576 * !SD_OVERLAP domains can assume that child groups
6577 * span the current group.
6580 group = child->groups;
6582 capacity += group->sgc->capacity;
6583 group = group->next;
6584 } while (group != child->groups);
6587 sdg->sgc->capacity = capacity;
6591 * Check whether the capacity of the rq has been noticeably reduced by side
6592 * activity. The imbalance_pct is used for the threshold.
6593 * Return true is the capacity is reduced
6596 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6598 return ((rq->cpu_capacity * sd->imbalance_pct) <
6599 (rq->cpu_capacity_orig * 100));
6603 * Group imbalance indicates (and tries to solve) the problem where balancing
6604 * groups is inadequate due to tsk_cpus_allowed() constraints.
6606 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6607 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6610 * { 0 1 2 3 } { 4 5 6 7 }
6613 * If we were to balance group-wise we'd place two tasks in the first group and
6614 * two tasks in the second group. Clearly this is undesired as it will overload
6615 * cpu 3 and leave one of the cpus in the second group unused.
6617 * The current solution to this issue is detecting the skew in the first group
6618 * by noticing the lower domain failed to reach balance and had difficulty
6619 * moving tasks due to affinity constraints.
6621 * When this is so detected; this group becomes a candidate for busiest; see
6622 * update_sd_pick_busiest(). And calculate_imbalance() and
6623 * find_busiest_group() avoid some of the usual balance conditions to allow it
6624 * to create an effective group imbalance.
6626 * This is a somewhat tricky proposition since the next run might not find the
6627 * group imbalance and decide the groups need to be balanced again. A most
6628 * subtle and fragile situation.
6631 static inline int sg_imbalanced(struct sched_group *group)
6633 return group->sgc->imbalance;
6637 * group_has_capacity returns true if the group has spare capacity that could
6638 * be used by some tasks.
6639 * We consider that a group has spare capacity if the * number of task is
6640 * smaller than the number of CPUs or if the utilization is lower than the
6641 * available capacity for CFS tasks.
6642 * For the latter, we use a threshold to stabilize the state, to take into
6643 * account the variance of the tasks' load and to return true if the available
6644 * capacity in meaningful for the load balancer.
6645 * As an example, an available capacity of 1% can appear but it doesn't make
6646 * any benefit for the load balance.
6649 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6651 if (sgs->sum_nr_running < sgs->group_weight)
6654 if ((sgs->group_capacity * 100) >
6655 (sgs->group_util * env->sd->imbalance_pct))
6662 * group_is_overloaded returns true if the group has more tasks than it can
6664 * group_is_overloaded is not equals to !group_has_capacity because a group
6665 * with the exact right number of tasks, has no more spare capacity but is not
6666 * overloaded so both group_has_capacity and group_is_overloaded return
6670 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6672 if (sgs->sum_nr_running <= sgs->group_weight)
6675 if ((sgs->group_capacity * 100) <
6676 (sgs->group_util * env->sd->imbalance_pct))
6683 group_type group_classify(struct sched_group *group,
6684 struct sg_lb_stats *sgs)
6686 if (sgs->group_no_capacity)
6687 return group_overloaded;
6689 if (sg_imbalanced(group))
6690 return group_imbalanced;
6696 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6697 * @env: The load balancing environment.
6698 * @group: sched_group whose statistics are to be updated.
6699 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6700 * @local_group: Does group contain this_cpu.
6701 * @sgs: variable to hold the statistics for this group.
6702 * @overload: Indicate more than one runnable task for any CPU.
6704 static inline void update_sg_lb_stats(struct lb_env *env,
6705 struct sched_group *group, int load_idx,
6706 int local_group, struct sg_lb_stats *sgs,
6712 memset(sgs, 0, sizeof(*sgs));
6714 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6715 struct rq *rq = cpu_rq(i);
6717 /* Bias balancing toward cpus of our domain */
6719 load = target_load(i, load_idx);
6721 load = source_load(i, load_idx);
6723 sgs->group_load += load;
6724 sgs->group_util += cpu_util(i);
6725 sgs->sum_nr_running += rq->cfs.h_nr_running;
6727 nr_running = rq->nr_running;
6731 #ifdef CONFIG_NUMA_BALANCING
6732 sgs->nr_numa_running += rq->nr_numa_running;
6733 sgs->nr_preferred_running += rq->nr_preferred_running;
6735 sgs->sum_weighted_load += weighted_cpuload(i);
6737 * No need to call idle_cpu() if nr_running is not 0
6739 if (!nr_running && idle_cpu(i))
6743 /* Adjust by relative CPU capacity of the group */
6744 sgs->group_capacity = group->sgc->capacity;
6745 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6747 if (sgs->sum_nr_running)
6748 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6750 sgs->group_weight = group->group_weight;
6752 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6753 sgs->group_type = group_classify(group, sgs);
6757 * update_sd_pick_busiest - return 1 on busiest group
6758 * @env: The load balancing environment.
6759 * @sds: sched_domain statistics
6760 * @sg: sched_group candidate to be checked for being the busiest
6761 * @sgs: sched_group statistics
6763 * Determine if @sg is a busier group than the previously selected
6766 * Return: %true if @sg is a busier group than the previously selected
6767 * busiest group. %false otherwise.
6769 static bool update_sd_pick_busiest(struct lb_env *env,
6770 struct sd_lb_stats *sds,
6771 struct sched_group *sg,
6772 struct sg_lb_stats *sgs)
6774 struct sg_lb_stats *busiest = &sds->busiest_stat;
6776 if (sgs->group_type > busiest->group_type)
6779 if (sgs->group_type < busiest->group_type)
6782 if (sgs->avg_load <= busiest->avg_load)
6785 /* This is the busiest node in its class. */
6786 if (!(env->sd->flags & SD_ASYM_PACKING))
6789 /* No ASYM_PACKING if target cpu is already busy */
6790 if (env->idle == CPU_NOT_IDLE)
6793 * ASYM_PACKING needs to move all the work to the lowest
6794 * numbered CPUs in the group, therefore mark all groups
6795 * higher than ourself as busy.
6797 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6801 /* Prefer to move from highest possible cpu's work */
6802 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6809 #ifdef CONFIG_NUMA_BALANCING
6810 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6812 if (sgs->sum_nr_running > sgs->nr_numa_running)
6814 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6819 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6821 if (rq->nr_running > rq->nr_numa_running)
6823 if (rq->nr_running > rq->nr_preferred_running)
6828 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6833 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6837 #endif /* CONFIG_NUMA_BALANCING */
6840 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6841 * @env: The load balancing environment.
6842 * @sds: variable to hold the statistics for this sched_domain.
6844 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6846 struct sched_domain *child = env->sd->child;
6847 struct sched_group *sg = env->sd->groups;
6848 struct sg_lb_stats tmp_sgs;
6849 int load_idx, prefer_sibling = 0;
6850 bool overload = false;
6852 if (child && child->flags & SD_PREFER_SIBLING)
6855 load_idx = get_sd_load_idx(env->sd, env->idle);
6858 struct sg_lb_stats *sgs = &tmp_sgs;
6861 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6864 sgs = &sds->local_stat;
6866 if (env->idle != CPU_NEWLY_IDLE ||
6867 time_after_eq(jiffies, sg->sgc->next_update))
6868 update_group_capacity(env->sd, env->dst_cpu);
6871 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6878 * In case the child domain prefers tasks go to siblings
6879 * first, lower the sg capacity so that we'll try
6880 * and move all the excess tasks away. We lower the capacity
6881 * of a group only if the local group has the capacity to fit
6882 * these excess tasks. The extra check prevents the case where
6883 * you always pull from the heaviest group when it is already
6884 * under-utilized (possible with a large weight task outweighs
6885 * the tasks on the system).
6887 if (prefer_sibling && sds->local &&
6888 group_has_capacity(env, &sds->local_stat) &&
6889 (sgs->sum_nr_running > 1)) {
6890 sgs->group_no_capacity = 1;
6891 sgs->group_type = group_classify(sg, sgs);
6894 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6896 sds->busiest_stat = *sgs;
6900 /* Now, start updating sd_lb_stats */
6901 sds->total_load += sgs->group_load;
6902 sds->total_capacity += sgs->group_capacity;
6905 } while (sg != env->sd->groups);
6907 if (env->sd->flags & SD_NUMA)
6908 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6910 if (!env->sd->parent) {
6911 /* update overload indicator if we are at root domain */
6912 if (env->dst_rq->rd->overload != overload)
6913 env->dst_rq->rd->overload = overload;
6919 * check_asym_packing - Check to see if the group is packed into the
6922 * This is primarily intended to used at the sibling level. Some
6923 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6924 * case of POWER7, it can move to lower SMT modes only when higher
6925 * threads are idle. When in lower SMT modes, the threads will
6926 * perform better since they share less core resources. Hence when we
6927 * have idle threads, we want them to be the higher ones.
6929 * This packing function is run on idle threads. It checks to see if
6930 * the busiest CPU in this domain (core in the P7 case) has a higher
6931 * CPU number than the packing function is being run on. Here we are
6932 * assuming lower CPU number will be equivalent to lower a SMT thread
6935 * Return: 1 when packing is required and a task should be moved to
6936 * this CPU. The amount of the imbalance is returned in *imbalance.
6938 * @env: The load balancing environment.
6939 * @sds: Statistics of the sched_domain which is to be packed
6941 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6945 if (!(env->sd->flags & SD_ASYM_PACKING))
6948 if (env->idle == CPU_NOT_IDLE)
6954 busiest_cpu = group_first_cpu(sds->busiest);
6955 if (env->dst_cpu > busiest_cpu)
6958 env->imbalance = DIV_ROUND_CLOSEST(
6959 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6960 SCHED_CAPACITY_SCALE);
6966 * fix_small_imbalance - Calculate the minor imbalance that exists
6967 * amongst the groups of a sched_domain, during
6969 * @env: The load balancing environment.
6970 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6973 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6975 unsigned long tmp, capa_now = 0, capa_move = 0;
6976 unsigned int imbn = 2;
6977 unsigned long scaled_busy_load_per_task;
6978 struct sg_lb_stats *local, *busiest;
6980 local = &sds->local_stat;
6981 busiest = &sds->busiest_stat;
6983 if (!local->sum_nr_running)
6984 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6985 else if (busiest->load_per_task > local->load_per_task)
6988 scaled_busy_load_per_task =
6989 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6990 busiest->group_capacity;
6992 if (busiest->avg_load + scaled_busy_load_per_task >=
6993 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6994 env->imbalance = busiest->load_per_task;
6999 * OK, we don't have enough imbalance to justify moving tasks,
7000 * however we may be able to increase total CPU capacity used by
7004 capa_now += busiest->group_capacity *
7005 min(busiest->load_per_task, busiest->avg_load);
7006 capa_now += local->group_capacity *
7007 min(local->load_per_task, local->avg_load);
7008 capa_now /= SCHED_CAPACITY_SCALE;
7010 /* Amount of load we'd subtract */
7011 if (busiest->avg_load > scaled_busy_load_per_task) {
7012 capa_move += busiest->group_capacity *
7013 min(busiest->load_per_task,
7014 busiest->avg_load - scaled_busy_load_per_task);
7017 /* Amount of load we'd add */
7018 if (busiest->avg_load * busiest->group_capacity <
7019 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7020 tmp = (busiest->avg_load * busiest->group_capacity) /
7021 local->group_capacity;
7023 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7024 local->group_capacity;
7026 capa_move += local->group_capacity *
7027 min(local->load_per_task, local->avg_load + tmp);
7028 capa_move /= SCHED_CAPACITY_SCALE;
7030 /* Move if we gain throughput */
7031 if (capa_move > capa_now)
7032 env->imbalance = busiest->load_per_task;
7036 * calculate_imbalance - Calculate the amount of imbalance present within the
7037 * groups of a given sched_domain during load balance.
7038 * @env: load balance environment
7039 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7041 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7043 unsigned long max_pull, load_above_capacity = ~0UL;
7044 struct sg_lb_stats *local, *busiest;
7046 local = &sds->local_stat;
7047 busiest = &sds->busiest_stat;
7049 if (busiest->group_type == group_imbalanced) {
7051 * In the group_imb case we cannot rely on group-wide averages
7052 * to ensure cpu-load equilibrium, look at wider averages. XXX
7054 busiest->load_per_task =
7055 min(busiest->load_per_task, sds->avg_load);
7059 * Avg load of busiest sg can be less and avg load of local sg can
7060 * be greater than avg load across all sgs of sd because avg load
7061 * factors in sg capacity and sgs with smaller group_type are
7062 * skipped when updating the busiest sg:
7064 if (busiest->avg_load <= sds->avg_load ||
7065 local->avg_load >= sds->avg_load) {
7067 return fix_small_imbalance(env, sds);
7071 * If there aren't any idle cpus, avoid creating some.
7073 if (busiest->group_type == group_overloaded &&
7074 local->group_type == group_overloaded) {
7075 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7076 if (load_above_capacity > busiest->group_capacity) {
7077 load_above_capacity -= busiest->group_capacity;
7078 load_above_capacity *= NICE_0_LOAD;
7079 load_above_capacity /= busiest->group_capacity;
7081 load_above_capacity = ~0UL;
7085 * We're trying to get all the cpus to the average_load, so we don't
7086 * want to push ourselves above the average load, nor do we wish to
7087 * reduce the max loaded cpu below the average load. At the same time,
7088 * we also don't want to reduce the group load below the group
7089 * capacity. Thus we look for the minimum possible imbalance.
7091 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7093 /* How much load to actually move to equalise the imbalance */
7094 env->imbalance = min(
7095 max_pull * busiest->group_capacity,
7096 (sds->avg_load - local->avg_load) * local->group_capacity
7097 ) / SCHED_CAPACITY_SCALE;
7100 * if *imbalance is less than the average load per runnable task
7101 * there is no guarantee that any tasks will be moved so we'll have
7102 * a think about bumping its value to force at least one task to be
7105 if (env->imbalance < busiest->load_per_task)
7106 return fix_small_imbalance(env, sds);
7109 /******* find_busiest_group() helpers end here *********************/
7112 * find_busiest_group - Returns the busiest group within the sched_domain
7113 * if there is an imbalance.
7115 * Also calculates the amount of weighted load which should be moved
7116 * to restore balance.
7118 * @env: The load balancing environment.
7120 * Return: - The busiest group if imbalance exists.
7122 static struct sched_group *find_busiest_group(struct lb_env *env)
7124 struct sg_lb_stats *local, *busiest;
7125 struct sd_lb_stats sds;
7127 init_sd_lb_stats(&sds);
7130 * Compute the various statistics relavent for load balancing at
7133 update_sd_lb_stats(env, &sds);
7134 local = &sds.local_stat;
7135 busiest = &sds.busiest_stat;
7137 /* ASYM feature bypasses nice load balance check */
7138 if (check_asym_packing(env, &sds))
7141 /* There is no busy sibling group to pull tasks from */
7142 if (!sds.busiest || busiest->sum_nr_running == 0)
7145 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7146 / sds.total_capacity;
7149 * If the busiest group is imbalanced the below checks don't
7150 * work because they assume all things are equal, which typically
7151 * isn't true due to cpus_allowed constraints and the like.
7153 if (busiest->group_type == group_imbalanced)
7156 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7157 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7158 busiest->group_no_capacity)
7162 * If the local group is busier than the selected busiest group
7163 * don't try and pull any tasks.
7165 if (local->avg_load >= busiest->avg_load)
7169 * Don't pull any tasks if this group is already above the domain
7172 if (local->avg_load >= sds.avg_load)
7175 if (env->idle == CPU_IDLE) {
7177 * This cpu is idle. If the busiest group is not overloaded
7178 * and there is no imbalance between this and busiest group
7179 * wrt idle cpus, it is balanced. The imbalance becomes
7180 * significant if the diff is greater than 1 otherwise we
7181 * might end up to just move the imbalance on another group
7183 if ((busiest->group_type != group_overloaded) &&
7184 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7188 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7189 * imbalance_pct to be conservative.
7191 if (100 * busiest->avg_load <=
7192 env->sd->imbalance_pct * local->avg_load)
7197 /* Looks like there is an imbalance. Compute it */
7198 calculate_imbalance(env, &sds);
7207 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7209 static struct rq *find_busiest_queue(struct lb_env *env,
7210 struct sched_group *group)
7212 struct rq *busiest = NULL, *rq;
7213 unsigned long busiest_load = 0, busiest_capacity = 1;
7216 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7217 unsigned long capacity, wl;
7221 rt = fbq_classify_rq(rq);
7224 * We classify groups/runqueues into three groups:
7225 * - regular: there are !numa tasks
7226 * - remote: there are numa tasks that run on the 'wrong' node
7227 * - all: there is no distinction
7229 * In order to avoid migrating ideally placed numa tasks,
7230 * ignore those when there's better options.
7232 * If we ignore the actual busiest queue to migrate another
7233 * task, the next balance pass can still reduce the busiest
7234 * queue by moving tasks around inside the node.
7236 * If we cannot move enough load due to this classification
7237 * the next pass will adjust the group classification and
7238 * allow migration of more tasks.
7240 * Both cases only affect the total convergence complexity.
7242 if (rt > env->fbq_type)
7245 capacity = capacity_of(i);
7247 wl = weighted_cpuload(i);
7250 * When comparing with imbalance, use weighted_cpuload()
7251 * which is not scaled with the cpu capacity.
7254 if (rq->nr_running == 1 && wl > env->imbalance &&
7255 !check_cpu_capacity(rq, env->sd))
7259 * For the load comparisons with the other cpu's, consider
7260 * the weighted_cpuload() scaled with the cpu capacity, so
7261 * that the load can be moved away from the cpu that is
7262 * potentially running at a lower capacity.
7264 * Thus we're looking for max(wl_i / capacity_i), crosswise
7265 * multiplication to rid ourselves of the division works out
7266 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7267 * our previous maximum.
7269 if (wl * busiest_capacity > busiest_load * capacity) {
7271 busiest_capacity = capacity;
7280 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7281 * so long as it is large enough.
7283 #define MAX_PINNED_INTERVAL 512
7285 /* Working cpumask for load_balance and load_balance_newidle. */
7286 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7288 static int need_active_balance(struct lb_env *env)
7290 struct sched_domain *sd = env->sd;
7292 if (env->idle == CPU_NEWLY_IDLE) {
7295 * ASYM_PACKING needs to force migrate tasks from busy but
7296 * higher numbered CPUs in order to pack all tasks in the
7297 * lowest numbered CPUs.
7299 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7304 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7305 * It's worth migrating the task if the src_cpu's capacity is reduced
7306 * because of other sched_class or IRQs if more capacity stays
7307 * available on dst_cpu.
7309 if ((env->idle != CPU_NOT_IDLE) &&
7310 (env->src_rq->cfs.h_nr_running == 1)) {
7311 if ((check_cpu_capacity(env->src_rq, sd)) &&
7312 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7316 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7319 static int active_load_balance_cpu_stop(void *data);
7321 static int should_we_balance(struct lb_env *env)
7323 struct sched_group *sg = env->sd->groups;
7324 struct cpumask *sg_cpus, *sg_mask;
7325 int cpu, balance_cpu = -1;
7328 * In the newly idle case, we will allow all the cpu's
7329 * to do the newly idle load balance.
7331 if (env->idle == CPU_NEWLY_IDLE)
7334 sg_cpus = sched_group_cpus(sg);
7335 sg_mask = sched_group_mask(sg);
7336 /* Try to find first idle cpu */
7337 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7338 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7345 if (balance_cpu == -1)
7346 balance_cpu = group_balance_cpu(sg);
7349 * First idle cpu or the first cpu(busiest) in this sched group
7350 * is eligible for doing load balancing at this and above domains.
7352 return balance_cpu == env->dst_cpu;
7356 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7357 * tasks if there is an imbalance.
7359 static int load_balance(int this_cpu, struct rq *this_rq,
7360 struct sched_domain *sd, enum cpu_idle_type idle,
7361 int *continue_balancing)
7363 int ld_moved, cur_ld_moved, active_balance = 0;
7364 struct sched_domain *sd_parent = sd->parent;
7365 struct sched_group *group;
7367 unsigned long flags;
7368 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7370 struct lb_env env = {
7372 .dst_cpu = this_cpu,
7374 .dst_grpmask = sched_group_cpus(sd->groups),
7376 .loop_break = sched_nr_migrate_break,
7379 .tasks = LIST_HEAD_INIT(env.tasks),
7383 * For NEWLY_IDLE load_balancing, we don't need to consider
7384 * other cpus in our group
7386 if (idle == CPU_NEWLY_IDLE)
7387 env.dst_grpmask = NULL;
7389 cpumask_copy(cpus, cpu_active_mask);
7391 schedstat_inc(sd, lb_count[idle]);
7394 if (!should_we_balance(&env)) {
7395 *continue_balancing = 0;
7399 group = find_busiest_group(&env);
7401 schedstat_inc(sd, lb_nobusyg[idle]);
7405 busiest = find_busiest_queue(&env, group);
7407 schedstat_inc(sd, lb_nobusyq[idle]);
7411 BUG_ON(busiest == env.dst_rq);
7413 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7415 env.src_cpu = busiest->cpu;
7416 env.src_rq = busiest;
7419 if (busiest->nr_running > 1) {
7421 * Attempt to move tasks. If find_busiest_group has found
7422 * an imbalance but busiest->nr_running <= 1, the group is
7423 * still unbalanced. ld_moved simply stays zero, so it is
7424 * correctly treated as an imbalance.
7426 env.flags |= LBF_ALL_PINNED;
7427 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7430 raw_spin_lock_irqsave(&busiest->lock, flags);
7433 * cur_ld_moved - load moved in current iteration
7434 * ld_moved - cumulative load moved across iterations
7436 cur_ld_moved = detach_tasks(&env);
7439 * We've detached some tasks from busiest_rq. Every
7440 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7441 * unlock busiest->lock, and we are able to be sure
7442 * that nobody can manipulate the tasks in parallel.
7443 * See task_rq_lock() family for the details.
7446 raw_spin_unlock(&busiest->lock);
7450 ld_moved += cur_ld_moved;
7453 local_irq_restore(flags);
7455 if (env.flags & LBF_NEED_BREAK) {
7456 env.flags &= ~LBF_NEED_BREAK;
7461 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7462 * us and move them to an alternate dst_cpu in our sched_group
7463 * where they can run. The upper limit on how many times we
7464 * iterate on same src_cpu is dependent on number of cpus in our
7467 * This changes load balance semantics a bit on who can move
7468 * load to a given_cpu. In addition to the given_cpu itself
7469 * (or a ilb_cpu acting on its behalf where given_cpu is
7470 * nohz-idle), we now have balance_cpu in a position to move
7471 * load to given_cpu. In rare situations, this may cause
7472 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7473 * _independently_ and at _same_ time to move some load to
7474 * given_cpu) causing exceess load to be moved to given_cpu.
7475 * This however should not happen so much in practice and
7476 * moreover subsequent load balance cycles should correct the
7477 * excess load moved.
7479 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7481 /* Prevent to re-select dst_cpu via env's cpus */
7482 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7484 env.dst_rq = cpu_rq(env.new_dst_cpu);
7485 env.dst_cpu = env.new_dst_cpu;
7486 env.flags &= ~LBF_DST_PINNED;
7488 env.loop_break = sched_nr_migrate_break;
7491 * Go back to "more_balance" rather than "redo" since we
7492 * need to continue with same src_cpu.
7498 * We failed to reach balance because of affinity.
7501 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7503 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7504 *group_imbalance = 1;
7507 /* All tasks on this runqueue were pinned by CPU affinity */
7508 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7509 cpumask_clear_cpu(cpu_of(busiest), cpus);
7510 if (!cpumask_empty(cpus)) {
7512 env.loop_break = sched_nr_migrate_break;
7515 goto out_all_pinned;
7520 schedstat_inc(sd, lb_failed[idle]);
7522 * Increment the failure counter only on periodic balance.
7523 * We do not want newidle balance, which can be very
7524 * frequent, pollute the failure counter causing
7525 * excessive cache_hot migrations and active balances.
7527 if (idle != CPU_NEWLY_IDLE)
7528 sd->nr_balance_failed++;
7530 if (need_active_balance(&env)) {
7531 raw_spin_lock_irqsave(&busiest->lock, flags);
7533 /* don't kick the active_load_balance_cpu_stop,
7534 * if the curr task on busiest cpu can't be
7537 if (!cpumask_test_cpu(this_cpu,
7538 tsk_cpus_allowed(busiest->curr))) {
7539 raw_spin_unlock_irqrestore(&busiest->lock,
7541 env.flags |= LBF_ALL_PINNED;
7542 goto out_one_pinned;
7546 * ->active_balance synchronizes accesses to
7547 * ->active_balance_work. Once set, it's cleared
7548 * only after active load balance is finished.
7550 if (!busiest->active_balance) {
7551 busiest->active_balance = 1;
7552 busiest->push_cpu = this_cpu;
7555 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7557 if (active_balance) {
7558 stop_one_cpu_nowait(cpu_of(busiest),
7559 active_load_balance_cpu_stop, busiest,
7560 &busiest->active_balance_work);
7563 /* We've kicked active balancing, force task migration. */
7564 sd->nr_balance_failed = sd->cache_nice_tries+1;
7567 sd->nr_balance_failed = 0;
7569 if (likely(!active_balance)) {
7570 /* We were unbalanced, so reset the balancing interval */
7571 sd->balance_interval = sd->min_interval;
7574 * If we've begun active balancing, start to back off. This
7575 * case may not be covered by the all_pinned logic if there
7576 * is only 1 task on the busy runqueue (because we don't call
7579 if (sd->balance_interval < sd->max_interval)
7580 sd->balance_interval *= 2;
7587 * We reach balance although we may have faced some affinity
7588 * constraints. Clear the imbalance flag if it was set.
7591 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7593 if (*group_imbalance)
7594 *group_imbalance = 0;
7599 * We reach balance because all tasks are pinned at this level so
7600 * we can't migrate them. Let the imbalance flag set so parent level
7601 * can try to migrate them.
7603 schedstat_inc(sd, lb_balanced[idle]);
7605 sd->nr_balance_failed = 0;
7608 /* tune up the balancing interval */
7609 if (((env.flags & LBF_ALL_PINNED) &&
7610 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7611 (sd->balance_interval < sd->max_interval))
7612 sd->balance_interval *= 2;
7619 static inline unsigned long
7620 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7622 unsigned long interval = sd->balance_interval;
7625 interval *= sd->busy_factor;
7627 /* scale ms to jiffies */
7628 interval = msecs_to_jiffies(interval);
7629 interval = clamp(interval, 1UL, max_load_balance_interval);
7635 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7637 unsigned long interval, next;
7639 interval = get_sd_balance_interval(sd, cpu_busy);
7640 next = sd->last_balance + interval;
7642 if (time_after(*next_balance, next))
7643 *next_balance = next;
7647 * idle_balance is called by schedule() if this_cpu is about to become
7648 * idle. Attempts to pull tasks from other CPUs.
7650 static int idle_balance(struct rq *this_rq)
7652 unsigned long next_balance = jiffies + HZ;
7653 int this_cpu = this_rq->cpu;
7654 struct sched_domain *sd;
7655 int pulled_task = 0;
7659 * We must set idle_stamp _before_ calling idle_balance(), such that we
7660 * measure the duration of idle_balance() as idle time.
7662 this_rq->idle_stamp = rq_clock(this_rq);
7664 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7665 !this_rq->rd->overload) {
7667 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7669 update_next_balance(sd, 0, &next_balance);
7675 raw_spin_unlock(&this_rq->lock);
7677 update_blocked_averages(this_cpu);
7679 for_each_domain(this_cpu, sd) {
7680 int continue_balancing = 1;
7681 u64 t0, domain_cost;
7683 if (!(sd->flags & SD_LOAD_BALANCE))
7686 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7687 update_next_balance(sd, 0, &next_balance);
7691 if (sd->flags & SD_BALANCE_NEWIDLE) {
7692 t0 = sched_clock_cpu(this_cpu);
7694 pulled_task = load_balance(this_cpu, this_rq,
7696 &continue_balancing);
7698 domain_cost = sched_clock_cpu(this_cpu) - t0;
7699 if (domain_cost > sd->max_newidle_lb_cost)
7700 sd->max_newidle_lb_cost = domain_cost;
7702 curr_cost += domain_cost;
7705 update_next_balance(sd, 0, &next_balance);
7708 * Stop searching for tasks to pull if there are
7709 * now runnable tasks on this rq.
7711 if (pulled_task || this_rq->nr_running > 0)
7716 raw_spin_lock(&this_rq->lock);
7718 if (curr_cost > this_rq->max_idle_balance_cost)
7719 this_rq->max_idle_balance_cost = curr_cost;
7722 * While browsing the domains, we released the rq lock, a task could
7723 * have been enqueued in the meantime. Since we're not going idle,
7724 * pretend we pulled a task.
7726 if (this_rq->cfs.h_nr_running && !pulled_task)
7730 /* Move the next balance forward */
7731 if (time_after(this_rq->next_balance, next_balance))
7732 this_rq->next_balance = next_balance;
7734 /* Is there a task of a high priority class? */
7735 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7739 this_rq->idle_stamp = 0;
7745 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7746 * running tasks off the busiest CPU onto idle CPUs. It requires at
7747 * least 1 task to be running on each physical CPU where possible, and
7748 * avoids physical / logical imbalances.
7750 static int active_load_balance_cpu_stop(void *data)
7752 struct rq *busiest_rq = data;
7753 int busiest_cpu = cpu_of(busiest_rq);
7754 int target_cpu = busiest_rq->push_cpu;
7755 struct rq *target_rq = cpu_rq(target_cpu);
7756 struct sched_domain *sd;
7757 struct task_struct *p = NULL;
7759 raw_spin_lock_irq(&busiest_rq->lock);
7761 /* make sure the requested cpu hasn't gone down in the meantime */
7762 if (unlikely(busiest_cpu != smp_processor_id() ||
7763 !busiest_rq->active_balance))
7766 /* Is there any task to move? */
7767 if (busiest_rq->nr_running <= 1)
7771 * This condition is "impossible", if it occurs
7772 * we need to fix it. Originally reported by
7773 * Bjorn Helgaas on a 128-cpu setup.
7775 BUG_ON(busiest_rq == target_rq);
7777 /* Search for an sd spanning us and the target CPU. */
7779 for_each_domain(target_cpu, sd) {
7780 if ((sd->flags & SD_LOAD_BALANCE) &&
7781 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7786 struct lb_env env = {
7788 .dst_cpu = target_cpu,
7789 .dst_rq = target_rq,
7790 .src_cpu = busiest_rq->cpu,
7791 .src_rq = busiest_rq,
7795 schedstat_inc(sd, alb_count);
7797 p = detach_one_task(&env);
7799 schedstat_inc(sd, alb_pushed);
7800 /* Active balancing done, reset the failure counter. */
7801 sd->nr_balance_failed = 0;
7803 schedstat_inc(sd, alb_failed);
7808 busiest_rq->active_balance = 0;
7809 raw_spin_unlock(&busiest_rq->lock);
7812 attach_one_task(target_rq, p);
7819 static inline int on_null_domain(struct rq *rq)
7821 return unlikely(!rcu_dereference_sched(rq->sd));
7824 #ifdef CONFIG_NO_HZ_COMMON
7826 * idle load balancing details
7827 * - When one of the busy CPUs notice that there may be an idle rebalancing
7828 * needed, they will kick the idle load balancer, which then does idle
7829 * load balancing for all the idle CPUs.
7832 cpumask_var_t idle_cpus_mask;
7834 unsigned long next_balance; /* in jiffy units */
7835 } nohz ____cacheline_aligned;
7837 static inline int find_new_ilb(void)
7839 int ilb = cpumask_first(nohz.idle_cpus_mask);
7841 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7848 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7849 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7850 * CPU (if there is one).
7852 static void nohz_balancer_kick(void)
7856 nohz.next_balance++;
7858 ilb_cpu = find_new_ilb();
7860 if (ilb_cpu >= nr_cpu_ids)
7863 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7866 * Use smp_send_reschedule() instead of resched_cpu().
7867 * This way we generate a sched IPI on the target cpu which
7868 * is idle. And the softirq performing nohz idle load balance
7869 * will be run before returning from the IPI.
7871 smp_send_reschedule(ilb_cpu);
7875 void nohz_balance_exit_idle(unsigned int cpu)
7877 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7879 * Completely isolated CPUs don't ever set, so we must test.
7881 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7882 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7883 atomic_dec(&nohz.nr_cpus);
7885 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7889 static inline void set_cpu_sd_state_busy(void)
7891 struct sched_domain *sd;
7892 int cpu = smp_processor_id();
7895 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7897 if (!sd || !sd->nohz_idle)
7901 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7906 void set_cpu_sd_state_idle(void)
7908 struct sched_domain *sd;
7909 int cpu = smp_processor_id();
7912 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7914 if (!sd || sd->nohz_idle)
7918 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7924 * This routine will record that the cpu is going idle with tick stopped.
7925 * This info will be used in performing idle load balancing in the future.
7927 void nohz_balance_enter_idle(int cpu)
7930 * If this cpu is going down, then nothing needs to be done.
7932 if (!cpu_active(cpu))
7935 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7939 * If we're a completely isolated CPU, we don't play.
7941 if (on_null_domain(cpu_rq(cpu)))
7944 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7945 atomic_inc(&nohz.nr_cpus);
7946 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7950 static DEFINE_SPINLOCK(balancing);
7953 * Scale the max load_balance interval with the number of CPUs in the system.
7954 * This trades load-balance latency on larger machines for less cross talk.
7956 void update_max_interval(void)
7958 max_load_balance_interval = HZ*num_online_cpus()/10;
7962 * It checks each scheduling domain to see if it is due to be balanced,
7963 * and initiates a balancing operation if so.
7965 * Balancing parameters are set up in init_sched_domains.
7967 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7969 int continue_balancing = 1;
7971 unsigned long interval;
7972 struct sched_domain *sd;
7973 /* Earliest time when we have to do rebalance again */
7974 unsigned long next_balance = jiffies + 60*HZ;
7975 int update_next_balance = 0;
7976 int need_serialize, need_decay = 0;
7979 update_blocked_averages(cpu);
7982 for_each_domain(cpu, sd) {
7984 * Decay the newidle max times here because this is a regular
7985 * visit to all the domains. Decay ~1% per second.
7987 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7988 sd->max_newidle_lb_cost =
7989 (sd->max_newidle_lb_cost * 253) / 256;
7990 sd->next_decay_max_lb_cost = jiffies + HZ;
7993 max_cost += sd->max_newidle_lb_cost;
7995 if (!(sd->flags & SD_LOAD_BALANCE))
7999 * Stop the load balance at this level. There is another
8000 * CPU in our sched group which is doing load balancing more
8003 if (!continue_balancing) {
8009 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8011 need_serialize = sd->flags & SD_SERIALIZE;
8012 if (need_serialize) {
8013 if (!spin_trylock(&balancing))
8017 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8018 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8020 * The LBF_DST_PINNED logic could have changed
8021 * env->dst_cpu, so we can't know our idle
8022 * state even if we migrated tasks. Update it.
8024 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8026 sd->last_balance = jiffies;
8027 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8030 spin_unlock(&balancing);
8032 if (time_after(next_balance, sd->last_balance + interval)) {
8033 next_balance = sd->last_balance + interval;
8034 update_next_balance = 1;
8039 * Ensure the rq-wide value also decays but keep it at a
8040 * reasonable floor to avoid funnies with rq->avg_idle.
8042 rq->max_idle_balance_cost =
8043 max((u64)sysctl_sched_migration_cost, max_cost);
8048 * next_balance will be updated only when there is a need.
8049 * When the cpu is attached to null domain for ex, it will not be
8052 if (likely(update_next_balance)) {
8053 rq->next_balance = next_balance;
8055 #ifdef CONFIG_NO_HZ_COMMON
8057 * If this CPU has been elected to perform the nohz idle
8058 * balance. Other idle CPUs have already rebalanced with
8059 * nohz_idle_balance() and nohz.next_balance has been
8060 * updated accordingly. This CPU is now running the idle load
8061 * balance for itself and we need to update the
8062 * nohz.next_balance accordingly.
8064 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8065 nohz.next_balance = rq->next_balance;
8070 #ifdef CONFIG_NO_HZ_COMMON
8072 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8073 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8075 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8077 int this_cpu = this_rq->cpu;
8080 /* Earliest time when we have to do rebalance again */
8081 unsigned long next_balance = jiffies + 60*HZ;
8082 int update_next_balance = 0;
8084 if (idle != CPU_IDLE ||
8085 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8088 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8089 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8093 * If this cpu gets work to do, stop the load balancing
8094 * work being done for other cpus. Next load
8095 * balancing owner will pick it up.
8100 rq = cpu_rq(balance_cpu);
8103 * If time for next balance is due,
8106 if (time_after_eq(jiffies, rq->next_balance)) {
8107 raw_spin_lock_irq(&rq->lock);
8108 update_rq_clock(rq);
8109 cpu_load_update_idle(rq);
8110 raw_spin_unlock_irq(&rq->lock);
8111 rebalance_domains(rq, CPU_IDLE);
8114 if (time_after(next_balance, rq->next_balance)) {
8115 next_balance = rq->next_balance;
8116 update_next_balance = 1;
8121 * next_balance will be updated only when there is a need.
8122 * When the CPU is attached to null domain for ex, it will not be
8125 if (likely(update_next_balance))
8126 nohz.next_balance = next_balance;
8128 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8132 * Current heuristic for kicking the idle load balancer in the presence
8133 * of an idle cpu in the system.
8134 * - This rq has more than one task.
8135 * - This rq has at least one CFS task and the capacity of the CPU is
8136 * significantly reduced because of RT tasks or IRQs.
8137 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8138 * multiple busy cpu.
8139 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8140 * domain span are idle.
8142 static inline bool nohz_kick_needed(struct rq *rq)
8144 unsigned long now = jiffies;
8145 struct sched_domain *sd;
8146 struct sched_group_capacity *sgc;
8147 int nr_busy, cpu = rq->cpu;
8150 if (unlikely(rq->idle_balance))
8154 * We may be recently in ticked or tickless idle mode. At the first
8155 * busy tick after returning from idle, we will update the busy stats.
8157 set_cpu_sd_state_busy();
8158 nohz_balance_exit_idle(cpu);
8161 * None are in tickless mode and hence no need for NOHZ idle load
8164 if (likely(!atomic_read(&nohz.nr_cpus)))
8167 if (time_before(now, nohz.next_balance))
8170 if (rq->nr_running >= 2)
8174 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8176 sgc = sd->groups->sgc;
8177 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8186 sd = rcu_dereference(rq->sd);
8188 if ((rq->cfs.h_nr_running >= 1) &&
8189 check_cpu_capacity(rq, sd)) {
8195 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8196 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8197 sched_domain_span(sd)) < cpu)) {
8207 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8211 * run_rebalance_domains is triggered when needed from the scheduler tick.
8212 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8214 static void run_rebalance_domains(struct softirq_action *h)
8216 struct rq *this_rq = this_rq();
8217 enum cpu_idle_type idle = this_rq->idle_balance ?
8218 CPU_IDLE : CPU_NOT_IDLE;
8221 * If this cpu has a pending nohz_balance_kick, then do the
8222 * balancing on behalf of the other idle cpus whose ticks are
8223 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8224 * give the idle cpus a chance to load balance. Else we may
8225 * load balance only within the local sched_domain hierarchy
8226 * and abort nohz_idle_balance altogether if we pull some load.
8228 nohz_idle_balance(this_rq, idle);
8229 rebalance_domains(this_rq, idle);
8233 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8235 void trigger_load_balance(struct rq *rq)
8237 /* Don't need to rebalance while attached to NULL domain */
8238 if (unlikely(on_null_domain(rq)))
8241 if (time_after_eq(jiffies, rq->next_balance))
8242 raise_softirq(SCHED_SOFTIRQ);
8243 #ifdef CONFIG_NO_HZ_COMMON
8244 if (nohz_kick_needed(rq))
8245 nohz_balancer_kick();
8249 static void rq_online_fair(struct rq *rq)
8253 update_runtime_enabled(rq);
8256 static void rq_offline_fair(struct rq *rq)
8260 /* Ensure any throttled groups are reachable by pick_next_task */
8261 unthrottle_offline_cfs_rqs(rq);
8264 #endif /* CONFIG_SMP */
8267 * scheduler tick hitting a task of our scheduling class:
8269 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8271 struct cfs_rq *cfs_rq;
8272 struct sched_entity *se = &curr->se;
8274 for_each_sched_entity(se) {
8275 cfs_rq = cfs_rq_of(se);
8276 entity_tick(cfs_rq, se, queued);
8279 if (static_branch_unlikely(&sched_numa_balancing))
8280 task_tick_numa(rq, curr);
8284 * called on fork with the child task as argument from the parent's context
8285 * - child not yet on the tasklist
8286 * - preemption disabled
8288 static void task_fork_fair(struct task_struct *p)
8290 struct cfs_rq *cfs_rq;
8291 struct sched_entity *se = &p->se, *curr;
8292 int this_cpu = smp_processor_id();
8293 struct rq *rq = this_rq();
8294 unsigned long flags;
8296 raw_spin_lock_irqsave(&rq->lock, flags);
8298 update_rq_clock(rq);
8300 cfs_rq = task_cfs_rq(current);
8301 curr = cfs_rq->curr;
8304 * Not only the cpu but also the task_group of the parent might have
8305 * been changed after parent->se.parent,cfs_rq were copied to
8306 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8307 * of child point to valid ones.
8310 __set_task_cpu(p, this_cpu);
8313 update_curr(cfs_rq);
8316 se->vruntime = curr->vruntime;
8317 place_entity(cfs_rq, se, 1);
8319 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8321 * Upon rescheduling, sched_class::put_prev_task() will place
8322 * 'current' within the tree based on its new key value.
8324 swap(curr->vruntime, se->vruntime);
8328 se->vruntime -= cfs_rq->min_vruntime;
8330 raw_spin_unlock_irqrestore(&rq->lock, flags);
8334 * Priority of the task has changed. Check to see if we preempt
8338 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8340 if (!task_on_rq_queued(p))
8344 * Reschedule if we are currently running on this runqueue and
8345 * our priority decreased, or if we are not currently running on
8346 * this runqueue and our priority is higher than the current's
8348 if (rq->curr == p) {
8349 if (p->prio > oldprio)
8352 check_preempt_curr(rq, p, 0);
8355 static inline bool vruntime_normalized(struct task_struct *p)
8357 struct sched_entity *se = &p->se;
8360 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8361 * the dequeue_entity(.flags=0) will already have normalized the
8368 * When !on_rq, vruntime of the task has usually NOT been normalized.
8369 * But there are some cases where it has already been normalized:
8371 * - A forked child which is waiting for being woken up by
8372 * wake_up_new_task().
8373 * - A task which has been woken up by try_to_wake_up() and
8374 * waiting for actually being woken up by sched_ttwu_pending().
8376 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8382 static void detach_task_cfs_rq(struct task_struct *p)
8384 struct sched_entity *se = &p->se;
8385 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8387 if (!vruntime_normalized(p)) {
8389 * Fix up our vruntime so that the current sleep doesn't
8390 * cause 'unlimited' sleep bonus.
8392 place_entity(cfs_rq, se, 0);
8393 se->vruntime -= cfs_rq->min_vruntime;
8396 /* Catch up with the cfs_rq and remove our load when we leave */
8397 detach_entity_load_avg(cfs_rq, se);
8400 static void attach_task_cfs_rq(struct task_struct *p)
8402 struct sched_entity *se = &p->se;
8403 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8405 #ifdef CONFIG_FAIR_GROUP_SCHED
8407 * Since the real-depth could have been changed (only FAIR
8408 * class maintain depth value), reset depth properly.
8410 se->depth = se->parent ? se->parent->depth + 1 : 0;
8413 /* Synchronize task with its cfs_rq */
8414 attach_entity_load_avg(cfs_rq, se);
8416 if (!vruntime_normalized(p))
8417 se->vruntime += cfs_rq->min_vruntime;
8420 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8422 detach_task_cfs_rq(p);
8425 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8427 attach_task_cfs_rq(p);
8429 if (task_on_rq_queued(p)) {
8431 * We were most likely switched from sched_rt, so
8432 * kick off the schedule if running, otherwise just see
8433 * if we can still preempt the current task.
8438 check_preempt_curr(rq, p, 0);
8442 /* Account for a task changing its policy or group.
8444 * This routine is mostly called to set cfs_rq->curr field when a task
8445 * migrates between groups/classes.
8447 static void set_curr_task_fair(struct rq *rq)
8449 struct sched_entity *se = &rq->curr->se;
8451 for_each_sched_entity(se) {
8452 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8454 set_next_entity(cfs_rq, se);
8455 /* ensure bandwidth has been allocated on our new cfs_rq */
8456 account_cfs_rq_runtime(cfs_rq, 0);
8460 void init_cfs_rq(struct cfs_rq *cfs_rq)
8462 cfs_rq->tasks_timeline = RB_ROOT;
8463 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8464 #ifndef CONFIG_64BIT
8465 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8468 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8469 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8473 #ifdef CONFIG_FAIR_GROUP_SCHED
8474 static void task_move_group_fair(struct task_struct *p)
8476 detach_task_cfs_rq(p);
8477 set_task_rq(p, task_cpu(p));
8480 /* Tell se's cfs_rq has been changed -- migrated */
8481 p->se.avg.last_update_time = 0;
8483 attach_task_cfs_rq(p);
8486 void free_fair_sched_group(struct task_group *tg)
8490 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8492 for_each_possible_cpu(i) {
8494 kfree(tg->cfs_rq[i]);
8503 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8505 struct sched_entity *se;
8506 struct cfs_rq *cfs_rq;
8510 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8513 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8517 tg->shares = NICE_0_LOAD;
8519 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8521 for_each_possible_cpu(i) {
8524 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8525 GFP_KERNEL, cpu_to_node(i));
8529 se = kzalloc_node(sizeof(struct sched_entity),
8530 GFP_KERNEL, cpu_to_node(i));
8534 init_cfs_rq(cfs_rq);
8535 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8536 init_entity_runnable_average(se);
8538 raw_spin_lock_irq(&rq->lock);
8539 post_init_entity_util_avg(se);
8540 raw_spin_unlock_irq(&rq->lock);
8551 void unregister_fair_sched_group(struct task_group *tg)
8553 unsigned long flags;
8557 for_each_possible_cpu(cpu) {
8559 remove_entity_load_avg(tg->se[cpu]);
8562 * Only empty task groups can be destroyed; so we can speculatively
8563 * check on_list without danger of it being re-added.
8565 if (!tg->cfs_rq[cpu]->on_list)
8570 raw_spin_lock_irqsave(&rq->lock, flags);
8571 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8572 raw_spin_unlock_irqrestore(&rq->lock, flags);
8576 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8577 struct sched_entity *se, int cpu,
8578 struct sched_entity *parent)
8580 struct rq *rq = cpu_rq(cpu);
8584 init_cfs_rq_runtime(cfs_rq);
8586 tg->cfs_rq[cpu] = cfs_rq;
8589 /* se could be NULL for root_task_group */
8594 se->cfs_rq = &rq->cfs;
8597 se->cfs_rq = parent->my_q;
8598 se->depth = parent->depth + 1;
8602 /* guarantee group entities always have weight */
8603 update_load_set(&se->load, NICE_0_LOAD);
8604 se->parent = parent;
8607 static DEFINE_MUTEX(shares_mutex);
8609 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8612 unsigned long flags;
8615 * We can't change the weight of the root cgroup.
8620 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8622 mutex_lock(&shares_mutex);
8623 if (tg->shares == shares)
8626 tg->shares = shares;
8627 for_each_possible_cpu(i) {
8628 struct rq *rq = cpu_rq(i);
8629 struct sched_entity *se;
8632 /* Propagate contribution to hierarchy */
8633 raw_spin_lock_irqsave(&rq->lock, flags);
8635 /* Possible calls to update_curr() need rq clock */
8636 update_rq_clock(rq);
8637 for_each_sched_entity(se)
8638 update_cfs_shares(group_cfs_rq(se));
8639 raw_spin_unlock_irqrestore(&rq->lock, flags);
8643 mutex_unlock(&shares_mutex);
8646 #else /* CONFIG_FAIR_GROUP_SCHED */
8648 void free_fair_sched_group(struct task_group *tg) { }
8650 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8655 void unregister_fair_sched_group(struct task_group *tg) { }
8657 #endif /* CONFIG_FAIR_GROUP_SCHED */
8660 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8662 struct sched_entity *se = &task->se;
8663 unsigned int rr_interval = 0;
8666 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8669 if (rq->cfs.load.weight)
8670 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8676 * All the scheduling class methods:
8678 const struct sched_class fair_sched_class = {
8679 .next = &idle_sched_class,
8680 .enqueue_task = enqueue_task_fair,
8681 .dequeue_task = dequeue_task_fair,
8682 .yield_task = yield_task_fair,
8683 .yield_to_task = yield_to_task_fair,
8685 .check_preempt_curr = check_preempt_wakeup,
8687 .pick_next_task = pick_next_task_fair,
8688 .put_prev_task = put_prev_task_fair,
8691 .select_task_rq = select_task_rq_fair,
8692 .migrate_task_rq = migrate_task_rq_fair,
8694 .rq_online = rq_online_fair,
8695 .rq_offline = rq_offline_fair,
8697 .task_dead = task_dead_fair,
8698 .set_cpus_allowed = set_cpus_allowed_common,
8701 .set_curr_task = set_curr_task_fair,
8702 .task_tick = task_tick_fair,
8703 .task_fork = task_fork_fair,
8705 .prio_changed = prio_changed_fair,
8706 .switched_from = switched_from_fair,
8707 .switched_to = switched_to_fair,
8709 .get_rr_interval = get_rr_interval_fair,
8711 .update_curr = update_curr_fair,
8713 #ifdef CONFIG_FAIR_GROUP_SCHED
8714 .task_move_group = task_move_group_fair,
8718 #ifdef CONFIG_SCHED_DEBUG
8719 void print_cfs_stats(struct seq_file *m, int cpu)
8721 struct cfs_rq *cfs_rq;
8724 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8725 print_cfs_rq(m, cpu, cfs_rq);
8729 #ifdef CONFIG_NUMA_BALANCING
8730 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8733 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8735 for_each_online_node(node) {
8736 if (p->numa_faults) {
8737 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8738 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8740 if (p->numa_group) {
8741 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8742 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8744 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8747 #endif /* CONFIG_NUMA_BALANCING */
8748 #endif /* CONFIG_SCHED_DEBUG */
8750 __init void init_sched_fair_class(void)
8753 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8755 #ifdef CONFIG_NO_HZ_COMMON
8756 nohz.next_balance = jiffies;
8757 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);