2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched.h>
24 #include <linux/latencytop.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 * At this point, util_avg won't be used in select_task_rq_fair anyway
690 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
694 * With new tasks being created, their initial util_avgs are extrapolated
695 * based on the cfs_rq's current util_avg:
697 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
699 * However, in many cases, the above util_avg does not give a desired
700 * value. Moreover, the sum of the util_avgs may be divergent, such
701 * as when the series is a harmonic series.
703 * To solve this problem, we also cap the util_avg of successive tasks to
704 * only 1/2 of the left utilization budget:
706 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
708 * where n denotes the nth task.
710 * For example, a simplest series from the beginning would be like:
712 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
713 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
715 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
716 * if util_avg > util_avg_cap.
718 void post_init_entity_util_avg(struct sched_entity *se)
720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
721 struct sched_avg *sa = &se->avg;
722 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
725 if (cfs_rq->avg.util_avg != 0) {
726 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
727 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
729 if (sa->util_avg > cap)
734 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
739 void init_entity_runnable_average(struct sched_entity *se)
742 void post_init_entity_util_avg(struct sched_entity *se)
748 * Update the current task's runtime statistics.
750 static void update_curr(struct cfs_rq *cfs_rq)
752 struct sched_entity *curr = cfs_rq->curr;
753 u64 now = rq_clock_task(rq_of(cfs_rq));
759 delta_exec = now - curr->exec_start;
760 if (unlikely((s64)delta_exec <= 0))
763 curr->exec_start = now;
765 schedstat_set(curr->statistics.exec_max,
766 max(delta_exec, curr->statistics.exec_max));
768 curr->sum_exec_runtime += delta_exec;
769 schedstat_add(cfs_rq, exec_clock, delta_exec);
771 curr->vruntime += calc_delta_fair(delta_exec, curr);
772 update_min_vruntime(cfs_rq);
774 if (entity_is_task(curr)) {
775 struct task_struct *curtask = task_of(curr);
777 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
778 cpuacct_charge(curtask, delta_exec);
779 account_group_exec_runtime(curtask, delta_exec);
782 account_cfs_rq_runtime(cfs_rq, delta_exec);
785 static void update_curr_fair(struct rq *rq)
787 update_curr(cfs_rq_of(&rq->curr->se));
790 #ifdef CONFIG_SCHEDSTATS
792 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
794 u64 wait_start = rq_clock(rq_of(cfs_rq));
796 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
797 likely(wait_start > se->statistics.wait_start))
798 wait_start -= se->statistics.wait_start;
800 se->statistics.wait_start = wait_start;
804 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
806 struct task_struct *p;
809 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
811 if (entity_is_task(se)) {
813 if (task_on_rq_migrating(p)) {
815 * Preserve migrating task's wait time so wait_start
816 * time stamp can be adjusted to accumulate wait time
817 * prior to migration.
819 se->statistics.wait_start = delta;
822 trace_sched_stat_wait(p, delta);
825 se->statistics.wait_max = max(se->statistics.wait_max, delta);
826 se->statistics.wait_count++;
827 se->statistics.wait_sum += delta;
828 se->statistics.wait_start = 0;
832 * Task is being enqueued - update stats:
835 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
838 * Are we enqueueing a waiting task? (for current tasks
839 * a dequeue/enqueue event is a NOP)
841 if (se != cfs_rq->curr)
842 update_stats_wait_start(cfs_rq, se);
846 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
849 * Mark the end of the wait period if dequeueing a
852 if (se != cfs_rq->curr)
853 update_stats_wait_end(cfs_rq, se);
855 if (flags & DEQUEUE_SLEEP) {
856 if (entity_is_task(se)) {
857 struct task_struct *tsk = task_of(se);
859 if (tsk->state & TASK_INTERRUPTIBLE)
860 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
861 if (tsk->state & TASK_UNINTERRUPTIBLE)
862 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
869 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
874 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
879 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
890 * We are picking a new current task - update its stats:
893 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
896 * We are starting a new run period:
898 se->exec_start = rq_clock_task(rq_of(cfs_rq));
901 /**************************************************
902 * Scheduling class queueing methods:
905 #ifdef CONFIG_NUMA_BALANCING
907 * Approximate time to scan a full NUMA task in ms. The task scan period is
908 * calculated based on the tasks virtual memory size and
909 * numa_balancing_scan_size.
911 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
912 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
914 /* Portion of address space to scan in MB */
915 unsigned int sysctl_numa_balancing_scan_size = 256;
917 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
918 unsigned int sysctl_numa_balancing_scan_delay = 1000;
920 static unsigned int task_nr_scan_windows(struct task_struct *p)
922 unsigned long rss = 0;
923 unsigned long nr_scan_pages;
926 * Calculations based on RSS as non-present and empty pages are skipped
927 * by the PTE scanner and NUMA hinting faults should be trapped based
930 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
931 rss = get_mm_rss(p->mm);
935 rss = round_up(rss, nr_scan_pages);
936 return rss / nr_scan_pages;
939 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
940 #define MAX_SCAN_WINDOW 2560
942 static unsigned int task_scan_min(struct task_struct *p)
944 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
945 unsigned int scan, floor;
946 unsigned int windows = 1;
948 if (scan_size < MAX_SCAN_WINDOW)
949 windows = MAX_SCAN_WINDOW / scan_size;
950 floor = 1000 / windows;
952 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
953 return max_t(unsigned int, floor, scan);
956 static unsigned int task_scan_max(struct task_struct *p)
958 unsigned int smin = task_scan_min(p);
961 /* Watch for min being lower than max due to floor calculations */
962 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
963 return max(smin, smax);
966 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
968 rq->nr_numa_running += (p->numa_preferred_nid != -1);
969 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
972 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
974 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
975 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
981 spinlock_t lock; /* nr_tasks, tasks */
987 unsigned long total_faults;
988 unsigned long max_faults_cpu;
990 * Faults_cpu is used to decide whether memory should move
991 * towards the CPU. As a consequence, these stats are weighted
992 * more by CPU use than by memory faults.
994 unsigned long *faults_cpu;
995 unsigned long faults[0];
998 /* Shared or private faults. */
999 #define NR_NUMA_HINT_FAULT_TYPES 2
1001 /* Memory and CPU locality */
1002 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1004 /* Averaged statistics, and temporary buffers. */
1005 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1007 pid_t task_numa_group_id(struct task_struct *p)
1009 return p->numa_group ? p->numa_group->gid : 0;
1013 * The averaged statistics, shared & private, memory & cpu,
1014 * occupy the first half of the array. The second half of the
1015 * array is for current counters, which are averaged into the
1016 * first set by task_numa_placement.
1018 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1020 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1023 static inline unsigned long task_faults(struct task_struct *p, int nid)
1025 if (!p->numa_faults)
1028 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1029 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1032 static inline unsigned long group_faults(struct task_struct *p, int nid)
1037 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1038 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1041 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1043 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1044 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1048 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1049 * considered part of a numa group's pseudo-interleaving set. Migrations
1050 * between these nodes are slowed down, to allow things to settle down.
1052 #define ACTIVE_NODE_FRACTION 3
1054 static bool numa_is_active_node(int nid, struct numa_group *ng)
1056 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1059 /* Handle placement on systems where not all nodes are directly connected. */
1060 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1061 int maxdist, bool task)
1063 unsigned long score = 0;
1067 * All nodes are directly connected, and the same distance
1068 * from each other. No need for fancy placement algorithms.
1070 if (sched_numa_topology_type == NUMA_DIRECT)
1074 * This code is called for each node, introducing N^2 complexity,
1075 * which should be ok given the number of nodes rarely exceeds 8.
1077 for_each_online_node(node) {
1078 unsigned long faults;
1079 int dist = node_distance(nid, node);
1082 * The furthest away nodes in the system are not interesting
1083 * for placement; nid was already counted.
1085 if (dist == sched_max_numa_distance || node == nid)
1089 * On systems with a backplane NUMA topology, compare groups
1090 * of nodes, and move tasks towards the group with the most
1091 * memory accesses. When comparing two nodes at distance
1092 * "hoplimit", only nodes closer by than "hoplimit" are part
1093 * of each group. Skip other nodes.
1095 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1099 /* Add up the faults from nearby nodes. */
1101 faults = task_faults(p, node);
1103 faults = group_faults(p, node);
1106 * On systems with a glueless mesh NUMA topology, there are
1107 * no fixed "groups of nodes". Instead, nodes that are not
1108 * directly connected bounce traffic through intermediate
1109 * nodes; a numa_group can occupy any set of nodes.
1110 * The further away a node is, the less the faults count.
1111 * This seems to result in good task placement.
1113 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1114 faults *= (sched_max_numa_distance - dist);
1115 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1125 * These return the fraction of accesses done by a particular task, or
1126 * task group, on a particular numa node. The group weight is given a
1127 * larger multiplier, in order to group tasks together that are almost
1128 * evenly spread out between numa nodes.
1130 static inline unsigned long task_weight(struct task_struct *p, int nid,
1133 unsigned long faults, total_faults;
1135 if (!p->numa_faults)
1138 total_faults = p->total_numa_faults;
1143 faults = task_faults(p, nid);
1144 faults += score_nearby_nodes(p, nid, dist, true);
1146 return 1000 * faults / total_faults;
1149 static inline unsigned long group_weight(struct task_struct *p, int nid,
1152 unsigned long faults, total_faults;
1157 total_faults = p->numa_group->total_faults;
1162 faults = group_faults(p, nid);
1163 faults += score_nearby_nodes(p, nid, dist, false);
1165 return 1000 * faults / total_faults;
1168 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1169 int src_nid, int dst_cpu)
1171 struct numa_group *ng = p->numa_group;
1172 int dst_nid = cpu_to_node(dst_cpu);
1173 int last_cpupid, this_cpupid;
1175 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1178 * Multi-stage node selection is used in conjunction with a periodic
1179 * migration fault to build a temporal task<->page relation. By using
1180 * a two-stage filter we remove short/unlikely relations.
1182 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1183 * a task's usage of a particular page (n_p) per total usage of this
1184 * page (n_t) (in a given time-span) to a probability.
1186 * Our periodic faults will sample this probability and getting the
1187 * same result twice in a row, given these samples are fully
1188 * independent, is then given by P(n)^2, provided our sample period
1189 * is sufficiently short compared to the usage pattern.
1191 * This quadric squishes small probabilities, making it less likely we
1192 * act on an unlikely task<->page relation.
1194 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1195 if (!cpupid_pid_unset(last_cpupid) &&
1196 cpupid_to_nid(last_cpupid) != dst_nid)
1199 /* Always allow migrate on private faults */
1200 if (cpupid_match_pid(p, last_cpupid))
1203 /* A shared fault, but p->numa_group has not been set up yet. */
1208 * Destination node is much more heavily used than the source
1209 * node? Allow migration.
1211 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1212 ACTIVE_NODE_FRACTION)
1216 * Distribute memory according to CPU & memory use on each node,
1217 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1219 * faults_cpu(dst) 3 faults_cpu(src)
1220 * --------------- * - > ---------------
1221 * faults_mem(dst) 4 faults_mem(src)
1223 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1224 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1227 static unsigned long weighted_cpuload(const int cpu);
1228 static unsigned long source_load(int cpu, int type);
1229 static unsigned long target_load(int cpu, int type);
1230 static unsigned long capacity_of(int cpu);
1231 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1233 /* Cached statistics for all CPUs within a node */
1235 unsigned long nr_running;
1238 /* Total compute capacity of CPUs on a node */
1239 unsigned long compute_capacity;
1241 /* Approximate capacity in terms of runnable tasks on a node */
1242 unsigned long task_capacity;
1243 int has_free_capacity;
1247 * XXX borrowed from update_sg_lb_stats
1249 static void update_numa_stats(struct numa_stats *ns, int nid)
1251 int smt, cpu, cpus = 0;
1252 unsigned long capacity;
1254 memset(ns, 0, sizeof(*ns));
1255 for_each_cpu(cpu, cpumask_of_node(nid)) {
1256 struct rq *rq = cpu_rq(cpu);
1258 ns->nr_running += rq->nr_running;
1259 ns->load += weighted_cpuload(cpu);
1260 ns->compute_capacity += capacity_of(cpu);
1266 * If we raced with hotplug and there are no CPUs left in our mask
1267 * the @ns structure is NULL'ed and task_numa_compare() will
1268 * not find this node attractive.
1270 * We'll either bail at !has_free_capacity, or we'll detect a huge
1271 * imbalance and bail there.
1276 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1277 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1278 capacity = cpus / smt; /* cores */
1280 ns->task_capacity = min_t(unsigned, capacity,
1281 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1282 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1285 struct task_numa_env {
1286 struct task_struct *p;
1288 int src_cpu, src_nid;
1289 int dst_cpu, dst_nid;
1291 struct numa_stats src_stats, dst_stats;
1296 struct task_struct *best_task;
1301 static void task_numa_assign(struct task_numa_env *env,
1302 struct task_struct *p, long imp)
1305 put_task_struct(env->best_task);
1308 env->best_imp = imp;
1309 env->best_cpu = env->dst_cpu;
1312 static bool load_too_imbalanced(long src_load, long dst_load,
1313 struct task_numa_env *env)
1316 long orig_src_load, orig_dst_load;
1317 long src_capacity, dst_capacity;
1320 * The load is corrected for the CPU capacity available on each node.
1323 * ------------ vs ---------
1324 * src_capacity dst_capacity
1326 src_capacity = env->src_stats.compute_capacity;
1327 dst_capacity = env->dst_stats.compute_capacity;
1329 /* We care about the slope of the imbalance, not the direction. */
1330 if (dst_load < src_load)
1331 swap(dst_load, src_load);
1333 /* Is the difference below the threshold? */
1334 imb = dst_load * src_capacity * 100 -
1335 src_load * dst_capacity * env->imbalance_pct;
1340 * The imbalance is above the allowed threshold.
1341 * Compare it with the old imbalance.
1343 orig_src_load = env->src_stats.load;
1344 orig_dst_load = env->dst_stats.load;
1346 if (orig_dst_load < orig_src_load)
1347 swap(orig_dst_load, orig_src_load);
1349 old_imb = orig_dst_load * src_capacity * 100 -
1350 orig_src_load * dst_capacity * env->imbalance_pct;
1352 /* Would this change make things worse? */
1353 return (imb > old_imb);
1357 * This checks if the overall compute and NUMA accesses of the system would
1358 * be improved if the source tasks was migrated to the target dst_cpu taking
1359 * into account that it might be best if task running on the dst_cpu should
1360 * be exchanged with the source task
1362 static void task_numa_compare(struct task_numa_env *env,
1363 long taskimp, long groupimp)
1365 struct rq *src_rq = cpu_rq(env->src_cpu);
1366 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1367 struct task_struct *cur;
1368 long src_load, dst_load;
1370 long imp = env->p->numa_group ? groupimp : taskimp;
1372 int dist = env->dist;
1373 bool assigned = false;
1377 raw_spin_lock_irq(&dst_rq->lock);
1380 * No need to move the exiting task or idle task.
1382 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1386 * The task_struct must be protected here to protect the
1387 * p->numa_faults access in the task_weight since the
1388 * numa_faults could already be freed in the following path:
1389 * finish_task_switch()
1390 * --> put_task_struct()
1391 * --> __put_task_struct()
1392 * --> task_numa_free()
1394 get_task_struct(cur);
1397 raw_spin_unlock_irq(&dst_rq->lock);
1400 * Because we have preemption enabled we can get migrated around and
1401 * end try selecting ourselves (current == env->p) as a swap candidate.
1407 * "imp" is the fault differential for the source task between the
1408 * source and destination node. Calculate the total differential for
1409 * the source task and potential destination task. The more negative
1410 * the value is, the more rmeote accesses that would be expected to
1411 * be incurred if the tasks were swapped.
1414 /* Skip this swap candidate if cannot move to the source cpu */
1415 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1419 * If dst and source tasks are in the same NUMA group, or not
1420 * in any group then look only at task weights.
1422 if (cur->numa_group == env->p->numa_group) {
1423 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1424 task_weight(cur, env->dst_nid, dist);
1426 * Add some hysteresis to prevent swapping the
1427 * tasks within a group over tiny differences.
1429 if (cur->numa_group)
1433 * Compare the group weights. If a task is all by
1434 * itself (not part of a group), use the task weight
1437 if (cur->numa_group)
1438 imp += group_weight(cur, env->src_nid, dist) -
1439 group_weight(cur, env->dst_nid, dist);
1441 imp += task_weight(cur, env->src_nid, dist) -
1442 task_weight(cur, env->dst_nid, dist);
1446 if (imp <= env->best_imp && moveimp <= env->best_imp)
1450 /* Is there capacity at our destination? */
1451 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1452 !env->dst_stats.has_free_capacity)
1458 /* Balance doesn't matter much if we're running a task per cpu */
1459 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1460 dst_rq->nr_running == 1)
1464 * In the overloaded case, try and keep the load balanced.
1467 load = task_h_load(env->p);
1468 dst_load = env->dst_stats.load + load;
1469 src_load = env->src_stats.load - load;
1471 if (moveimp > imp && moveimp > env->best_imp) {
1473 * If the improvement from just moving env->p direction is
1474 * better than swapping tasks around, check if a move is
1475 * possible. Store a slightly smaller score than moveimp,
1476 * so an actually idle CPU will win.
1478 if (!load_too_imbalanced(src_load, dst_load, env)) {
1480 put_task_struct(cur);
1486 if (imp <= env->best_imp)
1490 load = task_h_load(cur);
1495 if (load_too_imbalanced(src_load, dst_load, env))
1499 * One idle CPU per node is evaluated for a task numa move.
1500 * Call select_idle_sibling to maybe find a better one.
1503 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1507 task_numa_assign(env, cur, imp);
1511 * The dst_rq->curr isn't assigned. The protection for task_struct is
1514 if (cur && !assigned)
1515 put_task_struct(cur);
1518 static void task_numa_find_cpu(struct task_numa_env *env,
1519 long taskimp, long groupimp)
1523 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1524 /* Skip this CPU if the source task cannot migrate */
1525 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1529 task_numa_compare(env, taskimp, groupimp);
1533 /* Only move tasks to a NUMA node less busy than the current node. */
1534 static bool numa_has_capacity(struct task_numa_env *env)
1536 struct numa_stats *src = &env->src_stats;
1537 struct numa_stats *dst = &env->dst_stats;
1539 if (src->has_free_capacity && !dst->has_free_capacity)
1543 * Only consider a task move if the source has a higher load
1544 * than the destination, corrected for CPU capacity on each node.
1546 * src->load dst->load
1547 * --------------------- vs ---------------------
1548 * src->compute_capacity dst->compute_capacity
1550 if (src->load * dst->compute_capacity * env->imbalance_pct >
1552 dst->load * src->compute_capacity * 100)
1558 static int task_numa_migrate(struct task_struct *p)
1560 struct task_numa_env env = {
1563 .src_cpu = task_cpu(p),
1564 .src_nid = task_node(p),
1566 .imbalance_pct = 112,
1572 struct sched_domain *sd;
1573 unsigned long taskweight, groupweight;
1575 long taskimp, groupimp;
1578 * Pick the lowest SD_NUMA domain, as that would have the smallest
1579 * imbalance and would be the first to start moving tasks about.
1581 * And we want to avoid any moving of tasks about, as that would create
1582 * random movement of tasks -- counter the numa conditions we're trying
1586 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1588 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1592 * Cpusets can break the scheduler domain tree into smaller
1593 * balance domains, some of which do not cross NUMA boundaries.
1594 * Tasks that are "trapped" in such domains cannot be migrated
1595 * elsewhere, so there is no point in (re)trying.
1597 if (unlikely(!sd)) {
1598 p->numa_preferred_nid = task_node(p);
1602 env.dst_nid = p->numa_preferred_nid;
1603 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1604 taskweight = task_weight(p, env.src_nid, dist);
1605 groupweight = group_weight(p, env.src_nid, dist);
1606 update_numa_stats(&env.src_stats, env.src_nid);
1607 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1608 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1609 update_numa_stats(&env.dst_stats, env.dst_nid);
1611 /* Try to find a spot on the preferred nid. */
1612 if (numa_has_capacity(&env))
1613 task_numa_find_cpu(&env, taskimp, groupimp);
1616 * Look at other nodes in these cases:
1617 * - there is no space available on the preferred_nid
1618 * - the task is part of a numa_group that is interleaved across
1619 * multiple NUMA nodes; in order to better consolidate the group,
1620 * we need to check other locations.
1622 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1623 for_each_online_node(nid) {
1624 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1627 dist = node_distance(env.src_nid, env.dst_nid);
1628 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1630 taskweight = task_weight(p, env.src_nid, dist);
1631 groupweight = group_weight(p, env.src_nid, dist);
1634 /* Only consider nodes where both task and groups benefit */
1635 taskimp = task_weight(p, nid, dist) - taskweight;
1636 groupimp = group_weight(p, nid, dist) - groupweight;
1637 if (taskimp < 0 && groupimp < 0)
1642 update_numa_stats(&env.dst_stats, env.dst_nid);
1643 if (numa_has_capacity(&env))
1644 task_numa_find_cpu(&env, taskimp, groupimp);
1649 * If the task is part of a workload that spans multiple NUMA nodes,
1650 * and is migrating into one of the workload's active nodes, remember
1651 * this node as the task's preferred numa node, so the workload can
1653 * A task that migrated to a second choice node will be better off
1654 * trying for a better one later. Do not set the preferred node here.
1656 if (p->numa_group) {
1657 struct numa_group *ng = p->numa_group;
1659 if (env.best_cpu == -1)
1664 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1665 sched_setnuma(p, env.dst_nid);
1668 /* No better CPU than the current one was found. */
1669 if (env.best_cpu == -1)
1673 * Reset the scan period if the task is being rescheduled on an
1674 * alternative node to recheck if the tasks is now properly placed.
1676 p->numa_scan_period = task_scan_min(p);
1678 if (env.best_task == NULL) {
1679 ret = migrate_task_to(p, env.best_cpu);
1681 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1685 ret = migrate_swap(p, env.best_task);
1687 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1688 put_task_struct(env.best_task);
1692 /* Attempt to migrate a task to a CPU on the preferred node. */
1693 static void numa_migrate_preferred(struct task_struct *p)
1695 unsigned long interval = HZ;
1697 /* This task has no NUMA fault statistics yet */
1698 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1701 /* Periodically retry migrating the task to the preferred node */
1702 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1703 p->numa_migrate_retry = jiffies + interval;
1705 /* Success if task is already running on preferred CPU */
1706 if (task_node(p) == p->numa_preferred_nid)
1709 /* Otherwise, try migrate to a CPU on the preferred node */
1710 task_numa_migrate(p);
1714 * Find out how many nodes on the workload is actively running on. Do this by
1715 * tracking the nodes from which NUMA hinting faults are triggered. This can
1716 * be different from the set of nodes where the workload's memory is currently
1719 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1721 unsigned long faults, max_faults = 0;
1722 int nid, active_nodes = 0;
1724 for_each_online_node(nid) {
1725 faults = group_faults_cpu(numa_group, nid);
1726 if (faults > max_faults)
1727 max_faults = faults;
1730 for_each_online_node(nid) {
1731 faults = group_faults_cpu(numa_group, nid);
1732 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1736 numa_group->max_faults_cpu = max_faults;
1737 numa_group->active_nodes = active_nodes;
1741 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1742 * increments. The more local the fault statistics are, the higher the scan
1743 * period will be for the next scan window. If local/(local+remote) ratio is
1744 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1745 * the scan period will decrease. Aim for 70% local accesses.
1747 #define NUMA_PERIOD_SLOTS 10
1748 #define NUMA_PERIOD_THRESHOLD 7
1751 * Increase the scan period (slow down scanning) if the majority of
1752 * our memory is already on our local node, or if the majority of
1753 * the page accesses are shared with other processes.
1754 * Otherwise, decrease the scan period.
1756 static void update_task_scan_period(struct task_struct *p,
1757 unsigned long shared, unsigned long private)
1759 unsigned int period_slot;
1763 unsigned long remote = p->numa_faults_locality[0];
1764 unsigned long local = p->numa_faults_locality[1];
1767 * If there were no record hinting faults then either the task is
1768 * completely idle or all activity is areas that are not of interest
1769 * to automatic numa balancing. Related to that, if there were failed
1770 * migration then it implies we are migrating too quickly or the local
1771 * node is overloaded. In either case, scan slower
1773 if (local + shared == 0 || p->numa_faults_locality[2]) {
1774 p->numa_scan_period = min(p->numa_scan_period_max,
1775 p->numa_scan_period << 1);
1777 p->mm->numa_next_scan = jiffies +
1778 msecs_to_jiffies(p->numa_scan_period);
1784 * Prepare to scale scan period relative to the current period.
1785 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1786 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1787 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1789 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1790 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1791 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1792 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1795 diff = slot * period_slot;
1797 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1800 * Scale scan rate increases based on sharing. There is an
1801 * inverse relationship between the degree of sharing and
1802 * the adjustment made to the scanning period. Broadly
1803 * speaking the intent is that there is little point
1804 * scanning faster if shared accesses dominate as it may
1805 * simply bounce migrations uselessly
1807 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1808 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1811 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1812 task_scan_min(p), task_scan_max(p));
1813 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1817 * Get the fraction of time the task has been running since the last
1818 * NUMA placement cycle. The scheduler keeps similar statistics, but
1819 * decays those on a 32ms period, which is orders of magnitude off
1820 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1821 * stats only if the task is so new there are no NUMA statistics yet.
1823 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1825 u64 runtime, delta, now;
1826 /* Use the start of this time slice to avoid calculations. */
1827 now = p->se.exec_start;
1828 runtime = p->se.sum_exec_runtime;
1830 if (p->last_task_numa_placement) {
1831 delta = runtime - p->last_sum_exec_runtime;
1832 *period = now - p->last_task_numa_placement;
1834 delta = p->se.avg.load_sum / p->se.load.weight;
1835 *period = LOAD_AVG_MAX;
1838 p->last_sum_exec_runtime = runtime;
1839 p->last_task_numa_placement = now;
1845 * Determine the preferred nid for a task in a numa_group. This needs to
1846 * be done in a way that produces consistent results with group_weight,
1847 * otherwise workloads might not converge.
1849 static int preferred_group_nid(struct task_struct *p, int nid)
1854 /* Direct connections between all NUMA nodes. */
1855 if (sched_numa_topology_type == NUMA_DIRECT)
1859 * On a system with glueless mesh NUMA topology, group_weight
1860 * scores nodes according to the number of NUMA hinting faults on
1861 * both the node itself, and on nearby nodes.
1863 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1864 unsigned long score, max_score = 0;
1865 int node, max_node = nid;
1867 dist = sched_max_numa_distance;
1869 for_each_online_node(node) {
1870 score = group_weight(p, node, dist);
1871 if (score > max_score) {
1880 * Finding the preferred nid in a system with NUMA backplane
1881 * interconnect topology is more involved. The goal is to locate
1882 * tasks from numa_groups near each other in the system, and
1883 * untangle workloads from different sides of the system. This requires
1884 * searching down the hierarchy of node groups, recursively searching
1885 * inside the highest scoring group of nodes. The nodemask tricks
1886 * keep the complexity of the search down.
1888 nodes = node_online_map;
1889 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1890 unsigned long max_faults = 0;
1891 nodemask_t max_group = NODE_MASK_NONE;
1894 /* Are there nodes at this distance from each other? */
1895 if (!find_numa_distance(dist))
1898 for_each_node_mask(a, nodes) {
1899 unsigned long faults = 0;
1900 nodemask_t this_group;
1901 nodes_clear(this_group);
1903 /* Sum group's NUMA faults; includes a==b case. */
1904 for_each_node_mask(b, nodes) {
1905 if (node_distance(a, b) < dist) {
1906 faults += group_faults(p, b);
1907 node_set(b, this_group);
1908 node_clear(b, nodes);
1912 /* Remember the top group. */
1913 if (faults > max_faults) {
1914 max_faults = faults;
1915 max_group = this_group;
1917 * subtle: at the smallest distance there is
1918 * just one node left in each "group", the
1919 * winner is the preferred nid.
1924 /* Next round, evaluate the nodes within max_group. */
1932 static void task_numa_placement(struct task_struct *p)
1934 int seq, nid, max_nid = -1, max_group_nid = -1;
1935 unsigned long max_faults = 0, max_group_faults = 0;
1936 unsigned long fault_types[2] = { 0, 0 };
1937 unsigned long total_faults;
1938 u64 runtime, period;
1939 spinlock_t *group_lock = NULL;
1942 * The p->mm->numa_scan_seq field gets updated without
1943 * exclusive access. Use READ_ONCE() here to ensure
1944 * that the field is read in a single access:
1946 seq = READ_ONCE(p->mm->numa_scan_seq);
1947 if (p->numa_scan_seq == seq)
1949 p->numa_scan_seq = seq;
1950 p->numa_scan_period_max = task_scan_max(p);
1952 total_faults = p->numa_faults_locality[0] +
1953 p->numa_faults_locality[1];
1954 runtime = numa_get_avg_runtime(p, &period);
1956 /* If the task is part of a group prevent parallel updates to group stats */
1957 if (p->numa_group) {
1958 group_lock = &p->numa_group->lock;
1959 spin_lock_irq(group_lock);
1962 /* Find the node with the highest number of faults */
1963 for_each_online_node(nid) {
1964 /* Keep track of the offsets in numa_faults array */
1965 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1966 unsigned long faults = 0, group_faults = 0;
1969 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1970 long diff, f_diff, f_weight;
1972 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1973 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1974 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1975 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1977 /* Decay existing window, copy faults since last scan */
1978 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1979 fault_types[priv] += p->numa_faults[membuf_idx];
1980 p->numa_faults[membuf_idx] = 0;
1983 * Normalize the faults_from, so all tasks in a group
1984 * count according to CPU use, instead of by the raw
1985 * number of faults. Tasks with little runtime have
1986 * little over-all impact on throughput, and thus their
1987 * faults are less important.
1989 f_weight = div64_u64(runtime << 16, period + 1);
1990 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1992 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1993 p->numa_faults[cpubuf_idx] = 0;
1995 p->numa_faults[mem_idx] += diff;
1996 p->numa_faults[cpu_idx] += f_diff;
1997 faults += p->numa_faults[mem_idx];
1998 p->total_numa_faults += diff;
1999 if (p->numa_group) {
2001 * safe because we can only change our own group
2003 * mem_idx represents the offset for a given
2004 * nid and priv in a specific region because it
2005 * is at the beginning of the numa_faults array.
2007 p->numa_group->faults[mem_idx] += diff;
2008 p->numa_group->faults_cpu[mem_idx] += f_diff;
2009 p->numa_group->total_faults += diff;
2010 group_faults += p->numa_group->faults[mem_idx];
2014 if (faults > max_faults) {
2015 max_faults = faults;
2019 if (group_faults > max_group_faults) {
2020 max_group_faults = group_faults;
2021 max_group_nid = nid;
2025 update_task_scan_period(p, fault_types[0], fault_types[1]);
2027 if (p->numa_group) {
2028 numa_group_count_active_nodes(p->numa_group);
2029 spin_unlock_irq(group_lock);
2030 max_nid = preferred_group_nid(p, max_group_nid);
2034 /* Set the new preferred node */
2035 if (max_nid != p->numa_preferred_nid)
2036 sched_setnuma(p, max_nid);
2038 if (task_node(p) != p->numa_preferred_nid)
2039 numa_migrate_preferred(p);
2043 static inline int get_numa_group(struct numa_group *grp)
2045 return atomic_inc_not_zero(&grp->refcount);
2048 static inline void put_numa_group(struct numa_group *grp)
2050 if (atomic_dec_and_test(&grp->refcount))
2051 kfree_rcu(grp, rcu);
2054 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2057 struct numa_group *grp, *my_grp;
2058 struct task_struct *tsk;
2060 int cpu = cpupid_to_cpu(cpupid);
2063 if (unlikely(!p->numa_group)) {
2064 unsigned int size = sizeof(struct numa_group) +
2065 4*nr_node_ids*sizeof(unsigned long);
2067 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2071 atomic_set(&grp->refcount, 1);
2072 grp->active_nodes = 1;
2073 grp->max_faults_cpu = 0;
2074 spin_lock_init(&grp->lock);
2076 /* Second half of the array tracks nids where faults happen */
2077 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2080 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2081 grp->faults[i] = p->numa_faults[i];
2083 grp->total_faults = p->total_numa_faults;
2086 rcu_assign_pointer(p->numa_group, grp);
2090 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2092 if (!cpupid_match_pid(tsk, cpupid))
2095 grp = rcu_dereference(tsk->numa_group);
2099 my_grp = p->numa_group;
2104 * Only join the other group if its bigger; if we're the bigger group,
2105 * the other task will join us.
2107 if (my_grp->nr_tasks > grp->nr_tasks)
2111 * Tie-break on the grp address.
2113 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2116 /* Always join threads in the same process. */
2117 if (tsk->mm == current->mm)
2120 /* Simple filter to avoid false positives due to PID collisions */
2121 if (flags & TNF_SHARED)
2124 /* Update priv based on whether false sharing was detected */
2127 if (join && !get_numa_group(grp))
2135 BUG_ON(irqs_disabled());
2136 double_lock_irq(&my_grp->lock, &grp->lock);
2138 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2139 my_grp->faults[i] -= p->numa_faults[i];
2140 grp->faults[i] += p->numa_faults[i];
2142 my_grp->total_faults -= p->total_numa_faults;
2143 grp->total_faults += p->total_numa_faults;
2148 spin_unlock(&my_grp->lock);
2149 spin_unlock_irq(&grp->lock);
2151 rcu_assign_pointer(p->numa_group, grp);
2153 put_numa_group(my_grp);
2161 void task_numa_free(struct task_struct *p)
2163 struct numa_group *grp = p->numa_group;
2164 void *numa_faults = p->numa_faults;
2165 unsigned long flags;
2169 spin_lock_irqsave(&grp->lock, flags);
2170 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2171 grp->faults[i] -= p->numa_faults[i];
2172 grp->total_faults -= p->total_numa_faults;
2175 spin_unlock_irqrestore(&grp->lock, flags);
2176 RCU_INIT_POINTER(p->numa_group, NULL);
2177 put_numa_group(grp);
2180 p->numa_faults = NULL;
2185 * Got a PROT_NONE fault for a page on @node.
2187 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2189 struct task_struct *p = current;
2190 bool migrated = flags & TNF_MIGRATED;
2191 int cpu_node = task_node(current);
2192 int local = !!(flags & TNF_FAULT_LOCAL);
2193 struct numa_group *ng;
2196 if (!static_branch_likely(&sched_numa_balancing))
2199 /* for example, ksmd faulting in a user's mm */
2203 /* Allocate buffer to track faults on a per-node basis */
2204 if (unlikely(!p->numa_faults)) {
2205 int size = sizeof(*p->numa_faults) *
2206 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2208 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2209 if (!p->numa_faults)
2212 p->total_numa_faults = 0;
2213 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2217 * First accesses are treated as private, otherwise consider accesses
2218 * to be private if the accessing pid has not changed
2220 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2223 priv = cpupid_match_pid(p, last_cpupid);
2224 if (!priv && !(flags & TNF_NO_GROUP))
2225 task_numa_group(p, last_cpupid, flags, &priv);
2229 * If a workload spans multiple NUMA nodes, a shared fault that
2230 * occurs wholly within the set of nodes that the workload is
2231 * actively using should be counted as local. This allows the
2232 * scan rate to slow down when a workload has settled down.
2235 if (!priv && !local && ng && ng->active_nodes > 1 &&
2236 numa_is_active_node(cpu_node, ng) &&
2237 numa_is_active_node(mem_node, ng))
2240 task_numa_placement(p);
2243 * Retry task to preferred node migration periodically, in case it
2244 * case it previously failed, or the scheduler moved us.
2246 if (time_after(jiffies, p->numa_migrate_retry))
2247 numa_migrate_preferred(p);
2250 p->numa_pages_migrated += pages;
2251 if (flags & TNF_MIGRATE_FAIL)
2252 p->numa_faults_locality[2] += pages;
2254 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2255 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2256 p->numa_faults_locality[local] += pages;
2259 static void reset_ptenuma_scan(struct task_struct *p)
2262 * We only did a read acquisition of the mmap sem, so
2263 * p->mm->numa_scan_seq is written to without exclusive access
2264 * and the update is not guaranteed to be atomic. That's not
2265 * much of an issue though, since this is just used for
2266 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2267 * expensive, to avoid any form of compiler optimizations:
2269 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2270 p->mm->numa_scan_offset = 0;
2274 * The expensive part of numa migration is done from task_work context.
2275 * Triggered from task_tick_numa().
2277 void task_numa_work(struct callback_head *work)
2279 unsigned long migrate, next_scan, now = jiffies;
2280 struct task_struct *p = current;
2281 struct mm_struct *mm = p->mm;
2282 u64 runtime = p->se.sum_exec_runtime;
2283 struct vm_area_struct *vma;
2284 unsigned long start, end;
2285 unsigned long nr_pte_updates = 0;
2286 long pages, virtpages;
2288 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2290 work->next = work; /* protect against double add */
2292 * Who cares about NUMA placement when they're dying.
2294 * NOTE: make sure not to dereference p->mm before this check,
2295 * exit_task_work() happens _after_ exit_mm() so we could be called
2296 * without p->mm even though we still had it when we enqueued this
2299 if (p->flags & PF_EXITING)
2302 if (!mm->numa_next_scan) {
2303 mm->numa_next_scan = now +
2304 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2308 * Enforce maximal scan/migration frequency..
2310 migrate = mm->numa_next_scan;
2311 if (time_before(now, migrate))
2314 if (p->numa_scan_period == 0) {
2315 p->numa_scan_period_max = task_scan_max(p);
2316 p->numa_scan_period = task_scan_min(p);
2319 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2320 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2324 * Delay this task enough that another task of this mm will likely win
2325 * the next time around.
2327 p->node_stamp += 2 * TICK_NSEC;
2329 start = mm->numa_scan_offset;
2330 pages = sysctl_numa_balancing_scan_size;
2331 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2332 virtpages = pages * 8; /* Scan up to this much virtual space */
2337 down_read(&mm->mmap_sem);
2338 vma = find_vma(mm, start);
2340 reset_ptenuma_scan(p);
2344 for (; vma; vma = vma->vm_next) {
2345 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2346 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2351 * Shared library pages mapped by multiple processes are not
2352 * migrated as it is expected they are cache replicated. Avoid
2353 * hinting faults in read-only file-backed mappings or the vdso
2354 * as migrating the pages will be of marginal benefit.
2357 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2361 * Skip inaccessible VMAs to avoid any confusion between
2362 * PROT_NONE and NUMA hinting ptes
2364 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2368 start = max(start, vma->vm_start);
2369 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2370 end = min(end, vma->vm_end);
2371 nr_pte_updates = change_prot_numa(vma, start, end);
2374 * Try to scan sysctl_numa_balancing_size worth of
2375 * hpages that have at least one present PTE that
2376 * is not already pte-numa. If the VMA contains
2377 * areas that are unused or already full of prot_numa
2378 * PTEs, scan up to virtpages, to skip through those
2382 pages -= (end - start) >> PAGE_SHIFT;
2383 virtpages -= (end - start) >> PAGE_SHIFT;
2386 if (pages <= 0 || virtpages <= 0)
2390 } while (end != vma->vm_end);
2395 * It is possible to reach the end of the VMA list but the last few
2396 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2397 * would find the !migratable VMA on the next scan but not reset the
2398 * scanner to the start so check it now.
2401 mm->numa_scan_offset = start;
2403 reset_ptenuma_scan(p);
2404 up_read(&mm->mmap_sem);
2407 * Make sure tasks use at least 32x as much time to run other code
2408 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2409 * Usually update_task_scan_period slows down scanning enough; on an
2410 * overloaded system we need to limit overhead on a per task basis.
2412 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2413 u64 diff = p->se.sum_exec_runtime - runtime;
2414 p->node_stamp += 32 * diff;
2419 * Drive the periodic memory faults..
2421 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2423 struct callback_head *work = &curr->numa_work;
2427 * We don't care about NUMA placement if we don't have memory.
2429 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2433 * Using runtime rather than walltime has the dual advantage that
2434 * we (mostly) drive the selection from busy threads and that the
2435 * task needs to have done some actual work before we bother with
2438 now = curr->se.sum_exec_runtime;
2439 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2441 if (now > curr->node_stamp + period) {
2442 if (!curr->node_stamp)
2443 curr->numa_scan_period = task_scan_min(curr);
2444 curr->node_stamp += period;
2446 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2447 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2448 task_work_add(curr, work, true);
2453 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2457 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2461 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2464 #endif /* CONFIG_NUMA_BALANCING */
2467 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2469 update_load_add(&cfs_rq->load, se->load.weight);
2470 if (!parent_entity(se))
2471 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2473 if (entity_is_task(se)) {
2474 struct rq *rq = rq_of(cfs_rq);
2476 account_numa_enqueue(rq, task_of(se));
2477 list_add(&se->group_node, &rq->cfs_tasks);
2480 cfs_rq->nr_running++;
2484 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2486 update_load_sub(&cfs_rq->load, se->load.weight);
2487 if (!parent_entity(se))
2488 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2490 if (entity_is_task(se)) {
2491 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2492 list_del_init(&se->group_node);
2495 cfs_rq->nr_running--;
2498 #ifdef CONFIG_FAIR_GROUP_SCHED
2500 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2505 * Use this CPU's real-time load instead of the last load contribution
2506 * as the updating of the contribution is delayed, and we will use the
2507 * the real-time load to calc the share. See update_tg_load_avg().
2509 tg_weight = atomic_long_read(&tg->load_avg);
2510 tg_weight -= cfs_rq->tg_load_avg_contrib;
2511 tg_weight += cfs_rq->load.weight;
2516 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2518 long tg_weight, load, shares;
2520 tg_weight = calc_tg_weight(tg, cfs_rq);
2521 load = cfs_rq->load.weight;
2523 shares = (tg->shares * load);
2525 shares /= tg_weight;
2527 if (shares < MIN_SHARES)
2528 shares = MIN_SHARES;
2529 if (shares > tg->shares)
2530 shares = tg->shares;
2534 # else /* CONFIG_SMP */
2535 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2539 # endif /* CONFIG_SMP */
2540 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2541 unsigned long weight)
2544 /* commit outstanding execution time */
2545 if (cfs_rq->curr == se)
2546 update_curr(cfs_rq);
2547 account_entity_dequeue(cfs_rq, se);
2550 update_load_set(&se->load, weight);
2553 account_entity_enqueue(cfs_rq, se);
2556 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2558 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2560 struct task_group *tg;
2561 struct sched_entity *se;
2565 se = tg->se[cpu_of(rq_of(cfs_rq))];
2566 if (!se || throttled_hierarchy(cfs_rq))
2569 if (likely(se->load.weight == tg->shares))
2572 shares = calc_cfs_shares(cfs_rq, tg);
2574 reweight_entity(cfs_rq_of(se), se, shares);
2576 #else /* CONFIG_FAIR_GROUP_SCHED */
2577 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2580 #endif /* CONFIG_FAIR_GROUP_SCHED */
2583 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2584 static const u32 runnable_avg_yN_inv[] = {
2585 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2586 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2587 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2588 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2589 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2590 0x85aac367, 0x82cd8698,
2594 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2595 * over-estimates when re-combining.
2597 static const u32 runnable_avg_yN_sum[] = {
2598 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2599 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2600 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2604 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2605 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2608 static const u32 __accumulated_sum_N32[] = {
2609 0, 23371, 35056, 40899, 43820, 45281,
2610 46011, 46376, 46559, 46650, 46696, 46719,
2615 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2617 static __always_inline u64 decay_load(u64 val, u64 n)
2619 unsigned int local_n;
2623 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2626 /* after bounds checking we can collapse to 32-bit */
2630 * As y^PERIOD = 1/2, we can combine
2631 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2632 * With a look-up table which covers y^n (n<PERIOD)
2634 * To achieve constant time decay_load.
2636 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2637 val >>= local_n / LOAD_AVG_PERIOD;
2638 local_n %= LOAD_AVG_PERIOD;
2641 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2646 * For updates fully spanning n periods, the contribution to runnable
2647 * average will be: \Sum 1024*y^n
2649 * We can compute this reasonably efficiently by combining:
2650 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2652 static u32 __compute_runnable_contrib(u64 n)
2656 if (likely(n <= LOAD_AVG_PERIOD))
2657 return runnable_avg_yN_sum[n];
2658 else if (unlikely(n >= LOAD_AVG_MAX_N))
2659 return LOAD_AVG_MAX;
2661 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2662 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2663 n %= LOAD_AVG_PERIOD;
2664 contrib = decay_load(contrib, n);
2665 return contrib + runnable_avg_yN_sum[n];
2668 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2671 * We can represent the historical contribution to runnable average as the
2672 * coefficients of a geometric series. To do this we sub-divide our runnable
2673 * history into segments of approximately 1ms (1024us); label the segment that
2674 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2676 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2678 * (now) (~1ms ago) (~2ms ago)
2680 * Let u_i denote the fraction of p_i that the entity was runnable.
2682 * We then designate the fractions u_i as our co-efficients, yielding the
2683 * following representation of historical load:
2684 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2686 * We choose y based on the with of a reasonably scheduling period, fixing:
2689 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2690 * approximately half as much as the contribution to load within the last ms
2693 * When a period "rolls over" and we have new u_0`, multiplying the previous
2694 * sum again by y is sufficient to update:
2695 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2696 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2698 static __always_inline int
2699 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2700 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2702 u64 delta, scaled_delta, periods;
2704 unsigned int delta_w, scaled_delta_w, decayed = 0;
2705 unsigned long scale_freq, scale_cpu;
2707 delta = now - sa->last_update_time;
2709 * This should only happen when time goes backwards, which it
2710 * unfortunately does during sched clock init when we swap over to TSC.
2712 if ((s64)delta < 0) {
2713 sa->last_update_time = now;
2718 * Use 1024ns as the unit of measurement since it's a reasonable
2719 * approximation of 1us and fast to compute.
2724 sa->last_update_time = now;
2726 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2727 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2729 /* delta_w is the amount already accumulated against our next period */
2730 delta_w = sa->period_contrib;
2731 if (delta + delta_w >= 1024) {
2734 /* how much left for next period will start over, we don't know yet */
2735 sa->period_contrib = 0;
2738 * Now that we know we're crossing a period boundary, figure
2739 * out how much from delta we need to complete the current
2740 * period and accrue it.
2742 delta_w = 1024 - delta_w;
2743 scaled_delta_w = cap_scale(delta_w, scale_freq);
2745 sa->load_sum += weight * scaled_delta_w;
2747 cfs_rq->runnable_load_sum +=
2748 weight * scaled_delta_w;
2752 sa->util_sum += scaled_delta_w * scale_cpu;
2756 /* Figure out how many additional periods this update spans */
2757 periods = delta / 1024;
2760 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2762 cfs_rq->runnable_load_sum =
2763 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2765 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2767 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2768 contrib = __compute_runnable_contrib(periods);
2769 contrib = cap_scale(contrib, scale_freq);
2771 sa->load_sum += weight * contrib;
2773 cfs_rq->runnable_load_sum += weight * contrib;
2776 sa->util_sum += contrib * scale_cpu;
2779 /* Remainder of delta accrued against u_0` */
2780 scaled_delta = cap_scale(delta, scale_freq);
2782 sa->load_sum += weight * scaled_delta;
2784 cfs_rq->runnable_load_sum += weight * scaled_delta;
2787 sa->util_sum += scaled_delta * scale_cpu;
2789 sa->period_contrib += delta;
2792 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2794 cfs_rq->runnable_load_avg =
2795 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2797 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2803 #ifdef CONFIG_FAIR_GROUP_SCHED
2805 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2806 * and effective_load (which is not done because it is too costly).
2808 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2810 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2813 * No need to update load_avg for root_task_group as it is not used.
2815 if (cfs_rq->tg == &root_task_group)
2818 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2819 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2820 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2825 * Called within set_task_rq() right before setting a task's cpu. The
2826 * caller only guarantees p->pi_lock is held; no other assumptions,
2827 * including the state of rq->lock, should be made.
2829 void set_task_rq_fair(struct sched_entity *se,
2830 struct cfs_rq *prev, struct cfs_rq *next)
2832 if (!sched_feat(ATTACH_AGE_LOAD))
2836 * We are supposed to update the task to "current" time, then its up to
2837 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2838 * getting what current time is, so simply throw away the out-of-date
2839 * time. This will result in the wakee task is less decayed, but giving
2840 * the wakee more load sounds not bad.
2842 if (se->avg.last_update_time && prev) {
2843 u64 p_last_update_time;
2844 u64 n_last_update_time;
2846 #ifndef CONFIG_64BIT
2847 u64 p_last_update_time_copy;
2848 u64 n_last_update_time_copy;
2851 p_last_update_time_copy = prev->load_last_update_time_copy;
2852 n_last_update_time_copy = next->load_last_update_time_copy;
2856 p_last_update_time = prev->avg.last_update_time;
2857 n_last_update_time = next->avg.last_update_time;
2859 } while (p_last_update_time != p_last_update_time_copy ||
2860 n_last_update_time != n_last_update_time_copy);
2862 p_last_update_time = prev->avg.last_update_time;
2863 n_last_update_time = next->avg.last_update_time;
2865 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2866 &se->avg, 0, 0, NULL);
2867 se->avg.last_update_time = n_last_update_time;
2870 #else /* CONFIG_FAIR_GROUP_SCHED */
2871 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2872 #endif /* CONFIG_FAIR_GROUP_SCHED */
2874 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2876 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2878 struct rq *rq = rq_of(cfs_rq);
2879 int cpu = cpu_of(rq);
2881 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2882 unsigned long max = rq->cpu_capacity_orig;
2885 * There are a few boundary cases this might miss but it should
2886 * get called often enough that that should (hopefully) not be
2887 * a real problem -- added to that it only calls on the local
2888 * CPU, so if we enqueue remotely we'll miss an update, but
2889 * the next tick/schedule should update.
2891 * It will not get called when we go idle, because the idle
2892 * thread is a different class (!fair), nor will the utilization
2893 * number include things like RT tasks.
2895 * As is, the util number is not freq-invariant (we'd have to
2896 * implement arch_scale_freq_capacity() for that).
2900 cpufreq_update_util(rq_clock(rq),
2901 min(cfs_rq->avg.util_avg, max), max);
2906 * Unsigned subtract and clamp on underflow.
2908 * Explicitly do a load-store to ensure the intermediate value never hits
2909 * memory. This allows lockless observations without ever seeing the negative
2912 #define sub_positive(_ptr, _val) do { \
2913 typeof(_ptr) ptr = (_ptr); \
2914 typeof(*ptr) val = (_val); \
2915 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2919 WRITE_ONCE(*ptr, res); \
2922 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2924 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2926 struct sched_avg *sa = &cfs_rq->avg;
2927 int decayed, removed_load = 0, removed_util = 0;
2929 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2930 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2931 sub_positive(&sa->load_avg, r);
2932 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2936 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2937 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2938 sub_positive(&sa->util_avg, r);
2939 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2943 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2944 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2946 #ifndef CONFIG_64BIT
2948 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2951 if (update_freq && (decayed || removed_util))
2952 cfs_rq_util_change(cfs_rq);
2954 return decayed || removed_load;
2957 /* Update task and its cfs_rq load average */
2958 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2960 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2961 u64 now = cfs_rq_clock_task(cfs_rq);
2962 struct rq *rq = rq_of(cfs_rq);
2963 int cpu = cpu_of(rq);
2966 * Track task load average for carrying it to new CPU after migrated, and
2967 * track group sched_entity load average for task_h_load calc in migration
2969 __update_load_avg(now, cpu, &se->avg,
2970 se->on_rq * scale_load_down(se->load.weight),
2971 cfs_rq->curr == se, NULL);
2973 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2974 update_tg_load_avg(cfs_rq, 0);
2977 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2979 if (!sched_feat(ATTACH_AGE_LOAD))
2983 * If we got migrated (either between CPUs or between cgroups) we'll
2984 * have aged the average right before clearing @last_update_time.
2986 if (se->avg.last_update_time) {
2987 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2988 &se->avg, 0, 0, NULL);
2991 * XXX: we could have just aged the entire load away if we've been
2992 * absent from the fair class for too long.
2997 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2998 cfs_rq->avg.load_avg += se->avg.load_avg;
2999 cfs_rq->avg.load_sum += se->avg.load_sum;
3000 cfs_rq->avg.util_avg += se->avg.util_avg;
3001 cfs_rq->avg.util_sum += se->avg.util_sum;
3003 cfs_rq_util_change(cfs_rq);
3006 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3008 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3009 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3010 cfs_rq->curr == se, NULL);
3012 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3013 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3014 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3015 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3017 cfs_rq_util_change(cfs_rq);
3020 /* Add the load generated by se into cfs_rq's load average */
3022 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3024 struct sched_avg *sa = &se->avg;
3025 u64 now = cfs_rq_clock_task(cfs_rq);
3026 int migrated, decayed;
3028 migrated = !sa->last_update_time;
3030 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3031 se->on_rq * scale_load_down(se->load.weight),
3032 cfs_rq->curr == se, NULL);
3035 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3037 cfs_rq->runnable_load_avg += sa->load_avg;
3038 cfs_rq->runnable_load_sum += sa->load_sum;
3041 attach_entity_load_avg(cfs_rq, se);
3043 if (decayed || migrated)
3044 update_tg_load_avg(cfs_rq, 0);
3047 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3049 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3051 update_load_avg(se, 1);
3053 cfs_rq->runnable_load_avg =
3054 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3055 cfs_rq->runnable_load_sum =
3056 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3059 #ifndef CONFIG_64BIT
3060 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3062 u64 last_update_time_copy;
3063 u64 last_update_time;
3066 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3068 last_update_time = cfs_rq->avg.last_update_time;
3069 } while (last_update_time != last_update_time_copy);
3071 return last_update_time;
3074 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3076 return cfs_rq->avg.last_update_time;
3081 * Task first catches up with cfs_rq, and then subtract
3082 * itself from the cfs_rq (task must be off the queue now).
3084 void remove_entity_load_avg(struct sched_entity *se)
3086 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3087 u64 last_update_time;
3090 * Newly created task or never used group entity should not be removed
3091 * from its (source) cfs_rq
3093 if (se->avg.last_update_time == 0)
3096 last_update_time = cfs_rq_last_update_time(cfs_rq);
3098 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3099 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3100 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3103 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3105 return cfs_rq->runnable_load_avg;
3108 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3110 return cfs_rq->avg.load_avg;
3113 static int idle_balance(struct rq *this_rq);
3115 #else /* CONFIG_SMP */
3117 static inline void update_load_avg(struct sched_entity *se, int not_used)
3119 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3120 struct rq *rq = rq_of(cfs_rq);
3122 cpufreq_trigger_update(rq_clock(rq));
3126 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3128 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3129 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3132 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3134 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3136 static inline int idle_balance(struct rq *rq)
3141 #endif /* CONFIG_SMP */
3143 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3145 #ifdef CONFIG_SCHEDSTATS
3146 struct task_struct *tsk = NULL;
3148 if (entity_is_task(se))
3151 if (se->statistics.sleep_start) {
3152 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3157 if (unlikely(delta > se->statistics.sleep_max))
3158 se->statistics.sleep_max = delta;
3160 se->statistics.sleep_start = 0;
3161 se->statistics.sum_sleep_runtime += delta;
3164 account_scheduler_latency(tsk, delta >> 10, 1);
3165 trace_sched_stat_sleep(tsk, delta);
3168 if (se->statistics.block_start) {
3169 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3174 if (unlikely(delta > se->statistics.block_max))
3175 se->statistics.block_max = delta;
3177 se->statistics.block_start = 0;
3178 se->statistics.sum_sleep_runtime += delta;
3181 if (tsk->in_iowait) {
3182 se->statistics.iowait_sum += delta;
3183 se->statistics.iowait_count++;
3184 trace_sched_stat_iowait(tsk, delta);
3187 trace_sched_stat_blocked(tsk, delta);
3190 * Blocking time is in units of nanosecs, so shift by
3191 * 20 to get a milliseconds-range estimation of the
3192 * amount of time that the task spent sleeping:
3194 if (unlikely(prof_on == SLEEP_PROFILING)) {
3195 profile_hits(SLEEP_PROFILING,
3196 (void *)get_wchan(tsk),
3199 account_scheduler_latency(tsk, delta >> 10, 0);
3205 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3207 #ifdef CONFIG_SCHED_DEBUG
3208 s64 d = se->vruntime - cfs_rq->min_vruntime;
3213 if (d > 3*sysctl_sched_latency)
3214 schedstat_inc(cfs_rq, nr_spread_over);
3219 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3221 u64 vruntime = cfs_rq->min_vruntime;
3224 * The 'current' period is already promised to the current tasks,
3225 * however the extra weight of the new task will slow them down a
3226 * little, place the new task so that it fits in the slot that
3227 * stays open at the end.
3229 if (initial && sched_feat(START_DEBIT))
3230 vruntime += sched_vslice(cfs_rq, se);
3232 /* sleeps up to a single latency don't count. */
3234 unsigned long thresh = sysctl_sched_latency;
3237 * Halve their sleep time's effect, to allow
3238 * for a gentler effect of sleepers:
3240 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3246 /* ensure we never gain time by being placed backwards. */
3247 se->vruntime = max_vruntime(se->vruntime, vruntime);
3250 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3252 static inline void check_schedstat_required(void)
3254 #ifdef CONFIG_SCHEDSTATS
3255 if (schedstat_enabled())
3258 /* Force schedstat enabled if a dependent tracepoint is active */
3259 if (trace_sched_stat_wait_enabled() ||
3260 trace_sched_stat_sleep_enabled() ||
3261 trace_sched_stat_iowait_enabled() ||
3262 trace_sched_stat_blocked_enabled() ||
3263 trace_sched_stat_runtime_enabled()) {
3264 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3265 "stat_blocked and stat_runtime require the "
3266 "kernel parameter schedstats=enabled or "
3267 "kernel.sched_schedstats=1\n");
3278 * update_min_vruntime()
3279 * vruntime -= min_vruntime
3283 * update_min_vruntime()
3284 * vruntime += min_vruntime
3286 * this way the vruntime transition between RQs is done when both
3287 * min_vruntime are up-to-date.
3291 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3292 * vruntime -= min_vruntime
3296 * update_min_vruntime()
3297 * vruntime += min_vruntime
3299 * this way we don't have the most up-to-date min_vruntime on the originating
3300 * CPU and an up-to-date min_vruntime on the destination CPU.
3304 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3306 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3307 bool curr = cfs_rq->curr == se;
3310 * If we're the current task, we must renormalise before calling
3314 se->vruntime += cfs_rq->min_vruntime;
3316 update_curr(cfs_rq);
3319 * Otherwise, renormalise after, such that we're placed at the current
3320 * moment in time, instead of some random moment in the past. Being
3321 * placed in the past could significantly boost this task to the
3322 * fairness detriment of existing tasks.
3324 if (renorm && !curr)
3325 se->vruntime += cfs_rq->min_vruntime;
3327 enqueue_entity_load_avg(cfs_rq, se);
3328 account_entity_enqueue(cfs_rq, se);
3329 update_cfs_shares(cfs_rq);
3331 if (flags & ENQUEUE_WAKEUP) {
3332 place_entity(cfs_rq, se, 0);
3333 if (schedstat_enabled())
3334 enqueue_sleeper(cfs_rq, se);
3337 check_schedstat_required();
3338 if (schedstat_enabled()) {
3339 update_stats_enqueue(cfs_rq, se);
3340 check_spread(cfs_rq, se);
3343 __enqueue_entity(cfs_rq, se);
3346 if (cfs_rq->nr_running == 1) {
3347 list_add_leaf_cfs_rq(cfs_rq);
3348 check_enqueue_throttle(cfs_rq);
3352 static void __clear_buddies_last(struct sched_entity *se)
3354 for_each_sched_entity(se) {
3355 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3356 if (cfs_rq->last != se)
3359 cfs_rq->last = NULL;
3363 static void __clear_buddies_next(struct sched_entity *se)
3365 for_each_sched_entity(se) {
3366 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3367 if (cfs_rq->next != se)
3370 cfs_rq->next = NULL;
3374 static void __clear_buddies_skip(struct sched_entity *se)
3376 for_each_sched_entity(se) {
3377 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3378 if (cfs_rq->skip != se)
3381 cfs_rq->skip = NULL;
3385 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3387 if (cfs_rq->last == se)
3388 __clear_buddies_last(se);
3390 if (cfs_rq->next == se)
3391 __clear_buddies_next(se);
3393 if (cfs_rq->skip == se)
3394 __clear_buddies_skip(se);
3397 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3400 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3403 * Update run-time statistics of the 'current'.
3405 update_curr(cfs_rq);
3406 dequeue_entity_load_avg(cfs_rq, se);
3408 if (schedstat_enabled())
3409 update_stats_dequeue(cfs_rq, se, flags);
3411 clear_buddies(cfs_rq, se);
3413 if (se != cfs_rq->curr)
3414 __dequeue_entity(cfs_rq, se);
3416 account_entity_dequeue(cfs_rq, se);
3419 * Normalize the entity after updating the min_vruntime because the
3420 * update can refer to the ->curr item and we need to reflect this
3421 * movement in our normalized position.
3423 if (!(flags & DEQUEUE_SLEEP))
3424 se->vruntime -= cfs_rq->min_vruntime;
3426 /* return excess runtime on last dequeue */
3427 return_cfs_rq_runtime(cfs_rq);
3429 update_min_vruntime(cfs_rq);
3430 update_cfs_shares(cfs_rq);
3434 * Preempt the current task with a newly woken task if needed:
3437 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3439 unsigned long ideal_runtime, delta_exec;
3440 struct sched_entity *se;
3443 ideal_runtime = sched_slice(cfs_rq, curr);
3444 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3445 if (delta_exec > ideal_runtime) {
3446 resched_curr(rq_of(cfs_rq));
3448 * The current task ran long enough, ensure it doesn't get
3449 * re-elected due to buddy favours.
3451 clear_buddies(cfs_rq, curr);
3456 * Ensure that a task that missed wakeup preemption by a
3457 * narrow margin doesn't have to wait for a full slice.
3458 * This also mitigates buddy induced latencies under load.
3460 if (delta_exec < sysctl_sched_min_granularity)
3463 se = __pick_first_entity(cfs_rq);
3464 delta = curr->vruntime - se->vruntime;
3469 if (delta > ideal_runtime)
3470 resched_curr(rq_of(cfs_rq));
3474 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3476 /* 'current' is not kept within the tree. */
3479 * Any task has to be enqueued before it get to execute on
3480 * a CPU. So account for the time it spent waiting on the
3483 if (schedstat_enabled())
3484 update_stats_wait_end(cfs_rq, se);
3485 __dequeue_entity(cfs_rq, se);
3486 update_load_avg(se, 1);
3489 update_stats_curr_start(cfs_rq, se);
3491 #ifdef CONFIG_SCHEDSTATS
3493 * Track our maximum slice length, if the CPU's load is at
3494 * least twice that of our own weight (i.e. dont track it
3495 * when there are only lesser-weight tasks around):
3497 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3498 se->statistics.slice_max = max(se->statistics.slice_max,
3499 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3502 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3506 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3509 * Pick the next process, keeping these things in mind, in this order:
3510 * 1) keep things fair between processes/task groups
3511 * 2) pick the "next" process, since someone really wants that to run
3512 * 3) pick the "last" process, for cache locality
3513 * 4) do not run the "skip" process, if something else is available
3515 static struct sched_entity *
3516 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3518 struct sched_entity *left = __pick_first_entity(cfs_rq);
3519 struct sched_entity *se;
3522 * If curr is set we have to see if its left of the leftmost entity
3523 * still in the tree, provided there was anything in the tree at all.
3525 if (!left || (curr && entity_before(curr, left)))
3528 se = left; /* ideally we run the leftmost entity */
3531 * Avoid running the skip buddy, if running something else can
3532 * be done without getting too unfair.
3534 if (cfs_rq->skip == se) {
3535 struct sched_entity *second;
3538 second = __pick_first_entity(cfs_rq);
3540 second = __pick_next_entity(se);
3541 if (!second || (curr && entity_before(curr, second)))
3545 if (second && wakeup_preempt_entity(second, left) < 1)
3550 * Prefer last buddy, try to return the CPU to a preempted task.
3552 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3556 * Someone really wants this to run. If it's not unfair, run it.
3558 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3561 clear_buddies(cfs_rq, se);
3566 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3568 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3571 * If still on the runqueue then deactivate_task()
3572 * was not called and update_curr() has to be done:
3575 update_curr(cfs_rq);
3577 /* throttle cfs_rqs exceeding runtime */
3578 check_cfs_rq_runtime(cfs_rq);
3580 if (schedstat_enabled()) {
3581 check_spread(cfs_rq, prev);
3583 update_stats_wait_start(cfs_rq, prev);
3587 /* Put 'current' back into the tree. */
3588 __enqueue_entity(cfs_rq, prev);
3589 /* in !on_rq case, update occurred at dequeue */
3590 update_load_avg(prev, 0);
3592 cfs_rq->curr = NULL;
3596 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3599 * Update run-time statistics of the 'current'.
3601 update_curr(cfs_rq);
3604 * Ensure that runnable average is periodically updated.
3606 update_load_avg(curr, 1);
3607 update_cfs_shares(cfs_rq);
3609 #ifdef CONFIG_SCHED_HRTICK
3611 * queued ticks are scheduled to match the slice, so don't bother
3612 * validating it and just reschedule.
3615 resched_curr(rq_of(cfs_rq));
3619 * don't let the period tick interfere with the hrtick preemption
3621 if (!sched_feat(DOUBLE_TICK) &&
3622 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3626 if (cfs_rq->nr_running > 1)
3627 check_preempt_tick(cfs_rq, curr);
3631 /**************************************************
3632 * CFS bandwidth control machinery
3635 #ifdef CONFIG_CFS_BANDWIDTH
3637 #ifdef HAVE_JUMP_LABEL
3638 static struct static_key __cfs_bandwidth_used;
3640 static inline bool cfs_bandwidth_used(void)
3642 return static_key_false(&__cfs_bandwidth_used);
3645 void cfs_bandwidth_usage_inc(void)
3647 static_key_slow_inc(&__cfs_bandwidth_used);
3650 void cfs_bandwidth_usage_dec(void)
3652 static_key_slow_dec(&__cfs_bandwidth_used);
3654 #else /* HAVE_JUMP_LABEL */
3655 static bool cfs_bandwidth_used(void)
3660 void cfs_bandwidth_usage_inc(void) {}
3661 void cfs_bandwidth_usage_dec(void) {}
3662 #endif /* HAVE_JUMP_LABEL */
3665 * default period for cfs group bandwidth.
3666 * default: 0.1s, units: nanoseconds
3668 static inline u64 default_cfs_period(void)
3670 return 100000000ULL;
3673 static inline u64 sched_cfs_bandwidth_slice(void)
3675 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3679 * Replenish runtime according to assigned quota and update expiration time.
3680 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3681 * additional synchronization around rq->lock.
3683 * requires cfs_b->lock
3685 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3689 if (cfs_b->quota == RUNTIME_INF)
3692 now = sched_clock_cpu(smp_processor_id());
3693 cfs_b->runtime = cfs_b->quota;
3694 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3697 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3699 return &tg->cfs_bandwidth;
3702 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3703 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3705 if (unlikely(cfs_rq->throttle_count))
3706 return cfs_rq->throttled_clock_task;
3708 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3711 /* returns 0 on failure to allocate runtime */
3712 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3714 struct task_group *tg = cfs_rq->tg;
3715 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3716 u64 amount = 0, min_amount, expires;
3718 /* note: this is a positive sum as runtime_remaining <= 0 */
3719 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3721 raw_spin_lock(&cfs_b->lock);
3722 if (cfs_b->quota == RUNTIME_INF)
3723 amount = min_amount;
3725 start_cfs_bandwidth(cfs_b);
3727 if (cfs_b->runtime > 0) {
3728 amount = min(cfs_b->runtime, min_amount);
3729 cfs_b->runtime -= amount;
3733 expires = cfs_b->runtime_expires;
3734 raw_spin_unlock(&cfs_b->lock);
3736 cfs_rq->runtime_remaining += amount;
3738 * we may have advanced our local expiration to account for allowed
3739 * spread between our sched_clock and the one on which runtime was
3742 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3743 cfs_rq->runtime_expires = expires;
3745 return cfs_rq->runtime_remaining > 0;
3749 * Note: This depends on the synchronization provided by sched_clock and the
3750 * fact that rq->clock snapshots this value.
3752 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3754 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3756 /* if the deadline is ahead of our clock, nothing to do */
3757 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3760 if (cfs_rq->runtime_remaining < 0)
3764 * If the local deadline has passed we have to consider the
3765 * possibility that our sched_clock is 'fast' and the global deadline
3766 * has not truly expired.
3768 * Fortunately we can check determine whether this the case by checking
3769 * whether the global deadline has advanced. It is valid to compare
3770 * cfs_b->runtime_expires without any locks since we only care about
3771 * exact equality, so a partial write will still work.
3774 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3775 /* extend local deadline, drift is bounded above by 2 ticks */
3776 cfs_rq->runtime_expires += TICK_NSEC;
3778 /* global deadline is ahead, expiration has passed */
3779 cfs_rq->runtime_remaining = 0;
3783 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3785 /* dock delta_exec before expiring quota (as it could span periods) */
3786 cfs_rq->runtime_remaining -= delta_exec;
3787 expire_cfs_rq_runtime(cfs_rq);
3789 if (likely(cfs_rq->runtime_remaining > 0))
3793 * if we're unable to extend our runtime we resched so that the active
3794 * hierarchy can be throttled
3796 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3797 resched_curr(rq_of(cfs_rq));
3800 static __always_inline
3801 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3803 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3806 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3809 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3811 return cfs_bandwidth_used() && cfs_rq->throttled;
3814 /* check whether cfs_rq, or any parent, is throttled */
3815 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3817 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3821 * Ensure that neither of the group entities corresponding to src_cpu or
3822 * dest_cpu are members of a throttled hierarchy when performing group
3823 * load-balance operations.
3825 static inline int throttled_lb_pair(struct task_group *tg,
3826 int src_cpu, int dest_cpu)
3828 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3830 src_cfs_rq = tg->cfs_rq[src_cpu];
3831 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3833 return throttled_hierarchy(src_cfs_rq) ||
3834 throttled_hierarchy(dest_cfs_rq);
3837 /* updated child weight may affect parent so we have to do this bottom up */
3838 static int tg_unthrottle_up(struct task_group *tg, void *data)
3840 struct rq *rq = data;
3841 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3843 cfs_rq->throttle_count--;
3845 if (!cfs_rq->throttle_count) {
3846 /* adjust cfs_rq_clock_task() */
3847 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3848 cfs_rq->throttled_clock_task;
3855 static int tg_throttle_down(struct task_group *tg, void *data)
3857 struct rq *rq = data;
3858 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3860 /* group is entering throttled state, stop time */
3861 if (!cfs_rq->throttle_count)
3862 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3863 cfs_rq->throttle_count++;
3868 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3870 struct rq *rq = rq_of(cfs_rq);
3871 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3872 struct sched_entity *se;
3873 long task_delta, dequeue = 1;
3876 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3878 /* freeze hierarchy runnable averages while throttled */
3880 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3883 task_delta = cfs_rq->h_nr_running;
3884 for_each_sched_entity(se) {
3885 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3886 /* throttled entity or throttle-on-deactivate */
3891 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3892 qcfs_rq->h_nr_running -= task_delta;
3894 if (qcfs_rq->load.weight)
3899 sub_nr_running(rq, task_delta);
3901 cfs_rq->throttled = 1;
3902 cfs_rq->throttled_clock = rq_clock(rq);
3903 raw_spin_lock(&cfs_b->lock);
3904 empty = list_empty(&cfs_b->throttled_cfs_rq);
3907 * Add to the _head_ of the list, so that an already-started
3908 * distribute_cfs_runtime will not see us
3910 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3913 * If we're the first throttled task, make sure the bandwidth
3917 start_cfs_bandwidth(cfs_b);
3919 raw_spin_unlock(&cfs_b->lock);
3922 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3924 struct rq *rq = rq_of(cfs_rq);
3925 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3926 struct sched_entity *se;
3930 se = cfs_rq->tg->se[cpu_of(rq)];
3932 cfs_rq->throttled = 0;
3934 update_rq_clock(rq);
3936 raw_spin_lock(&cfs_b->lock);
3937 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3938 list_del_rcu(&cfs_rq->throttled_list);
3939 raw_spin_unlock(&cfs_b->lock);
3941 /* update hierarchical throttle state */
3942 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3944 if (!cfs_rq->load.weight)
3947 task_delta = cfs_rq->h_nr_running;
3948 for_each_sched_entity(se) {
3952 cfs_rq = cfs_rq_of(se);
3954 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3955 cfs_rq->h_nr_running += task_delta;
3957 if (cfs_rq_throttled(cfs_rq))
3962 add_nr_running(rq, task_delta);
3964 /* determine whether we need to wake up potentially idle cpu */
3965 if (rq->curr == rq->idle && rq->cfs.nr_running)
3969 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3970 u64 remaining, u64 expires)
3972 struct cfs_rq *cfs_rq;
3974 u64 starting_runtime = remaining;
3977 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3979 struct rq *rq = rq_of(cfs_rq);
3981 raw_spin_lock(&rq->lock);
3982 if (!cfs_rq_throttled(cfs_rq))
3985 runtime = -cfs_rq->runtime_remaining + 1;
3986 if (runtime > remaining)
3987 runtime = remaining;
3988 remaining -= runtime;
3990 cfs_rq->runtime_remaining += runtime;
3991 cfs_rq->runtime_expires = expires;
3993 /* we check whether we're throttled above */
3994 if (cfs_rq->runtime_remaining > 0)
3995 unthrottle_cfs_rq(cfs_rq);
3998 raw_spin_unlock(&rq->lock);
4005 return starting_runtime - remaining;
4009 * Responsible for refilling a task_group's bandwidth and unthrottling its
4010 * cfs_rqs as appropriate. If there has been no activity within the last
4011 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4012 * used to track this state.
4014 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4016 u64 runtime, runtime_expires;
4019 /* no need to continue the timer with no bandwidth constraint */
4020 if (cfs_b->quota == RUNTIME_INF)
4021 goto out_deactivate;
4023 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4024 cfs_b->nr_periods += overrun;
4027 * idle depends on !throttled (for the case of a large deficit), and if
4028 * we're going inactive then everything else can be deferred
4030 if (cfs_b->idle && !throttled)
4031 goto out_deactivate;
4033 __refill_cfs_bandwidth_runtime(cfs_b);
4036 /* mark as potentially idle for the upcoming period */
4041 /* account preceding periods in which throttling occurred */
4042 cfs_b->nr_throttled += overrun;
4044 runtime_expires = cfs_b->runtime_expires;
4047 * This check is repeated as we are holding onto the new bandwidth while
4048 * we unthrottle. This can potentially race with an unthrottled group
4049 * trying to acquire new bandwidth from the global pool. This can result
4050 * in us over-using our runtime if it is all used during this loop, but
4051 * only by limited amounts in that extreme case.
4053 while (throttled && cfs_b->runtime > 0) {
4054 runtime = cfs_b->runtime;
4055 raw_spin_unlock(&cfs_b->lock);
4056 /* we can't nest cfs_b->lock while distributing bandwidth */
4057 runtime = distribute_cfs_runtime(cfs_b, runtime,
4059 raw_spin_lock(&cfs_b->lock);
4061 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4063 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4067 * While we are ensured activity in the period following an
4068 * unthrottle, this also covers the case in which the new bandwidth is
4069 * insufficient to cover the existing bandwidth deficit. (Forcing the
4070 * timer to remain active while there are any throttled entities.)
4080 /* a cfs_rq won't donate quota below this amount */
4081 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4082 /* minimum remaining period time to redistribute slack quota */
4083 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4084 /* how long we wait to gather additional slack before distributing */
4085 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4088 * Are we near the end of the current quota period?
4090 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4091 * hrtimer base being cleared by hrtimer_start. In the case of
4092 * migrate_hrtimers, base is never cleared, so we are fine.
4094 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4096 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4099 /* if the call-back is running a quota refresh is already occurring */
4100 if (hrtimer_callback_running(refresh_timer))
4103 /* is a quota refresh about to occur? */
4104 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4105 if (remaining < min_expire)
4111 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4113 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4115 /* if there's a quota refresh soon don't bother with slack */
4116 if (runtime_refresh_within(cfs_b, min_left))
4119 hrtimer_start(&cfs_b->slack_timer,
4120 ns_to_ktime(cfs_bandwidth_slack_period),
4124 /* we know any runtime found here is valid as update_curr() precedes return */
4125 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4127 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4128 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4130 if (slack_runtime <= 0)
4133 raw_spin_lock(&cfs_b->lock);
4134 if (cfs_b->quota != RUNTIME_INF &&
4135 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4136 cfs_b->runtime += slack_runtime;
4138 /* we are under rq->lock, defer unthrottling using a timer */
4139 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4140 !list_empty(&cfs_b->throttled_cfs_rq))
4141 start_cfs_slack_bandwidth(cfs_b);
4143 raw_spin_unlock(&cfs_b->lock);
4145 /* even if it's not valid for return we don't want to try again */
4146 cfs_rq->runtime_remaining -= slack_runtime;
4149 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4151 if (!cfs_bandwidth_used())
4154 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4157 __return_cfs_rq_runtime(cfs_rq);
4161 * This is done with a timer (instead of inline with bandwidth return) since
4162 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4164 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4166 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4169 /* confirm we're still not at a refresh boundary */
4170 raw_spin_lock(&cfs_b->lock);
4171 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4172 raw_spin_unlock(&cfs_b->lock);
4176 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4177 runtime = cfs_b->runtime;
4179 expires = cfs_b->runtime_expires;
4180 raw_spin_unlock(&cfs_b->lock);
4185 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4187 raw_spin_lock(&cfs_b->lock);
4188 if (expires == cfs_b->runtime_expires)
4189 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4190 raw_spin_unlock(&cfs_b->lock);
4194 * When a group wakes up we want to make sure that its quota is not already
4195 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4196 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4198 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4200 if (!cfs_bandwidth_used())
4203 /* Synchronize hierarchical throttle counter: */
4204 if (unlikely(!cfs_rq->throttle_uptodate)) {
4205 struct rq *rq = rq_of(cfs_rq);
4206 struct cfs_rq *pcfs_rq;
4207 struct task_group *tg;
4209 cfs_rq->throttle_uptodate = 1;
4211 /* Get closest up-to-date node, because leaves go first: */
4212 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4213 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4214 if (pcfs_rq->throttle_uptodate)
4218 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4219 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4223 /* an active group must be handled by the update_curr()->put() path */
4224 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4227 /* ensure the group is not already throttled */
4228 if (cfs_rq_throttled(cfs_rq))
4231 /* update runtime allocation */
4232 account_cfs_rq_runtime(cfs_rq, 0);
4233 if (cfs_rq->runtime_remaining <= 0)
4234 throttle_cfs_rq(cfs_rq);
4237 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4238 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4240 if (!cfs_bandwidth_used())
4243 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4247 * it's possible for a throttled entity to be forced into a running
4248 * state (e.g. set_curr_task), in this case we're finished.
4250 if (cfs_rq_throttled(cfs_rq))
4253 throttle_cfs_rq(cfs_rq);
4257 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4259 struct cfs_bandwidth *cfs_b =
4260 container_of(timer, struct cfs_bandwidth, slack_timer);
4262 do_sched_cfs_slack_timer(cfs_b);
4264 return HRTIMER_NORESTART;
4267 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4269 struct cfs_bandwidth *cfs_b =
4270 container_of(timer, struct cfs_bandwidth, period_timer);
4274 raw_spin_lock(&cfs_b->lock);
4276 overrun = hrtimer_forward_now(timer, cfs_b->period);
4280 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4283 cfs_b->period_active = 0;
4284 raw_spin_unlock(&cfs_b->lock);
4286 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4289 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4291 raw_spin_lock_init(&cfs_b->lock);
4293 cfs_b->quota = RUNTIME_INF;
4294 cfs_b->period = ns_to_ktime(default_cfs_period());
4296 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4297 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4298 cfs_b->period_timer.function = sched_cfs_period_timer;
4299 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4300 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4303 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4305 cfs_rq->runtime_enabled = 0;
4306 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4309 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4311 lockdep_assert_held(&cfs_b->lock);
4313 if (!cfs_b->period_active) {
4314 cfs_b->period_active = 1;
4315 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4316 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4320 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4322 /* init_cfs_bandwidth() was not called */
4323 if (!cfs_b->throttled_cfs_rq.next)
4326 hrtimer_cancel(&cfs_b->period_timer);
4327 hrtimer_cancel(&cfs_b->slack_timer);
4330 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4332 struct cfs_rq *cfs_rq;
4334 for_each_leaf_cfs_rq(rq, cfs_rq) {
4335 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4337 raw_spin_lock(&cfs_b->lock);
4338 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4339 raw_spin_unlock(&cfs_b->lock);
4343 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4345 struct cfs_rq *cfs_rq;
4347 for_each_leaf_cfs_rq(rq, cfs_rq) {
4348 if (!cfs_rq->runtime_enabled)
4352 * clock_task is not advancing so we just need to make sure
4353 * there's some valid quota amount
4355 cfs_rq->runtime_remaining = 1;
4357 * Offline rq is schedulable till cpu is completely disabled
4358 * in take_cpu_down(), so we prevent new cfs throttling here.
4360 cfs_rq->runtime_enabled = 0;
4362 if (cfs_rq_throttled(cfs_rq))
4363 unthrottle_cfs_rq(cfs_rq);
4367 #else /* CONFIG_CFS_BANDWIDTH */
4368 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4370 return rq_clock_task(rq_of(cfs_rq));
4373 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4374 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4375 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4376 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4378 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4383 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4388 static inline int throttled_lb_pair(struct task_group *tg,
4389 int src_cpu, int dest_cpu)
4394 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4396 #ifdef CONFIG_FAIR_GROUP_SCHED
4397 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4400 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4404 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4405 static inline void update_runtime_enabled(struct rq *rq) {}
4406 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4408 #endif /* CONFIG_CFS_BANDWIDTH */
4410 /**************************************************
4411 * CFS operations on tasks:
4414 #ifdef CONFIG_SCHED_HRTICK
4415 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4417 struct sched_entity *se = &p->se;
4418 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4420 WARN_ON(task_rq(p) != rq);
4422 if (cfs_rq->nr_running > 1) {
4423 u64 slice = sched_slice(cfs_rq, se);
4424 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4425 s64 delta = slice - ran;
4432 hrtick_start(rq, delta);
4437 * called from enqueue/dequeue and updates the hrtick when the
4438 * current task is from our class and nr_running is low enough
4441 static void hrtick_update(struct rq *rq)
4443 struct task_struct *curr = rq->curr;
4445 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4448 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4449 hrtick_start_fair(rq, curr);
4451 #else /* !CONFIG_SCHED_HRTICK */
4453 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4457 static inline void hrtick_update(struct rq *rq)
4463 * The enqueue_task method is called before nr_running is
4464 * increased. Here we update the fair scheduling stats and
4465 * then put the task into the rbtree:
4468 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4470 struct cfs_rq *cfs_rq;
4471 struct sched_entity *se = &p->se;
4473 for_each_sched_entity(se) {
4476 cfs_rq = cfs_rq_of(se);
4477 enqueue_entity(cfs_rq, se, flags);
4480 * end evaluation on encountering a throttled cfs_rq
4482 * note: in the case of encountering a throttled cfs_rq we will
4483 * post the final h_nr_running increment below.
4485 if (cfs_rq_throttled(cfs_rq))
4487 cfs_rq->h_nr_running++;
4489 flags = ENQUEUE_WAKEUP;
4492 for_each_sched_entity(se) {
4493 cfs_rq = cfs_rq_of(se);
4494 cfs_rq->h_nr_running++;
4496 if (cfs_rq_throttled(cfs_rq))
4499 update_load_avg(se, 1);
4500 update_cfs_shares(cfs_rq);
4504 add_nr_running(rq, 1);
4509 static void set_next_buddy(struct sched_entity *se);
4512 * The dequeue_task method is called before nr_running is
4513 * decreased. We remove the task from the rbtree and
4514 * update the fair scheduling stats:
4516 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4518 struct cfs_rq *cfs_rq;
4519 struct sched_entity *se = &p->se;
4520 int task_sleep = flags & DEQUEUE_SLEEP;
4522 for_each_sched_entity(se) {
4523 cfs_rq = cfs_rq_of(se);
4524 dequeue_entity(cfs_rq, se, flags);
4527 * end evaluation on encountering a throttled cfs_rq
4529 * note: in the case of encountering a throttled cfs_rq we will
4530 * post the final h_nr_running decrement below.
4532 if (cfs_rq_throttled(cfs_rq))
4534 cfs_rq->h_nr_running--;
4536 /* Don't dequeue parent if it has other entities besides us */
4537 if (cfs_rq->load.weight) {
4538 /* Avoid re-evaluating load for this entity: */
4539 se = parent_entity(se);
4541 * Bias pick_next to pick a task from this cfs_rq, as
4542 * p is sleeping when it is within its sched_slice.
4544 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4548 flags |= DEQUEUE_SLEEP;
4551 for_each_sched_entity(se) {
4552 cfs_rq = cfs_rq_of(se);
4553 cfs_rq->h_nr_running--;
4555 if (cfs_rq_throttled(cfs_rq))
4558 update_load_avg(se, 1);
4559 update_cfs_shares(cfs_rq);
4563 sub_nr_running(rq, 1);
4569 #ifdef CONFIG_NO_HZ_COMMON
4571 * per rq 'load' arrray crap; XXX kill this.
4575 * The exact cpuload calculated at every tick would be:
4577 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4579 * If a cpu misses updates for n ticks (as it was idle) and update gets
4580 * called on the n+1-th tick when cpu may be busy, then we have:
4582 * load_n = (1 - 1/2^i)^n * load_0
4583 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4585 * decay_load_missed() below does efficient calculation of
4587 * load' = (1 - 1/2^i)^n * load
4589 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4590 * This allows us to precompute the above in said factors, thereby allowing the
4591 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4592 * fixed_power_int())
4594 * The calculation is approximated on a 128 point scale.
4596 #define DEGRADE_SHIFT 7
4598 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4599 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4600 { 0, 0, 0, 0, 0, 0, 0, 0 },
4601 { 64, 32, 8, 0, 0, 0, 0, 0 },
4602 { 96, 72, 40, 12, 1, 0, 0, 0 },
4603 { 112, 98, 75, 43, 15, 1, 0, 0 },
4604 { 120, 112, 98, 76, 45, 16, 2, 0 }
4608 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4609 * would be when CPU is idle and so we just decay the old load without
4610 * adding any new load.
4612 static unsigned long
4613 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4617 if (!missed_updates)
4620 if (missed_updates >= degrade_zero_ticks[idx])
4624 return load >> missed_updates;
4626 while (missed_updates) {
4627 if (missed_updates % 2)
4628 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4630 missed_updates >>= 1;
4635 #endif /* CONFIG_NO_HZ_COMMON */
4638 * __cpu_load_update - update the rq->cpu_load[] statistics
4639 * @this_rq: The rq to update statistics for
4640 * @this_load: The current load
4641 * @pending_updates: The number of missed updates
4643 * Update rq->cpu_load[] statistics. This function is usually called every
4644 * scheduler tick (TICK_NSEC).
4646 * This function computes a decaying average:
4648 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4650 * Because of NOHZ it might not get called on every tick which gives need for
4651 * the @pending_updates argument.
4653 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4654 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4655 * = A * (A * load[i]_n-2 + B) + B
4656 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4657 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4658 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4659 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4660 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4662 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4663 * any change in load would have resulted in the tick being turned back on.
4665 * For regular NOHZ, this reduces to:
4667 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4669 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4672 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4673 unsigned long pending_updates)
4675 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4678 this_rq->nr_load_updates++;
4680 /* Update our load: */
4681 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4682 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4683 unsigned long old_load, new_load;
4685 /* scale is effectively 1 << i now, and >> i divides by scale */
4687 old_load = this_rq->cpu_load[i];
4688 #ifdef CONFIG_NO_HZ_COMMON
4689 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4690 if (tickless_load) {
4691 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4693 * old_load can never be a negative value because a
4694 * decayed tickless_load cannot be greater than the
4695 * original tickless_load.
4697 old_load += tickless_load;
4700 new_load = this_load;
4702 * Round up the averaging division if load is increasing. This
4703 * prevents us from getting stuck on 9 if the load is 10, for
4706 if (new_load > old_load)
4707 new_load += scale - 1;
4709 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4712 sched_avg_update(this_rq);
4715 /* Used instead of source_load when we know the type == 0 */
4716 static unsigned long weighted_cpuload(const int cpu)
4718 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4721 #ifdef CONFIG_NO_HZ_COMMON
4723 * There is no sane way to deal with nohz on smp when using jiffies because the
4724 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4725 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4727 * Therefore we need to avoid the delta approach from the regular tick when
4728 * possible since that would seriously skew the load calculation. This is why we
4729 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4730 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4731 * loop exit, nohz_idle_balance, nohz full exit...)
4733 * This means we might still be one tick off for nohz periods.
4736 static void cpu_load_update_nohz(struct rq *this_rq,
4737 unsigned long curr_jiffies,
4740 unsigned long pending_updates;
4742 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4743 if (pending_updates) {
4744 this_rq->last_load_update_tick = curr_jiffies;
4746 * In the regular NOHZ case, we were idle, this means load 0.
4747 * In the NOHZ_FULL case, we were non-idle, we should consider
4748 * its weighted load.
4750 cpu_load_update(this_rq, load, pending_updates);
4755 * Called from nohz_idle_balance() to update the load ratings before doing the
4758 static void cpu_load_update_idle(struct rq *this_rq)
4761 * bail if there's load or we're actually up-to-date.
4763 if (weighted_cpuload(cpu_of(this_rq)))
4766 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4770 * Record CPU load on nohz entry so we know the tickless load to account
4771 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4772 * than other cpu_load[idx] but it should be fine as cpu_load readers
4773 * shouldn't rely into synchronized cpu_load[*] updates.
4775 void cpu_load_update_nohz_start(void)
4777 struct rq *this_rq = this_rq();
4780 * This is all lockless but should be fine. If weighted_cpuload changes
4781 * concurrently we'll exit nohz. And cpu_load write can race with
4782 * cpu_load_update_idle() but both updater would be writing the same.
4784 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4788 * Account the tickless load in the end of a nohz frame.
4790 void cpu_load_update_nohz_stop(void)
4792 unsigned long curr_jiffies = READ_ONCE(jiffies);
4793 struct rq *this_rq = this_rq();
4796 if (curr_jiffies == this_rq->last_load_update_tick)
4799 load = weighted_cpuload(cpu_of(this_rq));
4800 raw_spin_lock(&this_rq->lock);
4801 update_rq_clock(this_rq);
4802 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4803 raw_spin_unlock(&this_rq->lock);
4805 #else /* !CONFIG_NO_HZ_COMMON */
4806 static inline void cpu_load_update_nohz(struct rq *this_rq,
4807 unsigned long curr_jiffies,
4808 unsigned long load) { }
4809 #endif /* CONFIG_NO_HZ_COMMON */
4811 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4813 #ifdef CONFIG_NO_HZ_COMMON
4814 /* See the mess around cpu_load_update_nohz(). */
4815 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4817 cpu_load_update(this_rq, load, 1);
4821 * Called from scheduler_tick()
4823 void cpu_load_update_active(struct rq *this_rq)
4825 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4827 if (tick_nohz_tick_stopped())
4828 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4830 cpu_load_update_periodic(this_rq, load);
4834 * Return a low guess at the load of a migration-source cpu weighted
4835 * according to the scheduling class and "nice" value.
4837 * We want to under-estimate the load of migration sources, to
4838 * balance conservatively.
4840 static unsigned long source_load(int cpu, int type)
4842 struct rq *rq = cpu_rq(cpu);
4843 unsigned long total = weighted_cpuload(cpu);
4845 if (type == 0 || !sched_feat(LB_BIAS))
4848 return min(rq->cpu_load[type-1], total);
4852 * Return a high guess at the load of a migration-target cpu weighted
4853 * according to the scheduling class and "nice" value.
4855 static unsigned long target_load(int cpu, int type)
4857 struct rq *rq = cpu_rq(cpu);
4858 unsigned long total = weighted_cpuload(cpu);
4860 if (type == 0 || !sched_feat(LB_BIAS))
4863 return max(rq->cpu_load[type-1], total);
4866 static unsigned long capacity_of(int cpu)
4868 return cpu_rq(cpu)->cpu_capacity;
4871 static unsigned long capacity_orig_of(int cpu)
4873 return cpu_rq(cpu)->cpu_capacity_orig;
4876 static unsigned long cpu_avg_load_per_task(int cpu)
4878 struct rq *rq = cpu_rq(cpu);
4879 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4880 unsigned long load_avg = weighted_cpuload(cpu);
4883 return load_avg / nr_running;
4888 #ifdef CONFIG_FAIR_GROUP_SCHED
4890 * effective_load() calculates the load change as seen from the root_task_group
4892 * Adding load to a group doesn't make a group heavier, but can cause movement
4893 * of group shares between cpus. Assuming the shares were perfectly aligned one
4894 * can calculate the shift in shares.
4896 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4897 * on this @cpu and results in a total addition (subtraction) of @wg to the
4898 * total group weight.
4900 * Given a runqueue weight distribution (rw_i) we can compute a shares
4901 * distribution (s_i) using:
4903 * s_i = rw_i / \Sum rw_j (1)
4905 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4906 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4907 * shares distribution (s_i):
4909 * rw_i = { 2, 4, 1, 0 }
4910 * s_i = { 2/7, 4/7, 1/7, 0 }
4912 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4913 * task used to run on and the CPU the waker is running on), we need to
4914 * compute the effect of waking a task on either CPU and, in case of a sync
4915 * wakeup, compute the effect of the current task going to sleep.
4917 * So for a change of @wl to the local @cpu with an overall group weight change
4918 * of @wl we can compute the new shares distribution (s'_i) using:
4920 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4922 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4923 * differences in waking a task to CPU 0. The additional task changes the
4924 * weight and shares distributions like:
4926 * rw'_i = { 3, 4, 1, 0 }
4927 * s'_i = { 3/8, 4/8, 1/8, 0 }
4929 * We can then compute the difference in effective weight by using:
4931 * dw_i = S * (s'_i - s_i) (3)
4933 * Where 'S' is the group weight as seen by its parent.
4935 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4936 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4937 * 4/7) times the weight of the group.
4939 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4941 struct sched_entity *se = tg->se[cpu];
4943 if (!tg->parent) /* the trivial, non-cgroup case */
4946 for_each_sched_entity(se) {
4947 struct cfs_rq *cfs_rq = se->my_q;
4948 long W, w = cfs_rq_load_avg(cfs_rq);
4953 * W = @wg + \Sum rw_j
4955 W = wg + atomic_long_read(&tg->load_avg);
4957 /* Ensure \Sum rw_j >= rw_i */
4958 W -= cfs_rq->tg_load_avg_contrib;
4967 * wl = S * s'_i; see (2)
4970 wl = (w * (long)tg->shares) / W;
4975 * Per the above, wl is the new se->load.weight value; since
4976 * those are clipped to [MIN_SHARES, ...) do so now. See
4977 * calc_cfs_shares().
4979 if (wl < MIN_SHARES)
4983 * wl = dw_i = S * (s'_i - s_i); see (3)
4985 wl -= se->avg.load_avg;
4988 * Recursively apply this logic to all parent groups to compute
4989 * the final effective load change on the root group. Since
4990 * only the @tg group gets extra weight, all parent groups can
4991 * only redistribute existing shares. @wl is the shift in shares
4992 * resulting from this level per the above.
5001 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5008 static void record_wakee(struct task_struct *p)
5011 * Only decay a single time; tasks that have less then 1 wakeup per
5012 * jiffy will not have built up many flips.
5014 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5015 current->wakee_flips >>= 1;
5016 current->wakee_flip_decay_ts = jiffies;
5019 if (current->last_wakee != p) {
5020 current->last_wakee = p;
5021 current->wakee_flips++;
5026 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5028 * A waker of many should wake a different task than the one last awakened
5029 * at a frequency roughly N times higher than one of its wakees.
5031 * In order to determine whether we should let the load spread vs consolidating
5032 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5033 * partner, and a factor of lls_size higher frequency in the other.
5035 * With both conditions met, we can be relatively sure that the relationship is
5036 * non-monogamous, with partner count exceeding socket size.
5038 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5039 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5042 static int wake_wide(struct task_struct *p)
5044 unsigned int master = current->wakee_flips;
5045 unsigned int slave = p->wakee_flips;
5046 int factor = this_cpu_read(sd_llc_size);
5049 swap(master, slave);
5050 if (slave < factor || master < slave * factor)
5055 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5057 s64 this_load, load;
5058 s64 this_eff_load, prev_eff_load;
5059 int idx, this_cpu, prev_cpu;
5060 struct task_group *tg;
5061 unsigned long weight;
5065 this_cpu = smp_processor_id();
5066 prev_cpu = task_cpu(p);
5067 load = source_load(prev_cpu, idx);
5068 this_load = target_load(this_cpu, idx);
5071 * If sync wakeup then subtract the (maximum possible)
5072 * effect of the currently running task from the load
5073 * of the current CPU:
5076 tg = task_group(current);
5077 weight = current->se.avg.load_avg;
5079 this_load += effective_load(tg, this_cpu, -weight, -weight);
5080 load += effective_load(tg, prev_cpu, 0, -weight);
5084 weight = p->se.avg.load_avg;
5087 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5088 * due to the sync cause above having dropped this_load to 0, we'll
5089 * always have an imbalance, but there's really nothing you can do
5090 * about that, so that's good too.
5092 * Otherwise check if either cpus are near enough in load to allow this
5093 * task to be woken on this_cpu.
5095 this_eff_load = 100;
5096 this_eff_load *= capacity_of(prev_cpu);
5098 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5099 prev_eff_load *= capacity_of(this_cpu);
5101 if (this_load > 0) {
5102 this_eff_load *= this_load +
5103 effective_load(tg, this_cpu, weight, weight);
5105 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5108 balanced = this_eff_load <= prev_eff_load;
5110 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5115 schedstat_inc(sd, ttwu_move_affine);
5116 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5122 * find_idlest_group finds and returns the least busy CPU group within the
5125 static struct sched_group *
5126 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5127 int this_cpu, int sd_flag)
5129 struct sched_group *idlest = NULL, *group = sd->groups;
5130 unsigned long min_load = ULONG_MAX, this_load = 0;
5131 int load_idx = sd->forkexec_idx;
5132 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5134 if (sd_flag & SD_BALANCE_WAKE)
5135 load_idx = sd->wake_idx;
5138 unsigned long load, avg_load;
5142 /* Skip over this group if it has no CPUs allowed */
5143 if (!cpumask_intersects(sched_group_cpus(group),
5144 tsk_cpus_allowed(p)))
5147 local_group = cpumask_test_cpu(this_cpu,
5148 sched_group_cpus(group));
5150 /* Tally up the load of all CPUs in the group */
5153 for_each_cpu(i, sched_group_cpus(group)) {
5154 /* Bias balancing toward cpus of our domain */
5156 load = source_load(i, load_idx);
5158 load = target_load(i, load_idx);
5163 /* Adjust by relative CPU capacity of the group */
5164 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5167 this_load = avg_load;
5168 } else if (avg_load < min_load) {
5169 min_load = avg_load;
5172 } while (group = group->next, group != sd->groups);
5174 if (!idlest || 100*this_load < imbalance*min_load)
5180 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5183 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5185 unsigned long load, min_load = ULONG_MAX;
5186 unsigned int min_exit_latency = UINT_MAX;
5187 u64 latest_idle_timestamp = 0;
5188 int least_loaded_cpu = this_cpu;
5189 int shallowest_idle_cpu = -1;
5192 /* Traverse only the allowed CPUs */
5193 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5195 struct rq *rq = cpu_rq(i);
5196 struct cpuidle_state *idle = idle_get_state(rq);
5197 if (idle && idle->exit_latency < min_exit_latency) {
5199 * We give priority to a CPU whose idle state
5200 * has the smallest exit latency irrespective
5201 * of any idle timestamp.
5203 min_exit_latency = idle->exit_latency;
5204 latest_idle_timestamp = rq->idle_stamp;
5205 shallowest_idle_cpu = i;
5206 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5207 rq->idle_stamp > latest_idle_timestamp) {
5209 * If equal or no active idle state, then
5210 * the most recently idled CPU might have
5213 latest_idle_timestamp = rq->idle_stamp;
5214 shallowest_idle_cpu = i;
5216 } else if (shallowest_idle_cpu == -1) {
5217 load = weighted_cpuload(i);
5218 if (load < min_load || (load == min_load && i == this_cpu)) {
5220 least_loaded_cpu = i;
5225 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5229 * Try and locate an idle CPU in the sched_domain.
5231 static int select_idle_sibling(struct task_struct *p, int target)
5233 struct sched_domain *sd;
5234 struct sched_group *sg;
5235 int i = task_cpu(p);
5237 if (idle_cpu(target))
5241 * If the prevous cpu is cache affine and idle, don't be stupid.
5243 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5247 * Otherwise, iterate the domains and find an eligible idle cpu.
5249 * A completely idle sched group at higher domains is more
5250 * desirable than an idle group at a lower level, because lower
5251 * domains have smaller groups and usually share hardware
5252 * resources which causes tasks to contend on them, e.g. x86
5253 * hyperthread siblings in the lowest domain (SMT) can contend
5254 * on the shared cpu pipeline.
5256 * However, while we prefer idle groups at higher domains
5257 * finding an idle cpu at the lowest domain is still better than
5258 * returning 'target', which we've already established, isn't
5261 sd = rcu_dereference(per_cpu(sd_llc, target));
5262 for_each_lower_domain(sd) {
5265 if (!cpumask_intersects(sched_group_cpus(sg),
5266 tsk_cpus_allowed(p)))
5269 /* Ensure the entire group is idle */
5270 for_each_cpu(i, sched_group_cpus(sg)) {
5271 if (i == target || !idle_cpu(i))
5276 * It doesn't matter which cpu we pick, the
5277 * whole group is idle.
5279 target = cpumask_first_and(sched_group_cpus(sg),
5280 tsk_cpus_allowed(p));
5284 } while (sg != sd->groups);
5291 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5292 * tasks. The unit of the return value must be the one of capacity so we can
5293 * compare the utilization with the capacity of the CPU that is available for
5294 * CFS task (ie cpu_capacity).
5296 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5297 * recent utilization of currently non-runnable tasks on a CPU. It represents
5298 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5299 * capacity_orig is the cpu_capacity available at the highest frequency
5300 * (arch_scale_freq_capacity()).
5301 * The utilization of a CPU converges towards a sum equal to or less than the
5302 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5303 * the running time on this CPU scaled by capacity_curr.
5305 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5306 * higher than capacity_orig because of unfortunate rounding in
5307 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5308 * the average stabilizes with the new running time. We need to check that the
5309 * utilization stays within the range of [0..capacity_orig] and cap it if
5310 * necessary. Without utilization capping, a group could be seen as overloaded
5311 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5312 * available capacity. We allow utilization to overshoot capacity_curr (but not
5313 * capacity_orig) as it useful for predicting the capacity required after task
5314 * migrations (scheduler-driven DVFS).
5316 static int cpu_util(int cpu)
5318 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5319 unsigned long capacity = capacity_orig_of(cpu);
5321 return (util >= capacity) ? capacity : util;
5325 * select_task_rq_fair: Select target runqueue for the waking task in domains
5326 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5327 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5329 * Balances load by selecting the idlest cpu in the idlest group, or under
5330 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5332 * Returns the target cpu number.
5334 * preempt must be disabled.
5337 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5339 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5340 int cpu = smp_processor_id();
5341 int new_cpu = prev_cpu;
5342 int want_affine = 0;
5343 int sync = wake_flags & WF_SYNC;
5345 if (sd_flag & SD_BALANCE_WAKE) {
5347 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5351 for_each_domain(cpu, tmp) {
5352 if (!(tmp->flags & SD_LOAD_BALANCE))
5356 * If both cpu and prev_cpu are part of this domain,
5357 * cpu is a valid SD_WAKE_AFFINE target.
5359 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5360 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5365 if (tmp->flags & sd_flag)
5367 else if (!want_affine)
5372 sd = NULL; /* Prefer wake_affine over balance flags */
5373 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5378 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5379 new_cpu = select_idle_sibling(p, new_cpu);
5382 struct sched_group *group;
5385 if (!(sd->flags & sd_flag)) {
5390 group = find_idlest_group(sd, p, cpu, sd_flag);
5396 new_cpu = find_idlest_cpu(group, p, cpu);
5397 if (new_cpu == -1 || new_cpu == cpu) {
5398 /* Now try balancing at a lower domain level of cpu */
5403 /* Now try balancing at a lower domain level of new_cpu */
5405 weight = sd->span_weight;
5407 for_each_domain(cpu, tmp) {
5408 if (weight <= tmp->span_weight)
5410 if (tmp->flags & sd_flag)
5413 /* while loop will break here if sd == NULL */
5421 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5422 * cfs_rq_of(p) references at time of call are still valid and identify the
5423 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5425 static void migrate_task_rq_fair(struct task_struct *p)
5428 * As blocked tasks retain absolute vruntime the migration needs to
5429 * deal with this by subtracting the old and adding the new
5430 * min_vruntime -- the latter is done by enqueue_entity() when placing
5431 * the task on the new runqueue.
5433 if (p->state == TASK_WAKING) {
5434 struct sched_entity *se = &p->se;
5435 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5438 #ifndef CONFIG_64BIT
5439 u64 min_vruntime_copy;
5442 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5444 min_vruntime = cfs_rq->min_vruntime;
5445 } while (min_vruntime != min_vruntime_copy);
5447 min_vruntime = cfs_rq->min_vruntime;
5450 se->vruntime -= min_vruntime;
5454 * We are supposed to update the task to "current" time, then its up to date
5455 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5456 * what current time is, so simply throw away the out-of-date time. This
5457 * will result in the wakee task is less decayed, but giving the wakee more
5458 * load sounds not bad.
5460 remove_entity_load_avg(&p->se);
5462 /* Tell new CPU we are migrated */
5463 p->se.avg.last_update_time = 0;
5465 /* We have migrated, no longer consider this task hot */
5466 p->se.exec_start = 0;
5469 static void task_dead_fair(struct task_struct *p)
5471 remove_entity_load_avg(&p->se);
5473 #endif /* CONFIG_SMP */
5475 static unsigned long
5476 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5478 unsigned long gran = sysctl_sched_wakeup_granularity;
5481 * Since its curr running now, convert the gran from real-time
5482 * to virtual-time in his units.
5484 * By using 'se' instead of 'curr' we penalize light tasks, so
5485 * they get preempted easier. That is, if 'se' < 'curr' then
5486 * the resulting gran will be larger, therefore penalizing the
5487 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5488 * be smaller, again penalizing the lighter task.
5490 * This is especially important for buddies when the leftmost
5491 * task is higher priority than the buddy.
5493 return calc_delta_fair(gran, se);
5497 * Should 'se' preempt 'curr'.
5511 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5513 s64 gran, vdiff = curr->vruntime - se->vruntime;
5518 gran = wakeup_gran(curr, se);
5525 static void set_last_buddy(struct sched_entity *se)
5527 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5530 for_each_sched_entity(se)
5531 cfs_rq_of(se)->last = se;
5534 static void set_next_buddy(struct sched_entity *se)
5536 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5539 for_each_sched_entity(se)
5540 cfs_rq_of(se)->next = se;
5543 static void set_skip_buddy(struct sched_entity *se)
5545 for_each_sched_entity(se)
5546 cfs_rq_of(se)->skip = se;
5550 * Preempt the current task with a newly woken task if needed:
5552 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5554 struct task_struct *curr = rq->curr;
5555 struct sched_entity *se = &curr->se, *pse = &p->se;
5556 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5557 int scale = cfs_rq->nr_running >= sched_nr_latency;
5558 int next_buddy_marked = 0;
5560 if (unlikely(se == pse))
5564 * This is possible from callers such as attach_tasks(), in which we
5565 * unconditionally check_prempt_curr() after an enqueue (which may have
5566 * lead to a throttle). This both saves work and prevents false
5567 * next-buddy nomination below.
5569 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5572 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5573 set_next_buddy(pse);
5574 next_buddy_marked = 1;
5578 * We can come here with TIF_NEED_RESCHED already set from new task
5581 * Note: this also catches the edge-case of curr being in a throttled
5582 * group (e.g. via set_curr_task), since update_curr() (in the
5583 * enqueue of curr) will have resulted in resched being set. This
5584 * prevents us from potentially nominating it as a false LAST_BUDDY
5587 if (test_tsk_need_resched(curr))
5590 /* Idle tasks are by definition preempted by non-idle tasks. */
5591 if (unlikely(curr->policy == SCHED_IDLE) &&
5592 likely(p->policy != SCHED_IDLE))
5596 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5597 * is driven by the tick):
5599 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5602 find_matching_se(&se, &pse);
5603 update_curr(cfs_rq_of(se));
5605 if (wakeup_preempt_entity(se, pse) == 1) {
5607 * Bias pick_next to pick the sched entity that is
5608 * triggering this preemption.
5610 if (!next_buddy_marked)
5611 set_next_buddy(pse);
5620 * Only set the backward buddy when the current task is still
5621 * on the rq. This can happen when a wakeup gets interleaved
5622 * with schedule on the ->pre_schedule() or idle_balance()
5623 * point, either of which can * drop the rq lock.
5625 * Also, during early boot the idle thread is in the fair class,
5626 * for obvious reasons its a bad idea to schedule back to it.
5628 if (unlikely(!se->on_rq || curr == rq->idle))
5631 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5635 static struct task_struct *
5636 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5638 struct cfs_rq *cfs_rq = &rq->cfs;
5639 struct sched_entity *se;
5640 struct task_struct *p;
5644 #ifdef CONFIG_FAIR_GROUP_SCHED
5645 if (!cfs_rq->nr_running)
5648 if (prev->sched_class != &fair_sched_class)
5652 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5653 * likely that a next task is from the same cgroup as the current.
5655 * Therefore attempt to avoid putting and setting the entire cgroup
5656 * hierarchy, only change the part that actually changes.
5660 struct sched_entity *curr = cfs_rq->curr;
5663 * Since we got here without doing put_prev_entity() we also
5664 * have to consider cfs_rq->curr. If it is still a runnable
5665 * entity, update_curr() will update its vruntime, otherwise
5666 * forget we've ever seen it.
5670 update_curr(cfs_rq);
5675 * This call to check_cfs_rq_runtime() will do the
5676 * throttle and dequeue its entity in the parent(s).
5677 * Therefore the 'simple' nr_running test will indeed
5680 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5684 se = pick_next_entity(cfs_rq, curr);
5685 cfs_rq = group_cfs_rq(se);
5691 * Since we haven't yet done put_prev_entity and if the selected task
5692 * is a different task than we started out with, try and touch the
5693 * least amount of cfs_rqs.
5696 struct sched_entity *pse = &prev->se;
5698 while (!(cfs_rq = is_same_group(se, pse))) {
5699 int se_depth = se->depth;
5700 int pse_depth = pse->depth;
5702 if (se_depth <= pse_depth) {
5703 put_prev_entity(cfs_rq_of(pse), pse);
5704 pse = parent_entity(pse);
5706 if (se_depth >= pse_depth) {
5707 set_next_entity(cfs_rq_of(se), se);
5708 se = parent_entity(se);
5712 put_prev_entity(cfs_rq, pse);
5713 set_next_entity(cfs_rq, se);
5716 if (hrtick_enabled(rq))
5717 hrtick_start_fair(rq, p);
5724 if (!cfs_rq->nr_running)
5727 put_prev_task(rq, prev);
5730 se = pick_next_entity(cfs_rq, NULL);
5731 set_next_entity(cfs_rq, se);
5732 cfs_rq = group_cfs_rq(se);
5737 if (hrtick_enabled(rq))
5738 hrtick_start_fair(rq, p);
5744 * This is OK, because current is on_cpu, which avoids it being picked
5745 * for load-balance and preemption/IRQs are still disabled avoiding
5746 * further scheduler activity on it and we're being very careful to
5747 * re-start the picking loop.
5749 lockdep_unpin_lock(&rq->lock, cookie);
5750 new_tasks = idle_balance(rq);
5751 lockdep_repin_lock(&rq->lock, cookie);
5753 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5754 * possible for any higher priority task to appear. In that case we
5755 * must re-start the pick_next_entity() loop.
5767 * Account for a descheduled task:
5769 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5771 struct sched_entity *se = &prev->se;
5772 struct cfs_rq *cfs_rq;
5774 for_each_sched_entity(se) {
5775 cfs_rq = cfs_rq_of(se);
5776 put_prev_entity(cfs_rq, se);
5781 * sched_yield() is very simple
5783 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5785 static void yield_task_fair(struct rq *rq)
5787 struct task_struct *curr = rq->curr;
5788 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5789 struct sched_entity *se = &curr->se;
5792 * Are we the only task in the tree?
5794 if (unlikely(rq->nr_running == 1))
5797 clear_buddies(cfs_rq, se);
5799 if (curr->policy != SCHED_BATCH) {
5800 update_rq_clock(rq);
5802 * Update run-time statistics of the 'current'.
5804 update_curr(cfs_rq);
5806 * Tell update_rq_clock() that we've just updated,
5807 * so we don't do microscopic update in schedule()
5808 * and double the fastpath cost.
5810 rq_clock_skip_update(rq, true);
5816 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5818 struct sched_entity *se = &p->se;
5820 /* throttled hierarchies are not runnable */
5821 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5824 /* Tell the scheduler that we'd really like pse to run next. */
5827 yield_task_fair(rq);
5833 /**************************************************
5834 * Fair scheduling class load-balancing methods.
5838 * The purpose of load-balancing is to achieve the same basic fairness the
5839 * per-cpu scheduler provides, namely provide a proportional amount of compute
5840 * time to each task. This is expressed in the following equation:
5842 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5844 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5845 * W_i,0 is defined as:
5847 * W_i,0 = \Sum_j w_i,j (2)
5849 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5850 * is derived from the nice value as per sched_prio_to_weight[].
5852 * The weight average is an exponential decay average of the instantaneous
5855 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5857 * C_i is the compute capacity of cpu i, typically it is the
5858 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5859 * can also include other factors [XXX].
5861 * To achieve this balance we define a measure of imbalance which follows
5862 * directly from (1):
5864 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5866 * We them move tasks around to minimize the imbalance. In the continuous
5867 * function space it is obvious this converges, in the discrete case we get
5868 * a few fun cases generally called infeasible weight scenarios.
5871 * - infeasible weights;
5872 * - local vs global optima in the discrete case. ]
5877 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5878 * for all i,j solution, we create a tree of cpus that follows the hardware
5879 * topology where each level pairs two lower groups (or better). This results
5880 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5881 * tree to only the first of the previous level and we decrease the frequency
5882 * of load-balance at each level inv. proportional to the number of cpus in
5888 * \Sum { --- * --- * 2^i } = O(n) (5)
5890 * `- size of each group
5891 * | | `- number of cpus doing load-balance
5893 * `- sum over all levels
5895 * Coupled with a limit on how many tasks we can migrate every balance pass,
5896 * this makes (5) the runtime complexity of the balancer.
5898 * An important property here is that each CPU is still (indirectly) connected
5899 * to every other cpu in at most O(log n) steps:
5901 * The adjacency matrix of the resulting graph is given by:
5904 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5907 * And you'll find that:
5909 * A^(log_2 n)_i,j != 0 for all i,j (7)
5911 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5912 * The task movement gives a factor of O(m), giving a convergence complexity
5915 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5920 * In order to avoid CPUs going idle while there's still work to do, new idle
5921 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5922 * tree itself instead of relying on other CPUs to bring it work.
5924 * This adds some complexity to both (5) and (8) but it reduces the total idle
5932 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5935 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5940 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5942 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5944 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5947 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5948 * rewrite all of this once again.]
5951 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5953 enum fbq_type { regular, remote, all };
5955 #define LBF_ALL_PINNED 0x01
5956 #define LBF_NEED_BREAK 0x02
5957 #define LBF_DST_PINNED 0x04
5958 #define LBF_SOME_PINNED 0x08
5961 struct sched_domain *sd;
5969 struct cpumask *dst_grpmask;
5971 enum cpu_idle_type idle;
5973 /* The set of CPUs under consideration for load-balancing */
5974 struct cpumask *cpus;
5979 unsigned int loop_break;
5980 unsigned int loop_max;
5982 enum fbq_type fbq_type;
5983 struct list_head tasks;
5987 * Is this task likely cache-hot:
5989 static int task_hot(struct task_struct *p, struct lb_env *env)
5993 lockdep_assert_held(&env->src_rq->lock);
5995 if (p->sched_class != &fair_sched_class)
5998 if (unlikely(p->policy == SCHED_IDLE))
6002 * Buddy candidates are cache hot:
6004 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6005 (&p->se == cfs_rq_of(&p->se)->next ||
6006 &p->se == cfs_rq_of(&p->se)->last))
6009 if (sysctl_sched_migration_cost == -1)
6011 if (sysctl_sched_migration_cost == 0)
6014 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6016 return delta < (s64)sysctl_sched_migration_cost;
6019 #ifdef CONFIG_NUMA_BALANCING
6021 * Returns 1, if task migration degrades locality
6022 * Returns 0, if task migration improves locality i.e migration preferred.
6023 * Returns -1, if task migration is not affected by locality.
6025 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6027 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6028 unsigned long src_faults, dst_faults;
6029 int src_nid, dst_nid;
6031 if (!static_branch_likely(&sched_numa_balancing))
6034 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6037 src_nid = cpu_to_node(env->src_cpu);
6038 dst_nid = cpu_to_node(env->dst_cpu);
6040 if (src_nid == dst_nid)
6043 /* Migrating away from the preferred node is always bad. */
6044 if (src_nid == p->numa_preferred_nid) {
6045 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6051 /* Encourage migration to the preferred node. */
6052 if (dst_nid == p->numa_preferred_nid)
6056 src_faults = group_faults(p, src_nid);
6057 dst_faults = group_faults(p, dst_nid);
6059 src_faults = task_faults(p, src_nid);
6060 dst_faults = task_faults(p, dst_nid);
6063 return dst_faults < src_faults;
6067 static inline int migrate_degrades_locality(struct task_struct *p,
6075 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6078 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6082 lockdep_assert_held(&env->src_rq->lock);
6085 * We do not migrate tasks that are:
6086 * 1) throttled_lb_pair, or
6087 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6088 * 3) running (obviously), or
6089 * 4) are cache-hot on their current CPU.
6091 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6094 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6097 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6099 env->flags |= LBF_SOME_PINNED;
6102 * Remember if this task can be migrated to any other cpu in
6103 * our sched_group. We may want to revisit it if we couldn't
6104 * meet load balance goals by pulling other tasks on src_cpu.
6106 * Also avoid computing new_dst_cpu if we have already computed
6107 * one in current iteration.
6109 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6112 /* Prevent to re-select dst_cpu via env's cpus */
6113 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6114 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6115 env->flags |= LBF_DST_PINNED;
6116 env->new_dst_cpu = cpu;
6124 /* Record that we found atleast one task that could run on dst_cpu */
6125 env->flags &= ~LBF_ALL_PINNED;
6127 if (task_running(env->src_rq, p)) {
6128 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6133 * Aggressive migration if:
6134 * 1) destination numa is preferred
6135 * 2) task is cache cold, or
6136 * 3) too many balance attempts have failed.
6138 tsk_cache_hot = migrate_degrades_locality(p, env);
6139 if (tsk_cache_hot == -1)
6140 tsk_cache_hot = task_hot(p, env);
6142 if (tsk_cache_hot <= 0 ||
6143 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6144 if (tsk_cache_hot == 1) {
6145 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6146 schedstat_inc(p, se.statistics.nr_forced_migrations);
6151 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6156 * detach_task() -- detach the task for the migration specified in env
6158 static void detach_task(struct task_struct *p, struct lb_env *env)
6160 lockdep_assert_held(&env->src_rq->lock);
6162 p->on_rq = TASK_ON_RQ_MIGRATING;
6163 deactivate_task(env->src_rq, p, 0);
6164 set_task_cpu(p, env->dst_cpu);
6168 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6169 * part of active balancing operations within "domain".
6171 * Returns a task if successful and NULL otherwise.
6173 static struct task_struct *detach_one_task(struct lb_env *env)
6175 struct task_struct *p, *n;
6177 lockdep_assert_held(&env->src_rq->lock);
6179 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6180 if (!can_migrate_task(p, env))
6183 detach_task(p, env);
6186 * Right now, this is only the second place where
6187 * lb_gained[env->idle] is updated (other is detach_tasks)
6188 * so we can safely collect stats here rather than
6189 * inside detach_tasks().
6191 schedstat_inc(env->sd, lb_gained[env->idle]);
6197 static const unsigned int sched_nr_migrate_break = 32;
6200 * detach_tasks() -- tries to detach up to imbalance weighted load from
6201 * busiest_rq, as part of a balancing operation within domain "sd".
6203 * Returns number of detached tasks if successful and 0 otherwise.
6205 static int detach_tasks(struct lb_env *env)
6207 struct list_head *tasks = &env->src_rq->cfs_tasks;
6208 struct task_struct *p;
6212 lockdep_assert_held(&env->src_rq->lock);
6214 if (env->imbalance <= 0)
6217 while (!list_empty(tasks)) {
6219 * We don't want to steal all, otherwise we may be treated likewise,
6220 * which could at worst lead to a livelock crash.
6222 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6225 p = list_first_entry(tasks, struct task_struct, se.group_node);
6228 /* We've more or less seen every task there is, call it quits */
6229 if (env->loop > env->loop_max)
6232 /* take a breather every nr_migrate tasks */
6233 if (env->loop > env->loop_break) {
6234 env->loop_break += sched_nr_migrate_break;
6235 env->flags |= LBF_NEED_BREAK;
6239 if (!can_migrate_task(p, env))
6242 load = task_h_load(p);
6244 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6247 if ((load / 2) > env->imbalance)
6250 detach_task(p, env);
6251 list_add(&p->se.group_node, &env->tasks);
6254 env->imbalance -= load;
6256 #ifdef CONFIG_PREEMPT
6258 * NEWIDLE balancing is a source of latency, so preemptible
6259 * kernels will stop after the first task is detached to minimize
6260 * the critical section.
6262 if (env->idle == CPU_NEWLY_IDLE)
6267 * We only want to steal up to the prescribed amount of
6270 if (env->imbalance <= 0)
6275 list_move_tail(&p->se.group_node, tasks);
6279 * Right now, this is one of only two places we collect this stat
6280 * so we can safely collect detach_one_task() stats here rather
6281 * than inside detach_one_task().
6283 schedstat_add(env->sd, lb_gained[env->idle], detached);
6289 * attach_task() -- attach the task detached by detach_task() to its new rq.
6291 static void attach_task(struct rq *rq, struct task_struct *p)
6293 lockdep_assert_held(&rq->lock);
6295 BUG_ON(task_rq(p) != rq);
6296 activate_task(rq, p, 0);
6297 p->on_rq = TASK_ON_RQ_QUEUED;
6298 check_preempt_curr(rq, p, 0);
6302 * attach_one_task() -- attaches the task returned from detach_one_task() to
6305 static void attach_one_task(struct rq *rq, struct task_struct *p)
6307 raw_spin_lock(&rq->lock);
6309 raw_spin_unlock(&rq->lock);
6313 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6316 static void attach_tasks(struct lb_env *env)
6318 struct list_head *tasks = &env->tasks;
6319 struct task_struct *p;
6321 raw_spin_lock(&env->dst_rq->lock);
6323 while (!list_empty(tasks)) {
6324 p = list_first_entry(tasks, struct task_struct, se.group_node);
6325 list_del_init(&p->se.group_node);
6327 attach_task(env->dst_rq, p);
6330 raw_spin_unlock(&env->dst_rq->lock);
6333 #ifdef CONFIG_FAIR_GROUP_SCHED
6334 static void update_blocked_averages(int cpu)
6336 struct rq *rq = cpu_rq(cpu);
6337 struct cfs_rq *cfs_rq;
6338 unsigned long flags;
6340 raw_spin_lock_irqsave(&rq->lock, flags);
6341 update_rq_clock(rq);
6344 * Iterates the task_group tree in a bottom up fashion, see
6345 * list_add_leaf_cfs_rq() for details.
6347 for_each_leaf_cfs_rq(rq, cfs_rq) {
6348 /* throttled entities do not contribute to load */
6349 if (throttled_hierarchy(cfs_rq))
6352 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6353 update_tg_load_avg(cfs_rq, 0);
6355 raw_spin_unlock_irqrestore(&rq->lock, flags);
6359 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6360 * This needs to be done in a top-down fashion because the load of a child
6361 * group is a fraction of its parents load.
6363 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6365 struct rq *rq = rq_of(cfs_rq);
6366 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6367 unsigned long now = jiffies;
6370 if (cfs_rq->last_h_load_update == now)
6373 cfs_rq->h_load_next = NULL;
6374 for_each_sched_entity(se) {
6375 cfs_rq = cfs_rq_of(se);
6376 cfs_rq->h_load_next = se;
6377 if (cfs_rq->last_h_load_update == now)
6382 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6383 cfs_rq->last_h_load_update = now;
6386 while ((se = cfs_rq->h_load_next) != NULL) {
6387 load = cfs_rq->h_load;
6388 load = div64_ul(load * se->avg.load_avg,
6389 cfs_rq_load_avg(cfs_rq) + 1);
6390 cfs_rq = group_cfs_rq(se);
6391 cfs_rq->h_load = load;
6392 cfs_rq->last_h_load_update = now;
6396 static unsigned long task_h_load(struct task_struct *p)
6398 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6400 update_cfs_rq_h_load(cfs_rq);
6401 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6402 cfs_rq_load_avg(cfs_rq) + 1);
6405 static inline void update_blocked_averages(int cpu)
6407 struct rq *rq = cpu_rq(cpu);
6408 struct cfs_rq *cfs_rq = &rq->cfs;
6409 unsigned long flags;
6411 raw_spin_lock_irqsave(&rq->lock, flags);
6412 update_rq_clock(rq);
6413 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6414 raw_spin_unlock_irqrestore(&rq->lock, flags);
6417 static unsigned long task_h_load(struct task_struct *p)
6419 return p->se.avg.load_avg;
6423 /********** Helpers for find_busiest_group ************************/
6432 * sg_lb_stats - stats of a sched_group required for load_balancing
6434 struct sg_lb_stats {
6435 unsigned long avg_load; /*Avg load across the CPUs of the group */
6436 unsigned long group_load; /* Total load over the CPUs of the group */
6437 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6438 unsigned long load_per_task;
6439 unsigned long group_capacity;
6440 unsigned long group_util; /* Total utilization of the group */
6441 unsigned int sum_nr_running; /* Nr tasks running in the group */
6442 unsigned int idle_cpus;
6443 unsigned int group_weight;
6444 enum group_type group_type;
6445 int group_no_capacity;
6446 #ifdef CONFIG_NUMA_BALANCING
6447 unsigned int nr_numa_running;
6448 unsigned int nr_preferred_running;
6453 * sd_lb_stats - Structure to store the statistics of a sched_domain
6454 * during load balancing.
6456 struct sd_lb_stats {
6457 struct sched_group *busiest; /* Busiest group in this sd */
6458 struct sched_group *local; /* Local group in this sd */
6459 unsigned long total_load; /* Total load of all groups in sd */
6460 unsigned long total_capacity; /* Total capacity of all groups in sd */
6461 unsigned long avg_load; /* Average load across all groups in sd */
6463 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6464 struct sg_lb_stats local_stat; /* Statistics of the local group */
6467 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6470 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6471 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6472 * We must however clear busiest_stat::avg_load because
6473 * update_sd_pick_busiest() reads this before assignment.
6475 *sds = (struct sd_lb_stats){
6479 .total_capacity = 0UL,
6482 .sum_nr_running = 0,
6483 .group_type = group_other,
6489 * get_sd_load_idx - Obtain the load index for a given sched domain.
6490 * @sd: The sched_domain whose load_idx is to be obtained.
6491 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6493 * Return: The load index.
6495 static inline int get_sd_load_idx(struct sched_domain *sd,
6496 enum cpu_idle_type idle)
6502 load_idx = sd->busy_idx;
6505 case CPU_NEWLY_IDLE:
6506 load_idx = sd->newidle_idx;
6509 load_idx = sd->idle_idx;
6516 static unsigned long scale_rt_capacity(int cpu)
6518 struct rq *rq = cpu_rq(cpu);
6519 u64 total, used, age_stamp, avg;
6523 * Since we're reading these variables without serialization make sure
6524 * we read them once before doing sanity checks on them.
6526 age_stamp = READ_ONCE(rq->age_stamp);
6527 avg = READ_ONCE(rq->rt_avg);
6528 delta = __rq_clock_broken(rq) - age_stamp;
6530 if (unlikely(delta < 0))
6533 total = sched_avg_period() + delta;
6535 used = div_u64(avg, total);
6537 if (likely(used < SCHED_CAPACITY_SCALE))
6538 return SCHED_CAPACITY_SCALE - used;
6543 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6545 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6546 struct sched_group *sdg = sd->groups;
6548 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6550 capacity *= scale_rt_capacity(cpu);
6551 capacity >>= SCHED_CAPACITY_SHIFT;
6556 cpu_rq(cpu)->cpu_capacity = capacity;
6557 sdg->sgc->capacity = capacity;
6560 void update_group_capacity(struct sched_domain *sd, int cpu)
6562 struct sched_domain *child = sd->child;
6563 struct sched_group *group, *sdg = sd->groups;
6564 unsigned long capacity;
6565 unsigned long interval;
6567 interval = msecs_to_jiffies(sd->balance_interval);
6568 interval = clamp(interval, 1UL, max_load_balance_interval);
6569 sdg->sgc->next_update = jiffies + interval;
6572 update_cpu_capacity(sd, cpu);
6578 if (child->flags & SD_OVERLAP) {
6580 * SD_OVERLAP domains cannot assume that child groups
6581 * span the current group.
6584 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6585 struct sched_group_capacity *sgc;
6586 struct rq *rq = cpu_rq(cpu);
6589 * build_sched_domains() -> init_sched_groups_capacity()
6590 * gets here before we've attached the domains to the
6593 * Use capacity_of(), which is set irrespective of domains
6594 * in update_cpu_capacity().
6596 * This avoids capacity from being 0 and
6597 * causing divide-by-zero issues on boot.
6599 if (unlikely(!rq->sd)) {
6600 capacity += capacity_of(cpu);
6604 sgc = rq->sd->groups->sgc;
6605 capacity += sgc->capacity;
6609 * !SD_OVERLAP domains can assume that child groups
6610 * span the current group.
6613 group = child->groups;
6615 capacity += group->sgc->capacity;
6616 group = group->next;
6617 } while (group != child->groups);
6620 sdg->sgc->capacity = capacity;
6624 * Check whether the capacity of the rq has been noticeably reduced by side
6625 * activity. The imbalance_pct is used for the threshold.
6626 * Return true is the capacity is reduced
6629 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6631 return ((rq->cpu_capacity * sd->imbalance_pct) <
6632 (rq->cpu_capacity_orig * 100));
6636 * Group imbalance indicates (and tries to solve) the problem where balancing
6637 * groups is inadequate due to tsk_cpus_allowed() constraints.
6639 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6640 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6643 * { 0 1 2 3 } { 4 5 6 7 }
6646 * If we were to balance group-wise we'd place two tasks in the first group and
6647 * two tasks in the second group. Clearly this is undesired as it will overload
6648 * cpu 3 and leave one of the cpus in the second group unused.
6650 * The current solution to this issue is detecting the skew in the first group
6651 * by noticing the lower domain failed to reach balance and had difficulty
6652 * moving tasks due to affinity constraints.
6654 * When this is so detected; this group becomes a candidate for busiest; see
6655 * update_sd_pick_busiest(). And calculate_imbalance() and
6656 * find_busiest_group() avoid some of the usual balance conditions to allow it
6657 * to create an effective group imbalance.
6659 * This is a somewhat tricky proposition since the next run might not find the
6660 * group imbalance and decide the groups need to be balanced again. A most
6661 * subtle and fragile situation.
6664 static inline int sg_imbalanced(struct sched_group *group)
6666 return group->sgc->imbalance;
6670 * group_has_capacity returns true if the group has spare capacity that could
6671 * be used by some tasks.
6672 * We consider that a group has spare capacity if the * number of task is
6673 * smaller than the number of CPUs or if the utilization is lower than the
6674 * available capacity for CFS tasks.
6675 * For the latter, we use a threshold to stabilize the state, to take into
6676 * account the variance of the tasks' load and to return true if the available
6677 * capacity in meaningful for the load balancer.
6678 * As an example, an available capacity of 1% can appear but it doesn't make
6679 * any benefit for the load balance.
6682 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6684 if (sgs->sum_nr_running < sgs->group_weight)
6687 if ((sgs->group_capacity * 100) >
6688 (sgs->group_util * env->sd->imbalance_pct))
6695 * group_is_overloaded returns true if the group has more tasks than it can
6697 * group_is_overloaded is not equals to !group_has_capacity because a group
6698 * with the exact right number of tasks, has no more spare capacity but is not
6699 * overloaded so both group_has_capacity and group_is_overloaded return
6703 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6705 if (sgs->sum_nr_running <= sgs->group_weight)
6708 if ((sgs->group_capacity * 100) <
6709 (sgs->group_util * env->sd->imbalance_pct))
6716 group_type group_classify(struct sched_group *group,
6717 struct sg_lb_stats *sgs)
6719 if (sgs->group_no_capacity)
6720 return group_overloaded;
6722 if (sg_imbalanced(group))
6723 return group_imbalanced;
6729 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6730 * @env: The load balancing environment.
6731 * @group: sched_group whose statistics are to be updated.
6732 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6733 * @local_group: Does group contain this_cpu.
6734 * @sgs: variable to hold the statistics for this group.
6735 * @overload: Indicate more than one runnable task for any CPU.
6737 static inline void update_sg_lb_stats(struct lb_env *env,
6738 struct sched_group *group, int load_idx,
6739 int local_group, struct sg_lb_stats *sgs,
6745 memset(sgs, 0, sizeof(*sgs));
6747 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6748 struct rq *rq = cpu_rq(i);
6750 /* Bias balancing toward cpus of our domain */
6752 load = target_load(i, load_idx);
6754 load = source_load(i, load_idx);
6756 sgs->group_load += load;
6757 sgs->group_util += cpu_util(i);
6758 sgs->sum_nr_running += rq->cfs.h_nr_running;
6760 nr_running = rq->nr_running;
6764 #ifdef CONFIG_NUMA_BALANCING
6765 sgs->nr_numa_running += rq->nr_numa_running;
6766 sgs->nr_preferred_running += rq->nr_preferred_running;
6768 sgs->sum_weighted_load += weighted_cpuload(i);
6770 * No need to call idle_cpu() if nr_running is not 0
6772 if (!nr_running && idle_cpu(i))
6776 /* Adjust by relative CPU capacity of the group */
6777 sgs->group_capacity = group->sgc->capacity;
6778 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6780 if (sgs->sum_nr_running)
6781 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6783 sgs->group_weight = group->group_weight;
6785 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6786 sgs->group_type = group_classify(group, sgs);
6790 * update_sd_pick_busiest - return 1 on busiest group
6791 * @env: The load balancing environment.
6792 * @sds: sched_domain statistics
6793 * @sg: sched_group candidate to be checked for being the busiest
6794 * @sgs: sched_group statistics
6796 * Determine if @sg is a busier group than the previously selected
6799 * Return: %true if @sg is a busier group than the previously selected
6800 * busiest group. %false otherwise.
6802 static bool update_sd_pick_busiest(struct lb_env *env,
6803 struct sd_lb_stats *sds,
6804 struct sched_group *sg,
6805 struct sg_lb_stats *sgs)
6807 struct sg_lb_stats *busiest = &sds->busiest_stat;
6809 if (sgs->group_type > busiest->group_type)
6812 if (sgs->group_type < busiest->group_type)
6815 if (sgs->avg_load <= busiest->avg_load)
6818 /* This is the busiest node in its class. */
6819 if (!(env->sd->flags & SD_ASYM_PACKING))
6822 /* No ASYM_PACKING if target cpu is already busy */
6823 if (env->idle == CPU_NOT_IDLE)
6826 * ASYM_PACKING needs to move all the work to the lowest
6827 * numbered CPUs in the group, therefore mark all groups
6828 * higher than ourself as busy.
6830 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6834 /* Prefer to move from highest possible cpu's work */
6835 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6842 #ifdef CONFIG_NUMA_BALANCING
6843 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6845 if (sgs->sum_nr_running > sgs->nr_numa_running)
6847 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6852 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6854 if (rq->nr_running > rq->nr_numa_running)
6856 if (rq->nr_running > rq->nr_preferred_running)
6861 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6866 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6870 #endif /* CONFIG_NUMA_BALANCING */
6873 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6874 * @env: The load balancing environment.
6875 * @sds: variable to hold the statistics for this sched_domain.
6877 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6879 struct sched_domain *child = env->sd->child;
6880 struct sched_group *sg = env->sd->groups;
6881 struct sg_lb_stats tmp_sgs;
6882 int load_idx, prefer_sibling = 0;
6883 bool overload = false;
6885 if (child && child->flags & SD_PREFER_SIBLING)
6888 load_idx = get_sd_load_idx(env->sd, env->idle);
6891 struct sg_lb_stats *sgs = &tmp_sgs;
6894 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6897 sgs = &sds->local_stat;
6899 if (env->idle != CPU_NEWLY_IDLE ||
6900 time_after_eq(jiffies, sg->sgc->next_update))
6901 update_group_capacity(env->sd, env->dst_cpu);
6904 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6911 * In case the child domain prefers tasks go to siblings
6912 * first, lower the sg capacity so that we'll try
6913 * and move all the excess tasks away. We lower the capacity
6914 * of a group only if the local group has the capacity to fit
6915 * these excess tasks. The extra check prevents the case where
6916 * you always pull from the heaviest group when it is already
6917 * under-utilized (possible with a large weight task outweighs
6918 * the tasks on the system).
6920 if (prefer_sibling && sds->local &&
6921 group_has_capacity(env, &sds->local_stat) &&
6922 (sgs->sum_nr_running > 1)) {
6923 sgs->group_no_capacity = 1;
6924 sgs->group_type = group_classify(sg, sgs);
6927 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6929 sds->busiest_stat = *sgs;
6933 /* Now, start updating sd_lb_stats */
6934 sds->total_load += sgs->group_load;
6935 sds->total_capacity += sgs->group_capacity;
6938 } while (sg != env->sd->groups);
6940 if (env->sd->flags & SD_NUMA)
6941 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6943 if (!env->sd->parent) {
6944 /* update overload indicator if we are at root domain */
6945 if (env->dst_rq->rd->overload != overload)
6946 env->dst_rq->rd->overload = overload;
6952 * check_asym_packing - Check to see if the group is packed into the
6955 * This is primarily intended to used at the sibling level. Some
6956 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6957 * case of POWER7, it can move to lower SMT modes only when higher
6958 * threads are idle. When in lower SMT modes, the threads will
6959 * perform better since they share less core resources. Hence when we
6960 * have idle threads, we want them to be the higher ones.
6962 * This packing function is run on idle threads. It checks to see if
6963 * the busiest CPU in this domain (core in the P7 case) has a higher
6964 * CPU number than the packing function is being run on. Here we are
6965 * assuming lower CPU number will be equivalent to lower a SMT thread
6968 * Return: 1 when packing is required and a task should be moved to
6969 * this CPU. The amount of the imbalance is returned in *imbalance.
6971 * @env: The load balancing environment.
6972 * @sds: Statistics of the sched_domain which is to be packed
6974 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6978 if (!(env->sd->flags & SD_ASYM_PACKING))
6981 if (env->idle == CPU_NOT_IDLE)
6987 busiest_cpu = group_first_cpu(sds->busiest);
6988 if (env->dst_cpu > busiest_cpu)
6991 env->imbalance = DIV_ROUND_CLOSEST(
6992 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6993 SCHED_CAPACITY_SCALE);
6999 * fix_small_imbalance - Calculate the minor imbalance that exists
7000 * amongst the groups of a sched_domain, during
7002 * @env: The load balancing environment.
7003 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7006 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7008 unsigned long tmp, capa_now = 0, capa_move = 0;
7009 unsigned int imbn = 2;
7010 unsigned long scaled_busy_load_per_task;
7011 struct sg_lb_stats *local, *busiest;
7013 local = &sds->local_stat;
7014 busiest = &sds->busiest_stat;
7016 if (!local->sum_nr_running)
7017 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7018 else if (busiest->load_per_task > local->load_per_task)
7021 scaled_busy_load_per_task =
7022 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7023 busiest->group_capacity;
7025 if (busiest->avg_load + scaled_busy_load_per_task >=
7026 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7027 env->imbalance = busiest->load_per_task;
7032 * OK, we don't have enough imbalance to justify moving tasks,
7033 * however we may be able to increase total CPU capacity used by
7037 capa_now += busiest->group_capacity *
7038 min(busiest->load_per_task, busiest->avg_load);
7039 capa_now += local->group_capacity *
7040 min(local->load_per_task, local->avg_load);
7041 capa_now /= SCHED_CAPACITY_SCALE;
7043 /* Amount of load we'd subtract */
7044 if (busiest->avg_load > scaled_busy_load_per_task) {
7045 capa_move += busiest->group_capacity *
7046 min(busiest->load_per_task,
7047 busiest->avg_load - scaled_busy_load_per_task);
7050 /* Amount of load we'd add */
7051 if (busiest->avg_load * busiest->group_capacity <
7052 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7053 tmp = (busiest->avg_load * busiest->group_capacity) /
7054 local->group_capacity;
7056 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7057 local->group_capacity;
7059 capa_move += local->group_capacity *
7060 min(local->load_per_task, local->avg_load + tmp);
7061 capa_move /= SCHED_CAPACITY_SCALE;
7063 /* Move if we gain throughput */
7064 if (capa_move > capa_now)
7065 env->imbalance = busiest->load_per_task;
7069 * calculate_imbalance - Calculate the amount of imbalance present within the
7070 * groups of a given sched_domain during load balance.
7071 * @env: load balance environment
7072 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7074 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7076 unsigned long max_pull, load_above_capacity = ~0UL;
7077 struct sg_lb_stats *local, *busiest;
7079 local = &sds->local_stat;
7080 busiest = &sds->busiest_stat;
7082 if (busiest->group_type == group_imbalanced) {
7084 * In the group_imb case we cannot rely on group-wide averages
7085 * to ensure cpu-load equilibrium, look at wider averages. XXX
7087 busiest->load_per_task =
7088 min(busiest->load_per_task, sds->avg_load);
7092 * Avg load of busiest sg can be less and avg load of local sg can
7093 * be greater than avg load across all sgs of sd because avg load
7094 * factors in sg capacity and sgs with smaller group_type are
7095 * skipped when updating the busiest sg:
7097 if (busiest->avg_load <= sds->avg_load ||
7098 local->avg_load >= sds->avg_load) {
7100 return fix_small_imbalance(env, sds);
7104 * If there aren't any idle cpus, avoid creating some.
7106 if (busiest->group_type == group_overloaded &&
7107 local->group_type == group_overloaded) {
7108 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7109 if (load_above_capacity > busiest->group_capacity) {
7110 load_above_capacity -= busiest->group_capacity;
7111 load_above_capacity *= NICE_0_LOAD;
7112 load_above_capacity /= busiest->group_capacity;
7114 load_above_capacity = ~0UL;
7118 * We're trying to get all the cpus to the average_load, so we don't
7119 * want to push ourselves above the average load, nor do we wish to
7120 * reduce the max loaded cpu below the average load. At the same time,
7121 * we also don't want to reduce the group load below the group
7122 * capacity. Thus we look for the minimum possible imbalance.
7124 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7126 /* How much load to actually move to equalise the imbalance */
7127 env->imbalance = min(
7128 max_pull * busiest->group_capacity,
7129 (sds->avg_load - local->avg_load) * local->group_capacity
7130 ) / SCHED_CAPACITY_SCALE;
7133 * if *imbalance is less than the average load per runnable task
7134 * there is no guarantee that any tasks will be moved so we'll have
7135 * a think about bumping its value to force at least one task to be
7138 if (env->imbalance < busiest->load_per_task)
7139 return fix_small_imbalance(env, sds);
7142 /******* find_busiest_group() helpers end here *********************/
7145 * find_busiest_group - Returns the busiest group within the sched_domain
7146 * if there is an imbalance.
7148 * Also calculates the amount of weighted load which should be moved
7149 * to restore balance.
7151 * @env: The load balancing environment.
7153 * Return: - The busiest group if imbalance exists.
7155 static struct sched_group *find_busiest_group(struct lb_env *env)
7157 struct sg_lb_stats *local, *busiest;
7158 struct sd_lb_stats sds;
7160 init_sd_lb_stats(&sds);
7163 * Compute the various statistics relavent for load balancing at
7166 update_sd_lb_stats(env, &sds);
7167 local = &sds.local_stat;
7168 busiest = &sds.busiest_stat;
7170 /* ASYM feature bypasses nice load balance check */
7171 if (check_asym_packing(env, &sds))
7174 /* There is no busy sibling group to pull tasks from */
7175 if (!sds.busiest || busiest->sum_nr_running == 0)
7178 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7179 / sds.total_capacity;
7182 * If the busiest group is imbalanced the below checks don't
7183 * work because they assume all things are equal, which typically
7184 * isn't true due to cpus_allowed constraints and the like.
7186 if (busiest->group_type == group_imbalanced)
7189 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7190 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7191 busiest->group_no_capacity)
7195 * If the local group is busier than the selected busiest group
7196 * don't try and pull any tasks.
7198 if (local->avg_load >= busiest->avg_load)
7202 * Don't pull any tasks if this group is already above the domain
7205 if (local->avg_load >= sds.avg_load)
7208 if (env->idle == CPU_IDLE) {
7210 * This cpu is idle. If the busiest group is not overloaded
7211 * and there is no imbalance between this and busiest group
7212 * wrt idle cpus, it is balanced. The imbalance becomes
7213 * significant if the diff is greater than 1 otherwise we
7214 * might end up to just move the imbalance on another group
7216 if ((busiest->group_type != group_overloaded) &&
7217 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7221 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7222 * imbalance_pct to be conservative.
7224 if (100 * busiest->avg_load <=
7225 env->sd->imbalance_pct * local->avg_load)
7230 /* Looks like there is an imbalance. Compute it */
7231 calculate_imbalance(env, &sds);
7240 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7242 static struct rq *find_busiest_queue(struct lb_env *env,
7243 struct sched_group *group)
7245 struct rq *busiest = NULL, *rq;
7246 unsigned long busiest_load = 0, busiest_capacity = 1;
7249 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7250 unsigned long capacity, wl;
7254 rt = fbq_classify_rq(rq);
7257 * We classify groups/runqueues into three groups:
7258 * - regular: there are !numa tasks
7259 * - remote: there are numa tasks that run on the 'wrong' node
7260 * - all: there is no distinction
7262 * In order to avoid migrating ideally placed numa tasks,
7263 * ignore those when there's better options.
7265 * If we ignore the actual busiest queue to migrate another
7266 * task, the next balance pass can still reduce the busiest
7267 * queue by moving tasks around inside the node.
7269 * If we cannot move enough load due to this classification
7270 * the next pass will adjust the group classification and
7271 * allow migration of more tasks.
7273 * Both cases only affect the total convergence complexity.
7275 if (rt > env->fbq_type)
7278 capacity = capacity_of(i);
7280 wl = weighted_cpuload(i);
7283 * When comparing with imbalance, use weighted_cpuload()
7284 * which is not scaled with the cpu capacity.
7287 if (rq->nr_running == 1 && wl > env->imbalance &&
7288 !check_cpu_capacity(rq, env->sd))
7292 * For the load comparisons with the other cpu's, consider
7293 * the weighted_cpuload() scaled with the cpu capacity, so
7294 * that the load can be moved away from the cpu that is
7295 * potentially running at a lower capacity.
7297 * Thus we're looking for max(wl_i / capacity_i), crosswise
7298 * multiplication to rid ourselves of the division works out
7299 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7300 * our previous maximum.
7302 if (wl * busiest_capacity > busiest_load * capacity) {
7304 busiest_capacity = capacity;
7313 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7314 * so long as it is large enough.
7316 #define MAX_PINNED_INTERVAL 512
7318 /* Working cpumask for load_balance and load_balance_newidle. */
7319 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7321 static int need_active_balance(struct lb_env *env)
7323 struct sched_domain *sd = env->sd;
7325 if (env->idle == CPU_NEWLY_IDLE) {
7328 * ASYM_PACKING needs to force migrate tasks from busy but
7329 * higher numbered CPUs in order to pack all tasks in the
7330 * lowest numbered CPUs.
7332 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7337 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7338 * It's worth migrating the task if the src_cpu's capacity is reduced
7339 * because of other sched_class or IRQs if more capacity stays
7340 * available on dst_cpu.
7342 if ((env->idle != CPU_NOT_IDLE) &&
7343 (env->src_rq->cfs.h_nr_running == 1)) {
7344 if ((check_cpu_capacity(env->src_rq, sd)) &&
7345 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7349 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7352 static int active_load_balance_cpu_stop(void *data);
7354 static int should_we_balance(struct lb_env *env)
7356 struct sched_group *sg = env->sd->groups;
7357 struct cpumask *sg_cpus, *sg_mask;
7358 int cpu, balance_cpu = -1;
7361 * In the newly idle case, we will allow all the cpu's
7362 * to do the newly idle load balance.
7364 if (env->idle == CPU_NEWLY_IDLE)
7367 sg_cpus = sched_group_cpus(sg);
7368 sg_mask = sched_group_mask(sg);
7369 /* Try to find first idle cpu */
7370 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7371 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7378 if (balance_cpu == -1)
7379 balance_cpu = group_balance_cpu(sg);
7382 * First idle cpu or the first cpu(busiest) in this sched group
7383 * is eligible for doing load balancing at this and above domains.
7385 return balance_cpu == env->dst_cpu;
7389 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7390 * tasks if there is an imbalance.
7392 static int load_balance(int this_cpu, struct rq *this_rq,
7393 struct sched_domain *sd, enum cpu_idle_type idle,
7394 int *continue_balancing)
7396 int ld_moved, cur_ld_moved, active_balance = 0;
7397 struct sched_domain *sd_parent = sd->parent;
7398 struct sched_group *group;
7400 unsigned long flags;
7401 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7403 struct lb_env env = {
7405 .dst_cpu = this_cpu,
7407 .dst_grpmask = sched_group_cpus(sd->groups),
7409 .loop_break = sched_nr_migrate_break,
7412 .tasks = LIST_HEAD_INIT(env.tasks),
7416 * For NEWLY_IDLE load_balancing, we don't need to consider
7417 * other cpus in our group
7419 if (idle == CPU_NEWLY_IDLE)
7420 env.dst_grpmask = NULL;
7422 cpumask_copy(cpus, cpu_active_mask);
7424 schedstat_inc(sd, lb_count[idle]);
7427 if (!should_we_balance(&env)) {
7428 *continue_balancing = 0;
7432 group = find_busiest_group(&env);
7434 schedstat_inc(sd, lb_nobusyg[idle]);
7438 busiest = find_busiest_queue(&env, group);
7440 schedstat_inc(sd, lb_nobusyq[idle]);
7444 BUG_ON(busiest == env.dst_rq);
7446 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7448 env.src_cpu = busiest->cpu;
7449 env.src_rq = busiest;
7452 if (busiest->nr_running > 1) {
7454 * Attempt to move tasks. If find_busiest_group has found
7455 * an imbalance but busiest->nr_running <= 1, the group is
7456 * still unbalanced. ld_moved simply stays zero, so it is
7457 * correctly treated as an imbalance.
7459 env.flags |= LBF_ALL_PINNED;
7460 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7463 raw_spin_lock_irqsave(&busiest->lock, flags);
7466 * cur_ld_moved - load moved in current iteration
7467 * ld_moved - cumulative load moved across iterations
7469 cur_ld_moved = detach_tasks(&env);
7472 * We've detached some tasks from busiest_rq. Every
7473 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7474 * unlock busiest->lock, and we are able to be sure
7475 * that nobody can manipulate the tasks in parallel.
7476 * See task_rq_lock() family for the details.
7479 raw_spin_unlock(&busiest->lock);
7483 ld_moved += cur_ld_moved;
7486 local_irq_restore(flags);
7488 if (env.flags & LBF_NEED_BREAK) {
7489 env.flags &= ~LBF_NEED_BREAK;
7494 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7495 * us and move them to an alternate dst_cpu in our sched_group
7496 * where they can run. The upper limit on how many times we
7497 * iterate on same src_cpu is dependent on number of cpus in our
7500 * This changes load balance semantics a bit on who can move
7501 * load to a given_cpu. In addition to the given_cpu itself
7502 * (or a ilb_cpu acting on its behalf where given_cpu is
7503 * nohz-idle), we now have balance_cpu in a position to move
7504 * load to given_cpu. In rare situations, this may cause
7505 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7506 * _independently_ and at _same_ time to move some load to
7507 * given_cpu) causing exceess load to be moved to given_cpu.
7508 * This however should not happen so much in practice and
7509 * moreover subsequent load balance cycles should correct the
7510 * excess load moved.
7512 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7514 /* Prevent to re-select dst_cpu via env's cpus */
7515 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7517 env.dst_rq = cpu_rq(env.new_dst_cpu);
7518 env.dst_cpu = env.new_dst_cpu;
7519 env.flags &= ~LBF_DST_PINNED;
7521 env.loop_break = sched_nr_migrate_break;
7524 * Go back to "more_balance" rather than "redo" since we
7525 * need to continue with same src_cpu.
7531 * We failed to reach balance because of affinity.
7534 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7536 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7537 *group_imbalance = 1;
7540 /* All tasks on this runqueue were pinned by CPU affinity */
7541 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7542 cpumask_clear_cpu(cpu_of(busiest), cpus);
7543 if (!cpumask_empty(cpus)) {
7545 env.loop_break = sched_nr_migrate_break;
7548 goto out_all_pinned;
7553 schedstat_inc(sd, lb_failed[idle]);
7555 * Increment the failure counter only on periodic balance.
7556 * We do not want newidle balance, which can be very
7557 * frequent, pollute the failure counter causing
7558 * excessive cache_hot migrations and active balances.
7560 if (idle != CPU_NEWLY_IDLE)
7561 sd->nr_balance_failed++;
7563 if (need_active_balance(&env)) {
7564 raw_spin_lock_irqsave(&busiest->lock, flags);
7566 /* don't kick the active_load_balance_cpu_stop,
7567 * if the curr task on busiest cpu can't be
7570 if (!cpumask_test_cpu(this_cpu,
7571 tsk_cpus_allowed(busiest->curr))) {
7572 raw_spin_unlock_irqrestore(&busiest->lock,
7574 env.flags |= LBF_ALL_PINNED;
7575 goto out_one_pinned;
7579 * ->active_balance synchronizes accesses to
7580 * ->active_balance_work. Once set, it's cleared
7581 * only after active load balance is finished.
7583 if (!busiest->active_balance) {
7584 busiest->active_balance = 1;
7585 busiest->push_cpu = this_cpu;
7588 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7590 if (active_balance) {
7591 stop_one_cpu_nowait(cpu_of(busiest),
7592 active_load_balance_cpu_stop, busiest,
7593 &busiest->active_balance_work);
7596 /* We've kicked active balancing, force task migration. */
7597 sd->nr_balance_failed = sd->cache_nice_tries+1;
7600 sd->nr_balance_failed = 0;
7602 if (likely(!active_balance)) {
7603 /* We were unbalanced, so reset the balancing interval */
7604 sd->balance_interval = sd->min_interval;
7607 * If we've begun active balancing, start to back off. This
7608 * case may not be covered by the all_pinned logic if there
7609 * is only 1 task on the busy runqueue (because we don't call
7612 if (sd->balance_interval < sd->max_interval)
7613 sd->balance_interval *= 2;
7620 * We reach balance although we may have faced some affinity
7621 * constraints. Clear the imbalance flag if it was set.
7624 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7626 if (*group_imbalance)
7627 *group_imbalance = 0;
7632 * We reach balance because all tasks are pinned at this level so
7633 * we can't migrate them. Let the imbalance flag set so parent level
7634 * can try to migrate them.
7636 schedstat_inc(sd, lb_balanced[idle]);
7638 sd->nr_balance_failed = 0;
7641 /* tune up the balancing interval */
7642 if (((env.flags & LBF_ALL_PINNED) &&
7643 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7644 (sd->balance_interval < sd->max_interval))
7645 sd->balance_interval *= 2;
7652 static inline unsigned long
7653 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7655 unsigned long interval = sd->balance_interval;
7658 interval *= sd->busy_factor;
7660 /* scale ms to jiffies */
7661 interval = msecs_to_jiffies(interval);
7662 interval = clamp(interval, 1UL, max_load_balance_interval);
7668 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7670 unsigned long interval, next;
7672 interval = get_sd_balance_interval(sd, cpu_busy);
7673 next = sd->last_balance + interval;
7675 if (time_after(*next_balance, next))
7676 *next_balance = next;
7680 * idle_balance is called by schedule() if this_cpu is about to become
7681 * idle. Attempts to pull tasks from other CPUs.
7683 static int idle_balance(struct rq *this_rq)
7685 unsigned long next_balance = jiffies + HZ;
7686 int this_cpu = this_rq->cpu;
7687 struct sched_domain *sd;
7688 int pulled_task = 0;
7692 * We must set idle_stamp _before_ calling idle_balance(), such that we
7693 * measure the duration of idle_balance() as idle time.
7695 this_rq->idle_stamp = rq_clock(this_rq);
7697 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7698 !this_rq->rd->overload) {
7700 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7702 update_next_balance(sd, 0, &next_balance);
7708 raw_spin_unlock(&this_rq->lock);
7710 update_blocked_averages(this_cpu);
7712 for_each_domain(this_cpu, sd) {
7713 int continue_balancing = 1;
7714 u64 t0, domain_cost;
7716 if (!(sd->flags & SD_LOAD_BALANCE))
7719 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7720 update_next_balance(sd, 0, &next_balance);
7724 if (sd->flags & SD_BALANCE_NEWIDLE) {
7725 t0 = sched_clock_cpu(this_cpu);
7727 pulled_task = load_balance(this_cpu, this_rq,
7729 &continue_balancing);
7731 domain_cost = sched_clock_cpu(this_cpu) - t0;
7732 if (domain_cost > sd->max_newidle_lb_cost)
7733 sd->max_newidle_lb_cost = domain_cost;
7735 curr_cost += domain_cost;
7738 update_next_balance(sd, 0, &next_balance);
7741 * Stop searching for tasks to pull if there are
7742 * now runnable tasks on this rq.
7744 if (pulled_task || this_rq->nr_running > 0)
7749 raw_spin_lock(&this_rq->lock);
7751 if (curr_cost > this_rq->max_idle_balance_cost)
7752 this_rq->max_idle_balance_cost = curr_cost;
7755 * While browsing the domains, we released the rq lock, a task could
7756 * have been enqueued in the meantime. Since we're not going idle,
7757 * pretend we pulled a task.
7759 if (this_rq->cfs.h_nr_running && !pulled_task)
7763 /* Move the next balance forward */
7764 if (time_after(this_rq->next_balance, next_balance))
7765 this_rq->next_balance = next_balance;
7767 /* Is there a task of a high priority class? */
7768 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7772 this_rq->idle_stamp = 0;
7778 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7779 * running tasks off the busiest CPU onto idle CPUs. It requires at
7780 * least 1 task to be running on each physical CPU where possible, and
7781 * avoids physical / logical imbalances.
7783 static int active_load_balance_cpu_stop(void *data)
7785 struct rq *busiest_rq = data;
7786 int busiest_cpu = cpu_of(busiest_rq);
7787 int target_cpu = busiest_rq->push_cpu;
7788 struct rq *target_rq = cpu_rq(target_cpu);
7789 struct sched_domain *sd;
7790 struct task_struct *p = NULL;
7792 raw_spin_lock_irq(&busiest_rq->lock);
7794 /* make sure the requested cpu hasn't gone down in the meantime */
7795 if (unlikely(busiest_cpu != smp_processor_id() ||
7796 !busiest_rq->active_balance))
7799 /* Is there any task to move? */
7800 if (busiest_rq->nr_running <= 1)
7804 * This condition is "impossible", if it occurs
7805 * we need to fix it. Originally reported by
7806 * Bjorn Helgaas on a 128-cpu setup.
7808 BUG_ON(busiest_rq == target_rq);
7810 /* Search for an sd spanning us and the target CPU. */
7812 for_each_domain(target_cpu, sd) {
7813 if ((sd->flags & SD_LOAD_BALANCE) &&
7814 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7819 struct lb_env env = {
7821 .dst_cpu = target_cpu,
7822 .dst_rq = target_rq,
7823 .src_cpu = busiest_rq->cpu,
7824 .src_rq = busiest_rq,
7828 schedstat_inc(sd, alb_count);
7830 p = detach_one_task(&env);
7832 schedstat_inc(sd, alb_pushed);
7833 /* Active balancing done, reset the failure counter. */
7834 sd->nr_balance_failed = 0;
7836 schedstat_inc(sd, alb_failed);
7841 busiest_rq->active_balance = 0;
7842 raw_spin_unlock(&busiest_rq->lock);
7845 attach_one_task(target_rq, p);
7852 static inline int on_null_domain(struct rq *rq)
7854 return unlikely(!rcu_dereference_sched(rq->sd));
7857 #ifdef CONFIG_NO_HZ_COMMON
7859 * idle load balancing details
7860 * - When one of the busy CPUs notice that there may be an idle rebalancing
7861 * needed, they will kick the idle load balancer, which then does idle
7862 * load balancing for all the idle CPUs.
7865 cpumask_var_t idle_cpus_mask;
7867 unsigned long next_balance; /* in jiffy units */
7868 } nohz ____cacheline_aligned;
7870 static inline int find_new_ilb(void)
7872 int ilb = cpumask_first(nohz.idle_cpus_mask);
7874 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7881 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7882 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7883 * CPU (if there is one).
7885 static void nohz_balancer_kick(void)
7889 nohz.next_balance++;
7891 ilb_cpu = find_new_ilb();
7893 if (ilb_cpu >= nr_cpu_ids)
7896 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7899 * Use smp_send_reschedule() instead of resched_cpu().
7900 * This way we generate a sched IPI on the target cpu which
7901 * is idle. And the softirq performing nohz idle load balance
7902 * will be run before returning from the IPI.
7904 smp_send_reschedule(ilb_cpu);
7908 void nohz_balance_exit_idle(unsigned int cpu)
7910 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7912 * Completely isolated CPUs don't ever set, so we must test.
7914 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7915 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7916 atomic_dec(&nohz.nr_cpus);
7918 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7922 static inline void set_cpu_sd_state_busy(void)
7924 struct sched_domain *sd;
7925 int cpu = smp_processor_id();
7928 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7930 if (!sd || !sd->nohz_idle)
7934 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7939 void set_cpu_sd_state_idle(void)
7941 struct sched_domain *sd;
7942 int cpu = smp_processor_id();
7945 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7947 if (!sd || sd->nohz_idle)
7951 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7957 * This routine will record that the cpu is going idle with tick stopped.
7958 * This info will be used in performing idle load balancing in the future.
7960 void nohz_balance_enter_idle(int cpu)
7963 * If this cpu is going down, then nothing needs to be done.
7965 if (!cpu_active(cpu))
7968 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7972 * If we're a completely isolated CPU, we don't play.
7974 if (on_null_domain(cpu_rq(cpu)))
7977 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7978 atomic_inc(&nohz.nr_cpus);
7979 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7983 static DEFINE_SPINLOCK(balancing);
7986 * Scale the max load_balance interval with the number of CPUs in the system.
7987 * This trades load-balance latency on larger machines for less cross talk.
7989 void update_max_interval(void)
7991 max_load_balance_interval = HZ*num_online_cpus()/10;
7995 * It checks each scheduling domain to see if it is due to be balanced,
7996 * and initiates a balancing operation if so.
7998 * Balancing parameters are set up in init_sched_domains.
8000 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8002 int continue_balancing = 1;
8004 unsigned long interval;
8005 struct sched_domain *sd;
8006 /* Earliest time when we have to do rebalance again */
8007 unsigned long next_balance = jiffies + 60*HZ;
8008 int update_next_balance = 0;
8009 int need_serialize, need_decay = 0;
8012 update_blocked_averages(cpu);
8015 for_each_domain(cpu, sd) {
8017 * Decay the newidle max times here because this is a regular
8018 * visit to all the domains. Decay ~1% per second.
8020 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8021 sd->max_newidle_lb_cost =
8022 (sd->max_newidle_lb_cost * 253) / 256;
8023 sd->next_decay_max_lb_cost = jiffies + HZ;
8026 max_cost += sd->max_newidle_lb_cost;
8028 if (!(sd->flags & SD_LOAD_BALANCE))
8032 * Stop the load balance at this level. There is another
8033 * CPU in our sched group which is doing load balancing more
8036 if (!continue_balancing) {
8042 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8044 need_serialize = sd->flags & SD_SERIALIZE;
8045 if (need_serialize) {
8046 if (!spin_trylock(&balancing))
8050 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8051 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8053 * The LBF_DST_PINNED logic could have changed
8054 * env->dst_cpu, so we can't know our idle
8055 * state even if we migrated tasks. Update it.
8057 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8059 sd->last_balance = jiffies;
8060 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8063 spin_unlock(&balancing);
8065 if (time_after(next_balance, sd->last_balance + interval)) {
8066 next_balance = sd->last_balance + interval;
8067 update_next_balance = 1;
8072 * Ensure the rq-wide value also decays but keep it at a
8073 * reasonable floor to avoid funnies with rq->avg_idle.
8075 rq->max_idle_balance_cost =
8076 max((u64)sysctl_sched_migration_cost, max_cost);
8081 * next_balance will be updated only when there is a need.
8082 * When the cpu is attached to null domain for ex, it will not be
8085 if (likely(update_next_balance)) {
8086 rq->next_balance = next_balance;
8088 #ifdef CONFIG_NO_HZ_COMMON
8090 * If this CPU has been elected to perform the nohz idle
8091 * balance. Other idle CPUs have already rebalanced with
8092 * nohz_idle_balance() and nohz.next_balance has been
8093 * updated accordingly. This CPU is now running the idle load
8094 * balance for itself and we need to update the
8095 * nohz.next_balance accordingly.
8097 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8098 nohz.next_balance = rq->next_balance;
8103 #ifdef CONFIG_NO_HZ_COMMON
8105 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8106 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8108 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8110 int this_cpu = this_rq->cpu;
8113 /* Earliest time when we have to do rebalance again */
8114 unsigned long next_balance = jiffies + 60*HZ;
8115 int update_next_balance = 0;
8117 if (idle != CPU_IDLE ||
8118 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8121 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8122 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8126 * If this cpu gets work to do, stop the load balancing
8127 * work being done for other cpus. Next load
8128 * balancing owner will pick it up.
8133 rq = cpu_rq(balance_cpu);
8136 * If time for next balance is due,
8139 if (time_after_eq(jiffies, rq->next_balance)) {
8140 raw_spin_lock_irq(&rq->lock);
8141 update_rq_clock(rq);
8142 cpu_load_update_idle(rq);
8143 raw_spin_unlock_irq(&rq->lock);
8144 rebalance_domains(rq, CPU_IDLE);
8147 if (time_after(next_balance, rq->next_balance)) {
8148 next_balance = rq->next_balance;
8149 update_next_balance = 1;
8154 * next_balance will be updated only when there is a need.
8155 * When the CPU is attached to null domain for ex, it will not be
8158 if (likely(update_next_balance))
8159 nohz.next_balance = next_balance;
8161 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8165 * Current heuristic for kicking the idle load balancer in the presence
8166 * of an idle cpu in the system.
8167 * - This rq has more than one task.
8168 * - This rq has at least one CFS task and the capacity of the CPU is
8169 * significantly reduced because of RT tasks or IRQs.
8170 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8171 * multiple busy cpu.
8172 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8173 * domain span are idle.
8175 static inline bool nohz_kick_needed(struct rq *rq)
8177 unsigned long now = jiffies;
8178 struct sched_domain *sd;
8179 struct sched_group_capacity *sgc;
8180 int nr_busy, cpu = rq->cpu;
8183 if (unlikely(rq->idle_balance))
8187 * We may be recently in ticked or tickless idle mode. At the first
8188 * busy tick after returning from idle, we will update the busy stats.
8190 set_cpu_sd_state_busy();
8191 nohz_balance_exit_idle(cpu);
8194 * None are in tickless mode and hence no need for NOHZ idle load
8197 if (likely(!atomic_read(&nohz.nr_cpus)))
8200 if (time_before(now, nohz.next_balance))
8203 if (rq->nr_running >= 2)
8207 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8209 sgc = sd->groups->sgc;
8210 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8219 sd = rcu_dereference(rq->sd);
8221 if ((rq->cfs.h_nr_running >= 1) &&
8222 check_cpu_capacity(rq, sd)) {
8228 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8229 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8230 sched_domain_span(sd)) < cpu)) {
8240 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8244 * run_rebalance_domains is triggered when needed from the scheduler tick.
8245 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8247 static void run_rebalance_domains(struct softirq_action *h)
8249 struct rq *this_rq = this_rq();
8250 enum cpu_idle_type idle = this_rq->idle_balance ?
8251 CPU_IDLE : CPU_NOT_IDLE;
8254 * If this cpu has a pending nohz_balance_kick, then do the
8255 * balancing on behalf of the other idle cpus whose ticks are
8256 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8257 * give the idle cpus a chance to load balance. Else we may
8258 * load balance only within the local sched_domain hierarchy
8259 * and abort nohz_idle_balance altogether if we pull some load.
8261 nohz_idle_balance(this_rq, idle);
8262 rebalance_domains(this_rq, idle);
8266 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8268 void trigger_load_balance(struct rq *rq)
8270 /* Don't need to rebalance while attached to NULL domain */
8271 if (unlikely(on_null_domain(rq)))
8274 if (time_after_eq(jiffies, rq->next_balance))
8275 raise_softirq(SCHED_SOFTIRQ);
8276 #ifdef CONFIG_NO_HZ_COMMON
8277 if (nohz_kick_needed(rq))
8278 nohz_balancer_kick();
8282 static void rq_online_fair(struct rq *rq)
8286 update_runtime_enabled(rq);
8289 static void rq_offline_fair(struct rq *rq)
8293 /* Ensure any throttled groups are reachable by pick_next_task */
8294 unthrottle_offline_cfs_rqs(rq);
8297 #endif /* CONFIG_SMP */
8300 * scheduler tick hitting a task of our scheduling class:
8302 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8304 struct cfs_rq *cfs_rq;
8305 struct sched_entity *se = &curr->se;
8307 for_each_sched_entity(se) {
8308 cfs_rq = cfs_rq_of(se);
8309 entity_tick(cfs_rq, se, queued);
8312 if (static_branch_unlikely(&sched_numa_balancing))
8313 task_tick_numa(rq, curr);
8317 * called on fork with the child task as argument from the parent's context
8318 * - child not yet on the tasklist
8319 * - preemption disabled
8321 static void task_fork_fair(struct task_struct *p)
8323 struct cfs_rq *cfs_rq;
8324 struct sched_entity *se = &p->se, *curr;
8325 int this_cpu = smp_processor_id();
8326 struct rq *rq = this_rq();
8327 unsigned long flags;
8329 raw_spin_lock_irqsave(&rq->lock, flags);
8331 update_rq_clock(rq);
8333 cfs_rq = task_cfs_rq(current);
8334 curr = cfs_rq->curr;
8337 * Not only the cpu but also the task_group of the parent might have
8338 * been changed after parent->se.parent,cfs_rq were copied to
8339 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8340 * of child point to valid ones.
8343 __set_task_cpu(p, this_cpu);
8346 update_curr(cfs_rq);
8349 se->vruntime = curr->vruntime;
8350 place_entity(cfs_rq, se, 1);
8352 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8354 * Upon rescheduling, sched_class::put_prev_task() will place
8355 * 'current' within the tree based on its new key value.
8357 swap(curr->vruntime, se->vruntime);
8361 se->vruntime -= cfs_rq->min_vruntime;
8363 raw_spin_unlock_irqrestore(&rq->lock, flags);
8367 * Priority of the task has changed. Check to see if we preempt
8371 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8373 if (!task_on_rq_queued(p))
8377 * Reschedule if we are currently running on this runqueue and
8378 * our priority decreased, or if we are not currently running on
8379 * this runqueue and our priority is higher than the current's
8381 if (rq->curr == p) {
8382 if (p->prio > oldprio)
8385 check_preempt_curr(rq, p, 0);
8388 static inline bool vruntime_normalized(struct task_struct *p)
8390 struct sched_entity *se = &p->se;
8393 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8394 * the dequeue_entity(.flags=0) will already have normalized the
8401 * When !on_rq, vruntime of the task has usually NOT been normalized.
8402 * But there are some cases where it has already been normalized:
8404 * - A forked child which is waiting for being woken up by
8405 * wake_up_new_task().
8406 * - A task which has been woken up by try_to_wake_up() and
8407 * waiting for actually being woken up by sched_ttwu_pending().
8409 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8415 static void detach_task_cfs_rq(struct task_struct *p)
8417 struct sched_entity *se = &p->se;
8418 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8420 if (!vruntime_normalized(p)) {
8422 * Fix up our vruntime so that the current sleep doesn't
8423 * cause 'unlimited' sleep bonus.
8425 place_entity(cfs_rq, se, 0);
8426 se->vruntime -= cfs_rq->min_vruntime;
8429 /* Catch up with the cfs_rq and remove our load when we leave */
8430 detach_entity_load_avg(cfs_rq, se);
8433 static void attach_task_cfs_rq(struct task_struct *p)
8435 struct sched_entity *se = &p->se;
8436 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8438 #ifdef CONFIG_FAIR_GROUP_SCHED
8440 * Since the real-depth could have been changed (only FAIR
8441 * class maintain depth value), reset depth properly.
8443 se->depth = se->parent ? se->parent->depth + 1 : 0;
8446 /* Synchronize task with its cfs_rq */
8447 attach_entity_load_avg(cfs_rq, se);
8449 if (!vruntime_normalized(p))
8450 se->vruntime += cfs_rq->min_vruntime;
8453 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8455 detach_task_cfs_rq(p);
8458 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8460 attach_task_cfs_rq(p);
8462 if (task_on_rq_queued(p)) {
8464 * We were most likely switched from sched_rt, so
8465 * kick off the schedule if running, otherwise just see
8466 * if we can still preempt the current task.
8471 check_preempt_curr(rq, p, 0);
8475 /* Account for a task changing its policy or group.
8477 * This routine is mostly called to set cfs_rq->curr field when a task
8478 * migrates between groups/classes.
8480 static void set_curr_task_fair(struct rq *rq)
8482 struct sched_entity *se = &rq->curr->se;
8484 for_each_sched_entity(se) {
8485 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8487 set_next_entity(cfs_rq, se);
8488 /* ensure bandwidth has been allocated on our new cfs_rq */
8489 account_cfs_rq_runtime(cfs_rq, 0);
8493 void init_cfs_rq(struct cfs_rq *cfs_rq)
8495 cfs_rq->tasks_timeline = RB_ROOT;
8496 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8497 #ifndef CONFIG_64BIT
8498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8501 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8502 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8506 #ifdef CONFIG_FAIR_GROUP_SCHED
8507 static void task_move_group_fair(struct task_struct *p)
8509 detach_task_cfs_rq(p);
8510 set_task_rq(p, task_cpu(p));
8513 /* Tell se's cfs_rq has been changed -- migrated */
8514 p->se.avg.last_update_time = 0;
8516 attach_task_cfs_rq(p);
8519 void free_fair_sched_group(struct task_group *tg)
8523 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8525 for_each_possible_cpu(i) {
8527 kfree(tg->cfs_rq[i]);
8536 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8538 struct sched_entity *se;
8539 struct cfs_rq *cfs_rq;
8543 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8546 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8550 tg->shares = NICE_0_LOAD;
8552 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8554 for_each_possible_cpu(i) {
8557 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8558 GFP_KERNEL, cpu_to_node(i));
8562 se = kzalloc_node(sizeof(struct sched_entity),
8563 GFP_KERNEL, cpu_to_node(i));
8567 init_cfs_rq(cfs_rq);
8568 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8569 init_entity_runnable_average(se);
8571 raw_spin_lock_irq(&rq->lock);
8572 post_init_entity_util_avg(se);
8573 raw_spin_unlock_irq(&rq->lock);
8584 void unregister_fair_sched_group(struct task_group *tg)
8586 unsigned long flags;
8590 for_each_possible_cpu(cpu) {
8592 remove_entity_load_avg(tg->se[cpu]);
8595 * Only empty task groups can be destroyed; so we can speculatively
8596 * check on_list without danger of it being re-added.
8598 if (!tg->cfs_rq[cpu]->on_list)
8603 raw_spin_lock_irqsave(&rq->lock, flags);
8604 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8605 raw_spin_unlock_irqrestore(&rq->lock, flags);
8609 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8610 struct sched_entity *se, int cpu,
8611 struct sched_entity *parent)
8613 struct rq *rq = cpu_rq(cpu);
8617 init_cfs_rq_runtime(cfs_rq);
8619 tg->cfs_rq[cpu] = cfs_rq;
8622 /* se could be NULL for root_task_group */
8627 se->cfs_rq = &rq->cfs;
8630 se->cfs_rq = parent->my_q;
8631 se->depth = parent->depth + 1;
8635 /* guarantee group entities always have weight */
8636 update_load_set(&se->load, NICE_0_LOAD);
8637 se->parent = parent;
8640 static DEFINE_MUTEX(shares_mutex);
8642 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8645 unsigned long flags;
8648 * We can't change the weight of the root cgroup.
8653 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8655 mutex_lock(&shares_mutex);
8656 if (tg->shares == shares)
8659 tg->shares = shares;
8660 for_each_possible_cpu(i) {
8661 struct rq *rq = cpu_rq(i);
8662 struct sched_entity *se;
8665 /* Propagate contribution to hierarchy */
8666 raw_spin_lock_irqsave(&rq->lock, flags);
8668 /* Possible calls to update_curr() need rq clock */
8669 update_rq_clock(rq);
8670 for_each_sched_entity(se)
8671 update_cfs_shares(group_cfs_rq(se));
8672 raw_spin_unlock_irqrestore(&rq->lock, flags);
8676 mutex_unlock(&shares_mutex);
8679 #else /* CONFIG_FAIR_GROUP_SCHED */
8681 void free_fair_sched_group(struct task_group *tg) { }
8683 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8688 void unregister_fair_sched_group(struct task_group *tg) { }
8690 #endif /* CONFIG_FAIR_GROUP_SCHED */
8693 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8695 struct sched_entity *se = &task->se;
8696 unsigned int rr_interval = 0;
8699 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8702 if (rq->cfs.load.weight)
8703 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8709 * All the scheduling class methods:
8711 const struct sched_class fair_sched_class = {
8712 .next = &idle_sched_class,
8713 .enqueue_task = enqueue_task_fair,
8714 .dequeue_task = dequeue_task_fair,
8715 .yield_task = yield_task_fair,
8716 .yield_to_task = yield_to_task_fair,
8718 .check_preempt_curr = check_preempt_wakeup,
8720 .pick_next_task = pick_next_task_fair,
8721 .put_prev_task = put_prev_task_fair,
8724 .select_task_rq = select_task_rq_fair,
8725 .migrate_task_rq = migrate_task_rq_fair,
8727 .rq_online = rq_online_fair,
8728 .rq_offline = rq_offline_fair,
8730 .task_dead = task_dead_fair,
8731 .set_cpus_allowed = set_cpus_allowed_common,
8734 .set_curr_task = set_curr_task_fair,
8735 .task_tick = task_tick_fair,
8736 .task_fork = task_fork_fair,
8738 .prio_changed = prio_changed_fair,
8739 .switched_from = switched_from_fair,
8740 .switched_to = switched_to_fair,
8742 .get_rr_interval = get_rr_interval_fair,
8744 .update_curr = update_curr_fair,
8746 #ifdef CONFIG_FAIR_GROUP_SCHED
8747 .task_move_group = task_move_group_fair,
8751 #ifdef CONFIG_SCHED_DEBUG
8752 void print_cfs_stats(struct seq_file *m, int cpu)
8754 struct cfs_rq *cfs_rq;
8757 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8758 print_cfs_rq(m, cpu, cfs_rq);
8762 #ifdef CONFIG_NUMA_BALANCING
8763 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8766 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8768 for_each_online_node(node) {
8769 if (p->numa_faults) {
8770 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8771 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8773 if (p->numa_group) {
8774 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8775 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8777 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8780 #endif /* CONFIG_NUMA_BALANCING */
8781 #endif /* CONFIG_SCHED_DEBUG */
8783 __init void init_sched_fair_class(void)
8786 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8788 #ifdef CONFIG_NO_HZ_COMMON
8789 nohz.next_balance = jiffies;
8790 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);