1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
24 #include <linux/sched/mm.h>
25 #include <linux/sched/topology.h>
27 #include <linux/latencytop.h>
28 #include <linux/cpumask.h>
29 #include <linux/cpuidle.h>
30 #include <linux/slab.h>
31 #include <linux/profile.h>
32 #include <linux/interrupt.h>
33 #include <linux/mempolicy.h>
34 #include <linux/migrate.h>
35 #include <linux/task_work.h>
37 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
44 * NOTE: this latency value is not the same as the concept of
45 * 'timeslice length' - timeslices in CFS are of variable length
46 * and have no persistent notion like in traditional, time-slice
47 * based scheduling concepts.
49 * (to see the precise effective timeslice length of your workload,
50 * run vmstat and monitor the context-switches (cs) field)
52 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
54 unsigned int sysctl_sched_latency = 6000000ULL;
55 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
58 * The initial- and re-scaling of tunables is configurable
62 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
63 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
64 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
66 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
68 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
71 * Minimal preemption granularity for CPU-bound tasks:
73 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
75 unsigned int sysctl_sched_min_granularity = 750000ULL;
76 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
79 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
81 static unsigned int sched_nr_latency = 8;
84 * After fork, child runs first. If set to 0 (default) then
85 * parent will (try to) run first.
87 unsigned int sysctl_sched_child_runs_first __read_mostly;
90 * SCHED_OTHER wake-up granularity.
92 * This option delays the preemption effects of decoupled workloads
93 * and reduces their over-scheduling. Synchronous workloads will still
94 * have immediate wakeup/sleep latencies.
96 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
98 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
99 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
101 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
105 * For asym packing, by default the lower numbered cpu has higher priority.
107 int __weak arch_asym_cpu_priority(int cpu)
113 #ifdef CONFIG_CFS_BANDWIDTH
115 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
116 * each time a cfs_rq requests quota.
118 * Note: in the case that the slice exceeds the runtime remaining (either due
119 * to consumption or the quota being specified to be smaller than the slice)
120 * we will always only issue the remaining available time.
122 * (default: 5 msec, units: microseconds)
124 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
128 * The margin used when comparing utilization with CPU capacity:
129 * util * margin < capacity * 1024
133 unsigned int capacity_margin = 1280;
135 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
141 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
147 static inline void update_load_set(struct load_weight *lw, unsigned long w)
154 * Increase the granularity value when there are more CPUs,
155 * because with more CPUs the 'effective latency' as visible
156 * to users decreases. But the relationship is not linear,
157 * so pick a second-best guess by going with the log2 of the
160 * This idea comes from the SD scheduler of Con Kolivas:
162 static unsigned int get_update_sysctl_factor(void)
164 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
167 switch (sysctl_sched_tunable_scaling) {
168 case SCHED_TUNABLESCALING_NONE:
171 case SCHED_TUNABLESCALING_LINEAR:
174 case SCHED_TUNABLESCALING_LOG:
176 factor = 1 + ilog2(cpus);
183 static void update_sysctl(void)
185 unsigned int factor = get_update_sysctl_factor();
187 #define SET_SYSCTL(name) \
188 (sysctl_##name = (factor) * normalized_sysctl_##name)
189 SET_SYSCTL(sched_min_granularity);
190 SET_SYSCTL(sched_latency);
191 SET_SYSCTL(sched_wakeup_granularity);
195 void sched_init_granularity(void)
200 #define WMULT_CONST (~0U)
201 #define WMULT_SHIFT 32
203 static void __update_inv_weight(struct load_weight *lw)
207 if (likely(lw->inv_weight))
210 w = scale_load_down(lw->weight);
212 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
214 else if (unlikely(!w))
215 lw->inv_weight = WMULT_CONST;
217 lw->inv_weight = WMULT_CONST / w;
221 * delta_exec * weight / lw.weight
223 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
225 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
226 * we're guaranteed shift stays positive because inv_weight is guaranteed to
227 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
229 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
230 * weight/lw.weight <= 1, and therefore our shift will also be positive.
232 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
234 u64 fact = scale_load_down(weight);
235 int shift = WMULT_SHIFT;
237 __update_inv_weight(lw);
239 if (unlikely(fact >> 32)) {
246 /* hint to use a 32x32->64 mul */
247 fact = (u64)(u32)fact * lw->inv_weight;
254 return mul_u64_u32_shr(delta_exec, fact, shift);
258 const struct sched_class fair_sched_class;
260 /**************************************************************
261 * CFS operations on generic schedulable entities:
264 #ifdef CONFIG_FAIR_GROUP_SCHED
266 /* cpu runqueue to which this cfs_rq is attached */
267 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
272 /* An entity is a task if it doesn't "own" a runqueue */
273 #define entity_is_task(se) (!se->my_q)
275 static inline struct task_struct *task_of(struct sched_entity *se)
277 SCHED_WARN_ON(!entity_is_task(se));
278 return container_of(se, struct task_struct, se);
281 /* Walk up scheduling entities hierarchy */
282 #define for_each_sched_entity(se) \
283 for (; se; se = se->parent)
285 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
290 /* runqueue on which this entity is (to be) queued */
291 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
296 /* runqueue "owned" by this group */
297 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
302 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
304 if (!cfs_rq->on_list) {
305 struct rq *rq = rq_of(cfs_rq);
306 int cpu = cpu_of(rq);
308 * Ensure we either appear before our parent (if already
309 * enqueued) or force our parent to appear after us when it is
310 * enqueued. The fact that we always enqueue bottom-up
311 * reduces this to two cases and a special case for the root
312 * cfs_rq. Furthermore, it also means that we will always reset
313 * tmp_alone_branch either when the branch is connected
314 * to a tree or when we reach the beg of the tree
316 if (cfs_rq->tg->parent &&
317 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
319 * If parent is already on the list, we add the child
320 * just before. Thanks to circular linked property of
321 * the list, this means to put the child at the tail
322 * of the list that starts by parent.
324 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
325 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
327 * The branch is now connected to its tree so we can
328 * reset tmp_alone_branch to the beginning of the
331 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
332 } else if (!cfs_rq->tg->parent) {
334 * cfs rq without parent should be put
335 * at the tail of the list.
337 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
338 &rq->leaf_cfs_rq_list);
340 * We have reach the beg of a tree so we can reset
341 * tmp_alone_branch to the beginning of the list.
343 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
346 * The parent has not already been added so we want to
347 * make sure that it will be put after us.
348 * tmp_alone_branch points to the beg of the branch
349 * where we will add parent.
351 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
352 rq->tmp_alone_branch);
354 * update tmp_alone_branch to points to the new beg
357 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
364 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
366 if (cfs_rq->on_list) {
367 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
372 /* Iterate through all leaf cfs_rq's on a runqueue: */
373 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
374 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq *
378 is_same_group(struct sched_entity *se, struct sched_entity *pse)
380 if (se->cfs_rq == pse->cfs_rq)
386 static inline struct sched_entity *parent_entity(struct sched_entity *se)
392 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
394 int se_depth, pse_depth;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth = (*se)->depth;
405 pse_depth = (*pse)->depth;
407 while (se_depth > pse_depth) {
409 *se = parent_entity(*se);
412 while (pse_depth > se_depth) {
414 *pse = parent_entity(*pse);
417 while (!is_same_group(*se, *pse)) {
418 *se = parent_entity(*se);
419 *pse = parent_entity(*pse);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct *task_of(struct sched_entity *se)
427 return container_of(se, struct task_struct, se);
430 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
432 return container_of(cfs_rq, struct rq, cfs);
435 #define entity_is_task(se) 1
437 #define for_each_sched_entity(se) \
438 for (; se; se = NULL)
440 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
442 return &task_rq(p)->cfs;
445 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
447 struct task_struct *p = task_of(se);
448 struct rq *rq = task_rq(p);
453 /* runqueue "owned" by this group */
454 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
459 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
467 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
468 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
470 static inline struct sched_entity *parent_entity(struct sched_entity *se)
476 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
480 #endif /* CONFIG_FAIR_GROUP_SCHED */
482 static __always_inline
483 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
485 /**************************************************************
486 * Scheduling class tree data structure manipulation methods:
489 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
491 s64 delta = (s64)(vruntime - max_vruntime);
493 max_vruntime = vruntime;
498 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
500 s64 delta = (s64)(vruntime - min_vruntime);
502 min_vruntime = vruntime;
507 static inline int entity_before(struct sched_entity *a,
508 struct sched_entity *b)
510 return (s64)(a->vruntime - b->vruntime) < 0;
513 static void update_min_vruntime(struct cfs_rq *cfs_rq)
515 struct sched_entity *curr = cfs_rq->curr;
516 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
518 u64 vruntime = cfs_rq->min_vruntime;
522 vruntime = curr->vruntime;
527 if (leftmost) { /* non-empty tree */
528 struct sched_entity *se;
529 se = rb_entry(leftmost, struct sched_entity, run_node);
532 vruntime = se->vruntime;
534 vruntime = min_vruntime(vruntime, se->vruntime);
537 /* ensure we never gain time by being placed backwards. */
538 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
541 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
546 * Enqueue an entity into the rb-tree:
548 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
550 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
551 struct rb_node *parent = NULL;
552 struct sched_entity *entry;
553 bool leftmost = true;
556 * Find the right place in the rbtree:
560 entry = rb_entry(parent, struct sched_entity, run_node);
562 * We dont care about collisions. Nodes with
563 * the same key stay together.
565 if (entity_before(se, entry)) {
566 link = &parent->rb_left;
568 link = &parent->rb_right;
573 rb_link_node(&se->run_node, parent, link);
574 rb_insert_color_cached(&se->run_node,
575 &cfs_rq->tasks_timeline, leftmost);
578 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
580 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
583 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
585 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
590 return rb_entry(left, struct sched_entity, run_node);
593 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
595 struct rb_node *next = rb_next(&se->run_node);
600 return rb_entry(next, struct sched_entity, run_node);
603 #ifdef CONFIG_SCHED_DEBUG
604 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
606 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
611 return rb_entry(last, struct sched_entity, run_node);
614 /**************************************************************
615 * Scheduling class statistics methods:
618 int sched_proc_update_handler(struct ctl_table *table, int write,
619 void __user *buffer, size_t *lenp,
622 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
623 unsigned int factor = get_update_sysctl_factor();
628 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
629 sysctl_sched_min_granularity);
631 #define WRT_SYSCTL(name) \
632 (normalized_sysctl_##name = sysctl_##name / (factor))
633 WRT_SYSCTL(sched_min_granularity);
634 WRT_SYSCTL(sched_latency);
635 WRT_SYSCTL(sched_wakeup_granularity);
645 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
647 if (unlikely(se->load.weight != NICE_0_LOAD))
648 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
654 * The idea is to set a period in which each task runs once.
656 * When there are too many tasks (sched_nr_latency) we have to stretch
657 * this period because otherwise the slices get too small.
659 * p = (nr <= nl) ? l : l*nr/nl
661 static u64 __sched_period(unsigned long nr_running)
663 if (unlikely(nr_running > sched_nr_latency))
664 return nr_running * sysctl_sched_min_granularity;
666 return sysctl_sched_latency;
670 * We calculate the wall-time slice from the period by taking a part
671 * proportional to the weight.
675 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
677 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
679 for_each_sched_entity(se) {
680 struct load_weight *load;
681 struct load_weight lw;
683 cfs_rq = cfs_rq_of(se);
684 load = &cfs_rq->load;
686 if (unlikely(!se->on_rq)) {
689 update_load_add(&lw, se->load.weight);
692 slice = __calc_delta(slice, se->load.weight, load);
698 * We calculate the vruntime slice of a to-be-inserted task.
702 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
704 return calc_delta_fair(sched_slice(cfs_rq, se), se);
709 #include "sched-pelt.h"
711 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
712 static unsigned long task_h_load(struct task_struct *p);
714 /* Give new sched_entity start runnable values to heavy its load in infant time */
715 void init_entity_runnable_average(struct sched_entity *se)
717 struct sched_avg *sa = &se->avg;
719 sa->last_update_time = 0;
721 * sched_avg's period_contrib should be strictly less then 1024, so
722 * we give it 1023 to make sure it is almost a period (1024us), and
723 * will definitely be update (after enqueue).
725 sa->period_contrib = 1023;
727 * Tasks are intialized with full load to be seen as heavy tasks until
728 * they get a chance to stabilize to their real load level.
729 * Group entities are intialized with zero load to reflect the fact that
730 * nothing has been attached to the task group yet.
732 if (entity_is_task(se))
733 sa->load_avg = scale_load_down(se->load.weight);
734 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
736 * At this point, util_avg won't be used in select_task_rq_fair anyway
740 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
743 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
744 static void attach_entity_cfs_rq(struct sched_entity *se);
747 * With new tasks being created, their initial util_avgs are extrapolated
748 * based on the cfs_rq's current util_avg:
750 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
752 * However, in many cases, the above util_avg does not give a desired
753 * value. Moreover, the sum of the util_avgs may be divergent, such
754 * as when the series is a harmonic series.
756 * To solve this problem, we also cap the util_avg of successive tasks to
757 * only 1/2 of the left utilization budget:
759 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
761 * where n denotes the nth task and cpu_scale the CPU capacity.
763 * For example, for a CPU with 1024 of capacity, a simplest series from
764 * the beginning would be like:
766 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
767 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
769 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
770 * if util_avg > util_avg_cap.
772 void post_init_entity_util_avg(struct sched_entity *se)
774 struct cfs_rq *cfs_rq = cfs_rq_of(se);
775 struct sched_avg *sa = &se->avg;
776 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
777 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
780 if (cfs_rq->avg.util_avg != 0) {
781 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
782 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
784 if (sa->util_avg > cap)
789 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
792 if (entity_is_task(se)) {
793 struct task_struct *p = task_of(se);
794 if (p->sched_class != &fair_sched_class) {
796 * For !fair tasks do:
798 update_cfs_rq_load_avg(now, cfs_rq);
799 attach_entity_load_avg(cfs_rq, se);
800 switched_from_fair(rq, p);
802 * such that the next switched_to_fair() has the
805 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
810 attach_entity_cfs_rq(se);
813 #else /* !CONFIG_SMP */
814 void init_entity_runnable_average(struct sched_entity *se)
817 void post_init_entity_util_avg(struct sched_entity *se)
820 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
823 #endif /* CONFIG_SMP */
826 * Update the current task's runtime statistics.
828 static void update_curr(struct cfs_rq *cfs_rq)
830 struct sched_entity *curr = cfs_rq->curr;
831 u64 now = rq_clock_task(rq_of(cfs_rq));
837 delta_exec = now - curr->exec_start;
838 if (unlikely((s64)delta_exec <= 0))
841 curr->exec_start = now;
843 schedstat_set(curr->statistics.exec_max,
844 max(delta_exec, curr->statistics.exec_max));
846 curr->sum_exec_runtime += delta_exec;
847 schedstat_add(cfs_rq->exec_clock, delta_exec);
849 curr->vruntime += calc_delta_fair(delta_exec, curr);
850 update_min_vruntime(cfs_rq);
852 if (entity_is_task(curr)) {
853 struct task_struct *curtask = task_of(curr);
855 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
856 cpuacct_charge(curtask, delta_exec);
857 account_group_exec_runtime(curtask, delta_exec);
860 account_cfs_rq_runtime(cfs_rq, delta_exec);
863 static void update_curr_fair(struct rq *rq)
865 update_curr(cfs_rq_of(&rq->curr->se));
869 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
871 u64 wait_start, prev_wait_start;
873 if (!schedstat_enabled())
876 wait_start = rq_clock(rq_of(cfs_rq));
877 prev_wait_start = schedstat_val(se->statistics.wait_start);
879 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
880 likely(wait_start > prev_wait_start))
881 wait_start -= prev_wait_start;
883 schedstat_set(se->statistics.wait_start, wait_start);
887 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
889 struct task_struct *p;
892 if (!schedstat_enabled())
895 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
897 if (entity_is_task(se)) {
899 if (task_on_rq_migrating(p)) {
901 * Preserve migrating task's wait time so wait_start
902 * time stamp can be adjusted to accumulate wait time
903 * prior to migration.
905 schedstat_set(se->statistics.wait_start, delta);
908 trace_sched_stat_wait(p, delta);
911 schedstat_set(se->statistics.wait_max,
912 max(schedstat_val(se->statistics.wait_max), delta));
913 schedstat_inc(se->statistics.wait_count);
914 schedstat_add(se->statistics.wait_sum, delta);
915 schedstat_set(se->statistics.wait_start, 0);
919 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
921 struct task_struct *tsk = NULL;
922 u64 sleep_start, block_start;
924 if (!schedstat_enabled())
927 sleep_start = schedstat_val(se->statistics.sleep_start);
928 block_start = schedstat_val(se->statistics.block_start);
930 if (entity_is_task(se))
934 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
939 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
940 schedstat_set(se->statistics.sleep_max, delta);
942 schedstat_set(se->statistics.sleep_start, 0);
943 schedstat_add(se->statistics.sum_sleep_runtime, delta);
946 account_scheduler_latency(tsk, delta >> 10, 1);
947 trace_sched_stat_sleep(tsk, delta);
951 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
956 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
957 schedstat_set(se->statistics.block_max, delta);
959 schedstat_set(se->statistics.block_start, 0);
960 schedstat_add(se->statistics.sum_sleep_runtime, delta);
963 if (tsk->in_iowait) {
964 schedstat_add(se->statistics.iowait_sum, delta);
965 schedstat_inc(se->statistics.iowait_count);
966 trace_sched_stat_iowait(tsk, delta);
969 trace_sched_stat_blocked(tsk, delta);
972 * Blocking time is in units of nanosecs, so shift by
973 * 20 to get a milliseconds-range estimation of the
974 * amount of time that the task spent sleeping:
976 if (unlikely(prof_on == SLEEP_PROFILING)) {
977 profile_hits(SLEEP_PROFILING,
978 (void *)get_wchan(tsk),
981 account_scheduler_latency(tsk, delta >> 10, 0);
987 * Task is being enqueued - update stats:
990 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
992 if (!schedstat_enabled())
996 * Are we enqueueing a waiting task? (for current tasks
997 * a dequeue/enqueue event is a NOP)
999 if (se != cfs_rq->curr)
1000 update_stats_wait_start(cfs_rq, se);
1002 if (flags & ENQUEUE_WAKEUP)
1003 update_stats_enqueue_sleeper(cfs_rq, se);
1007 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1010 if (!schedstat_enabled())
1014 * Mark the end of the wait period if dequeueing a
1017 if (se != cfs_rq->curr)
1018 update_stats_wait_end(cfs_rq, se);
1020 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1021 struct task_struct *tsk = task_of(se);
1023 if (tsk->state & TASK_INTERRUPTIBLE)
1024 schedstat_set(se->statistics.sleep_start,
1025 rq_clock(rq_of(cfs_rq)));
1026 if (tsk->state & TASK_UNINTERRUPTIBLE)
1027 schedstat_set(se->statistics.block_start,
1028 rq_clock(rq_of(cfs_rq)));
1033 * We are picking a new current task - update its stats:
1036 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1039 * We are starting a new run period:
1041 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1044 /**************************************************
1045 * Scheduling class queueing methods:
1048 #ifdef CONFIG_NUMA_BALANCING
1050 * Approximate time to scan a full NUMA task in ms. The task scan period is
1051 * calculated based on the tasks virtual memory size and
1052 * numa_balancing_scan_size.
1054 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1055 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1057 /* Portion of address space to scan in MB */
1058 unsigned int sysctl_numa_balancing_scan_size = 256;
1060 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1061 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1066 spinlock_t lock; /* nr_tasks, tasks */
1071 struct rcu_head rcu;
1072 unsigned long total_faults;
1073 unsigned long max_faults_cpu;
1075 * Faults_cpu is used to decide whether memory should move
1076 * towards the CPU. As a consequence, these stats are weighted
1077 * more by CPU use than by memory faults.
1079 unsigned long *faults_cpu;
1080 unsigned long faults[0];
1083 static inline unsigned long group_faults_priv(struct numa_group *ng);
1084 static inline unsigned long group_faults_shared(struct numa_group *ng);
1086 static unsigned int task_nr_scan_windows(struct task_struct *p)
1088 unsigned long rss = 0;
1089 unsigned long nr_scan_pages;
1092 * Calculations based on RSS as non-present and empty pages are skipped
1093 * by the PTE scanner and NUMA hinting faults should be trapped based
1096 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1097 rss = get_mm_rss(p->mm);
1099 rss = nr_scan_pages;
1101 rss = round_up(rss, nr_scan_pages);
1102 return rss / nr_scan_pages;
1105 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1106 #define MAX_SCAN_WINDOW 2560
1108 static unsigned int task_scan_min(struct task_struct *p)
1110 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1111 unsigned int scan, floor;
1112 unsigned int windows = 1;
1114 if (scan_size < MAX_SCAN_WINDOW)
1115 windows = MAX_SCAN_WINDOW / scan_size;
1116 floor = 1000 / windows;
1118 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1119 return max_t(unsigned int, floor, scan);
1122 static unsigned int task_scan_start(struct task_struct *p)
1124 unsigned long smin = task_scan_min(p);
1125 unsigned long period = smin;
1127 /* Scale the maximum scan period with the amount of shared memory. */
1128 if (p->numa_group) {
1129 struct numa_group *ng = p->numa_group;
1130 unsigned long shared = group_faults_shared(ng);
1131 unsigned long private = group_faults_priv(ng);
1133 period *= atomic_read(&ng->refcount);
1134 period *= shared + 1;
1135 period /= private + shared + 1;
1138 return max(smin, period);
1141 static unsigned int task_scan_max(struct task_struct *p)
1143 unsigned long smin = task_scan_min(p);
1146 /* Watch for min being lower than max due to floor calculations */
1147 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1149 /* Scale the maximum scan period with the amount of shared memory. */
1150 if (p->numa_group) {
1151 struct numa_group *ng = p->numa_group;
1152 unsigned long shared = group_faults_shared(ng);
1153 unsigned long private = group_faults_priv(ng);
1154 unsigned long period = smax;
1156 period *= atomic_read(&ng->refcount);
1157 period *= shared + 1;
1158 period /= private + shared + 1;
1160 smax = max(smax, period);
1163 return max(smin, smax);
1166 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1168 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1169 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1172 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1174 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1175 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1178 /* Shared or private faults. */
1179 #define NR_NUMA_HINT_FAULT_TYPES 2
1181 /* Memory and CPU locality */
1182 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1184 /* Averaged statistics, and temporary buffers. */
1185 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1187 pid_t task_numa_group_id(struct task_struct *p)
1189 return p->numa_group ? p->numa_group->gid : 0;
1193 * The averaged statistics, shared & private, memory & cpu,
1194 * occupy the first half of the array. The second half of the
1195 * array is for current counters, which are averaged into the
1196 * first set by task_numa_placement.
1198 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1200 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1203 static inline unsigned long task_faults(struct task_struct *p, int nid)
1205 if (!p->numa_faults)
1208 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1209 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1212 static inline unsigned long group_faults(struct task_struct *p, int nid)
1217 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1218 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1221 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1223 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1224 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1227 static inline unsigned long group_faults_priv(struct numa_group *ng)
1229 unsigned long faults = 0;
1232 for_each_online_node(node) {
1233 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1239 static inline unsigned long group_faults_shared(struct numa_group *ng)
1241 unsigned long faults = 0;
1244 for_each_online_node(node) {
1245 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1252 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1253 * considered part of a numa group's pseudo-interleaving set. Migrations
1254 * between these nodes are slowed down, to allow things to settle down.
1256 #define ACTIVE_NODE_FRACTION 3
1258 static bool numa_is_active_node(int nid, struct numa_group *ng)
1260 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1263 /* Handle placement on systems where not all nodes are directly connected. */
1264 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1265 int maxdist, bool task)
1267 unsigned long score = 0;
1271 * All nodes are directly connected, and the same distance
1272 * from each other. No need for fancy placement algorithms.
1274 if (sched_numa_topology_type == NUMA_DIRECT)
1278 * This code is called for each node, introducing N^2 complexity,
1279 * which should be ok given the number of nodes rarely exceeds 8.
1281 for_each_online_node(node) {
1282 unsigned long faults;
1283 int dist = node_distance(nid, node);
1286 * The furthest away nodes in the system are not interesting
1287 * for placement; nid was already counted.
1289 if (dist == sched_max_numa_distance || node == nid)
1293 * On systems with a backplane NUMA topology, compare groups
1294 * of nodes, and move tasks towards the group with the most
1295 * memory accesses. When comparing two nodes at distance
1296 * "hoplimit", only nodes closer by than "hoplimit" are part
1297 * of each group. Skip other nodes.
1299 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1303 /* Add up the faults from nearby nodes. */
1305 faults = task_faults(p, node);
1307 faults = group_faults(p, node);
1310 * On systems with a glueless mesh NUMA topology, there are
1311 * no fixed "groups of nodes". Instead, nodes that are not
1312 * directly connected bounce traffic through intermediate
1313 * nodes; a numa_group can occupy any set of nodes.
1314 * The further away a node is, the less the faults count.
1315 * This seems to result in good task placement.
1317 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1318 faults *= (sched_max_numa_distance - dist);
1319 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1329 * These return the fraction of accesses done by a particular task, or
1330 * task group, on a particular numa node. The group weight is given a
1331 * larger multiplier, in order to group tasks together that are almost
1332 * evenly spread out between numa nodes.
1334 static inline unsigned long task_weight(struct task_struct *p, int nid,
1337 unsigned long faults, total_faults;
1339 if (!p->numa_faults)
1342 total_faults = p->total_numa_faults;
1347 faults = task_faults(p, nid);
1348 faults += score_nearby_nodes(p, nid, dist, true);
1350 return 1000 * faults / total_faults;
1353 static inline unsigned long group_weight(struct task_struct *p, int nid,
1356 unsigned long faults, total_faults;
1361 total_faults = p->numa_group->total_faults;
1366 faults = group_faults(p, nid);
1367 faults += score_nearby_nodes(p, nid, dist, false);
1369 return 1000 * faults / total_faults;
1372 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1373 int src_nid, int dst_cpu)
1375 struct numa_group *ng = p->numa_group;
1376 int dst_nid = cpu_to_node(dst_cpu);
1377 int last_cpupid, this_cpupid;
1379 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1382 * Multi-stage node selection is used in conjunction with a periodic
1383 * migration fault to build a temporal task<->page relation. By using
1384 * a two-stage filter we remove short/unlikely relations.
1386 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1387 * a task's usage of a particular page (n_p) per total usage of this
1388 * page (n_t) (in a given time-span) to a probability.
1390 * Our periodic faults will sample this probability and getting the
1391 * same result twice in a row, given these samples are fully
1392 * independent, is then given by P(n)^2, provided our sample period
1393 * is sufficiently short compared to the usage pattern.
1395 * This quadric squishes small probabilities, making it less likely we
1396 * act on an unlikely task<->page relation.
1398 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1399 if (!cpupid_pid_unset(last_cpupid) &&
1400 cpupid_to_nid(last_cpupid) != dst_nid)
1403 /* Always allow migrate on private faults */
1404 if (cpupid_match_pid(p, last_cpupid))
1407 /* A shared fault, but p->numa_group has not been set up yet. */
1412 * Destination node is much more heavily used than the source
1413 * node? Allow migration.
1415 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1416 ACTIVE_NODE_FRACTION)
1420 * Distribute memory according to CPU & memory use on each node,
1421 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1423 * faults_cpu(dst) 3 faults_cpu(src)
1424 * --------------- * - > ---------------
1425 * faults_mem(dst) 4 faults_mem(src)
1427 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1428 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1431 static unsigned long weighted_cpuload(struct rq *rq);
1432 static unsigned long source_load(int cpu, int type);
1433 static unsigned long target_load(int cpu, int type);
1434 static unsigned long capacity_of(int cpu);
1436 /* Cached statistics for all CPUs within a node */
1438 unsigned long nr_running;
1441 /* Total compute capacity of CPUs on a node */
1442 unsigned long compute_capacity;
1444 /* Approximate capacity in terms of runnable tasks on a node */
1445 unsigned long task_capacity;
1446 int has_free_capacity;
1450 * XXX borrowed from update_sg_lb_stats
1452 static void update_numa_stats(struct numa_stats *ns, int nid)
1454 int smt, cpu, cpus = 0;
1455 unsigned long capacity;
1457 memset(ns, 0, sizeof(*ns));
1458 for_each_cpu(cpu, cpumask_of_node(nid)) {
1459 struct rq *rq = cpu_rq(cpu);
1461 ns->nr_running += rq->nr_running;
1462 ns->load += weighted_cpuload(rq);
1463 ns->compute_capacity += capacity_of(cpu);
1469 * If we raced with hotplug and there are no CPUs left in our mask
1470 * the @ns structure is NULL'ed and task_numa_compare() will
1471 * not find this node attractive.
1473 * We'll either bail at !has_free_capacity, or we'll detect a huge
1474 * imbalance and bail there.
1479 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1480 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1481 capacity = cpus / smt; /* cores */
1483 ns->task_capacity = min_t(unsigned, capacity,
1484 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1485 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1488 struct task_numa_env {
1489 struct task_struct *p;
1491 int src_cpu, src_nid;
1492 int dst_cpu, dst_nid;
1494 struct numa_stats src_stats, dst_stats;
1499 struct task_struct *best_task;
1504 static void task_numa_assign(struct task_numa_env *env,
1505 struct task_struct *p, long imp)
1508 put_task_struct(env->best_task);
1513 env->best_imp = imp;
1514 env->best_cpu = env->dst_cpu;
1517 static bool load_too_imbalanced(long src_load, long dst_load,
1518 struct task_numa_env *env)
1521 long orig_src_load, orig_dst_load;
1522 long src_capacity, dst_capacity;
1525 * The load is corrected for the CPU capacity available on each node.
1528 * ------------ vs ---------
1529 * src_capacity dst_capacity
1531 src_capacity = env->src_stats.compute_capacity;
1532 dst_capacity = env->dst_stats.compute_capacity;
1534 /* We care about the slope of the imbalance, not the direction. */
1535 if (dst_load < src_load)
1536 swap(dst_load, src_load);
1538 /* Is the difference below the threshold? */
1539 imb = dst_load * src_capacity * 100 -
1540 src_load * dst_capacity * env->imbalance_pct;
1545 * The imbalance is above the allowed threshold.
1546 * Compare it with the old imbalance.
1548 orig_src_load = env->src_stats.load;
1549 orig_dst_load = env->dst_stats.load;
1551 if (orig_dst_load < orig_src_load)
1552 swap(orig_dst_load, orig_src_load);
1554 old_imb = orig_dst_load * src_capacity * 100 -
1555 orig_src_load * dst_capacity * env->imbalance_pct;
1557 /* Would this change make things worse? */
1558 return (imb > old_imb);
1562 * This checks if the overall compute and NUMA accesses of the system would
1563 * be improved if the source tasks was migrated to the target dst_cpu taking
1564 * into account that it might be best if task running on the dst_cpu should
1565 * be exchanged with the source task
1567 static void task_numa_compare(struct task_numa_env *env,
1568 long taskimp, long groupimp)
1570 struct rq *src_rq = cpu_rq(env->src_cpu);
1571 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1572 struct task_struct *cur;
1573 long src_load, dst_load;
1575 long imp = env->p->numa_group ? groupimp : taskimp;
1577 int dist = env->dist;
1580 cur = task_rcu_dereference(&dst_rq->curr);
1581 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1585 * Because we have preemption enabled we can get migrated around and
1586 * end try selecting ourselves (current == env->p) as a swap candidate.
1592 * "imp" is the fault differential for the source task between the
1593 * source and destination node. Calculate the total differential for
1594 * the source task and potential destination task. The more negative
1595 * the value is, the more rmeote accesses that would be expected to
1596 * be incurred if the tasks were swapped.
1599 /* Skip this swap candidate if cannot move to the source cpu */
1600 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1604 * If dst and source tasks are in the same NUMA group, or not
1605 * in any group then look only at task weights.
1607 if (cur->numa_group == env->p->numa_group) {
1608 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1609 task_weight(cur, env->dst_nid, dist);
1611 * Add some hysteresis to prevent swapping the
1612 * tasks within a group over tiny differences.
1614 if (cur->numa_group)
1618 * Compare the group weights. If a task is all by
1619 * itself (not part of a group), use the task weight
1622 if (cur->numa_group)
1623 imp += group_weight(cur, env->src_nid, dist) -
1624 group_weight(cur, env->dst_nid, dist);
1626 imp += task_weight(cur, env->src_nid, dist) -
1627 task_weight(cur, env->dst_nid, dist);
1631 if (imp <= env->best_imp && moveimp <= env->best_imp)
1635 /* Is there capacity at our destination? */
1636 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1637 !env->dst_stats.has_free_capacity)
1643 /* Balance doesn't matter much if we're running a task per cpu */
1644 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1645 dst_rq->nr_running == 1)
1649 * In the overloaded case, try and keep the load balanced.
1652 load = task_h_load(env->p);
1653 dst_load = env->dst_stats.load + load;
1654 src_load = env->src_stats.load - load;
1656 if (moveimp > imp && moveimp > env->best_imp) {
1658 * If the improvement from just moving env->p direction is
1659 * better than swapping tasks around, check if a move is
1660 * possible. Store a slightly smaller score than moveimp,
1661 * so an actually idle CPU will win.
1663 if (!load_too_imbalanced(src_load, dst_load, env)) {
1670 if (imp <= env->best_imp)
1674 load = task_h_load(cur);
1679 if (load_too_imbalanced(src_load, dst_load, env))
1683 * One idle CPU per node is evaluated for a task numa move.
1684 * Call select_idle_sibling to maybe find a better one.
1688 * select_idle_siblings() uses an per-cpu cpumask that
1689 * can be used from IRQ context.
1691 local_irq_disable();
1692 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1698 task_numa_assign(env, cur, imp);
1703 static void task_numa_find_cpu(struct task_numa_env *env,
1704 long taskimp, long groupimp)
1708 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1709 /* Skip this CPU if the source task cannot migrate */
1710 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1714 task_numa_compare(env, taskimp, groupimp);
1718 /* Only move tasks to a NUMA node less busy than the current node. */
1719 static bool numa_has_capacity(struct task_numa_env *env)
1721 struct numa_stats *src = &env->src_stats;
1722 struct numa_stats *dst = &env->dst_stats;
1724 if (src->has_free_capacity && !dst->has_free_capacity)
1728 * Only consider a task move if the source has a higher load
1729 * than the destination, corrected for CPU capacity on each node.
1731 * src->load dst->load
1732 * --------------------- vs ---------------------
1733 * src->compute_capacity dst->compute_capacity
1735 if (src->load * dst->compute_capacity * env->imbalance_pct >
1737 dst->load * src->compute_capacity * 100)
1743 static int task_numa_migrate(struct task_struct *p)
1745 struct task_numa_env env = {
1748 .src_cpu = task_cpu(p),
1749 .src_nid = task_node(p),
1751 .imbalance_pct = 112,
1757 struct sched_domain *sd;
1758 unsigned long taskweight, groupweight;
1760 long taskimp, groupimp;
1763 * Pick the lowest SD_NUMA domain, as that would have the smallest
1764 * imbalance and would be the first to start moving tasks about.
1766 * And we want to avoid any moving of tasks about, as that would create
1767 * random movement of tasks -- counter the numa conditions we're trying
1771 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1773 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1777 * Cpusets can break the scheduler domain tree into smaller
1778 * balance domains, some of which do not cross NUMA boundaries.
1779 * Tasks that are "trapped" in such domains cannot be migrated
1780 * elsewhere, so there is no point in (re)trying.
1782 if (unlikely(!sd)) {
1783 p->numa_preferred_nid = task_node(p);
1787 env.dst_nid = p->numa_preferred_nid;
1788 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1789 taskweight = task_weight(p, env.src_nid, dist);
1790 groupweight = group_weight(p, env.src_nid, dist);
1791 update_numa_stats(&env.src_stats, env.src_nid);
1792 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1793 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1794 update_numa_stats(&env.dst_stats, env.dst_nid);
1796 /* Try to find a spot on the preferred nid. */
1797 if (numa_has_capacity(&env))
1798 task_numa_find_cpu(&env, taskimp, groupimp);
1801 * Look at other nodes in these cases:
1802 * - there is no space available on the preferred_nid
1803 * - the task is part of a numa_group that is interleaved across
1804 * multiple NUMA nodes; in order to better consolidate the group,
1805 * we need to check other locations.
1807 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1808 for_each_online_node(nid) {
1809 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1812 dist = node_distance(env.src_nid, env.dst_nid);
1813 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1815 taskweight = task_weight(p, env.src_nid, dist);
1816 groupweight = group_weight(p, env.src_nid, dist);
1819 /* Only consider nodes where both task and groups benefit */
1820 taskimp = task_weight(p, nid, dist) - taskweight;
1821 groupimp = group_weight(p, nid, dist) - groupweight;
1822 if (taskimp < 0 && groupimp < 0)
1827 update_numa_stats(&env.dst_stats, env.dst_nid);
1828 if (numa_has_capacity(&env))
1829 task_numa_find_cpu(&env, taskimp, groupimp);
1834 * If the task is part of a workload that spans multiple NUMA nodes,
1835 * and is migrating into one of the workload's active nodes, remember
1836 * this node as the task's preferred numa node, so the workload can
1838 * A task that migrated to a second choice node will be better off
1839 * trying for a better one later. Do not set the preferred node here.
1841 if (p->numa_group) {
1842 struct numa_group *ng = p->numa_group;
1844 if (env.best_cpu == -1)
1849 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1850 sched_setnuma(p, env.dst_nid);
1853 /* No better CPU than the current one was found. */
1854 if (env.best_cpu == -1)
1858 * Reset the scan period if the task is being rescheduled on an
1859 * alternative node to recheck if the tasks is now properly placed.
1861 p->numa_scan_period = task_scan_start(p);
1863 if (env.best_task == NULL) {
1864 ret = migrate_task_to(p, env.best_cpu);
1866 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1870 ret = migrate_swap(p, env.best_task);
1872 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1873 put_task_struct(env.best_task);
1877 /* Attempt to migrate a task to a CPU on the preferred node. */
1878 static void numa_migrate_preferred(struct task_struct *p)
1880 unsigned long interval = HZ;
1882 /* This task has no NUMA fault statistics yet */
1883 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1886 /* Periodically retry migrating the task to the preferred node */
1887 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1888 p->numa_migrate_retry = jiffies + interval;
1890 /* Success if task is already running on preferred CPU */
1891 if (task_node(p) == p->numa_preferred_nid)
1894 /* Otherwise, try migrate to a CPU on the preferred node */
1895 task_numa_migrate(p);
1899 * Find out how many nodes on the workload is actively running on. Do this by
1900 * tracking the nodes from which NUMA hinting faults are triggered. This can
1901 * be different from the set of nodes where the workload's memory is currently
1904 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1906 unsigned long faults, max_faults = 0;
1907 int nid, active_nodes = 0;
1909 for_each_online_node(nid) {
1910 faults = group_faults_cpu(numa_group, nid);
1911 if (faults > max_faults)
1912 max_faults = faults;
1915 for_each_online_node(nid) {
1916 faults = group_faults_cpu(numa_group, nid);
1917 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1921 numa_group->max_faults_cpu = max_faults;
1922 numa_group->active_nodes = active_nodes;
1926 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1927 * increments. The more local the fault statistics are, the higher the scan
1928 * period will be for the next scan window. If local/(local+remote) ratio is
1929 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1930 * the scan period will decrease. Aim for 70% local accesses.
1932 #define NUMA_PERIOD_SLOTS 10
1933 #define NUMA_PERIOD_THRESHOLD 7
1936 * Increase the scan period (slow down scanning) if the majority of
1937 * our memory is already on our local node, or if the majority of
1938 * the page accesses are shared with other processes.
1939 * Otherwise, decrease the scan period.
1941 static void update_task_scan_period(struct task_struct *p,
1942 unsigned long shared, unsigned long private)
1944 unsigned int period_slot;
1945 int lr_ratio, ps_ratio;
1948 unsigned long remote = p->numa_faults_locality[0];
1949 unsigned long local = p->numa_faults_locality[1];
1952 * If there were no record hinting faults then either the task is
1953 * completely idle or all activity is areas that are not of interest
1954 * to automatic numa balancing. Related to that, if there were failed
1955 * migration then it implies we are migrating too quickly or the local
1956 * node is overloaded. In either case, scan slower
1958 if (local + shared == 0 || p->numa_faults_locality[2]) {
1959 p->numa_scan_period = min(p->numa_scan_period_max,
1960 p->numa_scan_period << 1);
1962 p->mm->numa_next_scan = jiffies +
1963 msecs_to_jiffies(p->numa_scan_period);
1969 * Prepare to scale scan period relative to the current period.
1970 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1971 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1972 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1974 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1975 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1976 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1978 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1980 * Most memory accesses are local. There is no need to
1981 * do fast NUMA scanning, since memory is already local.
1983 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1986 diff = slot * period_slot;
1987 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1989 * Most memory accesses are shared with other tasks.
1990 * There is no point in continuing fast NUMA scanning,
1991 * since other tasks may just move the memory elsewhere.
1993 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1996 diff = slot * period_slot;
1999 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2000 * yet they are not on the local NUMA node. Speed up
2001 * NUMA scanning to get the memory moved over.
2003 int ratio = max(lr_ratio, ps_ratio);
2004 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2007 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2008 task_scan_min(p), task_scan_max(p));
2009 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2013 * Get the fraction of time the task has been running since the last
2014 * NUMA placement cycle. The scheduler keeps similar statistics, but
2015 * decays those on a 32ms period, which is orders of magnitude off
2016 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2017 * stats only if the task is so new there are no NUMA statistics yet.
2019 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2021 u64 runtime, delta, now;
2022 /* Use the start of this time slice to avoid calculations. */
2023 now = p->se.exec_start;
2024 runtime = p->se.sum_exec_runtime;
2026 if (p->last_task_numa_placement) {
2027 delta = runtime - p->last_sum_exec_runtime;
2028 *period = now - p->last_task_numa_placement;
2030 /* Avoid time going backwards, prevent potential divide error: */
2031 if (unlikely((s64)*period < 0))
2034 delta = p->se.avg.load_sum / p->se.load.weight;
2035 *period = LOAD_AVG_MAX;
2038 p->last_sum_exec_runtime = runtime;
2039 p->last_task_numa_placement = now;
2045 * Determine the preferred nid for a task in a numa_group. This needs to
2046 * be done in a way that produces consistent results with group_weight,
2047 * otherwise workloads might not converge.
2049 static int preferred_group_nid(struct task_struct *p, int nid)
2054 /* Direct connections between all NUMA nodes. */
2055 if (sched_numa_topology_type == NUMA_DIRECT)
2059 * On a system with glueless mesh NUMA topology, group_weight
2060 * scores nodes according to the number of NUMA hinting faults on
2061 * both the node itself, and on nearby nodes.
2063 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2064 unsigned long score, max_score = 0;
2065 int node, max_node = nid;
2067 dist = sched_max_numa_distance;
2069 for_each_online_node(node) {
2070 score = group_weight(p, node, dist);
2071 if (score > max_score) {
2080 * Finding the preferred nid in a system with NUMA backplane
2081 * interconnect topology is more involved. The goal is to locate
2082 * tasks from numa_groups near each other in the system, and
2083 * untangle workloads from different sides of the system. This requires
2084 * searching down the hierarchy of node groups, recursively searching
2085 * inside the highest scoring group of nodes. The nodemask tricks
2086 * keep the complexity of the search down.
2088 nodes = node_online_map;
2089 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2090 unsigned long max_faults = 0;
2091 nodemask_t max_group = NODE_MASK_NONE;
2094 /* Are there nodes at this distance from each other? */
2095 if (!find_numa_distance(dist))
2098 for_each_node_mask(a, nodes) {
2099 unsigned long faults = 0;
2100 nodemask_t this_group;
2101 nodes_clear(this_group);
2103 /* Sum group's NUMA faults; includes a==b case. */
2104 for_each_node_mask(b, nodes) {
2105 if (node_distance(a, b) < dist) {
2106 faults += group_faults(p, b);
2107 node_set(b, this_group);
2108 node_clear(b, nodes);
2112 /* Remember the top group. */
2113 if (faults > max_faults) {
2114 max_faults = faults;
2115 max_group = this_group;
2117 * subtle: at the smallest distance there is
2118 * just one node left in each "group", the
2119 * winner is the preferred nid.
2124 /* Next round, evaluate the nodes within max_group. */
2132 static void task_numa_placement(struct task_struct *p)
2134 int seq, nid, max_nid = -1, max_group_nid = -1;
2135 unsigned long max_faults = 0, max_group_faults = 0;
2136 unsigned long fault_types[2] = { 0, 0 };
2137 unsigned long total_faults;
2138 u64 runtime, period;
2139 spinlock_t *group_lock = NULL;
2142 * The p->mm->numa_scan_seq field gets updated without
2143 * exclusive access. Use READ_ONCE() here to ensure
2144 * that the field is read in a single access:
2146 seq = READ_ONCE(p->mm->numa_scan_seq);
2147 if (p->numa_scan_seq == seq)
2149 p->numa_scan_seq = seq;
2150 p->numa_scan_period_max = task_scan_max(p);
2152 total_faults = p->numa_faults_locality[0] +
2153 p->numa_faults_locality[1];
2154 runtime = numa_get_avg_runtime(p, &period);
2156 /* If the task is part of a group prevent parallel updates to group stats */
2157 if (p->numa_group) {
2158 group_lock = &p->numa_group->lock;
2159 spin_lock_irq(group_lock);
2162 /* Find the node with the highest number of faults */
2163 for_each_online_node(nid) {
2164 /* Keep track of the offsets in numa_faults array */
2165 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2166 unsigned long faults = 0, group_faults = 0;
2169 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2170 long diff, f_diff, f_weight;
2172 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2173 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2174 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2175 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2177 /* Decay existing window, copy faults since last scan */
2178 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2179 fault_types[priv] += p->numa_faults[membuf_idx];
2180 p->numa_faults[membuf_idx] = 0;
2183 * Normalize the faults_from, so all tasks in a group
2184 * count according to CPU use, instead of by the raw
2185 * number of faults. Tasks with little runtime have
2186 * little over-all impact on throughput, and thus their
2187 * faults are less important.
2189 f_weight = div64_u64(runtime << 16, period + 1);
2190 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2192 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2193 p->numa_faults[cpubuf_idx] = 0;
2195 p->numa_faults[mem_idx] += diff;
2196 p->numa_faults[cpu_idx] += f_diff;
2197 faults += p->numa_faults[mem_idx];
2198 p->total_numa_faults += diff;
2199 if (p->numa_group) {
2201 * safe because we can only change our own group
2203 * mem_idx represents the offset for a given
2204 * nid and priv in a specific region because it
2205 * is at the beginning of the numa_faults array.
2207 p->numa_group->faults[mem_idx] += diff;
2208 p->numa_group->faults_cpu[mem_idx] += f_diff;
2209 p->numa_group->total_faults += diff;
2210 group_faults += p->numa_group->faults[mem_idx];
2214 if (faults > max_faults) {
2215 max_faults = faults;
2219 if (group_faults > max_group_faults) {
2220 max_group_faults = group_faults;
2221 max_group_nid = nid;
2225 update_task_scan_period(p, fault_types[0], fault_types[1]);
2227 if (p->numa_group) {
2228 numa_group_count_active_nodes(p->numa_group);
2229 spin_unlock_irq(group_lock);
2230 max_nid = preferred_group_nid(p, max_group_nid);
2234 /* Set the new preferred node */
2235 if (max_nid != p->numa_preferred_nid)
2236 sched_setnuma(p, max_nid);
2238 if (task_node(p) != p->numa_preferred_nid)
2239 numa_migrate_preferred(p);
2243 static inline int get_numa_group(struct numa_group *grp)
2245 return atomic_inc_not_zero(&grp->refcount);
2248 static inline void put_numa_group(struct numa_group *grp)
2250 if (atomic_dec_and_test(&grp->refcount))
2251 kfree_rcu(grp, rcu);
2254 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2257 struct numa_group *grp, *my_grp;
2258 struct task_struct *tsk;
2260 int cpu = cpupid_to_cpu(cpupid);
2263 if (unlikely(!p->numa_group)) {
2264 unsigned int size = sizeof(struct numa_group) +
2265 4*nr_node_ids*sizeof(unsigned long);
2267 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2271 atomic_set(&grp->refcount, 1);
2272 grp->active_nodes = 1;
2273 grp->max_faults_cpu = 0;
2274 spin_lock_init(&grp->lock);
2276 /* Second half of the array tracks nids where faults happen */
2277 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2280 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2281 grp->faults[i] = p->numa_faults[i];
2283 grp->total_faults = p->total_numa_faults;
2286 rcu_assign_pointer(p->numa_group, grp);
2290 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2292 if (!cpupid_match_pid(tsk, cpupid))
2295 grp = rcu_dereference(tsk->numa_group);
2299 my_grp = p->numa_group;
2304 * Only join the other group if its bigger; if we're the bigger group,
2305 * the other task will join us.
2307 if (my_grp->nr_tasks > grp->nr_tasks)
2311 * Tie-break on the grp address.
2313 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2316 /* Always join threads in the same process. */
2317 if (tsk->mm == current->mm)
2320 /* Simple filter to avoid false positives due to PID collisions */
2321 if (flags & TNF_SHARED)
2324 /* Update priv based on whether false sharing was detected */
2327 if (join && !get_numa_group(grp))
2335 BUG_ON(irqs_disabled());
2336 double_lock_irq(&my_grp->lock, &grp->lock);
2338 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2339 my_grp->faults[i] -= p->numa_faults[i];
2340 grp->faults[i] += p->numa_faults[i];
2342 my_grp->total_faults -= p->total_numa_faults;
2343 grp->total_faults += p->total_numa_faults;
2348 spin_unlock(&my_grp->lock);
2349 spin_unlock_irq(&grp->lock);
2351 rcu_assign_pointer(p->numa_group, grp);
2353 put_numa_group(my_grp);
2362 * Get rid of NUMA staticstics associated with a task (either current or dead).
2363 * If @final is set, the task is dead and has reached refcount zero, so we can
2364 * safely free all relevant data structures. Otherwise, there might be
2365 * concurrent reads from places like load balancing and procfs, and we should
2366 * reset the data back to default state without freeing ->numa_faults.
2368 void task_numa_free(struct task_struct *p, bool final)
2370 struct numa_group *grp = p->numa_group;
2371 unsigned long *numa_faults = p->numa_faults;
2372 unsigned long flags;
2379 spin_lock_irqsave(&grp->lock, flags);
2380 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2381 grp->faults[i] -= p->numa_faults[i];
2382 grp->total_faults -= p->total_numa_faults;
2385 spin_unlock_irqrestore(&grp->lock, flags);
2386 RCU_INIT_POINTER(p->numa_group, NULL);
2387 put_numa_group(grp);
2391 p->numa_faults = NULL;
2394 p->total_numa_faults = 0;
2395 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2401 * Got a PROT_NONE fault for a page on @node.
2403 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2405 struct task_struct *p = current;
2406 bool migrated = flags & TNF_MIGRATED;
2407 int cpu_node = task_node(current);
2408 int local = !!(flags & TNF_FAULT_LOCAL);
2409 struct numa_group *ng;
2412 if (!static_branch_likely(&sched_numa_balancing))
2415 /* for example, ksmd faulting in a user's mm */
2419 /* Allocate buffer to track faults on a per-node basis */
2420 if (unlikely(!p->numa_faults)) {
2421 int size = sizeof(*p->numa_faults) *
2422 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2424 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2425 if (!p->numa_faults)
2428 p->total_numa_faults = 0;
2429 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2433 * First accesses are treated as private, otherwise consider accesses
2434 * to be private if the accessing pid has not changed
2436 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2439 priv = cpupid_match_pid(p, last_cpupid);
2440 if (!priv && !(flags & TNF_NO_GROUP))
2441 task_numa_group(p, last_cpupid, flags, &priv);
2445 * If a workload spans multiple NUMA nodes, a shared fault that
2446 * occurs wholly within the set of nodes that the workload is
2447 * actively using should be counted as local. This allows the
2448 * scan rate to slow down when a workload has settled down.
2451 if (!priv && !local && ng && ng->active_nodes > 1 &&
2452 numa_is_active_node(cpu_node, ng) &&
2453 numa_is_active_node(mem_node, ng))
2456 task_numa_placement(p);
2459 * Retry task to preferred node migration periodically, in case it
2460 * case it previously failed, or the scheduler moved us.
2462 if (time_after(jiffies, p->numa_migrate_retry))
2463 numa_migrate_preferred(p);
2466 p->numa_pages_migrated += pages;
2467 if (flags & TNF_MIGRATE_FAIL)
2468 p->numa_faults_locality[2] += pages;
2470 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2471 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2472 p->numa_faults_locality[local] += pages;
2475 static void reset_ptenuma_scan(struct task_struct *p)
2478 * We only did a read acquisition of the mmap sem, so
2479 * p->mm->numa_scan_seq is written to without exclusive access
2480 * and the update is not guaranteed to be atomic. That's not
2481 * much of an issue though, since this is just used for
2482 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2483 * expensive, to avoid any form of compiler optimizations:
2485 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2486 p->mm->numa_scan_offset = 0;
2490 * The expensive part of numa migration is done from task_work context.
2491 * Triggered from task_tick_numa().
2493 void task_numa_work(struct callback_head *work)
2495 unsigned long migrate, next_scan, now = jiffies;
2496 struct task_struct *p = current;
2497 struct mm_struct *mm = p->mm;
2498 u64 runtime = p->se.sum_exec_runtime;
2499 struct vm_area_struct *vma;
2500 unsigned long start, end;
2501 unsigned long nr_pte_updates = 0;
2502 long pages, virtpages;
2504 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2506 work->next = work; /* protect against double add */
2508 * Who cares about NUMA placement when they're dying.
2510 * NOTE: make sure not to dereference p->mm before this check,
2511 * exit_task_work() happens _after_ exit_mm() so we could be called
2512 * without p->mm even though we still had it when we enqueued this
2515 if (p->flags & PF_EXITING)
2518 if (!mm->numa_next_scan) {
2519 mm->numa_next_scan = now +
2520 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2524 * Enforce maximal scan/migration frequency..
2526 migrate = mm->numa_next_scan;
2527 if (time_before(now, migrate))
2530 if (p->numa_scan_period == 0) {
2531 p->numa_scan_period_max = task_scan_max(p);
2532 p->numa_scan_period = task_scan_start(p);
2535 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2536 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2540 * Delay this task enough that another task of this mm will likely win
2541 * the next time around.
2543 p->node_stamp += 2 * TICK_NSEC;
2545 start = mm->numa_scan_offset;
2546 pages = sysctl_numa_balancing_scan_size;
2547 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2548 virtpages = pages * 8; /* Scan up to this much virtual space */
2553 if (!down_read_trylock(&mm->mmap_sem))
2555 vma = find_vma(mm, start);
2557 reset_ptenuma_scan(p);
2561 for (; vma; vma = vma->vm_next) {
2562 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2563 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2568 * Shared library pages mapped by multiple processes are not
2569 * migrated as it is expected they are cache replicated. Avoid
2570 * hinting faults in read-only file-backed mappings or the vdso
2571 * as migrating the pages will be of marginal benefit.
2574 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2578 * Skip inaccessible VMAs to avoid any confusion between
2579 * PROT_NONE and NUMA hinting ptes
2581 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2585 start = max(start, vma->vm_start);
2586 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2587 end = min(end, vma->vm_end);
2588 nr_pte_updates = change_prot_numa(vma, start, end);
2591 * Try to scan sysctl_numa_balancing_size worth of
2592 * hpages that have at least one present PTE that
2593 * is not already pte-numa. If the VMA contains
2594 * areas that are unused or already full of prot_numa
2595 * PTEs, scan up to virtpages, to skip through those
2599 pages -= (end - start) >> PAGE_SHIFT;
2600 virtpages -= (end - start) >> PAGE_SHIFT;
2603 if (pages <= 0 || virtpages <= 0)
2607 } while (end != vma->vm_end);
2612 * It is possible to reach the end of the VMA list but the last few
2613 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2614 * would find the !migratable VMA on the next scan but not reset the
2615 * scanner to the start so check it now.
2618 mm->numa_scan_offset = start;
2620 reset_ptenuma_scan(p);
2621 up_read(&mm->mmap_sem);
2624 * Make sure tasks use at least 32x as much time to run other code
2625 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2626 * Usually update_task_scan_period slows down scanning enough; on an
2627 * overloaded system we need to limit overhead on a per task basis.
2629 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2630 u64 diff = p->se.sum_exec_runtime - runtime;
2631 p->node_stamp += 32 * diff;
2636 * Drive the periodic memory faults..
2638 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2640 struct callback_head *work = &curr->numa_work;
2644 * We don't care about NUMA placement if we don't have memory.
2646 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2650 * Using runtime rather than walltime has the dual advantage that
2651 * we (mostly) drive the selection from busy threads and that the
2652 * task needs to have done some actual work before we bother with
2655 now = curr->se.sum_exec_runtime;
2656 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2658 if (now > curr->node_stamp + period) {
2659 if (!curr->node_stamp)
2660 curr->numa_scan_period = task_scan_start(curr);
2661 curr->node_stamp += period;
2663 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2664 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2665 task_work_add(curr, work, true);
2671 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2675 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2679 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2683 #endif /* CONFIG_NUMA_BALANCING */
2686 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2688 update_load_add(&cfs_rq->load, se->load.weight);
2689 if (!parent_entity(se))
2690 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2692 if (entity_is_task(se)) {
2693 struct rq *rq = rq_of(cfs_rq);
2695 account_numa_enqueue(rq, task_of(se));
2696 list_add(&se->group_node, &rq->cfs_tasks);
2699 cfs_rq->nr_running++;
2703 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2705 update_load_sub(&cfs_rq->load, se->load.weight);
2706 if (!parent_entity(se))
2707 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2709 if (entity_is_task(se)) {
2710 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2711 list_del_init(&se->group_node);
2714 cfs_rq->nr_running--;
2717 #ifdef CONFIG_FAIR_GROUP_SCHED
2719 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2721 long tg_weight, load, shares;
2724 * This really should be: cfs_rq->avg.load_avg, but instead we use
2725 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2726 * the shares for small weight interactive tasks.
2728 load = scale_load_down(cfs_rq->load.weight);
2730 tg_weight = atomic_long_read(&tg->load_avg);
2732 /* Ensure tg_weight >= load */
2733 tg_weight -= cfs_rq->tg_load_avg_contrib;
2736 shares = (tg->shares * load);
2738 shares /= tg_weight;
2741 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2742 * of a group with small tg->shares value. It is a floor value which is
2743 * assigned as a minimum load.weight to the sched_entity representing
2744 * the group on a CPU.
2746 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2747 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2748 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2749 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2752 if (shares < MIN_SHARES)
2753 shares = MIN_SHARES;
2754 if (shares > tg->shares)
2755 shares = tg->shares;
2759 # else /* CONFIG_SMP */
2760 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2764 # endif /* CONFIG_SMP */
2766 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2767 unsigned long weight)
2770 /* commit outstanding execution time */
2771 if (cfs_rq->curr == se)
2772 update_curr(cfs_rq);
2773 account_entity_dequeue(cfs_rq, se);
2776 update_load_set(&se->load, weight);
2779 account_entity_enqueue(cfs_rq, se);
2782 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2784 static void update_cfs_shares(struct sched_entity *se)
2786 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2787 struct task_group *tg;
2793 if (throttled_hierarchy(cfs_rq))
2799 if (likely(se->load.weight == tg->shares))
2802 shares = calc_cfs_shares(cfs_rq, tg);
2804 reweight_entity(cfs_rq_of(se), se, shares);
2807 #else /* CONFIG_FAIR_GROUP_SCHED */
2808 static inline void update_cfs_shares(struct sched_entity *se)
2811 #endif /* CONFIG_FAIR_GROUP_SCHED */
2813 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2815 struct rq *rq = rq_of(cfs_rq);
2817 if (&rq->cfs == cfs_rq) {
2819 * There are a few boundary cases this might miss but it should
2820 * get called often enough that that should (hopefully) not be
2821 * a real problem -- added to that it only calls on the local
2822 * CPU, so if we enqueue remotely we'll miss an update, but
2823 * the next tick/schedule should update.
2825 * It will not get called when we go idle, because the idle
2826 * thread is a different class (!fair), nor will the utilization
2827 * number include things like RT tasks.
2829 * As is, the util number is not freq-invariant (we'd have to
2830 * implement arch_scale_freq_capacity() for that).
2834 cpufreq_update_util(rq, 0);
2841 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2843 static u64 decay_load(u64 val, u64 n)
2845 unsigned int local_n;
2847 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2850 /* after bounds checking we can collapse to 32-bit */
2854 * As y^PERIOD = 1/2, we can combine
2855 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2856 * With a look-up table which covers y^n (n<PERIOD)
2858 * To achieve constant time decay_load.
2860 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2861 val >>= local_n / LOAD_AVG_PERIOD;
2862 local_n %= LOAD_AVG_PERIOD;
2865 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2869 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2871 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2876 c1 = decay_load((u64)d1, periods);
2880 * c2 = 1024 \Sum y^n
2884 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2887 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2889 return c1 + c2 + c3;
2892 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2895 * Accumulate the three separate parts of the sum; d1 the remainder
2896 * of the last (incomplete) period, d2 the span of full periods and d3
2897 * the remainder of the (incomplete) current period.
2902 * |<->|<----------------->|<--->|
2903 * ... |---x---|------| ... |------|-----x (now)
2906 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2909 * = u y^p + (Step 1)
2912 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2915 static __always_inline u32
2916 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2917 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2919 unsigned long scale_freq, scale_cpu;
2920 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2923 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2924 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2926 delta += sa->period_contrib;
2927 periods = delta / 1024; /* A period is 1024us (~1ms) */
2930 * Step 1: decay old *_sum if we crossed period boundaries.
2933 sa->load_sum = decay_load(sa->load_sum, periods);
2935 cfs_rq->runnable_load_sum =
2936 decay_load(cfs_rq->runnable_load_sum, periods);
2938 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2944 contrib = __accumulate_pelt_segments(periods,
2945 1024 - sa->period_contrib, delta);
2947 sa->period_contrib = delta;
2949 contrib = cap_scale(contrib, scale_freq);
2951 sa->load_sum += weight * contrib;
2953 cfs_rq->runnable_load_sum += weight * contrib;
2956 sa->util_sum += contrib * scale_cpu;
2962 * We can represent the historical contribution to runnable average as the
2963 * coefficients of a geometric series. To do this we sub-divide our runnable
2964 * history into segments of approximately 1ms (1024us); label the segment that
2965 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2967 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2969 * (now) (~1ms ago) (~2ms ago)
2971 * Let u_i denote the fraction of p_i that the entity was runnable.
2973 * We then designate the fractions u_i as our co-efficients, yielding the
2974 * following representation of historical load:
2975 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2977 * We choose y based on the with of a reasonably scheduling period, fixing:
2980 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2981 * approximately half as much as the contribution to load within the last ms
2984 * When a period "rolls over" and we have new u_0`, multiplying the previous
2985 * sum again by y is sufficient to update:
2986 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2987 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2989 static __always_inline int
2990 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2991 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2995 delta = now - sa->last_update_time;
2997 * This should only happen when time goes backwards, which it
2998 * unfortunately does during sched clock init when we swap over to TSC.
3000 if ((s64)delta < 0) {
3001 sa->last_update_time = now;
3006 * Use 1024ns as the unit of measurement since it's a reasonable
3007 * approximation of 1us and fast to compute.
3013 sa->last_update_time += delta << 10;
3016 * running is a subset of runnable (weight) so running can't be set if
3017 * runnable is clear. But there are some corner cases where the current
3018 * se has been already dequeued but cfs_rq->curr still points to it.
3019 * This means that weight will be 0 but not running for a sched_entity
3020 * but also for a cfs_rq if the latter becomes idle. As an example,
3021 * this happens during idle_balance() which calls
3022 * update_blocked_averages()
3028 * Now we know we crossed measurement unit boundaries. The *_avg
3029 * accrues by two steps:
3031 * Step 1: accumulate *_sum since last_update_time. If we haven't
3032 * crossed period boundaries, finish.
3034 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
3038 * Step 2: update *_avg.
3040 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3042 cfs_rq->runnable_load_avg =
3043 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3045 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
3051 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3053 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
3057 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3059 return ___update_load_avg(now, cpu, &se->avg,
3060 se->on_rq * scale_load_down(se->load.weight),
3061 cfs_rq->curr == se, NULL);
3065 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3067 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3068 scale_load_down(cfs_rq->load.weight),
3069 cfs_rq->curr != NULL, cfs_rq);
3073 * Signed add and clamp on underflow.
3075 * Explicitly do a load-store to ensure the intermediate value never hits
3076 * memory. This allows lockless observations without ever seeing the negative
3079 #define add_positive(_ptr, _val) do { \
3080 typeof(_ptr) ptr = (_ptr); \
3081 typeof(_val) val = (_val); \
3082 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3086 if (val < 0 && res > var) \
3089 WRITE_ONCE(*ptr, res); \
3092 #ifdef CONFIG_FAIR_GROUP_SCHED
3094 * update_tg_load_avg - update the tg's load avg
3095 * @cfs_rq: the cfs_rq whose avg changed
3096 * @force: update regardless of how small the difference
3098 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3099 * However, because tg->load_avg is a global value there are performance
3102 * In order to avoid having to look at the other cfs_rq's, we use a
3103 * differential update where we store the last value we propagated. This in
3104 * turn allows skipping updates if the differential is 'small'.
3106 * Updating tg's load_avg is necessary before update_cfs_share().
3108 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3110 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3113 * No need to update load_avg for root_task_group as it is not used.
3115 if (cfs_rq->tg == &root_task_group)
3118 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3119 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3120 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3125 * Called within set_task_rq() right before setting a task's cpu. The
3126 * caller only guarantees p->pi_lock is held; no other assumptions,
3127 * including the state of rq->lock, should be made.
3129 void set_task_rq_fair(struct sched_entity *se,
3130 struct cfs_rq *prev, struct cfs_rq *next)
3132 u64 p_last_update_time;
3133 u64 n_last_update_time;
3135 if (!sched_feat(ATTACH_AGE_LOAD))
3139 * We are supposed to update the task to "current" time, then its up to
3140 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3141 * getting what current time is, so simply throw away the out-of-date
3142 * time. This will result in the wakee task is less decayed, but giving
3143 * the wakee more load sounds not bad.
3145 if (!(se->avg.last_update_time && prev))
3148 #ifndef CONFIG_64BIT
3150 u64 p_last_update_time_copy;
3151 u64 n_last_update_time_copy;
3154 p_last_update_time_copy = prev->load_last_update_time_copy;
3155 n_last_update_time_copy = next->load_last_update_time_copy;
3159 p_last_update_time = prev->avg.last_update_time;
3160 n_last_update_time = next->avg.last_update_time;
3162 } while (p_last_update_time != p_last_update_time_copy ||
3163 n_last_update_time != n_last_update_time_copy);
3166 p_last_update_time = prev->avg.last_update_time;
3167 n_last_update_time = next->avg.last_update_time;
3169 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3170 se->avg.last_update_time = n_last_update_time;
3173 /* Take into account change of utilization of a child task group */
3175 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3177 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3178 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3180 /* Nothing to update */
3184 /* Set new sched_entity's utilization */
3185 se->avg.util_avg = gcfs_rq->avg.util_avg;
3186 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3188 /* Update parent cfs_rq utilization */
3189 add_positive(&cfs_rq->avg.util_avg, delta);
3190 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3193 /* Take into account change of load of a child task group */
3195 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3197 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3198 long delta, load = gcfs_rq->avg.load_avg;
3201 * If the load of group cfs_rq is null, the load of the
3202 * sched_entity will also be null so we can skip the formula
3207 /* Get tg's load and ensure tg_load > 0 */
3208 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3210 /* Ensure tg_load >= load and updated with current load*/
3211 tg_load -= gcfs_rq->tg_load_avg_contrib;
3215 * We need to compute a correction term in the case that the
3216 * task group is consuming more CPU than a task of equal
3217 * weight. A task with a weight equals to tg->shares will have
3218 * a load less or equal to scale_load_down(tg->shares).
3219 * Similarly, the sched_entities that represent the task group
3220 * at parent level, can't have a load higher than
3221 * scale_load_down(tg->shares). And the Sum of sched_entities'
3222 * load must be <= scale_load_down(tg->shares).
3224 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3225 /* scale gcfs_rq's load into tg's shares*/
3226 load *= scale_load_down(gcfs_rq->tg->shares);
3231 delta = load - se->avg.load_avg;
3233 /* Nothing to update */
3237 /* Set new sched_entity's load */
3238 se->avg.load_avg = load;
3239 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3241 /* Update parent cfs_rq load */
3242 add_positive(&cfs_rq->avg.load_avg, delta);
3243 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3246 * If the sched_entity is already enqueued, we also have to update the
3247 * runnable load avg.
3250 /* Update parent cfs_rq runnable_load_avg */
3251 add_positive(&cfs_rq->runnable_load_avg, delta);
3252 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3256 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3258 cfs_rq->propagate_avg = 1;
3261 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3263 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3265 if (!cfs_rq->propagate_avg)
3268 cfs_rq->propagate_avg = 0;
3272 /* Update task and its cfs_rq load average */
3273 static inline int propagate_entity_load_avg(struct sched_entity *se)
3275 struct cfs_rq *cfs_rq;
3277 if (entity_is_task(se))
3280 if (!test_and_clear_tg_cfs_propagate(se))
3283 cfs_rq = cfs_rq_of(se);
3285 set_tg_cfs_propagate(cfs_rq);
3287 update_tg_cfs_util(cfs_rq, se);
3288 update_tg_cfs_load(cfs_rq, se);
3294 * Check if we need to update the load and the utilization of a blocked
3297 static inline bool skip_blocked_update(struct sched_entity *se)
3299 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3302 * If sched_entity still have not zero load or utilization, we have to
3305 if (se->avg.load_avg || se->avg.util_avg)
3309 * If there is a pending propagation, we have to update the load and
3310 * the utilization of the sched_entity:
3312 if (gcfs_rq->propagate_avg)
3316 * Otherwise, the load and the utilization of the sched_entity is
3317 * already zero and there is no pending propagation, so it will be a
3318 * waste of time to try to decay it:
3323 #else /* CONFIG_FAIR_GROUP_SCHED */
3325 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3327 static inline int propagate_entity_load_avg(struct sched_entity *se)
3332 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3334 #endif /* CONFIG_FAIR_GROUP_SCHED */
3337 * Unsigned subtract and clamp on underflow.
3339 * Explicitly do a load-store to ensure the intermediate value never hits
3340 * memory. This allows lockless observations without ever seeing the negative
3343 #define sub_positive(_ptr, _val) do { \
3344 typeof(_ptr) ptr = (_ptr); \
3345 typeof(*ptr) val = (_val); \
3346 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3350 WRITE_ONCE(*ptr, res); \
3354 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3355 * @now: current time, as per cfs_rq_clock_task()
3356 * @cfs_rq: cfs_rq to update
3358 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3359 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3360 * post_init_entity_util_avg().
3362 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3364 * Returns true if the load decayed or we removed load.
3366 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3367 * call update_tg_load_avg() when this function returns true.
3370 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3372 struct sched_avg *sa = &cfs_rq->avg;
3373 int decayed, removed_load = 0, removed_util = 0;
3375 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3376 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3377 sub_positive(&sa->load_avg, r);
3378 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3380 set_tg_cfs_propagate(cfs_rq);
3383 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3384 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3385 sub_positive(&sa->util_avg, r);
3386 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3388 set_tg_cfs_propagate(cfs_rq);
3391 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3393 #ifndef CONFIG_64BIT
3395 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3398 if (decayed || removed_util)
3399 cfs_rq_util_change(cfs_rq);
3401 return decayed || removed_load;
3405 * Optional action to be done while updating the load average
3407 #define UPDATE_TG 0x1
3408 #define SKIP_AGE_LOAD 0x2
3410 /* Update task and its cfs_rq load average */
3411 static inline void update_load_avg(struct sched_entity *se, int flags)
3413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3414 u64 now = cfs_rq_clock_task(cfs_rq);
3415 struct rq *rq = rq_of(cfs_rq);
3416 int cpu = cpu_of(rq);
3420 * Track task load average for carrying it to new CPU after migrated, and
3421 * track group sched_entity load average for task_h_load calc in migration
3423 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3424 __update_load_avg_se(now, cpu, cfs_rq, se);
3426 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3427 decayed |= propagate_entity_load_avg(se);
3429 if (decayed && (flags & UPDATE_TG))
3430 update_tg_load_avg(cfs_rq, 0);
3434 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3435 * @cfs_rq: cfs_rq to attach to
3436 * @se: sched_entity to attach
3438 * Must call update_cfs_rq_load_avg() before this, since we rely on
3439 * cfs_rq->avg.last_update_time being current.
3441 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3443 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3444 cfs_rq->avg.load_avg += se->avg.load_avg;
3445 cfs_rq->avg.load_sum += se->avg.load_sum;
3446 cfs_rq->avg.util_avg += se->avg.util_avg;
3447 cfs_rq->avg.util_sum += se->avg.util_sum;
3448 set_tg_cfs_propagate(cfs_rq);
3450 cfs_rq_util_change(cfs_rq);
3454 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3455 * @cfs_rq: cfs_rq to detach from
3456 * @se: sched_entity to detach
3458 * Must call update_cfs_rq_load_avg() before this, since we rely on
3459 * cfs_rq->avg.last_update_time being current.
3461 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3464 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3465 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3466 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3467 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3468 set_tg_cfs_propagate(cfs_rq);
3470 cfs_rq_util_change(cfs_rq);
3473 /* Add the load generated by se into cfs_rq's load average */
3475 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3477 struct sched_avg *sa = &se->avg;
3479 cfs_rq->runnable_load_avg += sa->load_avg;
3480 cfs_rq->runnable_load_sum += sa->load_sum;
3482 if (!sa->last_update_time) {
3483 attach_entity_load_avg(cfs_rq, se);
3484 update_tg_load_avg(cfs_rq, 0);
3488 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3490 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3492 cfs_rq->runnable_load_avg =
3493 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3494 cfs_rq->runnable_load_sum =
3495 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3498 #ifndef CONFIG_64BIT
3499 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3501 u64 last_update_time_copy;
3502 u64 last_update_time;
3505 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3507 last_update_time = cfs_rq->avg.last_update_time;
3508 } while (last_update_time != last_update_time_copy);
3510 return last_update_time;
3513 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3515 return cfs_rq->avg.last_update_time;
3520 * Synchronize entity load avg of dequeued entity without locking
3523 void sync_entity_load_avg(struct sched_entity *se)
3525 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3526 u64 last_update_time;
3528 last_update_time = cfs_rq_last_update_time(cfs_rq);
3529 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3533 * Task first catches up with cfs_rq, and then subtract
3534 * itself from the cfs_rq (task must be off the queue now).
3536 void remove_entity_load_avg(struct sched_entity *se)
3538 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3541 * tasks cannot exit without having gone through wake_up_new_task() ->
3542 * post_init_entity_util_avg() which will have added things to the
3543 * cfs_rq, so we can remove unconditionally.
3545 * Similarly for groups, they will have passed through
3546 * post_init_entity_util_avg() before unregister_sched_fair_group()
3550 sync_entity_load_avg(se);
3551 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3552 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3555 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3557 return cfs_rq->runnable_load_avg;
3560 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3562 return cfs_rq->avg.load_avg;
3565 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3567 #else /* CONFIG_SMP */
3570 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3575 #define UPDATE_TG 0x0
3576 #define SKIP_AGE_LOAD 0x0
3578 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3580 cfs_rq_util_change(cfs_rq_of(se));
3584 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3586 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3587 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3590 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3592 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3594 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3599 #endif /* CONFIG_SMP */
3601 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3603 #ifdef CONFIG_SCHED_DEBUG
3604 s64 d = se->vruntime - cfs_rq->min_vruntime;
3609 if (d > 3*sysctl_sched_latency)
3610 schedstat_inc(cfs_rq->nr_spread_over);
3614 static inline bool entity_is_long_sleeper(struct sched_entity *se)
3616 struct cfs_rq *cfs_rq;
3619 if (se->exec_start == 0)
3622 cfs_rq = cfs_rq_of(se);
3624 sleep_time = rq_clock_task(rq_of(cfs_rq));
3626 /* Happen while migrating because of clock task divergence */
3627 if (sleep_time <= se->exec_start)
3630 sleep_time -= se->exec_start;
3631 if (sleep_time > ((1ULL << 63) / scale_load_down(NICE_0_LOAD)))
3638 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3640 u64 vruntime = cfs_rq->min_vruntime;
3643 * The 'current' period is already promised to the current tasks,
3644 * however the extra weight of the new task will slow them down a
3645 * little, place the new task so that it fits in the slot that
3646 * stays open at the end.
3648 if (initial && sched_feat(START_DEBIT))
3649 vruntime += sched_vslice(cfs_rq, se);
3651 /* sleeps up to a single latency don't count. */
3653 unsigned long thresh = sysctl_sched_latency;
3656 * Halve their sleep time's effect, to allow
3657 * for a gentler effect of sleepers:
3659 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3666 * Pull vruntime of the entity being placed to the base level of
3667 * cfs_rq, to prevent boosting it if placed backwards.
3668 * However, min_vruntime can advance much faster than real time, with
3669 * the extreme being when an entity with the minimal weight always runs
3670 * on the cfs_rq. If the waking entity slept for a long time, its
3671 * vruntime difference from min_vruntime may overflow s64 and their
3672 * comparison may get inversed, so ignore the entity's original
3673 * vruntime in that case.
3674 * The maximal vruntime speedup is given by the ratio of normal to
3675 * minimal weight: scale_load_down(NICE_0_LOAD) / MIN_SHARES.
3676 * When placing a migrated waking entity, its exec_start has been set
3677 * from a different rq. In order to take into account a possible
3678 * divergence between new and prev rq's clocks task because of irq and
3679 * stolen time, we take an additional margin.
3680 * So, cutting off on the sleep time of
3681 * 2^63 / scale_load_down(NICE_0_LOAD) ~ 104 days
3684 if (entity_is_long_sleeper(se))
3685 se->vruntime = vruntime;
3687 se->vruntime = max_vruntime(se->vruntime, vruntime);
3690 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3692 static inline void check_schedstat_required(void)
3694 #ifdef CONFIG_SCHEDSTATS
3695 if (schedstat_enabled())
3698 /* Force schedstat enabled if a dependent tracepoint is active */
3699 if (trace_sched_stat_wait_enabled() ||
3700 trace_sched_stat_sleep_enabled() ||
3701 trace_sched_stat_iowait_enabled() ||
3702 trace_sched_stat_blocked_enabled() ||
3703 trace_sched_stat_runtime_enabled()) {
3704 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3705 "stat_blocked and stat_runtime require the "
3706 "kernel parameter schedstats=enable or "
3707 "kernel.sched_schedstats=1\n");
3718 * update_min_vruntime()
3719 * vruntime -= min_vruntime
3723 * update_min_vruntime()
3724 * vruntime += min_vruntime
3726 * this way the vruntime transition between RQs is done when both
3727 * min_vruntime are up-to-date.
3731 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3732 * vruntime -= min_vruntime
3736 * update_min_vruntime()
3737 * vruntime += min_vruntime
3739 * this way we don't have the most up-to-date min_vruntime on the originating
3740 * CPU and an up-to-date min_vruntime on the destination CPU.
3744 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3746 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3747 bool curr = cfs_rq->curr == se;
3750 * If we're the current task, we must renormalise before calling
3754 se->vruntime += cfs_rq->min_vruntime;
3756 update_curr(cfs_rq);
3759 * Otherwise, renormalise after, such that we're placed at the current
3760 * moment in time, instead of some random moment in the past. Being
3761 * placed in the past could significantly boost this task to the
3762 * fairness detriment of existing tasks.
3764 if (renorm && !curr)
3765 se->vruntime += cfs_rq->min_vruntime;
3768 * When enqueuing a sched_entity, we must:
3769 * - Update loads to have both entity and cfs_rq synced with now.
3770 * - Add its load to cfs_rq->runnable_avg
3771 * - For group_entity, update its weight to reflect the new share of
3773 * - Add its new weight to cfs_rq->load.weight
3775 update_load_avg(se, UPDATE_TG);
3776 enqueue_entity_load_avg(cfs_rq, se);
3777 update_cfs_shares(se);
3778 account_entity_enqueue(cfs_rq, se);
3780 if (flags & ENQUEUE_WAKEUP)
3781 place_entity(cfs_rq, se, 0);
3782 /* Entity has migrated, no longer consider this task hot */
3783 if (flags & ENQUEUE_MIGRATED)
3786 check_schedstat_required();
3787 update_stats_enqueue(cfs_rq, se, flags);
3788 check_spread(cfs_rq, se);
3790 __enqueue_entity(cfs_rq, se);
3793 if (cfs_rq->nr_running == 1) {
3794 list_add_leaf_cfs_rq(cfs_rq);
3795 check_enqueue_throttle(cfs_rq);
3799 static void __clear_buddies_last(struct sched_entity *se)
3801 for_each_sched_entity(se) {
3802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3803 if (cfs_rq->last != se)
3806 cfs_rq->last = NULL;
3810 static void __clear_buddies_next(struct sched_entity *se)
3812 for_each_sched_entity(se) {
3813 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3814 if (cfs_rq->next != se)
3817 cfs_rq->next = NULL;
3821 static void __clear_buddies_skip(struct sched_entity *se)
3823 for_each_sched_entity(se) {
3824 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3825 if (cfs_rq->skip != se)
3828 cfs_rq->skip = NULL;
3832 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3834 if (cfs_rq->last == se)
3835 __clear_buddies_last(se);
3837 if (cfs_rq->next == se)
3838 __clear_buddies_next(se);
3840 if (cfs_rq->skip == se)
3841 __clear_buddies_skip(se);
3844 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3847 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3850 * Update run-time statistics of the 'current'.
3852 update_curr(cfs_rq);
3855 * When dequeuing a sched_entity, we must:
3856 * - Update loads to have both entity and cfs_rq synced with now.
3857 * - Substract its load from the cfs_rq->runnable_avg.
3858 * - Substract its previous weight from cfs_rq->load.weight.
3859 * - For group entity, update its weight to reflect the new share
3860 * of its group cfs_rq.
3862 update_load_avg(se, UPDATE_TG);
3863 dequeue_entity_load_avg(cfs_rq, se);
3865 update_stats_dequeue(cfs_rq, se, flags);
3867 clear_buddies(cfs_rq, se);
3869 if (se != cfs_rq->curr)
3870 __dequeue_entity(cfs_rq, se);
3872 account_entity_dequeue(cfs_rq, se);
3875 * Normalize after update_curr(); which will also have moved
3876 * min_vruntime if @se is the one holding it back. But before doing
3877 * update_min_vruntime() again, which will discount @se's position and
3878 * can move min_vruntime forward still more.
3880 if (!(flags & DEQUEUE_SLEEP))
3881 se->vruntime -= cfs_rq->min_vruntime;
3883 /* return excess runtime on last dequeue */
3884 return_cfs_rq_runtime(cfs_rq);
3886 update_cfs_shares(se);
3889 * Now advance min_vruntime if @se was the entity holding it back,
3890 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3891 * put back on, and if we advance min_vruntime, we'll be placed back
3892 * further than we started -- ie. we'll be penalized.
3894 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
3895 update_min_vruntime(cfs_rq);
3899 * Preempt the current task with a newly woken task if needed:
3902 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3904 unsigned long ideal_runtime, delta_exec;
3905 struct sched_entity *se;
3908 ideal_runtime = sched_slice(cfs_rq, curr);
3909 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3910 if (delta_exec > ideal_runtime) {
3911 resched_curr(rq_of(cfs_rq));
3913 * The current task ran long enough, ensure it doesn't get
3914 * re-elected due to buddy favours.
3916 clear_buddies(cfs_rq, curr);
3921 * Ensure that a task that missed wakeup preemption by a
3922 * narrow margin doesn't have to wait for a full slice.
3923 * This also mitigates buddy induced latencies under load.
3925 if (delta_exec < sysctl_sched_min_granularity)
3928 se = __pick_first_entity(cfs_rq);
3929 delta = curr->vruntime - se->vruntime;
3934 if (delta > ideal_runtime)
3935 resched_curr(rq_of(cfs_rq));
3939 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3941 /* 'current' is not kept within the tree. */
3944 * Any task has to be enqueued before it get to execute on
3945 * a CPU. So account for the time it spent waiting on the
3948 update_stats_wait_end(cfs_rq, se);
3949 __dequeue_entity(cfs_rq, se);
3950 update_load_avg(se, UPDATE_TG);
3953 update_stats_curr_start(cfs_rq, se);
3957 * Track our maximum slice length, if the CPU's load is at
3958 * least twice that of our own weight (i.e. dont track it
3959 * when there are only lesser-weight tasks around):
3961 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3962 schedstat_set(se->statistics.slice_max,
3963 max((u64)schedstat_val(se->statistics.slice_max),
3964 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3967 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3971 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3974 * Pick the next process, keeping these things in mind, in this order:
3975 * 1) keep things fair between processes/task groups
3976 * 2) pick the "next" process, since someone really wants that to run
3977 * 3) pick the "last" process, for cache locality
3978 * 4) do not run the "skip" process, if something else is available
3980 static struct sched_entity *
3981 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3983 struct sched_entity *left = __pick_first_entity(cfs_rq);
3984 struct sched_entity *se;
3987 * If curr is set we have to see if its left of the leftmost entity
3988 * still in the tree, provided there was anything in the tree at all.
3990 if (!left || (curr && entity_before(curr, left)))
3993 se = left; /* ideally we run the leftmost entity */
3996 * Avoid running the skip buddy, if running something else can
3997 * be done without getting too unfair.
3999 if (cfs_rq->skip == se) {
4000 struct sched_entity *second;
4003 second = __pick_first_entity(cfs_rq);
4005 second = __pick_next_entity(se);
4006 if (!second || (curr && entity_before(curr, second)))
4010 if (second && wakeup_preempt_entity(second, left) < 1)
4015 * Prefer last buddy, try to return the CPU to a preempted task.
4017 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4021 * Someone really wants this to run. If it's not unfair, run it.
4023 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4026 clear_buddies(cfs_rq, se);
4031 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4033 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4036 * If still on the runqueue then deactivate_task()
4037 * was not called and update_curr() has to be done:
4040 update_curr(cfs_rq);
4042 /* throttle cfs_rqs exceeding runtime */
4043 check_cfs_rq_runtime(cfs_rq);
4045 check_spread(cfs_rq, prev);
4048 update_stats_wait_start(cfs_rq, prev);
4049 /* Put 'current' back into the tree. */
4050 __enqueue_entity(cfs_rq, prev);
4051 /* in !on_rq case, update occurred at dequeue */
4052 update_load_avg(prev, 0);
4054 cfs_rq->curr = NULL;
4058 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4061 * Update run-time statistics of the 'current'.
4063 update_curr(cfs_rq);
4066 * Ensure that runnable average is periodically updated.
4068 update_load_avg(curr, UPDATE_TG);
4069 update_cfs_shares(curr);
4071 #ifdef CONFIG_SCHED_HRTICK
4073 * queued ticks are scheduled to match the slice, so don't bother
4074 * validating it and just reschedule.
4077 resched_curr(rq_of(cfs_rq));
4081 * don't let the period tick interfere with the hrtick preemption
4083 if (!sched_feat(DOUBLE_TICK) &&
4084 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4088 if (cfs_rq->nr_running > 1)
4089 check_preempt_tick(cfs_rq, curr);
4093 /**************************************************
4094 * CFS bandwidth control machinery
4097 #ifdef CONFIG_CFS_BANDWIDTH
4099 #ifdef HAVE_JUMP_LABEL
4100 static struct static_key __cfs_bandwidth_used;
4102 static inline bool cfs_bandwidth_used(void)
4104 return static_key_false(&__cfs_bandwidth_used);
4107 void cfs_bandwidth_usage_inc(void)
4109 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4112 void cfs_bandwidth_usage_dec(void)
4114 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4116 #else /* HAVE_JUMP_LABEL */
4117 static bool cfs_bandwidth_used(void)
4122 void cfs_bandwidth_usage_inc(void) {}
4123 void cfs_bandwidth_usage_dec(void) {}
4124 #endif /* HAVE_JUMP_LABEL */
4127 * default period for cfs group bandwidth.
4128 * default: 0.1s, units: nanoseconds
4130 static inline u64 default_cfs_period(void)
4132 return 100000000ULL;
4135 static inline u64 sched_cfs_bandwidth_slice(void)
4137 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4141 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4142 * directly instead of rq->clock to avoid adding additional synchronization
4145 * requires cfs_b->lock
4147 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4149 if (cfs_b->quota != RUNTIME_INF)
4150 cfs_b->runtime = cfs_b->quota;
4153 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4155 return &tg->cfs_bandwidth;
4158 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4159 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4161 if (unlikely(cfs_rq->throttle_count))
4162 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4164 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4167 /* returns 0 on failure to allocate runtime */
4168 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4170 struct task_group *tg = cfs_rq->tg;
4171 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4172 u64 amount = 0, min_amount;
4174 /* note: this is a positive sum as runtime_remaining <= 0 */
4175 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4177 raw_spin_lock(&cfs_b->lock);
4178 if (cfs_b->quota == RUNTIME_INF)
4179 amount = min_amount;
4181 start_cfs_bandwidth(cfs_b);
4183 if (cfs_b->runtime > 0) {
4184 amount = min(cfs_b->runtime, min_amount);
4185 cfs_b->runtime -= amount;
4189 raw_spin_unlock(&cfs_b->lock);
4191 cfs_rq->runtime_remaining += amount;
4193 return cfs_rq->runtime_remaining > 0;
4196 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4198 /* dock delta_exec before expiring quota (as it could span periods) */
4199 cfs_rq->runtime_remaining -= delta_exec;
4201 if (likely(cfs_rq->runtime_remaining > 0))
4204 if (cfs_rq->throttled)
4207 * if we're unable to extend our runtime we resched so that the active
4208 * hierarchy can be throttled
4210 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4211 resched_curr(rq_of(cfs_rq));
4214 static __always_inline
4215 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4217 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4220 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4223 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4225 return cfs_bandwidth_used() && cfs_rq->throttled;
4228 /* check whether cfs_rq, or any parent, is throttled */
4229 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4231 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4235 * Ensure that neither of the group entities corresponding to src_cpu or
4236 * dest_cpu are members of a throttled hierarchy when performing group
4237 * load-balance operations.
4239 static inline int throttled_lb_pair(struct task_group *tg,
4240 int src_cpu, int dest_cpu)
4242 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4244 src_cfs_rq = tg->cfs_rq[src_cpu];
4245 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4247 return throttled_hierarchy(src_cfs_rq) ||
4248 throttled_hierarchy(dest_cfs_rq);
4251 /* updated child weight may affect parent so we have to do this bottom up */
4252 static int tg_unthrottle_up(struct task_group *tg, void *data)
4254 struct rq *rq = data;
4255 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4257 cfs_rq->throttle_count--;
4258 if (!cfs_rq->throttle_count) {
4259 /* adjust cfs_rq_clock_task() */
4260 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4261 cfs_rq->throttled_clock_task;
4267 static int tg_throttle_down(struct task_group *tg, void *data)
4269 struct rq *rq = data;
4270 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4272 /* group is entering throttled state, stop time */
4273 if (!cfs_rq->throttle_count)
4274 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4275 cfs_rq->throttle_count++;
4280 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4282 struct rq *rq = rq_of(cfs_rq);
4283 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4284 struct sched_entity *se;
4285 long task_delta, dequeue = 1;
4288 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4290 /* freeze hierarchy runnable averages while throttled */
4292 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4295 task_delta = cfs_rq->h_nr_running;
4296 for_each_sched_entity(se) {
4297 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4298 /* throttled entity or throttle-on-deactivate */
4303 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4304 qcfs_rq->h_nr_running -= task_delta;
4306 if (qcfs_rq->load.weight)
4311 sub_nr_running(rq, task_delta);
4313 cfs_rq->throttled = 1;
4314 cfs_rq->throttled_clock = rq_clock(rq);
4315 raw_spin_lock(&cfs_b->lock);
4316 empty = list_empty(&cfs_b->throttled_cfs_rq);
4319 * Add to the _head_ of the list, so that an already-started
4320 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4321 * not running add to the tail so that later runqueues don't get starved.
4323 if (cfs_b->distribute_running)
4324 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4326 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4329 * If we're the first throttled task, make sure the bandwidth
4333 start_cfs_bandwidth(cfs_b);
4335 raw_spin_unlock(&cfs_b->lock);
4338 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4340 struct rq *rq = rq_of(cfs_rq);
4341 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4342 struct sched_entity *se;
4346 se = cfs_rq->tg->se[cpu_of(rq)];
4348 cfs_rq->throttled = 0;
4350 update_rq_clock(rq);
4352 raw_spin_lock(&cfs_b->lock);
4353 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4354 list_del_rcu(&cfs_rq->throttled_list);
4355 raw_spin_unlock(&cfs_b->lock);
4357 /* update hierarchical throttle state */
4358 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4360 if (!cfs_rq->load.weight)
4363 task_delta = cfs_rq->h_nr_running;
4364 for_each_sched_entity(se) {
4368 cfs_rq = cfs_rq_of(se);
4370 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4371 cfs_rq->h_nr_running += task_delta;
4373 if (cfs_rq_throttled(cfs_rq))
4378 add_nr_running(rq, task_delta);
4380 /* determine whether we need to wake up potentially idle cpu */
4381 if (rq->curr == rq->idle && rq->cfs.nr_running)
4385 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4387 struct cfs_rq *cfs_rq;
4389 u64 starting_runtime = remaining;
4392 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4394 struct rq *rq = rq_of(cfs_rq);
4398 if (!cfs_rq_throttled(cfs_rq))
4401 /* By the above check, this should never be true */
4402 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4404 runtime = -cfs_rq->runtime_remaining + 1;
4405 if (runtime > remaining)
4406 runtime = remaining;
4407 remaining -= runtime;
4409 cfs_rq->runtime_remaining += runtime;
4411 /* we check whether we're throttled above */
4412 if (cfs_rq->runtime_remaining > 0)
4413 unthrottle_cfs_rq(cfs_rq);
4423 return starting_runtime - remaining;
4427 * Responsible for refilling a task_group's bandwidth and unthrottling its
4428 * cfs_rqs as appropriate. If there has been no activity within the last
4429 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4430 * used to track this state.
4432 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4437 /* no need to continue the timer with no bandwidth constraint */
4438 if (cfs_b->quota == RUNTIME_INF)
4439 goto out_deactivate;
4441 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4442 cfs_b->nr_periods += overrun;
4445 * idle depends on !throttled (for the case of a large deficit), and if
4446 * we're going inactive then everything else can be deferred
4448 if (cfs_b->idle && !throttled)
4449 goto out_deactivate;
4451 __refill_cfs_bandwidth_runtime(cfs_b);
4454 /* mark as potentially idle for the upcoming period */
4459 /* account preceding periods in which throttling occurred */
4460 cfs_b->nr_throttled += overrun;
4463 * This check is repeated as we are holding onto the new bandwidth while
4464 * we unthrottle. This can potentially race with an unthrottled group
4465 * trying to acquire new bandwidth from the global pool. This can result
4466 * in us over-using our runtime if it is all used during this loop, but
4467 * only by limited amounts in that extreme case.
4469 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4470 runtime = cfs_b->runtime;
4471 cfs_b->distribute_running = 1;
4472 raw_spin_unlock(&cfs_b->lock);
4473 /* we can't nest cfs_b->lock while distributing bandwidth */
4474 runtime = distribute_cfs_runtime(cfs_b, runtime);
4475 raw_spin_lock(&cfs_b->lock);
4477 cfs_b->distribute_running = 0;
4478 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4480 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4484 * While we are ensured activity in the period following an
4485 * unthrottle, this also covers the case in which the new bandwidth is
4486 * insufficient to cover the existing bandwidth deficit. (Forcing the
4487 * timer to remain active while there are any throttled entities.)
4497 /* a cfs_rq won't donate quota below this amount */
4498 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4499 /* minimum remaining period time to redistribute slack quota */
4500 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4501 /* how long we wait to gather additional slack before distributing */
4502 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4505 * Are we near the end of the current quota period?
4507 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4508 * hrtimer base being cleared by hrtimer_start. In the case of
4509 * migrate_hrtimers, base is never cleared, so we are fine.
4511 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4513 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4516 /* if the call-back is running a quota refresh is already occurring */
4517 if (hrtimer_callback_running(refresh_timer))
4520 /* is a quota refresh about to occur? */
4521 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4522 if (remaining < (s64)min_expire)
4528 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4530 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4532 /* if there's a quota refresh soon don't bother with slack */
4533 if (runtime_refresh_within(cfs_b, min_left))
4536 hrtimer_start(&cfs_b->slack_timer,
4537 ns_to_ktime(cfs_bandwidth_slack_period),
4541 /* we know any runtime found here is valid as update_curr() precedes return */
4542 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4544 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4545 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4547 if (slack_runtime <= 0)
4550 raw_spin_lock(&cfs_b->lock);
4551 if (cfs_b->quota != RUNTIME_INF) {
4552 cfs_b->runtime += slack_runtime;
4554 /* we are under rq->lock, defer unthrottling using a timer */
4555 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4556 !list_empty(&cfs_b->throttled_cfs_rq))
4557 start_cfs_slack_bandwidth(cfs_b);
4559 raw_spin_unlock(&cfs_b->lock);
4561 /* even if it's not valid for return we don't want to try again */
4562 cfs_rq->runtime_remaining -= slack_runtime;
4565 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4567 if (!cfs_bandwidth_used())
4570 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4573 __return_cfs_rq_runtime(cfs_rq);
4577 * This is done with a timer (instead of inline with bandwidth return) since
4578 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4580 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4582 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4584 /* confirm we're still not at a refresh boundary */
4585 raw_spin_lock(&cfs_b->lock);
4586 if (cfs_b->distribute_running) {
4587 raw_spin_unlock(&cfs_b->lock);
4591 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4592 raw_spin_unlock(&cfs_b->lock);
4596 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4597 runtime = cfs_b->runtime;
4600 cfs_b->distribute_running = 1;
4602 raw_spin_unlock(&cfs_b->lock);
4607 runtime = distribute_cfs_runtime(cfs_b, runtime);
4609 raw_spin_lock(&cfs_b->lock);
4610 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4611 cfs_b->distribute_running = 0;
4612 raw_spin_unlock(&cfs_b->lock);
4616 * When a group wakes up we want to make sure that its quota is not already
4617 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4618 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4620 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4622 if (!cfs_bandwidth_used())
4625 /* an active group must be handled by the update_curr()->put() path */
4626 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4629 /* ensure the group is not already throttled */
4630 if (cfs_rq_throttled(cfs_rq))
4633 /* update runtime allocation */
4634 account_cfs_rq_runtime(cfs_rq, 0);
4635 if (cfs_rq->runtime_remaining <= 0)
4636 throttle_cfs_rq(cfs_rq);
4639 static void sync_throttle(struct task_group *tg, int cpu)
4641 struct cfs_rq *pcfs_rq, *cfs_rq;
4643 if (!cfs_bandwidth_used())
4649 cfs_rq = tg->cfs_rq[cpu];
4650 pcfs_rq = tg->parent->cfs_rq[cpu];
4652 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4653 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4656 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4657 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4659 if (!cfs_bandwidth_used())
4662 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4666 * it's possible for a throttled entity to be forced into a running
4667 * state (e.g. set_curr_task), in this case we're finished.
4669 if (cfs_rq_throttled(cfs_rq))
4672 throttle_cfs_rq(cfs_rq);
4676 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4678 struct cfs_bandwidth *cfs_b =
4679 container_of(timer, struct cfs_bandwidth, slack_timer);
4681 do_sched_cfs_slack_timer(cfs_b);
4683 return HRTIMER_NORESTART;
4686 extern const u64 max_cfs_quota_period;
4688 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4690 struct cfs_bandwidth *cfs_b =
4691 container_of(timer, struct cfs_bandwidth, period_timer);
4696 raw_spin_lock(&cfs_b->lock);
4698 overrun = hrtimer_forward_now(timer, cfs_b->period);
4703 u64 new, old = ktime_to_ns(cfs_b->period);
4706 * Grow period by a factor of 2 to avoid losing precision.
4707 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4711 if (new < max_cfs_quota_period) {
4712 cfs_b->period = ns_to_ktime(new);
4715 pr_warn_ratelimited(
4716 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4718 div_u64(new, NSEC_PER_USEC),
4719 div_u64(cfs_b->quota, NSEC_PER_USEC));
4721 pr_warn_ratelimited(
4722 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4724 div_u64(old, NSEC_PER_USEC),
4725 div_u64(cfs_b->quota, NSEC_PER_USEC));
4728 /* reset count so we don't come right back in here */
4732 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4735 cfs_b->period_active = 0;
4736 raw_spin_unlock(&cfs_b->lock);
4738 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4741 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4743 raw_spin_lock_init(&cfs_b->lock);
4745 cfs_b->quota = RUNTIME_INF;
4746 cfs_b->period = ns_to_ktime(default_cfs_period());
4748 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4749 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4750 cfs_b->period_timer.function = sched_cfs_period_timer;
4751 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4752 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4753 cfs_b->distribute_running = 0;
4756 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4758 cfs_rq->runtime_enabled = 0;
4759 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4762 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4764 lockdep_assert_held(&cfs_b->lock);
4766 if (!cfs_b->period_active) {
4767 cfs_b->period_active = 1;
4768 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4769 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4773 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4775 /* init_cfs_bandwidth() was not called */
4776 if (!cfs_b->throttled_cfs_rq.next)
4779 hrtimer_cancel(&cfs_b->period_timer);
4780 hrtimer_cancel(&cfs_b->slack_timer);
4784 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4786 * The race is harmless, since modifying bandwidth settings of unhooked group
4787 * bits doesn't do much.
4790 /* cpu online calback */
4791 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4793 struct task_group *tg;
4795 lockdep_assert_held(&rq->lock);
4798 list_for_each_entry_rcu(tg, &task_groups, list) {
4799 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4800 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4802 raw_spin_lock(&cfs_b->lock);
4803 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4804 raw_spin_unlock(&cfs_b->lock);
4809 /* cpu offline callback */
4810 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4812 struct task_group *tg;
4814 lockdep_assert_held(&rq->lock);
4817 list_for_each_entry_rcu(tg, &task_groups, list) {
4818 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4820 if (!cfs_rq->runtime_enabled)
4824 * clock_task is not advancing so we just need to make sure
4825 * there's some valid quota amount
4827 cfs_rq->runtime_remaining = 1;
4829 * Offline rq is schedulable till cpu is completely disabled
4830 * in take_cpu_down(), so we prevent new cfs throttling here.
4832 cfs_rq->runtime_enabled = 0;
4834 if (cfs_rq_throttled(cfs_rq))
4835 unthrottle_cfs_rq(cfs_rq);
4840 #else /* CONFIG_CFS_BANDWIDTH */
4841 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4843 return rq_clock_task(rq_of(cfs_rq));
4846 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4847 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4848 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4849 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4850 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4852 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4857 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4862 static inline int throttled_lb_pair(struct task_group *tg,
4863 int src_cpu, int dest_cpu)
4868 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4870 #ifdef CONFIG_FAIR_GROUP_SCHED
4871 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4874 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4878 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4879 static inline void update_runtime_enabled(struct rq *rq) {}
4880 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4882 #endif /* CONFIG_CFS_BANDWIDTH */
4884 /**************************************************
4885 * CFS operations on tasks:
4888 #ifdef CONFIG_SCHED_HRTICK
4889 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4891 struct sched_entity *se = &p->se;
4892 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4894 SCHED_WARN_ON(task_rq(p) != rq);
4896 if (rq->cfs.h_nr_running > 1) {
4897 u64 slice = sched_slice(cfs_rq, se);
4898 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4899 s64 delta = slice - ran;
4906 hrtick_start(rq, delta);
4911 * called from enqueue/dequeue and updates the hrtick when the
4912 * current task is from our class and nr_running is low enough
4915 static void hrtick_update(struct rq *rq)
4917 struct task_struct *curr = rq->curr;
4919 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4922 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4923 hrtick_start_fair(rq, curr);
4925 #else /* !CONFIG_SCHED_HRTICK */
4927 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4931 static inline void hrtick_update(struct rq *rq)
4937 * The enqueue_task method is called before nr_running is
4938 * increased. Here we update the fair scheduling stats and
4939 * then put the task into the rbtree:
4942 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4944 struct cfs_rq *cfs_rq;
4945 struct sched_entity *se = &p->se;
4948 * If in_iowait is set, the code below may not trigger any cpufreq
4949 * utilization updates, so do it here explicitly with the IOWAIT flag
4953 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
4955 for_each_sched_entity(se) {
4958 cfs_rq = cfs_rq_of(se);
4959 enqueue_entity(cfs_rq, se, flags);
4962 * end evaluation on encountering a throttled cfs_rq
4964 * note: in the case of encountering a throttled cfs_rq we will
4965 * post the final h_nr_running increment below.
4967 if (cfs_rq_throttled(cfs_rq))
4969 cfs_rq->h_nr_running++;
4971 flags = ENQUEUE_WAKEUP;
4974 for_each_sched_entity(se) {
4975 cfs_rq = cfs_rq_of(se);
4976 cfs_rq->h_nr_running++;
4978 if (cfs_rq_throttled(cfs_rq))
4981 update_load_avg(se, UPDATE_TG);
4982 update_cfs_shares(se);
4986 add_nr_running(rq, 1);
4991 static void set_next_buddy(struct sched_entity *se);
4994 * The dequeue_task method is called before nr_running is
4995 * decreased. We remove the task from the rbtree and
4996 * update the fair scheduling stats:
4998 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5000 struct cfs_rq *cfs_rq;
5001 struct sched_entity *se = &p->se;
5002 int task_sleep = flags & DEQUEUE_SLEEP;
5004 for_each_sched_entity(se) {
5005 cfs_rq = cfs_rq_of(se);
5006 dequeue_entity(cfs_rq, se, flags);
5009 * end evaluation on encountering a throttled cfs_rq
5011 * note: in the case of encountering a throttled cfs_rq we will
5012 * post the final h_nr_running decrement below.
5014 if (cfs_rq_throttled(cfs_rq))
5016 cfs_rq->h_nr_running--;
5018 /* Don't dequeue parent if it has other entities besides us */
5019 if (cfs_rq->load.weight) {
5020 /* Avoid re-evaluating load for this entity: */
5021 se = parent_entity(se);
5023 * Bias pick_next to pick a task from this cfs_rq, as
5024 * p is sleeping when it is within its sched_slice.
5026 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5030 flags |= DEQUEUE_SLEEP;
5033 for_each_sched_entity(se) {
5034 cfs_rq = cfs_rq_of(se);
5035 cfs_rq->h_nr_running--;
5037 if (cfs_rq_throttled(cfs_rq))
5040 update_load_avg(se, UPDATE_TG);
5041 update_cfs_shares(se);
5045 sub_nr_running(rq, 1);
5052 /* Working cpumask for: load_balance, load_balance_newidle. */
5053 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5054 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5056 #ifdef CONFIG_NO_HZ_COMMON
5058 * per rq 'load' arrray crap; XXX kill this.
5062 * The exact cpuload calculated at every tick would be:
5064 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5066 * If a cpu misses updates for n ticks (as it was idle) and update gets
5067 * called on the n+1-th tick when cpu may be busy, then we have:
5069 * load_n = (1 - 1/2^i)^n * load_0
5070 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5072 * decay_load_missed() below does efficient calculation of
5074 * load' = (1 - 1/2^i)^n * load
5076 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5077 * This allows us to precompute the above in said factors, thereby allowing the
5078 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5079 * fixed_power_int())
5081 * The calculation is approximated on a 128 point scale.
5083 #define DEGRADE_SHIFT 7
5085 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5086 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5087 { 0, 0, 0, 0, 0, 0, 0, 0 },
5088 { 64, 32, 8, 0, 0, 0, 0, 0 },
5089 { 96, 72, 40, 12, 1, 0, 0, 0 },
5090 { 112, 98, 75, 43, 15, 1, 0, 0 },
5091 { 120, 112, 98, 76, 45, 16, 2, 0 }
5095 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5096 * would be when CPU is idle and so we just decay the old load without
5097 * adding any new load.
5099 static unsigned long
5100 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5104 if (!missed_updates)
5107 if (missed_updates >= degrade_zero_ticks[idx])
5111 return load >> missed_updates;
5113 while (missed_updates) {
5114 if (missed_updates % 2)
5115 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5117 missed_updates >>= 1;
5122 #endif /* CONFIG_NO_HZ_COMMON */
5125 * __cpu_load_update - update the rq->cpu_load[] statistics
5126 * @this_rq: The rq to update statistics for
5127 * @this_load: The current load
5128 * @pending_updates: The number of missed updates
5130 * Update rq->cpu_load[] statistics. This function is usually called every
5131 * scheduler tick (TICK_NSEC).
5133 * This function computes a decaying average:
5135 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5137 * Because of NOHZ it might not get called on every tick which gives need for
5138 * the @pending_updates argument.
5140 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5141 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5142 * = A * (A * load[i]_n-2 + B) + B
5143 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5144 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5145 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5146 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5147 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5149 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5150 * any change in load would have resulted in the tick being turned back on.
5152 * For regular NOHZ, this reduces to:
5154 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5156 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5159 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5160 unsigned long pending_updates)
5162 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5165 this_rq->nr_load_updates++;
5167 /* Update our load: */
5168 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5169 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5170 unsigned long old_load, new_load;
5172 /* scale is effectively 1 << i now, and >> i divides by scale */
5174 old_load = this_rq->cpu_load[i];
5175 #ifdef CONFIG_NO_HZ_COMMON
5176 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5177 if (tickless_load) {
5178 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5180 * old_load can never be a negative value because a
5181 * decayed tickless_load cannot be greater than the
5182 * original tickless_load.
5184 old_load += tickless_load;
5187 new_load = this_load;
5189 * Round up the averaging division if load is increasing. This
5190 * prevents us from getting stuck on 9 if the load is 10, for
5193 if (new_load > old_load)
5194 new_load += scale - 1;
5196 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5199 sched_avg_update(this_rq);
5202 /* Used instead of source_load when we know the type == 0 */
5203 static unsigned long weighted_cpuload(struct rq *rq)
5205 return cfs_rq_runnable_load_avg(&rq->cfs);
5208 #ifdef CONFIG_NO_HZ_COMMON
5210 * There is no sane way to deal with nohz on smp when using jiffies because the
5211 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5212 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5214 * Therefore we need to avoid the delta approach from the regular tick when
5215 * possible since that would seriously skew the load calculation. This is why we
5216 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5217 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5218 * loop exit, nohz_idle_balance, nohz full exit...)
5220 * This means we might still be one tick off for nohz periods.
5223 static void cpu_load_update_nohz(struct rq *this_rq,
5224 unsigned long curr_jiffies,
5227 unsigned long pending_updates;
5229 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5230 if (pending_updates) {
5231 this_rq->last_load_update_tick = curr_jiffies;
5233 * In the regular NOHZ case, we were idle, this means load 0.
5234 * In the NOHZ_FULL case, we were non-idle, we should consider
5235 * its weighted load.
5237 cpu_load_update(this_rq, load, pending_updates);
5242 * Called from nohz_idle_balance() to update the load ratings before doing the
5245 static void cpu_load_update_idle(struct rq *this_rq)
5248 * bail if there's load or we're actually up-to-date.
5250 if (weighted_cpuload(this_rq))
5253 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5257 * Record CPU load on nohz entry so we know the tickless load to account
5258 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5259 * than other cpu_load[idx] but it should be fine as cpu_load readers
5260 * shouldn't rely into synchronized cpu_load[*] updates.
5262 void cpu_load_update_nohz_start(void)
5264 struct rq *this_rq = this_rq();
5267 * This is all lockless but should be fine. If weighted_cpuload changes
5268 * concurrently we'll exit nohz. And cpu_load write can race with
5269 * cpu_load_update_idle() but both updater would be writing the same.
5271 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5275 * Account the tickless load in the end of a nohz frame.
5277 void cpu_load_update_nohz_stop(void)
5279 unsigned long curr_jiffies = READ_ONCE(jiffies);
5280 struct rq *this_rq = this_rq();
5284 if (curr_jiffies == this_rq->last_load_update_tick)
5287 load = weighted_cpuload(this_rq);
5288 rq_lock(this_rq, &rf);
5289 update_rq_clock(this_rq);
5290 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5291 rq_unlock(this_rq, &rf);
5293 #else /* !CONFIG_NO_HZ_COMMON */
5294 static inline void cpu_load_update_nohz(struct rq *this_rq,
5295 unsigned long curr_jiffies,
5296 unsigned long load) { }
5297 #endif /* CONFIG_NO_HZ_COMMON */
5299 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5301 #ifdef CONFIG_NO_HZ_COMMON
5302 /* See the mess around cpu_load_update_nohz(). */
5303 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5305 cpu_load_update(this_rq, load, 1);
5309 * Called from scheduler_tick()
5311 void cpu_load_update_active(struct rq *this_rq)
5313 unsigned long load = weighted_cpuload(this_rq);
5315 if (tick_nohz_tick_stopped())
5316 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5318 cpu_load_update_periodic(this_rq, load);
5322 * Return a low guess at the load of a migration-source cpu weighted
5323 * according to the scheduling class and "nice" value.
5325 * We want to under-estimate the load of migration sources, to
5326 * balance conservatively.
5328 static unsigned long source_load(int cpu, int type)
5330 struct rq *rq = cpu_rq(cpu);
5331 unsigned long total = weighted_cpuload(rq);
5333 if (type == 0 || !sched_feat(LB_BIAS))
5336 return min(rq->cpu_load[type-1], total);
5340 * Return a high guess at the load of a migration-target cpu weighted
5341 * according to the scheduling class and "nice" value.
5343 static unsigned long target_load(int cpu, int type)
5345 struct rq *rq = cpu_rq(cpu);
5346 unsigned long total = weighted_cpuload(rq);
5348 if (type == 0 || !sched_feat(LB_BIAS))
5351 return max(rq->cpu_load[type-1], total);
5354 static unsigned long capacity_of(int cpu)
5356 return cpu_rq(cpu)->cpu_capacity;
5359 static unsigned long capacity_orig_of(int cpu)
5361 return cpu_rq(cpu)->cpu_capacity_orig;
5364 static unsigned long cpu_avg_load_per_task(int cpu)
5366 struct rq *rq = cpu_rq(cpu);
5367 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5368 unsigned long load_avg = weighted_cpuload(rq);
5371 return load_avg / nr_running;
5376 static void record_wakee(struct task_struct *p)
5379 * Only decay a single time; tasks that have less then 1 wakeup per
5380 * jiffy will not have built up many flips.
5382 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5383 current->wakee_flips >>= 1;
5384 current->wakee_flip_decay_ts = jiffies;
5387 if (current->last_wakee != p) {
5388 current->last_wakee = p;
5389 current->wakee_flips++;
5394 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5396 * A waker of many should wake a different task than the one last awakened
5397 * at a frequency roughly N times higher than one of its wakees.
5399 * In order to determine whether we should let the load spread vs consolidating
5400 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5401 * partner, and a factor of lls_size higher frequency in the other.
5403 * With both conditions met, we can be relatively sure that the relationship is
5404 * non-monogamous, with partner count exceeding socket size.
5406 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5407 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5410 static int wake_wide(struct task_struct *p)
5412 unsigned int master = current->wakee_flips;
5413 unsigned int slave = p->wakee_flips;
5414 int factor = this_cpu_read(sd_llc_size);
5417 swap(master, slave);
5418 if (slave < factor || master < slave * factor)
5424 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5425 * soonest. For the purpose of speed we only consider the waking and previous
5428 * wake_affine_idle() - only considers 'now', it check if the waking CPU is (or
5431 * wake_affine_weight() - considers the weight to reflect the average
5432 * scheduling latency of the CPUs. This seems to work
5433 * for the overloaded case.
5437 wake_affine_idle(struct sched_domain *sd, struct task_struct *p,
5438 int this_cpu, int prev_cpu, int sync)
5440 if (idle_cpu(this_cpu))
5443 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5450 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5451 int this_cpu, int prev_cpu, int sync)
5453 s64 this_eff_load, prev_eff_load;
5454 unsigned long task_load;
5456 this_eff_load = target_load(this_cpu, sd->wake_idx);
5457 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5460 unsigned long current_load = task_h_load(current);
5462 if (current_load > this_eff_load)
5465 this_eff_load -= current_load;
5468 task_load = task_h_load(p);
5470 this_eff_load += task_load;
5471 if (sched_feat(WA_BIAS))
5472 this_eff_load *= 100;
5473 this_eff_load *= capacity_of(prev_cpu);
5475 prev_eff_load -= task_load;
5476 if (sched_feat(WA_BIAS))
5477 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5478 prev_eff_load *= capacity_of(this_cpu);
5480 return this_eff_load <= prev_eff_load;
5483 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5484 int prev_cpu, int sync)
5486 int this_cpu = smp_processor_id();
5487 bool affine = false;
5489 if (sched_feat(WA_IDLE) && !affine)
5490 affine = wake_affine_idle(sd, p, this_cpu, prev_cpu, sync);
5492 if (sched_feat(WA_WEIGHT) && !affine)
5493 affine = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5495 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5497 schedstat_inc(sd->ttwu_move_affine);
5498 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5504 static inline int task_util(struct task_struct *p);
5505 static int cpu_util_wake(int cpu, struct task_struct *p);
5507 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5509 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5513 * find_idlest_group finds and returns the least busy CPU group within the
5516 static struct sched_group *
5517 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5518 int this_cpu, int sd_flag)
5520 struct sched_group *idlest = NULL, *group = sd->groups;
5521 struct sched_group *most_spare_sg = NULL;
5522 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5523 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5524 unsigned long most_spare = 0, this_spare = 0;
5525 int load_idx = sd->forkexec_idx;
5526 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5527 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5528 (sd->imbalance_pct-100) / 100;
5530 if (sd_flag & SD_BALANCE_WAKE)
5531 load_idx = sd->wake_idx;
5534 unsigned long load, avg_load, runnable_load;
5535 unsigned long spare_cap, max_spare_cap;
5539 /* Skip over this group if it has no CPUs allowed */
5540 if (!cpumask_intersects(sched_group_span(group),
5544 local_group = cpumask_test_cpu(this_cpu,
5545 sched_group_span(group));
5548 * Tally up the load of all CPUs in the group and find
5549 * the group containing the CPU with most spare capacity.
5555 for_each_cpu(i, sched_group_span(group)) {
5556 /* Bias balancing toward cpus of our domain */
5558 load = source_load(i, load_idx);
5560 load = target_load(i, load_idx);
5562 runnable_load += load;
5564 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5566 spare_cap = capacity_spare_wake(i, p);
5568 if (spare_cap > max_spare_cap)
5569 max_spare_cap = spare_cap;
5572 /* Adjust by relative CPU capacity of the group */
5573 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5574 group->sgc->capacity;
5575 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5576 group->sgc->capacity;
5579 this_runnable_load = runnable_load;
5580 this_avg_load = avg_load;
5581 this_spare = max_spare_cap;
5583 if (min_runnable_load > (runnable_load + imbalance)) {
5585 * The runnable load is significantly smaller
5586 * so we can pick this new cpu
5588 min_runnable_load = runnable_load;
5589 min_avg_load = avg_load;
5591 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5592 (100*min_avg_load > imbalance_scale*avg_load)) {
5594 * The runnable loads are close so take the
5595 * blocked load into account through avg_load.
5597 min_avg_load = avg_load;
5601 if (most_spare < max_spare_cap) {
5602 most_spare = max_spare_cap;
5603 most_spare_sg = group;
5606 } while (group = group->next, group != sd->groups);
5609 * The cross-over point between using spare capacity or least load
5610 * is too conservative for high utilization tasks on partially
5611 * utilized systems if we require spare_capacity > task_util(p),
5612 * so we allow for some task stuffing by using
5613 * spare_capacity > task_util(p)/2.
5615 * Spare capacity can't be used for fork because the utilization has
5616 * not been set yet, we must first select a rq to compute the initial
5619 if (sd_flag & SD_BALANCE_FORK)
5622 if (this_spare > task_util(p) / 2 &&
5623 imbalance_scale*this_spare > 100*most_spare)
5626 if (most_spare > task_util(p) / 2)
5627 return most_spare_sg;
5633 if (min_runnable_load > (this_runnable_load + imbalance))
5636 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5637 (100*this_avg_load < imbalance_scale*min_avg_load))
5644 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5647 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5649 unsigned long load, min_load = ULONG_MAX;
5650 unsigned int min_exit_latency = UINT_MAX;
5651 u64 latest_idle_timestamp = 0;
5652 int least_loaded_cpu = this_cpu;
5653 int shallowest_idle_cpu = -1;
5656 /* Check if we have any choice: */
5657 if (group->group_weight == 1)
5658 return cpumask_first(sched_group_span(group));
5660 /* Traverse only the allowed CPUs */
5661 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5663 struct rq *rq = cpu_rq(i);
5664 struct cpuidle_state *idle = idle_get_state(rq);
5665 if (idle && idle->exit_latency < min_exit_latency) {
5667 * We give priority to a CPU whose idle state
5668 * has the smallest exit latency irrespective
5669 * of any idle timestamp.
5671 min_exit_latency = idle->exit_latency;
5672 latest_idle_timestamp = rq->idle_stamp;
5673 shallowest_idle_cpu = i;
5674 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5675 rq->idle_stamp > latest_idle_timestamp) {
5677 * If equal or no active idle state, then
5678 * the most recently idled CPU might have
5681 latest_idle_timestamp = rq->idle_stamp;
5682 shallowest_idle_cpu = i;
5684 } else if (shallowest_idle_cpu == -1) {
5685 load = weighted_cpuload(cpu_rq(i));
5686 if (load < min_load || (load == min_load && i == this_cpu)) {
5688 least_loaded_cpu = i;
5693 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5696 #ifdef CONFIG_SCHED_SMT
5697 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5698 EXPORT_SYMBOL_GPL(sched_smt_present);
5700 static inline void set_idle_cores(int cpu, int val)
5702 struct sched_domain_shared *sds;
5704 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5706 WRITE_ONCE(sds->has_idle_cores, val);
5709 static inline bool test_idle_cores(int cpu, bool def)
5711 struct sched_domain_shared *sds;
5713 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5715 return READ_ONCE(sds->has_idle_cores);
5721 * Scans the local SMT mask to see if the entire core is idle, and records this
5722 * information in sd_llc_shared->has_idle_cores.
5724 * Since SMT siblings share all cache levels, inspecting this limited remote
5725 * state should be fairly cheap.
5727 void __update_idle_core(struct rq *rq)
5729 int core = cpu_of(rq);
5733 if (test_idle_cores(core, true))
5736 for_each_cpu(cpu, cpu_smt_mask(core)) {
5744 set_idle_cores(core, 1);
5750 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5751 * there are no idle cores left in the system; tracked through
5752 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5754 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5756 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5759 if (!static_branch_likely(&sched_smt_present))
5762 if (!test_idle_cores(target, false))
5765 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5767 for_each_cpu_wrap(core, cpus, target) {
5770 for_each_cpu(cpu, cpu_smt_mask(core)) {
5771 cpumask_clear_cpu(cpu, cpus);
5781 * Failed to find an idle core; stop looking for one.
5783 set_idle_cores(target, 0);
5789 * Scan the local SMT mask for idle CPUs.
5791 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5795 if (!static_branch_likely(&sched_smt_present))
5798 for_each_cpu(cpu, cpu_smt_mask(target)) {
5799 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5808 #else /* CONFIG_SCHED_SMT */
5810 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5815 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5820 #endif /* CONFIG_SCHED_SMT */
5823 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5824 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5825 * average idle time for this rq (as found in rq->avg_idle).
5827 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5829 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5830 struct sched_domain *this_sd;
5831 u64 avg_cost, avg_idle;
5834 int cpu, nr = INT_MAX;
5836 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5841 * Due to large variance we need a large fuzz factor; hackbench in
5842 * particularly is sensitive here.
5844 avg_idle = this_rq()->avg_idle / 512;
5845 avg_cost = this_sd->avg_scan_cost + 1;
5847 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5850 if (sched_feat(SIS_PROP)) {
5851 u64 span_avg = sd->span_weight * avg_idle;
5852 if (span_avg > 4*avg_cost)
5853 nr = div_u64(span_avg, avg_cost);
5858 time = local_clock();
5860 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5862 for_each_cpu_wrap(cpu, cpus, target) {
5869 time = local_clock() - time;
5870 cost = this_sd->avg_scan_cost;
5871 delta = (s64)(time - cost) / 8;
5872 this_sd->avg_scan_cost += delta;
5878 * Try and locate an idle core/thread in the LLC cache domain.
5880 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5882 struct sched_domain *sd;
5885 if (idle_cpu(target))
5889 * If the previous cpu is cache affine and idle, don't be stupid.
5891 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5894 sd = rcu_dereference(per_cpu(sd_llc, target));
5898 i = select_idle_core(p, sd, target);
5899 if ((unsigned)i < nr_cpumask_bits)
5902 i = select_idle_cpu(p, sd, target);
5903 if ((unsigned)i < nr_cpumask_bits)
5906 i = select_idle_smt(p, sd, target);
5907 if ((unsigned)i < nr_cpumask_bits)
5914 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5915 * tasks. The unit of the return value must be the one of capacity so we can
5916 * compare the utilization with the capacity of the CPU that is available for
5917 * CFS task (ie cpu_capacity).
5919 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5920 * recent utilization of currently non-runnable tasks on a CPU. It represents
5921 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5922 * capacity_orig is the cpu_capacity available at the highest frequency
5923 * (arch_scale_freq_capacity()).
5924 * The utilization of a CPU converges towards a sum equal to or less than the
5925 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5926 * the running time on this CPU scaled by capacity_curr.
5928 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5929 * higher than capacity_orig because of unfortunate rounding in
5930 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5931 * the average stabilizes with the new running time. We need to check that the
5932 * utilization stays within the range of [0..capacity_orig] and cap it if
5933 * necessary. Without utilization capping, a group could be seen as overloaded
5934 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5935 * available capacity. We allow utilization to overshoot capacity_curr (but not
5936 * capacity_orig) as it useful for predicting the capacity required after task
5937 * migrations (scheduler-driven DVFS).
5939 static int cpu_util(int cpu)
5941 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5942 unsigned long capacity = capacity_orig_of(cpu);
5944 return (util >= capacity) ? capacity : util;
5947 static inline int task_util(struct task_struct *p)
5949 return p->se.avg.util_avg;
5953 * cpu_util_wake: Compute cpu utilization with any contributions from
5954 * the waking task p removed.
5956 static int cpu_util_wake(int cpu, struct task_struct *p)
5958 unsigned long util, capacity;
5960 /* Task has no contribution or is new */
5961 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5962 return cpu_util(cpu);
5964 capacity = capacity_orig_of(cpu);
5965 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5967 return (util >= capacity) ? capacity : util;
5971 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5972 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5974 * In that case WAKE_AFFINE doesn't make sense and we'll let
5975 * BALANCE_WAKE sort things out.
5977 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5979 long min_cap, max_cap;
5981 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5982 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5984 /* Minimum capacity is close to max, no need to abort wake_affine */
5985 if (max_cap - min_cap < max_cap >> 3)
5988 /* Bring task utilization in sync with prev_cpu */
5989 sync_entity_load_avg(&p->se);
5991 return min_cap * 1024 < task_util(p) * capacity_margin;
5995 * select_task_rq_fair: Select target runqueue for the waking task in domains
5996 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5997 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5999 * Balances load by selecting the idlest cpu in the idlest group, or under
6000 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6002 * Returns the target cpu number.
6004 * preempt must be disabled.
6007 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6009 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6010 int cpu = smp_processor_id();
6011 int new_cpu = prev_cpu;
6012 int want_affine = 0;
6013 int sync = wake_flags & WF_SYNC;
6015 if (sd_flag & SD_BALANCE_WAKE) {
6017 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6018 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6022 for_each_domain(cpu, tmp) {
6023 if (!(tmp->flags & SD_LOAD_BALANCE))
6027 * If both cpu and prev_cpu are part of this domain,
6028 * cpu is a valid SD_WAKE_AFFINE target.
6030 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6031 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6036 if (tmp->flags & sd_flag)
6038 else if (!want_affine)
6043 sd = NULL; /* Prefer wake_affine over balance flags */
6044 if (cpu == prev_cpu)
6047 if (wake_affine(affine_sd, p, prev_cpu, sync))
6053 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6054 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6057 struct sched_group *group;
6060 if (!(sd->flags & sd_flag)) {
6065 group = find_idlest_group(sd, p, cpu, sd_flag);
6071 new_cpu = find_idlest_cpu(group, p, cpu);
6072 if (new_cpu == -1 || new_cpu == cpu) {
6073 /* Now try balancing at a lower domain level of cpu */
6078 /* Now try balancing at a lower domain level of new_cpu */
6080 weight = sd->span_weight;
6082 for_each_domain(cpu, tmp) {
6083 if (weight <= tmp->span_weight)
6085 if (tmp->flags & sd_flag)
6088 /* while loop will break here if sd == NULL */
6096 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6097 * cfs_rq_of(p) references at time of call are still valid and identify the
6098 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6100 static void migrate_task_rq_fair(struct task_struct *p)
6103 * As blocked tasks retain absolute vruntime the migration needs to
6104 * deal with this by subtracting the old and adding the new
6105 * min_vruntime -- the latter is done by enqueue_entity() when placing
6106 * the task on the new runqueue.
6108 if (p->state == TASK_WAKING) {
6109 struct sched_entity *se = &p->se;
6110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6113 #ifndef CONFIG_64BIT
6114 u64 min_vruntime_copy;
6117 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6119 min_vruntime = cfs_rq->min_vruntime;
6120 } while (min_vruntime != min_vruntime_copy);
6122 min_vruntime = cfs_rq->min_vruntime;
6125 se->vruntime -= min_vruntime;
6129 * We are supposed to update the task to "current" time, then its up to date
6130 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6131 * what current time is, so simply throw away the out-of-date time. This
6132 * will result in the wakee task is less decayed, but giving the wakee more
6133 * load sounds not bad.
6135 remove_entity_load_avg(&p->se);
6137 /* Tell new CPU we are migrated */
6138 p->se.avg.last_update_time = 0;
6141 static void task_dead_fair(struct task_struct *p)
6143 remove_entity_load_avg(&p->se);
6145 #endif /* CONFIG_SMP */
6147 static unsigned long
6148 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6150 unsigned long gran = sysctl_sched_wakeup_granularity;
6153 * Since its curr running now, convert the gran from real-time
6154 * to virtual-time in his units.
6156 * By using 'se' instead of 'curr' we penalize light tasks, so
6157 * they get preempted easier. That is, if 'se' < 'curr' then
6158 * the resulting gran will be larger, therefore penalizing the
6159 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6160 * be smaller, again penalizing the lighter task.
6162 * This is especially important for buddies when the leftmost
6163 * task is higher priority than the buddy.
6165 return calc_delta_fair(gran, se);
6169 * Should 'se' preempt 'curr'.
6183 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6185 s64 gran, vdiff = curr->vruntime - se->vruntime;
6190 gran = wakeup_gran(curr, se);
6197 static void set_last_buddy(struct sched_entity *se)
6199 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6202 for_each_sched_entity(se) {
6203 if (SCHED_WARN_ON(!se->on_rq))
6205 cfs_rq_of(se)->last = se;
6209 static void set_next_buddy(struct sched_entity *se)
6211 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6214 for_each_sched_entity(se) {
6215 if (SCHED_WARN_ON(!se->on_rq))
6217 cfs_rq_of(se)->next = se;
6221 static void set_skip_buddy(struct sched_entity *se)
6223 for_each_sched_entity(se)
6224 cfs_rq_of(se)->skip = se;
6228 * Preempt the current task with a newly woken task if needed:
6230 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6232 struct task_struct *curr = rq->curr;
6233 struct sched_entity *se = &curr->se, *pse = &p->se;
6234 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6235 int scale = cfs_rq->nr_running >= sched_nr_latency;
6236 int next_buddy_marked = 0;
6238 if (unlikely(se == pse))
6242 * This is possible from callers such as attach_tasks(), in which we
6243 * unconditionally check_prempt_curr() after an enqueue (which may have
6244 * lead to a throttle). This both saves work and prevents false
6245 * next-buddy nomination below.
6247 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6250 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6251 set_next_buddy(pse);
6252 next_buddy_marked = 1;
6256 * We can come here with TIF_NEED_RESCHED already set from new task
6259 * Note: this also catches the edge-case of curr being in a throttled
6260 * group (e.g. via set_curr_task), since update_curr() (in the
6261 * enqueue of curr) will have resulted in resched being set. This
6262 * prevents us from potentially nominating it as a false LAST_BUDDY
6265 if (test_tsk_need_resched(curr))
6268 /* Idle tasks are by definition preempted by non-idle tasks. */
6269 if (unlikely(curr->policy == SCHED_IDLE) &&
6270 likely(p->policy != SCHED_IDLE))
6274 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6275 * is driven by the tick):
6277 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6280 find_matching_se(&se, &pse);
6281 update_curr(cfs_rq_of(se));
6283 if (wakeup_preempt_entity(se, pse) == 1) {
6285 * Bias pick_next to pick the sched entity that is
6286 * triggering this preemption.
6288 if (!next_buddy_marked)
6289 set_next_buddy(pse);
6298 * Only set the backward buddy when the current task is still
6299 * on the rq. This can happen when a wakeup gets interleaved
6300 * with schedule on the ->pre_schedule() or idle_balance()
6301 * point, either of which can * drop the rq lock.
6303 * Also, during early boot the idle thread is in the fair class,
6304 * for obvious reasons its a bad idea to schedule back to it.
6306 if (unlikely(!se->on_rq || curr == rq->idle))
6309 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6313 static struct task_struct *
6314 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6316 struct cfs_rq *cfs_rq = &rq->cfs;
6317 struct sched_entity *se;
6318 struct task_struct *p;
6322 if (!cfs_rq->nr_running)
6325 #ifdef CONFIG_FAIR_GROUP_SCHED
6326 if (prev->sched_class != &fair_sched_class)
6330 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6331 * likely that a next task is from the same cgroup as the current.
6333 * Therefore attempt to avoid putting and setting the entire cgroup
6334 * hierarchy, only change the part that actually changes.
6338 struct sched_entity *curr = cfs_rq->curr;
6341 * Since we got here without doing put_prev_entity() we also
6342 * have to consider cfs_rq->curr. If it is still a runnable
6343 * entity, update_curr() will update its vruntime, otherwise
6344 * forget we've ever seen it.
6348 update_curr(cfs_rq);
6353 * This call to check_cfs_rq_runtime() will do the
6354 * throttle and dequeue its entity in the parent(s).
6355 * Therefore the nr_running test will indeed
6358 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6361 if (!cfs_rq->nr_running)
6368 se = pick_next_entity(cfs_rq, curr);
6369 cfs_rq = group_cfs_rq(se);
6375 * Since we haven't yet done put_prev_entity and if the selected task
6376 * is a different task than we started out with, try and touch the
6377 * least amount of cfs_rqs.
6380 struct sched_entity *pse = &prev->se;
6382 while (!(cfs_rq = is_same_group(se, pse))) {
6383 int se_depth = se->depth;
6384 int pse_depth = pse->depth;
6386 if (se_depth <= pse_depth) {
6387 put_prev_entity(cfs_rq_of(pse), pse);
6388 pse = parent_entity(pse);
6390 if (se_depth >= pse_depth) {
6391 set_next_entity(cfs_rq_of(se), se);
6392 se = parent_entity(se);
6396 put_prev_entity(cfs_rq, pse);
6397 set_next_entity(cfs_rq, se);
6400 if (hrtick_enabled(rq))
6401 hrtick_start_fair(rq, p);
6407 put_prev_task(rq, prev);
6410 se = pick_next_entity(cfs_rq, NULL);
6411 set_next_entity(cfs_rq, se);
6412 cfs_rq = group_cfs_rq(se);
6417 if (hrtick_enabled(rq))
6418 hrtick_start_fair(rq, p);
6423 new_tasks = idle_balance(rq, rf);
6426 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6427 * possible for any higher priority task to appear. In that case we
6428 * must re-start the pick_next_entity() loop.
6440 * Account for a descheduled task:
6442 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6444 struct sched_entity *se = &prev->se;
6445 struct cfs_rq *cfs_rq;
6447 for_each_sched_entity(se) {
6448 cfs_rq = cfs_rq_of(se);
6449 put_prev_entity(cfs_rq, se);
6454 * sched_yield() is very simple
6456 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6458 static void yield_task_fair(struct rq *rq)
6460 struct task_struct *curr = rq->curr;
6461 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6462 struct sched_entity *se = &curr->se;
6465 * Are we the only task in the tree?
6467 if (unlikely(rq->nr_running == 1))
6470 clear_buddies(cfs_rq, se);
6472 if (curr->policy != SCHED_BATCH) {
6473 update_rq_clock(rq);
6475 * Update run-time statistics of the 'current'.
6477 update_curr(cfs_rq);
6479 * Tell update_rq_clock() that we've just updated,
6480 * so we don't do microscopic update in schedule()
6481 * and double the fastpath cost.
6483 rq_clock_skip_update(rq, true);
6489 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6491 struct sched_entity *se = &p->se;
6493 /* throttled hierarchies are not runnable */
6494 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6497 /* Tell the scheduler that we'd really like pse to run next. */
6500 yield_task_fair(rq);
6506 /**************************************************
6507 * Fair scheduling class load-balancing methods.
6511 * The purpose of load-balancing is to achieve the same basic fairness the
6512 * per-cpu scheduler provides, namely provide a proportional amount of compute
6513 * time to each task. This is expressed in the following equation:
6515 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6517 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6518 * W_i,0 is defined as:
6520 * W_i,0 = \Sum_j w_i,j (2)
6522 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6523 * is derived from the nice value as per sched_prio_to_weight[].
6525 * The weight average is an exponential decay average of the instantaneous
6528 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6530 * C_i is the compute capacity of cpu i, typically it is the
6531 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6532 * can also include other factors [XXX].
6534 * To achieve this balance we define a measure of imbalance which follows
6535 * directly from (1):
6537 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6539 * We them move tasks around to minimize the imbalance. In the continuous
6540 * function space it is obvious this converges, in the discrete case we get
6541 * a few fun cases generally called infeasible weight scenarios.
6544 * - infeasible weights;
6545 * - local vs global optima in the discrete case. ]
6550 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6551 * for all i,j solution, we create a tree of cpus that follows the hardware
6552 * topology where each level pairs two lower groups (or better). This results
6553 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6554 * tree to only the first of the previous level and we decrease the frequency
6555 * of load-balance at each level inv. proportional to the number of cpus in
6561 * \Sum { --- * --- * 2^i } = O(n) (5)
6563 * `- size of each group
6564 * | | `- number of cpus doing load-balance
6566 * `- sum over all levels
6568 * Coupled with a limit on how many tasks we can migrate every balance pass,
6569 * this makes (5) the runtime complexity of the balancer.
6571 * An important property here is that each CPU is still (indirectly) connected
6572 * to every other cpu in at most O(log n) steps:
6574 * The adjacency matrix of the resulting graph is given by:
6577 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6580 * And you'll find that:
6582 * A^(log_2 n)_i,j != 0 for all i,j (7)
6584 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6585 * The task movement gives a factor of O(m), giving a convergence complexity
6588 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6593 * In order to avoid CPUs going idle while there's still work to do, new idle
6594 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6595 * tree itself instead of relying on other CPUs to bring it work.
6597 * This adds some complexity to both (5) and (8) but it reduces the total idle
6605 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6608 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6613 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6615 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6617 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6620 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6621 * rewrite all of this once again.]
6624 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6626 enum fbq_type { regular, remote, all };
6628 #define LBF_ALL_PINNED 0x01
6629 #define LBF_NEED_BREAK 0x02
6630 #define LBF_DST_PINNED 0x04
6631 #define LBF_SOME_PINNED 0x08
6634 struct sched_domain *sd;
6642 struct cpumask *dst_grpmask;
6644 enum cpu_idle_type idle;
6646 /* The set of CPUs under consideration for load-balancing */
6647 struct cpumask *cpus;
6652 unsigned int loop_break;
6653 unsigned int loop_max;
6655 enum fbq_type fbq_type;
6656 struct list_head tasks;
6660 * Is this task likely cache-hot:
6662 static int task_hot(struct task_struct *p, struct lb_env *env)
6666 lockdep_assert_held(&env->src_rq->lock);
6668 if (p->sched_class != &fair_sched_class)
6671 if (unlikely(p->policy == SCHED_IDLE))
6675 * Buddy candidates are cache hot:
6677 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6678 (&p->se == cfs_rq_of(&p->se)->next ||
6679 &p->se == cfs_rq_of(&p->se)->last))
6682 if (sysctl_sched_migration_cost == -1)
6684 if (sysctl_sched_migration_cost == 0)
6687 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6689 return delta < (s64)sysctl_sched_migration_cost;
6692 #ifdef CONFIG_NUMA_BALANCING
6694 * Returns 1, if task migration degrades locality
6695 * Returns 0, if task migration improves locality i.e migration preferred.
6696 * Returns -1, if task migration is not affected by locality.
6698 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6700 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6701 unsigned long src_faults, dst_faults;
6702 int src_nid, dst_nid;
6704 if (!static_branch_likely(&sched_numa_balancing))
6707 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6710 src_nid = cpu_to_node(env->src_cpu);
6711 dst_nid = cpu_to_node(env->dst_cpu);
6713 if (src_nid == dst_nid)
6716 /* Migrating away from the preferred node is always bad. */
6717 if (src_nid == p->numa_preferred_nid) {
6718 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6724 /* Encourage migration to the preferred node. */
6725 if (dst_nid == p->numa_preferred_nid)
6728 /* Leaving a core idle is often worse than degrading locality. */
6729 if (env->idle != CPU_NOT_IDLE)
6733 src_faults = group_faults(p, src_nid);
6734 dst_faults = group_faults(p, dst_nid);
6736 src_faults = task_faults(p, src_nid);
6737 dst_faults = task_faults(p, dst_nid);
6740 return dst_faults < src_faults;
6744 static inline int migrate_degrades_locality(struct task_struct *p,
6752 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6755 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6759 lockdep_assert_held(&env->src_rq->lock);
6762 * We do not migrate tasks that are:
6763 * 1) throttled_lb_pair, or
6764 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6765 * 3) running (obviously), or
6766 * 4) are cache-hot on their current CPU.
6768 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6771 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6774 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6776 env->flags |= LBF_SOME_PINNED;
6779 * Remember if this task can be migrated to any other cpu in
6780 * our sched_group. We may want to revisit it if we couldn't
6781 * meet load balance goals by pulling other tasks on src_cpu.
6783 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6784 * already computed one in current iteration.
6786 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6789 /* Prevent to re-select dst_cpu via env's cpus */
6790 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6791 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6792 env->flags |= LBF_DST_PINNED;
6793 env->new_dst_cpu = cpu;
6801 /* Record that we found atleast one task that could run on dst_cpu */
6802 env->flags &= ~LBF_ALL_PINNED;
6804 if (task_running(env->src_rq, p)) {
6805 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6810 * Aggressive migration if:
6811 * 1) destination numa is preferred
6812 * 2) task is cache cold, or
6813 * 3) too many balance attempts have failed.
6815 tsk_cache_hot = migrate_degrades_locality(p, env);
6816 if (tsk_cache_hot == -1)
6817 tsk_cache_hot = task_hot(p, env);
6819 if (tsk_cache_hot <= 0 ||
6820 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6821 if (tsk_cache_hot == 1) {
6822 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6823 schedstat_inc(p->se.statistics.nr_forced_migrations);
6828 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6833 * detach_task() -- detach the task for the migration specified in env
6835 static void detach_task(struct task_struct *p, struct lb_env *env)
6837 lockdep_assert_held(&env->src_rq->lock);
6839 p->on_rq = TASK_ON_RQ_MIGRATING;
6840 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6841 set_task_cpu(p, env->dst_cpu);
6845 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6846 * part of active balancing operations within "domain".
6848 * Returns a task if successful and NULL otherwise.
6850 static struct task_struct *detach_one_task(struct lb_env *env)
6852 struct task_struct *p, *n;
6854 lockdep_assert_held(&env->src_rq->lock);
6856 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6857 if (!can_migrate_task(p, env))
6860 detach_task(p, env);
6863 * Right now, this is only the second place where
6864 * lb_gained[env->idle] is updated (other is detach_tasks)
6865 * so we can safely collect stats here rather than
6866 * inside detach_tasks().
6868 schedstat_inc(env->sd->lb_gained[env->idle]);
6874 static const unsigned int sched_nr_migrate_break = 32;
6877 * detach_tasks() -- tries to detach up to imbalance weighted load from
6878 * busiest_rq, as part of a balancing operation within domain "sd".
6880 * Returns number of detached tasks if successful and 0 otherwise.
6882 static int detach_tasks(struct lb_env *env)
6884 struct list_head *tasks = &env->src_rq->cfs_tasks;
6885 struct task_struct *p;
6889 lockdep_assert_held(&env->src_rq->lock);
6891 if (env->imbalance <= 0)
6894 while (!list_empty(tasks)) {
6896 * We don't want to steal all, otherwise we may be treated likewise,
6897 * which could at worst lead to a livelock crash.
6899 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6902 p = list_first_entry(tasks, struct task_struct, se.group_node);
6905 /* We've more or less seen every task there is, call it quits */
6906 if (env->loop > env->loop_max)
6909 /* take a breather every nr_migrate tasks */
6910 if (env->loop > env->loop_break) {
6911 env->loop_break += sched_nr_migrate_break;
6912 env->flags |= LBF_NEED_BREAK;
6916 if (!can_migrate_task(p, env))
6920 * Depending of the number of CPUs and tasks and the
6921 * cgroup hierarchy, task_h_load() can return a null
6922 * value. Make sure that env->imbalance decreases
6923 * otherwise detach_tasks() will stop only after
6924 * detaching up to loop_max tasks.
6926 load = max_t(unsigned long, task_h_load(p), 1);
6929 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6932 if ((load / 2) > env->imbalance)
6935 detach_task(p, env);
6936 list_add(&p->se.group_node, &env->tasks);
6939 env->imbalance -= load;
6941 #ifdef CONFIG_PREEMPT
6943 * NEWIDLE balancing is a source of latency, so preemptible
6944 * kernels will stop after the first task is detached to minimize
6945 * the critical section.
6947 if (env->idle == CPU_NEWLY_IDLE)
6952 * We only want to steal up to the prescribed amount of
6955 if (env->imbalance <= 0)
6960 list_move_tail(&p->se.group_node, tasks);
6964 * Right now, this is one of only two places we collect this stat
6965 * so we can safely collect detach_one_task() stats here rather
6966 * than inside detach_one_task().
6968 schedstat_add(env->sd->lb_gained[env->idle], detached);
6974 * attach_task() -- attach the task detached by detach_task() to its new rq.
6976 static void attach_task(struct rq *rq, struct task_struct *p)
6978 lockdep_assert_held(&rq->lock);
6980 BUG_ON(task_rq(p) != rq);
6981 activate_task(rq, p, ENQUEUE_NOCLOCK);
6982 p->on_rq = TASK_ON_RQ_QUEUED;
6983 check_preempt_curr(rq, p, 0);
6987 * attach_one_task() -- attaches the task returned from detach_one_task() to
6990 static void attach_one_task(struct rq *rq, struct task_struct *p)
6995 update_rq_clock(rq);
7001 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7004 static void attach_tasks(struct lb_env *env)
7006 struct list_head *tasks = &env->tasks;
7007 struct task_struct *p;
7010 rq_lock(env->dst_rq, &rf);
7011 update_rq_clock(env->dst_rq);
7013 while (!list_empty(tasks)) {
7014 p = list_first_entry(tasks, struct task_struct, se.group_node);
7015 list_del_init(&p->se.group_node);
7017 attach_task(env->dst_rq, p);
7020 rq_unlock(env->dst_rq, &rf);
7023 #ifdef CONFIG_FAIR_GROUP_SCHED
7025 static void update_blocked_averages(int cpu)
7027 struct rq *rq = cpu_rq(cpu);
7028 struct cfs_rq *cfs_rq;
7031 rq_lock_irqsave(rq, &rf);
7032 update_rq_clock(rq);
7035 * Iterates the task_group tree in a bottom up fashion, see
7036 * list_add_leaf_cfs_rq() for details.
7038 for_each_leaf_cfs_rq(rq, cfs_rq) {
7039 struct sched_entity *se;
7041 /* throttled entities do not contribute to load */
7042 if (throttled_hierarchy(cfs_rq))
7045 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7046 update_tg_load_avg(cfs_rq, 0);
7048 /* Propagate pending load changes to the parent, if any: */
7049 se = cfs_rq->tg->se[cpu];
7050 if (se && !skip_blocked_update(se))
7051 update_load_avg(se, 0);
7053 rq_unlock_irqrestore(rq, &rf);
7057 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7058 * This needs to be done in a top-down fashion because the load of a child
7059 * group is a fraction of its parents load.
7061 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7063 struct rq *rq = rq_of(cfs_rq);
7064 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7065 unsigned long now = jiffies;
7068 if (cfs_rq->last_h_load_update == now)
7071 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7072 for_each_sched_entity(se) {
7073 cfs_rq = cfs_rq_of(se);
7074 WRITE_ONCE(cfs_rq->h_load_next, se);
7075 if (cfs_rq->last_h_load_update == now)
7080 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7081 cfs_rq->last_h_load_update = now;
7084 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7085 load = cfs_rq->h_load;
7086 load = div64_ul(load * se->avg.load_avg,
7087 cfs_rq_load_avg(cfs_rq) + 1);
7088 cfs_rq = group_cfs_rq(se);
7089 cfs_rq->h_load = load;
7090 cfs_rq->last_h_load_update = now;
7094 static unsigned long task_h_load(struct task_struct *p)
7096 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7098 update_cfs_rq_h_load(cfs_rq);
7099 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7100 cfs_rq_load_avg(cfs_rq) + 1);
7103 static inline void update_blocked_averages(int cpu)
7105 struct rq *rq = cpu_rq(cpu);
7106 struct cfs_rq *cfs_rq = &rq->cfs;
7109 rq_lock_irqsave(rq, &rf);
7110 update_rq_clock(rq);
7111 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7112 rq_unlock_irqrestore(rq, &rf);
7115 static unsigned long task_h_load(struct task_struct *p)
7117 return p->se.avg.load_avg;
7121 /********** Helpers for find_busiest_group ************************/
7130 * sg_lb_stats - stats of a sched_group required for load_balancing
7132 struct sg_lb_stats {
7133 unsigned long avg_load; /*Avg load across the CPUs of the group */
7134 unsigned long group_load; /* Total load over the CPUs of the group */
7135 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7136 unsigned long load_per_task;
7137 unsigned long group_capacity;
7138 unsigned long group_util; /* Total utilization of the group */
7139 unsigned int sum_nr_running; /* Nr tasks running in the group */
7140 unsigned int idle_cpus;
7141 unsigned int group_weight;
7142 enum group_type group_type;
7143 int group_no_capacity;
7144 #ifdef CONFIG_NUMA_BALANCING
7145 unsigned int nr_numa_running;
7146 unsigned int nr_preferred_running;
7151 * sd_lb_stats - Structure to store the statistics of a sched_domain
7152 * during load balancing.
7154 struct sd_lb_stats {
7155 struct sched_group *busiest; /* Busiest group in this sd */
7156 struct sched_group *local; /* Local group in this sd */
7157 unsigned long total_running;
7158 unsigned long total_load; /* Total load of all groups in sd */
7159 unsigned long total_capacity; /* Total capacity of all groups in sd */
7160 unsigned long avg_load; /* Average load across all groups in sd */
7162 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7163 struct sg_lb_stats local_stat; /* Statistics of the local group */
7166 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7169 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7170 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7171 * We must however clear busiest_stat::avg_load because
7172 * update_sd_pick_busiest() reads this before assignment.
7174 *sds = (struct sd_lb_stats){
7177 .total_running = 0UL,
7179 .total_capacity = 0UL,
7182 .sum_nr_running = 0,
7183 .group_type = group_other,
7189 * get_sd_load_idx - Obtain the load index for a given sched domain.
7190 * @sd: The sched_domain whose load_idx is to be obtained.
7191 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7193 * Return: The load index.
7195 static inline int get_sd_load_idx(struct sched_domain *sd,
7196 enum cpu_idle_type idle)
7202 load_idx = sd->busy_idx;
7205 case CPU_NEWLY_IDLE:
7206 load_idx = sd->newidle_idx;
7209 load_idx = sd->idle_idx;
7216 static unsigned long scale_rt_capacity(int cpu)
7218 struct rq *rq = cpu_rq(cpu);
7219 u64 total, used, age_stamp, avg;
7223 * Since we're reading these variables without serialization make sure
7224 * we read them once before doing sanity checks on them.
7226 age_stamp = READ_ONCE(rq->age_stamp);
7227 avg = READ_ONCE(rq->rt_avg);
7228 delta = __rq_clock_broken(rq) - age_stamp;
7230 if (unlikely(delta < 0))
7233 total = sched_avg_period() + delta;
7235 used = div_u64(avg, total);
7237 if (likely(used < SCHED_CAPACITY_SCALE))
7238 return SCHED_CAPACITY_SCALE - used;
7243 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7245 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7246 struct sched_group *sdg = sd->groups;
7248 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7250 capacity *= scale_rt_capacity(cpu);
7251 capacity >>= SCHED_CAPACITY_SHIFT;
7256 cpu_rq(cpu)->cpu_capacity = capacity;
7257 sdg->sgc->capacity = capacity;
7258 sdg->sgc->min_capacity = capacity;
7261 void update_group_capacity(struct sched_domain *sd, int cpu)
7263 struct sched_domain *child = sd->child;
7264 struct sched_group *group, *sdg = sd->groups;
7265 unsigned long capacity, min_capacity;
7266 unsigned long interval;
7268 interval = msecs_to_jiffies(sd->balance_interval);
7269 interval = clamp(interval, 1UL, max_load_balance_interval);
7270 sdg->sgc->next_update = jiffies + interval;
7273 update_cpu_capacity(sd, cpu);
7278 min_capacity = ULONG_MAX;
7280 if (child->flags & SD_OVERLAP) {
7282 * SD_OVERLAP domains cannot assume that child groups
7283 * span the current group.
7286 for_each_cpu(cpu, sched_group_span(sdg)) {
7287 struct sched_group_capacity *sgc;
7288 struct rq *rq = cpu_rq(cpu);
7291 * build_sched_domains() -> init_sched_groups_capacity()
7292 * gets here before we've attached the domains to the
7295 * Use capacity_of(), which is set irrespective of domains
7296 * in update_cpu_capacity().
7298 * This avoids capacity from being 0 and
7299 * causing divide-by-zero issues on boot.
7301 if (unlikely(!rq->sd)) {
7302 capacity += capacity_of(cpu);
7304 sgc = rq->sd->groups->sgc;
7305 capacity += sgc->capacity;
7308 min_capacity = min(capacity, min_capacity);
7312 * !SD_OVERLAP domains can assume that child groups
7313 * span the current group.
7316 group = child->groups;
7318 struct sched_group_capacity *sgc = group->sgc;
7320 capacity += sgc->capacity;
7321 min_capacity = min(sgc->min_capacity, min_capacity);
7322 group = group->next;
7323 } while (group != child->groups);
7326 sdg->sgc->capacity = capacity;
7327 sdg->sgc->min_capacity = min_capacity;
7331 * Check whether the capacity of the rq has been noticeably reduced by side
7332 * activity. The imbalance_pct is used for the threshold.
7333 * Return true is the capacity is reduced
7336 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7338 return ((rq->cpu_capacity * sd->imbalance_pct) <
7339 (rq->cpu_capacity_orig * 100));
7343 * Group imbalance indicates (and tries to solve) the problem where balancing
7344 * groups is inadequate due to ->cpus_allowed constraints.
7346 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7347 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7350 * { 0 1 2 3 } { 4 5 6 7 }
7353 * If we were to balance group-wise we'd place two tasks in the first group and
7354 * two tasks in the second group. Clearly this is undesired as it will overload
7355 * cpu 3 and leave one of the cpus in the second group unused.
7357 * The current solution to this issue is detecting the skew in the first group
7358 * by noticing the lower domain failed to reach balance and had difficulty
7359 * moving tasks due to affinity constraints.
7361 * When this is so detected; this group becomes a candidate for busiest; see
7362 * update_sd_pick_busiest(). And calculate_imbalance() and
7363 * find_busiest_group() avoid some of the usual balance conditions to allow it
7364 * to create an effective group imbalance.
7366 * This is a somewhat tricky proposition since the next run might not find the
7367 * group imbalance and decide the groups need to be balanced again. A most
7368 * subtle and fragile situation.
7371 static inline int sg_imbalanced(struct sched_group *group)
7373 return group->sgc->imbalance;
7377 * group_has_capacity returns true if the group has spare capacity that could
7378 * be used by some tasks.
7379 * We consider that a group has spare capacity if the * number of task is
7380 * smaller than the number of CPUs or if the utilization is lower than the
7381 * available capacity for CFS tasks.
7382 * For the latter, we use a threshold to stabilize the state, to take into
7383 * account the variance of the tasks' load and to return true if the available
7384 * capacity in meaningful for the load balancer.
7385 * As an example, an available capacity of 1% can appear but it doesn't make
7386 * any benefit for the load balance.
7389 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7391 if (sgs->sum_nr_running < sgs->group_weight)
7394 if ((sgs->group_capacity * 100) >
7395 (sgs->group_util * env->sd->imbalance_pct))
7402 * group_is_overloaded returns true if the group has more tasks than it can
7404 * group_is_overloaded is not equals to !group_has_capacity because a group
7405 * with the exact right number of tasks, has no more spare capacity but is not
7406 * overloaded so both group_has_capacity and group_is_overloaded return
7410 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7412 if (sgs->sum_nr_running <= sgs->group_weight)
7415 if ((sgs->group_capacity * 100) <
7416 (sgs->group_util * env->sd->imbalance_pct))
7423 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7424 * per-CPU capacity than sched_group ref.
7427 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7429 return sg->sgc->min_capacity * capacity_margin <
7430 ref->sgc->min_capacity * 1024;
7434 group_type group_classify(struct sched_group *group,
7435 struct sg_lb_stats *sgs)
7437 if (sgs->group_no_capacity)
7438 return group_overloaded;
7440 if (sg_imbalanced(group))
7441 return group_imbalanced;
7447 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7448 * @env: The load balancing environment.
7449 * @group: sched_group whose statistics are to be updated.
7450 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7451 * @local_group: Does group contain this_cpu.
7452 * @sgs: variable to hold the statistics for this group.
7453 * @overload: Indicate more than one runnable task for any CPU.
7455 static inline void update_sg_lb_stats(struct lb_env *env,
7456 struct sched_group *group, int load_idx,
7457 int local_group, struct sg_lb_stats *sgs,
7463 memset(sgs, 0, sizeof(*sgs));
7465 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7466 struct rq *rq = cpu_rq(i);
7468 /* Bias balancing toward cpus of our domain */
7470 load = target_load(i, load_idx);
7472 load = source_load(i, load_idx);
7474 sgs->group_load += load;
7475 sgs->group_util += cpu_util(i);
7476 sgs->sum_nr_running += rq->cfs.h_nr_running;
7478 nr_running = rq->nr_running;
7482 #ifdef CONFIG_NUMA_BALANCING
7483 sgs->nr_numa_running += rq->nr_numa_running;
7484 sgs->nr_preferred_running += rq->nr_preferred_running;
7486 sgs->sum_weighted_load += weighted_cpuload(rq);
7488 * No need to call idle_cpu() if nr_running is not 0
7490 if (!nr_running && idle_cpu(i))
7494 /* Adjust by relative CPU capacity of the group */
7495 sgs->group_capacity = group->sgc->capacity;
7496 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7498 if (sgs->sum_nr_running)
7499 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7501 sgs->group_weight = group->group_weight;
7503 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7504 sgs->group_type = group_classify(group, sgs);
7508 * update_sd_pick_busiest - return 1 on busiest group
7509 * @env: The load balancing environment.
7510 * @sds: sched_domain statistics
7511 * @sg: sched_group candidate to be checked for being the busiest
7512 * @sgs: sched_group statistics
7514 * Determine if @sg is a busier group than the previously selected
7517 * Return: %true if @sg is a busier group than the previously selected
7518 * busiest group. %false otherwise.
7520 static bool update_sd_pick_busiest(struct lb_env *env,
7521 struct sd_lb_stats *sds,
7522 struct sched_group *sg,
7523 struct sg_lb_stats *sgs)
7525 struct sg_lb_stats *busiest = &sds->busiest_stat;
7527 if (sgs->group_type > busiest->group_type)
7530 if (sgs->group_type < busiest->group_type)
7533 if (sgs->avg_load <= busiest->avg_load)
7536 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7540 * Candidate sg has no more than one task per CPU and
7541 * has higher per-CPU capacity. Migrating tasks to less
7542 * capable CPUs may harm throughput. Maximize throughput,
7543 * power/energy consequences are not considered.
7545 if (sgs->sum_nr_running <= sgs->group_weight &&
7546 group_smaller_cpu_capacity(sds->local, sg))
7550 /* This is the busiest node in its class. */
7551 if (!(env->sd->flags & SD_ASYM_PACKING))
7554 /* No ASYM_PACKING if target cpu is already busy */
7555 if (env->idle == CPU_NOT_IDLE)
7558 * ASYM_PACKING needs to move all the work to the highest
7559 * prority CPUs in the group, therefore mark all groups
7560 * of lower priority than ourself as busy.
7562 if (sgs->sum_nr_running &&
7563 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7567 /* Prefer to move from lowest priority cpu's work */
7568 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7569 sg->asym_prefer_cpu))
7576 #ifdef CONFIG_NUMA_BALANCING
7577 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7579 if (sgs->sum_nr_running > sgs->nr_numa_running)
7581 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7586 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7588 if (rq->nr_running > rq->nr_numa_running)
7590 if (rq->nr_running > rq->nr_preferred_running)
7595 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7600 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7604 #endif /* CONFIG_NUMA_BALANCING */
7607 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7608 * @env: The load balancing environment.
7609 * @sds: variable to hold the statistics for this sched_domain.
7611 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7613 struct sched_domain *child = env->sd->child;
7614 struct sched_group *sg = env->sd->groups;
7615 struct sg_lb_stats *local = &sds->local_stat;
7616 struct sg_lb_stats tmp_sgs;
7617 int load_idx, prefer_sibling = 0;
7618 bool overload = false;
7620 if (child && child->flags & SD_PREFER_SIBLING)
7623 load_idx = get_sd_load_idx(env->sd, env->idle);
7626 struct sg_lb_stats *sgs = &tmp_sgs;
7629 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7634 if (env->idle != CPU_NEWLY_IDLE ||
7635 time_after_eq(jiffies, sg->sgc->next_update))
7636 update_group_capacity(env->sd, env->dst_cpu);
7639 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7646 * In case the child domain prefers tasks go to siblings
7647 * first, lower the sg capacity so that we'll try
7648 * and move all the excess tasks away. We lower the capacity
7649 * of a group only if the local group has the capacity to fit
7650 * these excess tasks. The extra check prevents the case where
7651 * you always pull from the heaviest group when it is already
7652 * under-utilized (possible with a large weight task outweighs
7653 * the tasks on the system).
7655 if (prefer_sibling && sds->local &&
7656 group_has_capacity(env, local) &&
7657 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7658 sgs->group_no_capacity = 1;
7659 sgs->group_type = group_classify(sg, sgs);
7662 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7664 sds->busiest_stat = *sgs;
7668 /* Now, start updating sd_lb_stats */
7669 sds->total_running += sgs->sum_nr_running;
7670 sds->total_load += sgs->group_load;
7671 sds->total_capacity += sgs->group_capacity;
7674 } while (sg != env->sd->groups);
7676 if (env->sd->flags & SD_NUMA)
7677 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7679 if (!env->sd->parent) {
7680 /* update overload indicator if we are at root domain */
7681 if (env->dst_rq->rd->overload != overload)
7682 env->dst_rq->rd->overload = overload;
7687 * check_asym_packing - Check to see if the group is packed into the
7690 * This is primarily intended to used at the sibling level. Some
7691 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7692 * case of POWER7, it can move to lower SMT modes only when higher
7693 * threads are idle. When in lower SMT modes, the threads will
7694 * perform better since they share less core resources. Hence when we
7695 * have idle threads, we want them to be the higher ones.
7697 * This packing function is run on idle threads. It checks to see if
7698 * the busiest CPU in this domain (core in the P7 case) has a higher
7699 * CPU number than the packing function is being run on. Here we are
7700 * assuming lower CPU number will be equivalent to lower a SMT thread
7703 * Return: 1 when packing is required and a task should be moved to
7704 * this CPU. The amount of the imbalance is returned in env->imbalance.
7706 * @env: The load balancing environment.
7707 * @sds: Statistics of the sched_domain which is to be packed
7709 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7713 if (!(env->sd->flags & SD_ASYM_PACKING))
7716 if (env->idle == CPU_NOT_IDLE)
7722 busiest_cpu = sds->busiest->asym_prefer_cpu;
7723 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7726 env->imbalance = DIV_ROUND_CLOSEST(
7727 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7728 SCHED_CAPACITY_SCALE);
7734 * fix_small_imbalance - Calculate the minor imbalance that exists
7735 * amongst the groups of a sched_domain, during
7737 * @env: The load balancing environment.
7738 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7741 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7743 unsigned long tmp, capa_now = 0, capa_move = 0;
7744 unsigned int imbn = 2;
7745 unsigned long scaled_busy_load_per_task;
7746 struct sg_lb_stats *local, *busiest;
7748 local = &sds->local_stat;
7749 busiest = &sds->busiest_stat;
7751 if (!local->sum_nr_running)
7752 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7753 else if (busiest->load_per_task > local->load_per_task)
7756 scaled_busy_load_per_task =
7757 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7758 busiest->group_capacity;
7760 if (busiest->avg_load + scaled_busy_load_per_task >=
7761 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7762 env->imbalance = busiest->load_per_task;
7767 * OK, we don't have enough imbalance to justify moving tasks,
7768 * however we may be able to increase total CPU capacity used by
7772 capa_now += busiest->group_capacity *
7773 min(busiest->load_per_task, busiest->avg_load);
7774 capa_now += local->group_capacity *
7775 min(local->load_per_task, local->avg_load);
7776 capa_now /= SCHED_CAPACITY_SCALE;
7778 /* Amount of load we'd subtract */
7779 if (busiest->avg_load > scaled_busy_load_per_task) {
7780 capa_move += busiest->group_capacity *
7781 min(busiest->load_per_task,
7782 busiest->avg_load - scaled_busy_load_per_task);
7785 /* Amount of load we'd add */
7786 if (busiest->avg_load * busiest->group_capacity <
7787 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7788 tmp = (busiest->avg_load * busiest->group_capacity) /
7789 local->group_capacity;
7791 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7792 local->group_capacity;
7794 capa_move += local->group_capacity *
7795 min(local->load_per_task, local->avg_load + tmp);
7796 capa_move /= SCHED_CAPACITY_SCALE;
7798 /* Move if we gain throughput */
7799 if (capa_move > capa_now)
7800 env->imbalance = busiest->load_per_task;
7804 * calculate_imbalance - Calculate the amount of imbalance present within the
7805 * groups of a given sched_domain during load balance.
7806 * @env: load balance environment
7807 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7809 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7811 unsigned long max_pull, load_above_capacity = ~0UL;
7812 struct sg_lb_stats *local, *busiest;
7814 local = &sds->local_stat;
7815 busiest = &sds->busiest_stat;
7817 if (busiest->group_type == group_imbalanced) {
7819 * In the group_imb case we cannot rely on group-wide averages
7820 * to ensure cpu-load equilibrium, look at wider averages. XXX
7822 busiest->load_per_task =
7823 min(busiest->load_per_task, sds->avg_load);
7827 * Avg load of busiest sg can be less and avg load of local sg can
7828 * be greater than avg load across all sgs of sd because avg load
7829 * factors in sg capacity and sgs with smaller group_type are
7830 * skipped when updating the busiest sg:
7832 if (busiest->avg_load <= sds->avg_load ||
7833 local->avg_load >= sds->avg_load) {
7835 return fix_small_imbalance(env, sds);
7839 * If there aren't any idle cpus, avoid creating some.
7841 if (busiest->group_type == group_overloaded &&
7842 local->group_type == group_overloaded) {
7843 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7844 if (load_above_capacity > busiest->group_capacity) {
7845 load_above_capacity -= busiest->group_capacity;
7846 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7847 load_above_capacity /= busiest->group_capacity;
7849 load_above_capacity = ~0UL;
7853 * We're trying to get all the cpus to the average_load, so we don't
7854 * want to push ourselves above the average load, nor do we wish to
7855 * reduce the max loaded cpu below the average load. At the same time,
7856 * we also don't want to reduce the group load below the group
7857 * capacity. Thus we look for the minimum possible imbalance.
7859 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7861 /* How much load to actually move to equalise the imbalance */
7862 env->imbalance = min(
7863 max_pull * busiest->group_capacity,
7864 (sds->avg_load - local->avg_load) * local->group_capacity
7865 ) / SCHED_CAPACITY_SCALE;
7868 * if *imbalance is less than the average load per runnable task
7869 * there is no guarantee that any tasks will be moved so we'll have
7870 * a think about bumping its value to force at least one task to be
7873 if (env->imbalance < busiest->load_per_task)
7874 return fix_small_imbalance(env, sds);
7877 /******* find_busiest_group() helpers end here *********************/
7880 * find_busiest_group - Returns the busiest group within the sched_domain
7881 * if there is an imbalance.
7883 * Also calculates the amount of weighted load which should be moved
7884 * to restore balance.
7886 * @env: The load balancing environment.
7888 * Return: - The busiest group if imbalance exists.
7890 static struct sched_group *find_busiest_group(struct lb_env *env)
7892 struct sg_lb_stats *local, *busiest;
7893 struct sd_lb_stats sds;
7895 init_sd_lb_stats(&sds);
7898 * Compute the various statistics relavent for load balancing at
7901 update_sd_lb_stats(env, &sds);
7902 local = &sds.local_stat;
7903 busiest = &sds.busiest_stat;
7905 /* ASYM feature bypasses nice load balance check */
7906 if (check_asym_packing(env, &sds))
7909 /* There is no busy sibling group to pull tasks from */
7910 if (!sds.busiest || busiest->sum_nr_running == 0)
7913 /* XXX broken for overlapping NUMA groups */
7914 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7915 / sds.total_capacity;
7918 * If the busiest group is imbalanced the below checks don't
7919 * work because they assume all things are equal, which typically
7920 * isn't true due to cpus_allowed constraints and the like.
7922 if (busiest->group_type == group_imbalanced)
7925 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7926 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7927 busiest->group_no_capacity)
7931 * If the local group is busier than the selected busiest group
7932 * don't try and pull any tasks.
7934 if (local->avg_load >= busiest->avg_load)
7938 * Don't pull any tasks if this group is already above the domain
7941 if (local->avg_load >= sds.avg_load)
7944 if (env->idle == CPU_IDLE) {
7946 * This cpu is idle. If the busiest group is not overloaded
7947 * and there is no imbalance between this and busiest group
7948 * wrt idle cpus, it is balanced. The imbalance becomes
7949 * significant if the diff is greater than 1 otherwise we
7950 * might end up to just move the imbalance on another group
7952 if ((busiest->group_type != group_overloaded) &&
7953 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7957 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7958 * imbalance_pct to be conservative.
7960 if (100 * busiest->avg_load <=
7961 env->sd->imbalance_pct * local->avg_load)
7966 /* Looks like there is an imbalance. Compute it */
7967 calculate_imbalance(env, &sds);
7976 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7978 static struct rq *find_busiest_queue(struct lb_env *env,
7979 struct sched_group *group)
7981 struct rq *busiest = NULL, *rq;
7982 unsigned long busiest_load = 0, busiest_capacity = 1;
7985 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7986 unsigned long capacity, wl;
7990 rt = fbq_classify_rq(rq);
7993 * We classify groups/runqueues into three groups:
7994 * - regular: there are !numa tasks
7995 * - remote: there are numa tasks that run on the 'wrong' node
7996 * - all: there is no distinction
7998 * In order to avoid migrating ideally placed numa tasks,
7999 * ignore those when there's better options.
8001 * If we ignore the actual busiest queue to migrate another
8002 * task, the next balance pass can still reduce the busiest
8003 * queue by moving tasks around inside the node.
8005 * If we cannot move enough load due to this classification
8006 * the next pass will adjust the group classification and
8007 * allow migration of more tasks.
8009 * Both cases only affect the total convergence complexity.
8011 if (rt > env->fbq_type)
8014 capacity = capacity_of(i);
8016 wl = weighted_cpuload(rq);
8019 * When comparing with imbalance, use weighted_cpuload()
8020 * which is not scaled with the cpu capacity.
8023 if (rq->nr_running == 1 && wl > env->imbalance &&
8024 !check_cpu_capacity(rq, env->sd))
8028 * For the load comparisons with the other cpu's, consider
8029 * the weighted_cpuload() scaled with the cpu capacity, so
8030 * that the load can be moved away from the cpu that is
8031 * potentially running at a lower capacity.
8033 * Thus we're looking for max(wl_i / capacity_i), crosswise
8034 * multiplication to rid ourselves of the division works out
8035 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8036 * our previous maximum.
8038 if (wl * busiest_capacity > busiest_load * capacity) {
8040 busiest_capacity = capacity;
8049 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8050 * so long as it is large enough.
8052 #define MAX_PINNED_INTERVAL 512
8054 static int need_active_balance(struct lb_env *env)
8056 struct sched_domain *sd = env->sd;
8058 if (env->idle == CPU_NEWLY_IDLE) {
8061 * ASYM_PACKING needs to force migrate tasks from busy but
8062 * lower priority CPUs in order to pack all tasks in the
8063 * highest priority CPUs.
8065 if ((sd->flags & SD_ASYM_PACKING) &&
8066 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8071 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8072 * It's worth migrating the task if the src_cpu's capacity is reduced
8073 * because of other sched_class or IRQs if more capacity stays
8074 * available on dst_cpu.
8076 if ((env->idle != CPU_NOT_IDLE) &&
8077 (env->src_rq->cfs.h_nr_running == 1)) {
8078 if ((check_cpu_capacity(env->src_rq, sd)) &&
8079 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8083 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8086 static int active_load_balance_cpu_stop(void *data);
8088 static int should_we_balance(struct lb_env *env)
8090 struct sched_group *sg = env->sd->groups;
8091 int cpu, balance_cpu = -1;
8094 * Ensure the balancing environment is consistent; can happen
8095 * when the softirq triggers 'during' hotplug.
8097 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8101 * In the newly idle case, we will allow all the cpu's
8102 * to do the newly idle load balance.
8104 if (env->idle == CPU_NEWLY_IDLE)
8107 /* Try to find first idle cpu */
8108 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8116 if (balance_cpu == -1)
8117 balance_cpu = group_balance_cpu(sg);
8120 * First idle cpu or the first cpu(busiest) in this sched group
8121 * is eligible for doing load balancing at this and above domains.
8123 return balance_cpu == env->dst_cpu;
8127 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8128 * tasks if there is an imbalance.
8130 static int load_balance(int this_cpu, struct rq *this_rq,
8131 struct sched_domain *sd, enum cpu_idle_type idle,
8132 int *continue_balancing)
8134 int ld_moved, cur_ld_moved, active_balance = 0;
8135 struct sched_domain *sd_parent = sd->parent;
8136 struct sched_group *group;
8139 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8141 struct lb_env env = {
8143 .dst_cpu = this_cpu,
8145 .dst_grpmask = sched_group_span(sd->groups),
8147 .loop_break = sched_nr_migrate_break,
8150 .tasks = LIST_HEAD_INIT(env.tasks),
8153 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8155 schedstat_inc(sd->lb_count[idle]);
8158 if (!should_we_balance(&env)) {
8159 *continue_balancing = 0;
8163 group = find_busiest_group(&env);
8165 schedstat_inc(sd->lb_nobusyg[idle]);
8169 busiest = find_busiest_queue(&env, group);
8171 schedstat_inc(sd->lb_nobusyq[idle]);
8175 BUG_ON(busiest == env.dst_rq);
8177 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8179 env.src_cpu = busiest->cpu;
8180 env.src_rq = busiest;
8183 if (busiest->nr_running > 1) {
8185 * Attempt to move tasks. If find_busiest_group has found
8186 * an imbalance but busiest->nr_running <= 1, the group is
8187 * still unbalanced. ld_moved simply stays zero, so it is
8188 * correctly treated as an imbalance.
8190 env.flags |= LBF_ALL_PINNED;
8191 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8194 rq_lock_irqsave(busiest, &rf);
8195 update_rq_clock(busiest);
8198 * cur_ld_moved - load moved in current iteration
8199 * ld_moved - cumulative load moved across iterations
8201 cur_ld_moved = detach_tasks(&env);
8204 * We've detached some tasks from busiest_rq. Every
8205 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8206 * unlock busiest->lock, and we are able to be sure
8207 * that nobody can manipulate the tasks in parallel.
8208 * See task_rq_lock() family for the details.
8211 rq_unlock(busiest, &rf);
8215 ld_moved += cur_ld_moved;
8218 local_irq_restore(rf.flags);
8220 if (env.flags & LBF_NEED_BREAK) {
8221 env.flags &= ~LBF_NEED_BREAK;
8226 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8227 * us and move them to an alternate dst_cpu in our sched_group
8228 * where they can run. The upper limit on how many times we
8229 * iterate on same src_cpu is dependent on number of cpus in our
8232 * This changes load balance semantics a bit on who can move
8233 * load to a given_cpu. In addition to the given_cpu itself
8234 * (or a ilb_cpu acting on its behalf where given_cpu is
8235 * nohz-idle), we now have balance_cpu in a position to move
8236 * load to given_cpu. In rare situations, this may cause
8237 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8238 * _independently_ and at _same_ time to move some load to
8239 * given_cpu) causing exceess load to be moved to given_cpu.
8240 * This however should not happen so much in practice and
8241 * moreover subsequent load balance cycles should correct the
8242 * excess load moved.
8244 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8246 /* Prevent to re-select dst_cpu via env's cpus */
8247 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8249 env.dst_rq = cpu_rq(env.new_dst_cpu);
8250 env.dst_cpu = env.new_dst_cpu;
8251 env.flags &= ~LBF_DST_PINNED;
8253 env.loop_break = sched_nr_migrate_break;
8256 * Go back to "more_balance" rather than "redo" since we
8257 * need to continue with same src_cpu.
8263 * We failed to reach balance because of affinity.
8266 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8268 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8269 *group_imbalance = 1;
8272 /* All tasks on this runqueue were pinned by CPU affinity */
8273 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8274 cpumask_clear_cpu(cpu_of(busiest), cpus);
8276 * Attempting to continue load balancing at the current
8277 * sched_domain level only makes sense if there are
8278 * active CPUs remaining as possible busiest CPUs to
8279 * pull load from which are not contained within the
8280 * destination group that is receiving any migrated
8283 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8285 env.loop_break = sched_nr_migrate_break;
8288 goto out_all_pinned;
8293 schedstat_inc(sd->lb_failed[idle]);
8295 * Increment the failure counter only on periodic balance.
8296 * We do not want newidle balance, which can be very
8297 * frequent, pollute the failure counter causing
8298 * excessive cache_hot migrations and active balances.
8300 if (idle != CPU_NEWLY_IDLE)
8301 sd->nr_balance_failed++;
8303 if (need_active_balance(&env)) {
8304 unsigned long flags;
8306 raw_spin_lock_irqsave(&busiest->lock, flags);
8308 /* don't kick the active_load_balance_cpu_stop,
8309 * if the curr task on busiest cpu can't be
8312 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8313 raw_spin_unlock_irqrestore(&busiest->lock,
8315 env.flags |= LBF_ALL_PINNED;
8316 goto out_one_pinned;
8320 * ->active_balance synchronizes accesses to
8321 * ->active_balance_work. Once set, it's cleared
8322 * only after active load balance is finished.
8324 if (!busiest->active_balance) {
8325 busiest->active_balance = 1;
8326 busiest->push_cpu = this_cpu;
8329 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8331 if (active_balance) {
8332 stop_one_cpu_nowait(cpu_of(busiest),
8333 active_load_balance_cpu_stop, busiest,
8334 &busiest->active_balance_work);
8337 /* We've kicked active balancing, force task migration. */
8338 sd->nr_balance_failed = sd->cache_nice_tries+1;
8341 sd->nr_balance_failed = 0;
8343 if (likely(!active_balance)) {
8344 /* We were unbalanced, so reset the balancing interval */
8345 sd->balance_interval = sd->min_interval;
8348 * If we've begun active balancing, start to back off. This
8349 * case may not be covered by the all_pinned logic if there
8350 * is only 1 task on the busy runqueue (because we don't call
8353 if (sd->balance_interval < sd->max_interval)
8354 sd->balance_interval *= 2;
8361 * We reach balance although we may have faced some affinity
8362 * constraints. Clear the imbalance flag only if other tasks got
8363 * a chance to move and fix the imbalance.
8365 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
8366 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8368 if (*group_imbalance)
8369 *group_imbalance = 0;
8374 * We reach balance because all tasks are pinned at this level so
8375 * we can't migrate them. Let the imbalance flag set so parent level
8376 * can try to migrate them.
8378 schedstat_inc(sd->lb_balanced[idle]);
8380 sd->nr_balance_failed = 0;
8386 * idle_balance() disregards balance intervals, so we could repeatedly
8387 * reach this code, which would lead to balance_interval skyrocketting
8388 * in a short amount of time. Skip the balance_interval increase logic
8391 if (env.idle == CPU_NEWLY_IDLE)
8394 /* tune up the balancing interval */
8395 if (((env.flags & LBF_ALL_PINNED) &&
8396 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8397 (sd->balance_interval < sd->max_interval))
8398 sd->balance_interval *= 2;
8403 static inline unsigned long
8404 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8406 unsigned long interval = sd->balance_interval;
8409 interval *= sd->busy_factor;
8411 /* scale ms to jiffies */
8412 interval = msecs_to_jiffies(interval);
8413 interval = clamp(interval, 1UL, max_load_balance_interval);
8419 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8421 unsigned long interval, next;
8423 /* used by idle balance, so cpu_busy = 0 */
8424 interval = get_sd_balance_interval(sd, 0);
8425 next = sd->last_balance + interval;
8427 if (time_after(*next_balance, next))
8428 *next_balance = next;
8432 * idle_balance is called by schedule() if this_cpu is about to become
8433 * idle. Attempts to pull tasks from other CPUs.
8435 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8437 unsigned long next_balance = jiffies + HZ;
8438 int this_cpu = this_rq->cpu;
8439 struct sched_domain *sd;
8440 int pulled_task = 0;
8444 * We must set idle_stamp _before_ calling idle_balance(), such that we
8445 * measure the duration of idle_balance() as idle time.
8447 this_rq->idle_stamp = rq_clock(this_rq);
8450 * Do not pull tasks towards !active CPUs...
8452 if (!cpu_active(this_cpu))
8456 * This is OK, because current is on_cpu, which avoids it being picked
8457 * for load-balance and preemption/IRQs are still disabled avoiding
8458 * further scheduler activity on it and we're being very careful to
8459 * re-start the picking loop.
8461 rq_unpin_lock(this_rq, rf);
8463 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8464 !this_rq->rd->overload) {
8466 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8468 update_next_balance(sd, &next_balance);
8474 raw_spin_unlock(&this_rq->lock);
8476 update_blocked_averages(this_cpu);
8478 for_each_domain(this_cpu, sd) {
8479 int continue_balancing = 1;
8480 u64 t0, domain_cost;
8482 if (!(sd->flags & SD_LOAD_BALANCE))
8485 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8486 update_next_balance(sd, &next_balance);
8490 if (sd->flags & SD_BALANCE_NEWIDLE) {
8491 t0 = sched_clock_cpu(this_cpu);
8493 pulled_task = load_balance(this_cpu, this_rq,
8495 &continue_balancing);
8497 domain_cost = sched_clock_cpu(this_cpu) - t0;
8498 if (domain_cost > sd->max_newidle_lb_cost)
8499 sd->max_newidle_lb_cost = domain_cost;
8501 curr_cost += domain_cost;
8504 update_next_balance(sd, &next_balance);
8507 * Stop searching for tasks to pull if there are
8508 * now runnable tasks on this rq.
8510 if (pulled_task || this_rq->nr_running > 0)
8515 raw_spin_lock(&this_rq->lock);
8517 if (curr_cost > this_rq->max_idle_balance_cost)
8518 this_rq->max_idle_balance_cost = curr_cost;
8521 * While browsing the domains, we released the rq lock, a task could
8522 * have been enqueued in the meantime. Since we're not going idle,
8523 * pretend we pulled a task.
8525 if (this_rq->cfs.h_nr_running && !pulled_task)
8529 /* Move the next balance forward */
8530 if (time_after(this_rq->next_balance, next_balance))
8531 this_rq->next_balance = next_balance;
8533 /* Is there a task of a high priority class? */
8534 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8538 this_rq->idle_stamp = 0;
8540 rq_repin_lock(this_rq, rf);
8546 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8547 * running tasks off the busiest CPU onto idle CPUs. It requires at
8548 * least 1 task to be running on each physical CPU where possible, and
8549 * avoids physical / logical imbalances.
8551 static int active_load_balance_cpu_stop(void *data)
8553 struct rq *busiest_rq = data;
8554 int busiest_cpu = cpu_of(busiest_rq);
8555 int target_cpu = busiest_rq->push_cpu;
8556 struct rq *target_rq = cpu_rq(target_cpu);
8557 struct sched_domain *sd;
8558 struct task_struct *p = NULL;
8561 rq_lock_irq(busiest_rq, &rf);
8563 * Between queueing the stop-work and running it is a hole in which
8564 * CPUs can become inactive. We should not move tasks from or to
8567 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8570 /* make sure the requested cpu hasn't gone down in the meantime */
8571 if (unlikely(busiest_cpu != smp_processor_id() ||
8572 !busiest_rq->active_balance))
8575 /* Is there any task to move? */
8576 if (busiest_rq->nr_running <= 1)
8580 * This condition is "impossible", if it occurs
8581 * we need to fix it. Originally reported by
8582 * Bjorn Helgaas on a 128-cpu setup.
8584 BUG_ON(busiest_rq == target_rq);
8586 /* Search for an sd spanning us and the target CPU. */
8588 for_each_domain(target_cpu, sd) {
8589 if ((sd->flags & SD_LOAD_BALANCE) &&
8590 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8595 struct lb_env env = {
8597 .dst_cpu = target_cpu,
8598 .dst_rq = target_rq,
8599 .src_cpu = busiest_rq->cpu,
8600 .src_rq = busiest_rq,
8603 * can_migrate_task() doesn't need to compute new_dst_cpu
8604 * for active balancing. Since we have CPU_IDLE, but no
8605 * @dst_grpmask we need to make that test go away with lying
8608 .flags = LBF_DST_PINNED,
8611 schedstat_inc(sd->alb_count);
8612 update_rq_clock(busiest_rq);
8614 p = detach_one_task(&env);
8616 schedstat_inc(sd->alb_pushed);
8617 /* Active balancing done, reset the failure counter. */
8618 sd->nr_balance_failed = 0;
8620 schedstat_inc(sd->alb_failed);
8625 busiest_rq->active_balance = 0;
8626 rq_unlock(busiest_rq, &rf);
8629 attach_one_task(target_rq, p);
8636 static inline int on_null_domain(struct rq *rq)
8638 return unlikely(!rcu_dereference_sched(rq->sd));
8641 #ifdef CONFIG_NO_HZ_COMMON
8643 * idle load balancing details
8644 * - When one of the busy CPUs notice that there may be an idle rebalancing
8645 * needed, they will kick the idle load balancer, which then does idle
8646 * load balancing for all the idle CPUs.
8649 cpumask_var_t idle_cpus_mask;
8651 unsigned long next_balance; /* in jiffy units */
8652 } nohz ____cacheline_aligned;
8654 static inline int find_new_ilb(void)
8656 int ilb = cpumask_first(nohz.idle_cpus_mask);
8658 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8665 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8666 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8667 * CPU (if there is one).
8669 static void nohz_balancer_kick(void)
8673 nohz.next_balance++;
8675 ilb_cpu = find_new_ilb();
8677 if (ilb_cpu >= nr_cpu_ids)
8680 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8683 * Use smp_send_reschedule() instead of resched_cpu().
8684 * This way we generate a sched IPI on the target cpu which
8685 * is idle. And the softirq performing nohz idle load balance
8686 * will be run before returning from the IPI.
8688 smp_send_reschedule(ilb_cpu);
8692 void nohz_balance_exit_idle(unsigned int cpu)
8694 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8696 * Completely isolated CPUs don't ever set, so we must test.
8698 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8699 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8700 atomic_dec(&nohz.nr_cpus);
8702 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8706 static inline void set_cpu_sd_state_busy(void)
8708 struct sched_domain *sd;
8709 int cpu = smp_processor_id();
8712 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8714 if (!sd || !sd->nohz_idle)
8718 atomic_inc(&sd->shared->nr_busy_cpus);
8723 void set_cpu_sd_state_idle(void)
8725 struct sched_domain *sd;
8726 int cpu = smp_processor_id();
8729 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8731 if (!sd || sd->nohz_idle)
8735 atomic_dec(&sd->shared->nr_busy_cpus);
8741 * This routine will record that the cpu is going idle with tick stopped.
8742 * This info will be used in performing idle load balancing in the future.
8744 void nohz_balance_enter_idle(int cpu)
8747 * If this cpu is going down, then nothing needs to be done.
8749 if (!cpu_active(cpu))
8752 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8753 if (!is_housekeeping_cpu(cpu))
8756 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8760 * If we're a completely isolated CPU, we don't play.
8762 if (on_null_domain(cpu_rq(cpu)))
8765 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8766 atomic_inc(&nohz.nr_cpus);
8767 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8771 static DEFINE_SPINLOCK(balancing);
8774 * Scale the max load_balance interval with the number of CPUs in the system.
8775 * This trades load-balance latency on larger machines for less cross talk.
8777 void update_max_interval(void)
8779 max_load_balance_interval = HZ*num_online_cpus()/10;
8783 * It checks each scheduling domain to see if it is due to be balanced,
8784 * and initiates a balancing operation if so.
8786 * Balancing parameters are set up in init_sched_domains.
8788 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8790 int continue_balancing = 1;
8792 unsigned long interval;
8793 struct sched_domain *sd;
8794 /* Earliest time when we have to do rebalance again */
8795 unsigned long next_balance = jiffies + 60*HZ;
8796 int update_next_balance = 0;
8797 int need_serialize, need_decay = 0;
8800 update_blocked_averages(cpu);
8803 for_each_domain(cpu, sd) {
8805 * Decay the newidle max times here because this is a regular
8806 * visit to all the domains. Decay ~1% per second.
8808 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8809 sd->max_newidle_lb_cost =
8810 (sd->max_newidle_lb_cost * 253) / 256;
8811 sd->next_decay_max_lb_cost = jiffies + HZ;
8814 max_cost += sd->max_newidle_lb_cost;
8816 if (!(sd->flags & SD_LOAD_BALANCE))
8820 * Stop the load balance at this level. There is another
8821 * CPU in our sched group which is doing load balancing more
8824 if (!continue_balancing) {
8830 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8832 need_serialize = sd->flags & SD_SERIALIZE;
8833 if (need_serialize) {
8834 if (!spin_trylock(&balancing))
8838 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8839 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8841 * The LBF_DST_PINNED logic could have changed
8842 * env->dst_cpu, so we can't know our idle
8843 * state even if we migrated tasks. Update it.
8845 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8847 sd->last_balance = jiffies;
8848 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8851 spin_unlock(&balancing);
8853 if (time_after(next_balance, sd->last_balance + interval)) {
8854 next_balance = sd->last_balance + interval;
8855 update_next_balance = 1;
8860 * Ensure the rq-wide value also decays but keep it at a
8861 * reasonable floor to avoid funnies with rq->avg_idle.
8863 rq->max_idle_balance_cost =
8864 max((u64)sysctl_sched_migration_cost, max_cost);
8869 * next_balance will be updated only when there is a need.
8870 * When the cpu is attached to null domain for ex, it will not be
8873 if (likely(update_next_balance)) {
8874 rq->next_balance = next_balance;
8876 #ifdef CONFIG_NO_HZ_COMMON
8878 * If this CPU has been elected to perform the nohz idle
8879 * balance. Other idle CPUs have already rebalanced with
8880 * nohz_idle_balance() and nohz.next_balance has been
8881 * updated accordingly. This CPU is now running the idle load
8882 * balance for itself and we need to update the
8883 * nohz.next_balance accordingly.
8885 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8886 nohz.next_balance = rq->next_balance;
8891 #ifdef CONFIG_NO_HZ_COMMON
8893 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8894 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8896 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8898 int this_cpu = this_rq->cpu;
8901 /* Earliest time when we have to do rebalance again */
8902 unsigned long next_balance = jiffies + 60*HZ;
8903 int update_next_balance = 0;
8905 if (idle != CPU_IDLE ||
8906 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8909 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8910 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8914 * If this cpu gets work to do, stop the load balancing
8915 * work being done for other cpus. Next load
8916 * balancing owner will pick it up.
8921 rq = cpu_rq(balance_cpu);
8924 * If time for next balance is due,
8927 if (time_after_eq(jiffies, rq->next_balance)) {
8930 rq_lock_irq(rq, &rf);
8931 update_rq_clock(rq);
8932 cpu_load_update_idle(rq);
8933 rq_unlock_irq(rq, &rf);
8935 rebalance_domains(rq, CPU_IDLE);
8938 if (time_after(next_balance, rq->next_balance)) {
8939 next_balance = rq->next_balance;
8940 update_next_balance = 1;
8945 * next_balance will be updated only when there is a need.
8946 * When the CPU is attached to null domain for ex, it will not be
8949 if (likely(update_next_balance))
8950 nohz.next_balance = next_balance;
8952 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8956 * Current heuristic for kicking the idle load balancer in the presence
8957 * of an idle cpu in the system.
8958 * - This rq has more than one task.
8959 * - This rq has at least one CFS task and the capacity of the CPU is
8960 * significantly reduced because of RT tasks or IRQs.
8961 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8962 * multiple busy cpu.
8963 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8964 * domain span are idle.
8966 static inline bool nohz_kick_needed(struct rq *rq)
8968 unsigned long now = jiffies;
8969 struct sched_domain_shared *sds;
8970 struct sched_domain *sd;
8971 int nr_busy, i, cpu = rq->cpu;
8974 if (unlikely(rq->idle_balance))
8978 * We may be recently in ticked or tickless idle mode. At the first
8979 * busy tick after returning from idle, we will update the busy stats.
8981 set_cpu_sd_state_busy();
8982 nohz_balance_exit_idle(cpu);
8985 * None are in tickless mode and hence no need for NOHZ idle load
8988 if (likely(!atomic_read(&nohz.nr_cpus)))
8991 if (time_before(now, nohz.next_balance))
8994 if (rq->nr_running >= 2)
8998 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9001 * XXX: write a coherent comment on why we do this.
9002 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9004 nr_busy = atomic_read(&sds->nr_busy_cpus);
9012 sd = rcu_dereference(rq->sd);
9014 if ((rq->cfs.h_nr_running >= 1) &&
9015 check_cpu_capacity(rq, sd)) {
9021 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9023 for_each_cpu(i, sched_domain_span(sd)) {
9025 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9028 if (sched_asym_prefer(i, cpu)) {
9039 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9043 * run_rebalance_domains is triggered when needed from the scheduler tick.
9044 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9046 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9048 struct rq *this_rq = this_rq();
9049 enum cpu_idle_type idle = this_rq->idle_balance ?
9050 CPU_IDLE : CPU_NOT_IDLE;
9053 * If this cpu has a pending nohz_balance_kick, then do the
9054 * balancing on behalf of the other idle cpus whose ticks are
9055 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9056 * give the idle cpus a chance to load balance. Else we may
9057 * load balance only within the local sched_domain hierarchy
9058 * and abort nohz_idle_balance altogether if we pull some load.
9060 nohz_idle_balance(this_rq, idle);
9061 rebalance_domains(this_rq, idle);
9065 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9067 void trigger_load_balance(struct rq *rq)
9069 /* Don't need to rebalance while attached to NULL domain */
9070 if (unlikely(on_null_domain(rq)))
9073 if (time_after_eq(jiffies, rq->next_balance))
9074 raise_softirq(SCHED_SOFTIRQ);
9075 #ifdef CONFIG_NO_HZ_COMMON
9076 if (nohz_kick_needed(rq))
9077 nohz_balancer_kick();
9081 static void rq_online_fair(struct rq *rq)
9085 update_runtime_enabled(rq);
9088 static void rq_offline_fair(struct rq *rq)
9092 /* Ensure any throttled groups are reachable by pick_next_task */
9093 unthrottle_offline_cfs_rqs(rq);
9096 #endif /* CONFIG_SMP */
9099 * scheduler tick hitting a task of our scheduling class:
9101 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9103 struct cfs_rq *cfs_rq;
9104 struct sched_entity *se = &curr->se;
9106 for_each_sched_entity(se) {
9107 cfs_rq = cfs_rq_of(se);
9108 entity_tick(cfs_rq, se, queued);
9111 if (static_branch_unlikely(&sched_numa_balancing))
9112 task_tick_numa(rq, curr);
9116 * called on fork with the child task as argument from the parent's context
9117 * - child not yet on the tasklist
9118 * - preemption disabled
9120 static void task_fork_fair(struct task_struct *p)
9122 struct cfs_rq *cfs_rq;
9123 struct sched_entity *se = &p->se, *curr;
9124 struct rq *rq = this_rq();
9128 update_rq_clock(rq);
9130 cfs_rq = task_cfs_rq(current);
9131 curr = cfs_rq->curr;
9133 update_curr(cfs_rq);
9134 se->vruntime = curr->vruntime;
9136 place_entity(cfs_rq, se, 1);
9138 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9140 * Upon rescheduling, sched_class::put_prev_task() will place
9141 * 'current' within the tree based on its new key value.
9143 swap(curr->vruntime, se->vruntime);
9147 se->vruntime -= cfs_rq->min_vruntime;
9152 * Priority of the task has changed. Check to see if we preempt
9156 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9158 if (!task_on_rq_queued(p))
9162 * Reschedule if we are currently running on this runqueue and
9163 * our priority decreased, or if we are not currently running on
9164 * this runqueue and our priority is higher than the current's
9166 if (rq->curr == p) {
9167 if (p->prio > oldprio)
9170 check_preempt_curr(rq, p, 0);
9173 static inline bool vruntime_normalized(struct task_struct *p)
9175 struct sched_entity *se = &p->se;
9178 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9179 * the dequeue_entity(.flags=0) will already have normalized the
9186 * When !on_rq, vruntime of the task has usually NOT been normalized.
9187 * But there are some cases where it has already been normalized:
9189 * - A forked child which is waiting for being woken up by
9190 * wake_up_new_task().
9191 * - A task which has been woken up by try_to_wake_up() and
9192 * waiting for actually being woken up by sched_ttwu_pending().
9194 if (!se->sum_exec_runtime ||
9195 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9201 #ifdef CONFIG_FAIR_GROUP_SCHED
9203 * Propagate the changes of the sched_entity across the tg tree to make it
9204 * visible to the root
9206 static void propagate_entity_cfs_rq(struct sched_entity *se)
9208 struct cfs_rq *cfs_rq;
9210 /* Start to propagate at parent */
9213 for_each_sched_entity(se) {
9214 cfs_rq = cfs_rq_of(se);
9216 if (cfs_rq_throttled(cfs_rq))
9219 update_load_avg(se, UPDATE_TG);
9223 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9226 static void detach_entity_cfs_rq(struct sched_entity *se)
9228 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9230 /* Catch up with the cfs_rq and remove our load when we leave */
9231 update_load_avg(se, 0);
9232 detach_entity_load_avg(cfs_rq, se);
9233 update_tg_load_avg(cfs_rq, false);
9234 propagate_entity_cfs_rq(se);
9237 static void attach_entity_cfs_rq(struct sched_entity *se)
9239 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9241 #ifdef CONFIG_FAIR_GROUP_SCHED
9243 * Since the real-depth could have been changed (only FAIR
9244 * class maintain depth value), reset depth properly.
9246 se->depth = se->parent ? se->parent->depth + 1 : 0;
9249 /* Synchronize entity with its cfs_rq */
9250 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9251 attach_entity_load_avg(cfs_rq, se);
9252 update_tg_load_avg(cfs_rq, false);
9253 propagate_entity_cfs_rq(se);
9256 static void detach_task_cfs_rq(struct task_struct *p)
9258 struct sched_entity *se = &p->se;
9259 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9261 if (!vruntime_normalized(p)) {
9263 * Fix up our vruntime so that the current sleep doesn't
9264 * cause 'unlimited' sleep bonus.
9266 place_entity(cfs_rq, se, 0);
9267 se->vruntime -= cfs_rq->min_vruntime;
9270 detach_entity_cfs_rq(se);
9273 static void attach_task_cfs_rq(struct task_struct *p)
9275 struct sched_entity *se = &p->se;
9276 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9278 attach_entity_cfs_rq(se);
9280 if (!vruntime_normalized(p))
9281 se->vruntime += cfs_rq->min_vruntime;
9284 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9286 detach_task_cfs_rq(p);
9289 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9291 attach_task_cfs_rq(p);
9293 if (task_on_rq_queued(p)) {
9295 * We were most likely switched from sched_rt, so
9296 * kick off the schedule if running, otherwise just see
9297 * if we can still preempt the current task.
9302 check_preempt_curr(rq, p, 0);
9306 /* Account for a task changing its policy or group.
9308 * This routine is mostly called to set cfs_rq->curr field when a task
9309 * migrates between groups/classes.
9311 static void set_curr_task_fair(struct rq *rq)
9313 struct sched_entity *se = &rq->curr->se;
9315 for_each_sched_entity(se) {
9316 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9318 set_next_entity(cfs_rq, se);
9319 /* ensure bandwidth has been allocated on our new cfs_rq */
9320 account_cfs_rq_runtime(cfs_rq, 0);
9324 void init_cfs_rq(struct cfs_rq *cfs_rq)
9326 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9327 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9328 #ifndef CONFIG_64BIT
9329 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9332 #ifdef CONFIG_FAIR_GROUP_SCHED
9333 cfs_rq->propagate_avg = 0;
9335 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9336 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9340 #ifdef CONFIG_FAIR_GROUP_SCHED
9341 static void task_set_group_fair(struct task_struct *p)
9343 struct sched_entity *se = &p->se;
9345 set_task_rq(p, task_cpu(p));
9346 se->depth = se->parent ? se->parent->depth + 1 : 0;
9349 static void task_move_group_fair(struct task_struct *p)
9351 detach_task_cfs_rq(p);
9352 set_task_rq(p, task_cpu(p));
9355 /* Tell se's cfs_rq has been changed -- migrated */
9356 p->se.avg.last_update_time = 0;
9358 attach_task_cfs_rq(p);
9361 static void task_change_group_fair(struct task_struct *p, int type)
9364 case TASK_SET_GROUP:
9365 task_set_group_fair(p);
9368 case TASK_MOVE_GROUP:
9369 task_move_group_fair(p);
9374 void free_fair_sched_group(struct task_group *tg)
9378 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9380 for_each_possible_cpu(i) {
9382 kfree(tg->cfs_rq[i]);
9391 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9393 struct sched_entity *se;
9394 struct cfs_rq *cfs_rq;
9397 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9400 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9404 tg->shares = NICE_0_LOAD;
9406 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9408 for_each_possible_cpu(i) {
9409 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9410 GFP_KERNEL, cpu_to_node(i));
9414 se = kzalloc_node(sizeof(struct sched_entity),
9415 GFP_KERNEL, cpu_to_node(i));
9419 init_cfs_rq(cfs_rq);
9420 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9421 init_entity_runnable_average(se);
9432 void online_fair_sched_group(struct task_group *tg)
9434 struct sched_entity *se;
9439 for_each_possible_cpu(i) {
9442 rq_lock_irq(rq, &rf);
9443 update_rq_clock(rq);
9444 attach_entity_cfs_rq(se);
9445 sync_throttle(tg, i);
9446 rq_unlock_irq(rq, &rf);
9450 void unregister_fair_sched_group(struct task_group *tg)
9452 unsigned long flags;
9456 for_each_possible_cpu(cpu) {
9458 remove_entity_load_avg(tg->se[cpu]);
9461 * Only empty task groups can be destroyed; so we can speculatively
9462 * check on_list without danger of it being re-added.
9464 if (!tg->cfs_rq[cpu]->on_list)
9469 raw_spin_lock_irqsave(&rq->lock, flags);
9470 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9471 raw_spin_unlock_irqrestore(&rq->lock, flags);
9475 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9476 struct sched_entity *se, int cpu,
9477 struct sched_entity *parent)
9479 struct rq *rq = cpu_rq(cpu);
9483 init_cfs_rq_runtime(cfs_rq);
9485 tg->cfs_rq[cpu] = cfs_rq;
9488 /* se could be NULL for root_task_group */
9493 se->cfs_rq = &rq->cfs;
9496 se->cfs_rq = parent->my_q;
9497 se->depth = parent->depth + 1;
9501 /* guarantee group entities always have weight */
9502 update_load_set(&se->load, NICE_0_LOAD);
9503 se->parent = parent;
9506 static DEFINE_MUTEX(shares_mutex);
9508 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9513 * We can't change the weight of the root cgroup.
9518 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9520 mutex_lock(&shares_mutex);
9521 if (tg->shares == shares)
9524 tg->shares = shares;
9525 for_each_possible_cpu(i) {
9526 struct rq *rq = cpu_rq(i);
9527 struct sched_entity *se = tg->se[i];
9530 /* Propagate contribution to hierarchy */
9531 rq_lock_irqsave(rq, &rf);
9532 update_rq_clock(rq);
9533 for_each_sched_entity(se) {
9534 update_load_avg(se, UPDATE_TG);
9535 update_cfs_shares(se);
9537 rq_unlock_irqrestore(rq, &rf);
9541 mutex_unlock(&shares_mutex);
9544 #else /* CONFIG_FAIR_GROUP_SCHED */
9546 void free_fair_sched_group(struct task_group *tg) { }
9548 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9553 void online_fair_sched_group(struct task_group *tg) { }
9555 void unregister_fair_sched_group(struct task_group *tg) { }
9557 #endif /* CONFIG_FAIR_GROUP_SCHED */
9560 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9562 struct sched_entity *se = &task->se;
9563 unsigned int rr_interval = 0;
9566 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9569 if (rq->cfs.load.weight)
9570 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9576 * All the scheduling class methods:
9578 const struct sched_class fair_sched_class = {
9579 .next = &idle_sched_class,
9580 .enqueue_task = enqueue_task_fair,
9581 .dequeue_task = dequeue_task_fair,
9582 .yield_task = yield_task_fair,
9583 .yield_to_task = yield_to_task_fair,
9585 .check_preempt_curr = check_preempt_wakeup,
9587 .pick_next_task = pick_next_task_fair,
9588 .put_prev_task = put_prev_task_fair,
9591 .select_task_rq = select_task_rq_fair,
9592 .migrate_task_rq = migrate_task_rq_fair,
9594 .rq_online = rq_online_fair,
9595 .rq_offline = rq_offline_fair,
9597 .task_dead = task_dead_fair,
9598 .set_cpus_allowed = set_cpus_allowed_common,
9601 .set_curr_task = set_curr_task_fair,
9602 .task_tick = task_tick_fair,
9603 .task_fork = task_fork_fair,
9605 .prio_changed = prio_changed_fair,
9606 .switched_from = switched_from_fair,
9607 .switched_to = switched_to_fair,
9609 .get_rr_interval = get_rr_interval_fair,
9611 .update_curr = update_curr_fair,
9613 #ifdef CONFIG_FAIR_GROUP_SCHED
9614 .task_change_group = task_change_group_fair,
9618 #ifdef CONFIG_SCHED_DEBUG
9619 void print_cfs_stats(struct seq_file *m, int cpu)
9621 struct cfs_rq *cfs_rq;
9624 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9625 print_cfs_rq(m, cpu, cfs_rq);
9629 #ifdef CONFIG_NUMA_BALANCING
9630 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9633 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9635 for_each_online_node(node) {
9636 if (p->numa_faults) {
9637 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9638 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9640 if (p->numa_group) {
9641 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9642 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9644 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9647 #endif /* CONFIG_NUMA_BALANCING */
9648 #endif /* CONFIG_SCHED_DEBUG */
9650 __init void init_sched_fair_class(void)
9653 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9655 #ifdef CONFIG_NO_HZ_COMMON
9656 nohz.next_balance = jiffies;
9657 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);