2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched/mm.h>
24 #include <linux/sched/topology.h>
26 #include <linux/latencytop.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpuidle.h>
29 #include <linux/slab.h>
30 #include <linux/profile.h>
31 #include <linux/interrupt.h>
32 #include <linux/mempolicy.h>
33 #include <linux/migrate.h>
34 #include <linux/task_work.h>
36 #include <trace/events/sched.h>
41 * Targeted preemption latency for CPU-bound tasks:
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
57 * The initial- and re-scaling of tunables is configurable
61 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
62 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
65 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
67 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
70 * Minimal preemption granularity for CPU-bound tasks:
72 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
74 unsigned int sysctl_sched_min_granularity = 750000ULL;
75 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
78 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
80 static unsigned int sched_nr_latency = 8;
83 * After fork, child runs first. If set to 0 (default) then
84 * parent will (try to) run first.
86 unsigned int sysctl_sched_child_runs_first __read_mostly;
89 * SCHED_OTHER wake-up granularity.
91 * This option delays the preemption effects of decoupled workloads
92 * and reduces their over-scheduling. Synchronous workloads will still
93 * have immediate wakeup/sleep latencies.
95 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
97 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
100 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
104 * For asym packing, by default the lower numbered cpu has higher priority.
106 int __weak arch_asym_cpu_priority(int cpu)
112 #ifdef CONFIG_CFS_BANDWIDTH
114 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
115 * each time a cfs_rq requests quota.
117 * Note: in the case that the slice exceeds the runtime remaining (either due
118 * to consumption or the quota being specified to be smaller than the slice)
119 * we will always only issue the remaining available time.
121 * (default: 5 msec, units: microseconds)
123 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
127 * The margin used when comparing utilization with CPU capacity:
128 * util * margin < capacity * 1024
132 unsigned int capacity_margin = 1280;
134 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
140 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
146 static inline void update_load_set(struct load_weight *lw, unsigned long w)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling) {
167 case SCHED_TUNABLESCALING_NONE:
170 case SCHED_TUNABLESCALING_LINEAR:
173 case SCHED_TUNABLESCALING_LOG:
175 factor = 1 + ilog2(cpus);
182 static void update_sysctl(void)
184 unsigned int factor = get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity);
189 SET_SYSCTL(sched_latency);
190 SET_SYSCTL(sched_wakeup_granularity);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight *lw)
206 if (likely(lw->inv_weight))
209 w = scale_load_down(lw->weight);
211 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
213 else if (unlikely(!w))
214 lw->inv_weight = WMULT_CONST;
216 lw->inv_weight = WMULT_CONST / w;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
233 u64 fact = scale_load_down(weight);
234 int shift = WMULT_SHIFT;
236 __update_inv_weight(lw);
238 if (unlikely(fact >> 32)) {
245 /* hint to use a 32x32->64 mul */
246 fact = (u64)(u32)fact * lw->inv_weight;
253 return mul_u64_u32_shr(delta_exec, fact, shift);
257 const struct sched_class fair_sched_class;
259 /**************************************************************
260 * CFS operations on generic schedulable entities:
263 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* cpu runqueue to which this cfs_rq is attached */
266 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
271 /* An entity is a task if it doesn't "own" a runqueue */
272 #define entity_is_task(se) (!se->my_q)
274 static inline struct task_struct *task_of(struct sched_entity *se)
276 SCHED_WARN_ON(!entity_is_task(se));
277 return container_of(se, struct task_struct, se);
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
289 /* runqueue on which this entity is (to be) queued */
290 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
295 /* runqueue "owned" by this group */
296 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
301 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
303 if (!cfs_rq->on_list) {
304 struct rq *rq = rq_of(cfs_rq);
305 int cpu = cpu_of(rq);
307 * Ensure we either appear before our parent (if already
308 * enqueued) or force our parent to appear after us when it is
309 * enqueued. The fact that we always enqueue bottom-up
310 * reduces this to two cases and a special case for the root
311 * cfs_rq. Furthermore, it also means that we will always reset
312 * tmp_alone_branch either when the branch is connected
313 * to a tree or when we reach the beg of the tree
315 if (cfs_rq->tg->parent &&
316 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
318 * If parent is already on the list, we add the child
319 * just before. Thanks to circular linked property of
320 * the list, this means to put the child at the tail
321 * of the list that starts by parent.
323 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
324 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
326 * The branch is now connected to its tree so we can
327 * reset tmp_alone_branch to the beginning of the
330 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
331 } else if (!cfs_rq->tg->parent) {
333 * cfs rq without parent should be put
334 * at the tail of the list.
336 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
337 &rq->leaf_cfs_rq_list);
339 * We have reach the beg of a tree so we can reset
340 * tmp_alone_branch to the beginning of the list.
342 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the beg of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
351 rq->tmp_alone_branch);
353 * update tmp_alone_branch to points to the new beg
356 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
363 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
365 if (cfs_rq->on_list) {
366 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->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_safe(rq, cfs_rq, pos) \
468 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
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 memset(sa, 0, sizeof(*sa));
722 * Tasks are intialized with full load to be seen as heavy tasks until
723 * they get a chance to stabilize to their real load level.
724 * Group entities are intialized with zero load to reflect the fact that
725 * nothing has been attached to the task group yet.
727 if (entity_is_task(se))
728 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
730 se->runnable_weight = se->load.weight;
732 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
735 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
736 static void attach_entity_cfs_rq(struct sched_entity *se);
739 * With new tasks being created, their initial util_avgs are extrapolated
740 * based on the cfs_rq's current util_avg:
742 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
744 * However, in many cases, the above util_avg does not give a desired
745 * value. Moreover, the sum of the util_avgs may be divergent, such
746 * as when the series is a harmonic series.
748 * To solve this problem, we also cap the util_avg of successive tasks to
749 * only 1/2 of the left utilization budget:
751 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
753 * where n denotes the nth task.
755 * For example, a simplest series from the beginning would be like:
757 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
758 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
760 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
761 * if util_avg > util_avg_cap.
763 void post_init_entity_util_avg(struct sched_entity *se)
765 struct cfs_rq *cfs_rq = cfs_rq_of(se);
766 struct sched_avg *sa = &se->avg;
767 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
770 if (cfs_rq->avg.util_avg != 0) {
771 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
772 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
774 if (sa->util_avg > cap)
781 if (entity_is_task(se)) {
782 struct task_struct *p = task_of(se);
783 if (p->sched_class != &fair_sched_class) {
785 * For !fair tasks do:
787 update_cfs_rq_load_avg(now, cfs_rq);
788 attach_entity_load_avg(cfs_rq, se);
789 switched_from_fair(rq, p);
791 * such that the next switched_to_fair() has the
794 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
799 attach_entity_cfs_rq(se);
802 #else /* !CONFIG_SMP */
803 void init_entity_runnable_average(struct sched_entity *se)
806 void post_init_entity_util_avg(struct sched_entity *se)
809 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
812 #endif /* CONFIG_SMP */
815 * Update the current task's runtime statistics.
817 static void update_curr(struct cfs_rq *cfs_rq)
819 struct sched_entity *curr = cfs_rq->curr;
820 u64 now = rq_clock_task(rq_of(cfs_rq));
826 delta_exec = now - curr->exec_start;
827 if (unlikely((s64)delta_exec <= 0))
830 curr->exec_start = now;
832 schedstat_set(curr->statistics.exec_max,
833 max(delta_exec, curr->statistics.exec_max));
835 curr->sum_exec_runtime += delta_exec;
836 schedstat_add(cfs_rq->exec_clock, delta_exec);
838 curr->vruntime += calc_delta_fair(delta_exec, curr);
839 update_min_vruntime(cfs_rq);
841 if (entity_is_task(curr)) {
842 struct task_struct *curtask = task_of(curr);
844 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
845 cpuacct_charge(curtask, delta_exec);
846 account_group_exec_runtime(curtask, delta_exec);
849 account_cfs_rq_runtime(cfs_rq, delta_exec);
852 static void update_curr_fair(struct rq *rq)
854 update_curr(cfs_rq_of(&rq->curr->se));
858 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
860 u64 wait_start, prev_wait_start;
862 if (!schedstat_enabled())
865 wait_start = rq_clock(rq_of(cfs_rq));
866 prev_wait_start = schedstat_val(se->statistics.wait_start);
868 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
869 likely(wait_start > prev_wait_start))
870 wait_start -= prev_wait_start;
872 schedstat_set(se->statistics.wait_start, wait_start);
876 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
878 struct task_struct *p;
881 if (!schedstat_enabled())
884 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
886 if (entity_is_task(se)) {
888 if (task_on_rq_migrating(p)) {
890 * Preserve migrating task's wait time so wait_start
891 * time stamp can be adjusted to accumulate wait time
892 * prior to migration.
894 schedstat_set(se->statistics.wait_start, delta);
897 trace_sched_stat_wait(p, delta);
900 schedstat_set(se->statistics.wait_max,
901 max(schedstat_val(se->statistics.wait_max), delta));
902 schedstat_inc(se->statistics.wait_count);
903 schedstat_add(se->statistics.wait_sum, delta);
904 schedstat_set(se->statistics.wait_start, 0);
908 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
910 struct task_struct *tsk = NULL;
911 u64 sleep_start, block_start;
913 if (!schedstat_enabled())
916 sleep_start = schedstat_val(se->statistics.sleep_start);
917 block_start = schedstat_val(se->statistics.block_start);
919 if (entity_is_task(se))
923 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
928 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
929 schedstat_set(se->statistics.sleep_max, delta);
931 schedstat_set(se->statistics.sleep_start, 0);
932 schedstat_add(se->statistics.sum_sleep_runtime, delta);
935 account_scheduler_latency(tsk, delta >> 10, 1);
936 trace_sched_stat_sleep(tsk, delta);
940 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
945 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
946 schedstat_set(se->statistics.block_max, delta);
948 schedstat_set(se->statistics.block_start, 0);
949 schedstat_add(se->statistics.sum_sleep_runtime, delta);
952 if (tsk->in_iowait) {
953 schedstat_add(se->statistics.iowait_sum, delta);
954 schedstat_inc(se->statistics.iowait_count);
955 trace_sched_stat_iowait(tsk, delta);
958 trace_sched_stat_blocked(tsk, delta);
961 * Blocking time is in units of nanosecs, so shift by
962 * 20 to get a milliseconds-range estimation of the
963 * amount of time that the task spent sleeping:
965 if (unlikely(prof_on == SLEEP_PROFILING)) {
966 profile_hits(SLEEP_PROFILING,
967 (void *)get_wchan(tsk),
970 account_scheduler_latency(tsk, delta >> 10, 0);
976 * Task is being enqueued - update stats:
979 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
981 if (!schedstat_enabled())
985 * Are we enqueueing a waiting task? (for current tasks
986 * a dequeue/enqueue event is a NOP)
988 if (se != cfs_rq->curr)
989 update_stats_wait_start(cfs_rq, se);
991 if (flags & ENQUEUE_WAKEUP)
992 update_stats_enqueue_sleeper(cfs_rq, se);
996 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
999 if (!schedstat_enabled())
1003 * Mark the end of the wait period if dequeueing a
1006 if (se != cfs_rq->curr)
1007 update_stats_wait_end(cfs_rq, se);
1009 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1010 struct task_struct *tsk = task_of(se);
1012 if (tsk->state & TASK_INTERRUPTIBLE)
1013 schedstat_set(se->statistics.sleep_start,
1014 rq_clock(rq_of(cfs_rq)));
1015 if (tsk->state & TASK_UNINTERRUPTIBLE)
1016 schedstat_set(se->statistics.block_start,
1017 rq_clock(rq_of(cfs_rq)));
1022 * We are picking a new current task - update its stats:
1025 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1028 * We are starting a new run period:
1030 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1033 /**************************************************
1034 * Scheduling class queueing methods:
1037 #ifdef CONFIG_NUMA_BALANCING
1039 * Approximate time to scan a full NUMA task in ms. The task scan period is
1040 * calculated based on the tasks virtual memory size and
1041 * numa_balancing_scan_size.
1043 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1044 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1046 /* Portion of address space to scan in MB */
1047 unsigned int sysctl_numa_balancing_scan_size = 256;
1049 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1050 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1055 spinlock_t lock; /* nr_tasks, tasks */
1060 struct rcu_head rcu;
1061 unsigned long total_faults;
1062 unsigned long max_faults_cpu;
1064 * Faults_cpu is used to decide whether memory should move
1065 * towards the CPU. As a consequence, these stats are weighted
1066 * more by CPU use than by memory faults.
1068 unsigned long *faults_cpu;
1069 unsigned long faults[0];
1072 static inline unsigned long group_faults_priv(struct numa_group *ng);
1073 static inline unsigned long group_faults_shared(struct numa_group *ng);
1075 static unsigned int task_nr_scan_windows(struct task_struct *p)
1077 unsigned long rss = 0;
1078 unsigned long nr_scan_pages;
1081 * Calculations based on RSS as non-present and empty pages are skipped
1082 * by the PTE scanner and NUMA hinting faults should be trapped based
1085 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1086 rss = get_mm_rss(p->mm);
1088 rss = nr_scan_pages;
1090 rss = round_up(rss, nr_scan_pages);
1091 return rss / nr_scan_pages;
1094 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1095 #define MAX_SCAN_WINDOW 2560
1097 static unsigned int task_scan_min(struct task_struct *p)
1099 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1100 unsigned int scan, floor;
1101 unsigned int windows = 1;
1103 if (scan_size < MAX_SCAN_WINDOW)
1104 windows = MAX_SCAN_WINDOW / scan_size;
1105 floor = 1000 / windows;
1107 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1108 return max_t(unsigned int, floor, scan);
1111 static unsigned int task_scan_start(struct task_struct *p)
1113 unsigned long smin = task_scan_min(p);
1114 unsigned long period = smin;
1116 /* Scale the maximum scan period with the amount of shared memory. */
1117 if (p->numa_group) {
1118 struct numa_group *ng = p->numa_group;
1119 unsigned long shared = group_faults_shared(ng);
1120 unsigned long private = group_faults_priv(ng);
1122 period *= atomic_read(&ng->refcount);
1123 period *= shared + 1;
1124 period /= private + shared + 1;
1127 return max(smin, period);
1130 static unsigned int task_scan_max(struct task_struct *p)
1132 unsigned long smin = task_scan_min(p);
1135 /* Watch for min being lower than max due to floor calculations */
1136 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1138 /* Scale the maximum scan period with the amount of shared memory. */
1139 if (p->numa_group) {
1140 struct numa_group *ng = p->numa_group;
1141 unsigned long shared = group_faults_shared(ng);
1142 unsigned long private = group_faults_priv(ng);
1143 unsigned long period = smax;
1145 period *= atomic_read(&ng->refcount);
1146 period *= shared + 1;
1147 period /= private + shared + 1;
1149 smax = max(smax, period);
1152 return max(smin, smax);
1155 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1157 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1158 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1161 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1163 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1164 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1167 /* Shared or private faults. */
1168 #define NR_NUMA_HINT_FAULT_TYPES 2
1170 /* Memory and CPU locality */
1171 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1173 /* Averaged statistics, and temporary buffers. */
1174 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1176 pid_t task_numa_group_id(struct task_struct *p)
1178 return p->numa_group ? p->numa_group->gid : 0;
1182 * The averaged statistics, shared & private, memory & cpu,
1183 * occupy the first half of the array. The second half of the
1184 * array is for current counters, which are averaged into the
1185 * first set by task_numa_placement.
1187 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1189 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1192 static inline unsigned long task_faults(struct task_struct *p, int nid)
1194 if (!p->numa_faults)
1197 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1198 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1201 static inline unsigned long group_faults(struct task_struct *p, int nid)
1206 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1207 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1210 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1212 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1213 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1216 static inline unsigned long group_faults_priv(struct numa_group *ng)
1218 unsigned long faults = 0;
1221 for_each_online_node(node) {
1222 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1228 static inline unsigned long group_faults_shared(struct numa_group *ng)
1230 unsigned long faults = 0;
1233 for_each_online_node(node) {
1234 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1241 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1242 * considered part of a numa group's pseudo-interleaving set. Migrations
1243 * between these nodes are slowed down, to allow things to settle down.
1245 #define ACTIVE_NODE_FRACTION 3
1247 static bool numa_is_active_node(int nid, struct numa_group *ng)
1249 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1252 /* Handle placement on systems where not all nodes are directly connected. */
1253 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1254 int maxdist, bool task)
1256 unsigned long score = 0;
1260 * All nodes are directly connected, and the same distance
1261 * from each other. No need for fancy placement algorithms.
1263 if (sched_numa_topology_type == NUMA_DIRECT)
1267 * This code is called for each node, introducing N^2 complexity,
1268 * which should be ok given the number of nodes rarely exceeds 8.
1270 for_each_online_node(node) {
1271 unsigned long faults;
1272 int dist = node_distance(nid, node);
1275 * The furthest away nodes in the system are not interesting
1276 * for placement; nid was already counted.
1278 if (dist == sched_max_numa_distance || node == nid)
1282 * On systems with a backplane NUMA topology, compare groups
1283 * of nodes, and move tasks towards the group with the most
1284 * memory accesses. When comparing two nodes at distance
1285 * "hoplimit", only nodes closer by than "hoplimit" are part
1286 * of each group. Skip other nodes.
1288 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1292 /* Add up the faults from nearby nodes. */
1294 faults = task_faults(p, node);
1296 faults = group_faults(p, node);
1299 * On systems with a glueless mesh NUMA topology, there are
1300 * no fixed "groups of nodes". Instead, nodes that are not
1301 * directly connected bounce traffic through intermediate
1302 * nodes; a numa_group can occupy any set of nodes.
1303 * The further away a node is, the less the faults count.
1304 * This seems to result in good task placement.
1306 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1307 faults *= (sched_max_numa_distance - dist);
1308 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1318 * These return the fraction of accesses done by a particular task, or
1319 * task group, on a particular numa node. The group weight is given a
1320 * larger multiplier, in order to group tasks together that are almost
1321 * evenly spread out between numa nodes.
1323 static inline unsigned long task_weight(struct task_struct *p, int nid,
1326 unsigned long faults, total_faults;
1328 if (!p->numa_faults)
1331 total_faults = p->total_numa_faults;
1336 faults = task_faults(p, nid);
1337 faults += score_nearby_nodes(p, nid, dist, true);
1339 return 1000 * faults / total_faults;
1342 static inline unsigned long group_weight(struct task_struct *p, int nid,
1345 unsigned long faults, total_faults;
1350 total_faults = p->numa_group->total_faults;
1355 faults = group_faults(p, nid);
1356 faults += score_nearby_nodes(p, nid, dist, false);
1358 return 1000 * faults / total_faults;
1361 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1362 int src_nid, int dst_cpu)
1364 struct numa_group *ng = p->numa_group;
1365 int dst_nid = cpu_to_node(dst_cpu);
1366 int last_cpupid, this_cpupid;
1368 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1371 * Multi-stage node selection is used in conjunction with a periodic
1372 * migration fault to build a temporal task<->page relation. By using
1373 * a two-stage filter we remove short/unlikely relations.
1375 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1376 * a task's usage of a particular page (n_p) per total usage of this
1377 * page (n_t) (in a given time-span) to a probability.
1379 * Our periodic faults will sample this probability and getting the
1380 * same result twice in a row, given these samples are fully
1381 * independent, is then given by P(n)^2, provided our sample period
1382 * is sufficiently short compared to the usage pattern.
1384 * This quadric squishes small probabilities, making it less likely we
1385 * act on an unlikely task<->page relation.
1387 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1388 if (!cpupid_pid_unset(last_cpupid) &&
1389 cpupid_to_nid(last_cpupid) != dst_nid)
1392 /* Always allow migrate on private faults */
1393 if (cpupid_match_pid(p, last_cpupid))
1396 /* A shared fault, but p->numa_group has not been set up yet. */
1401 * Destination node is much more heavily used than the source
1402 * node? Allow migration.
1404 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1405 ACTIVE_NODE_FRACTION)
1409 * Distribute memory according to CPU & memory use on each node,
1410 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1412 * faults_cpu(dst) 3 faults_cpu(src)
1413 * --------------- * - > ---------------
1414 * faults_mem(dst) 4 faults_mem(src)
1416 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1417 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1420 static unsigned long weighted_cpuload(struct rq *rq);
1421 static unsigned long source_load(int cpu, int type);
1422 static unsigned long target_load(int cpu, int type);
1423 static unsigned long capacity_of(int cpu);
1425 /* Cached statistics for all CPUs within a node */
1427 unsigned long nr_running;
1430 /* Total compute capacity of CPUs on a node */
1431 unsigned long compute_capacity;
1433 /* Approximate capacity in terms of runnable tasks on a node */
1434 unsigned long task_capacity;
1435 int has_free_capacity;
1439 * XXX borrowed from update_sg_lb_stats
1441 static void update_numa_stats(struct numa_stats *ns, int nid)
1443 int smt, cpu, cpus = 0;
1444 unsigned long capacity;
1446 memset(ns, 0, sizeof(*ns));
1447 for_each_cpu(cpu, cpumask_of_node(nid)) {
1448 struct rq *rq = cpu_rq(cpu);
1450 ns->nr_running += rq->nr_running;
1451 ns->load += weighted_cpuload(rq);
1452 ns->compute_capacity += capacity_of(cpu);
1458 * If we raced with hotplug and there are no CPUs left in our mask
1459 * the @ns structure is NULL'ed and task_numa_compare() will
1460 * not find this node attractive.
1462 * We'll either bail at !has_free_capacity, or we'll detect a huge
1463 * imbalance and bail there.
1468 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1469 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1470 capacity = cpus / smt; /* cores */
1472 ns->task_capacity = min_t(unsigned, capacity,
1473 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1474 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1477 struct task_numa_env {
1478 struct task_struct *p;
1480 int src_cpu, src_nid;
1481 int dst_cpu, dst_nid;
1483 struct numa_stats src_stats, dst_stats;
1488 struct task_struct *best_task;
1493 static void task_numa_assign(struct task_numa_env *env,
1494 struct task_struct *p, long imp)
1497 put_task_struct(env->best_task);
1502 env->best_imp = imp;
1503 env->best_cpu = env->dst_cpu;
1506 static bool load_too_imbalanced(long src_load, long dst_load,
1507 struct task_numa_env *env)
1510 long orig_src_load, orig_dst_load;
1511 long src_capacity, dst_capacity;
1514 * The load is corrected for the CPU capacity available on each node.
1517 * ------------ vs ---------
1518 * src_capacity dst_capacity
1520 src_capacity = env->src_stats.compute_capacity;
1521 dst_capacity = env->dst_stats.compute_capacity;
1523 /* We care about the slope of the imbalance, not the direction. */
1524 if (dst_load < src_load)
1525 swap(dst_load, src_load);
1527 /* Is the difference below the threshold? */
1528 imb = dst_load * src_capacity * 100 -
1529 src_load * dst_capacity * env->imbalance_pct;
1534 * The imbalance is above the allowed threshold.
1535 * Compare it with the old imbalance.
1537 orig_src_load = env->src_stats.load;
1538 orig_dst_load = env->dst_stats.load;
1540 if (orig_dst_load < orig_src_load)
1541 swap(orig_dst_load, orig_src_load);
1543 old_imb = orig_dst_load * src_capacity * 100 -
1544 orig_src_load * dst_capacity * env->imbalance_pct;
1546 /* Would this change make things worse? */
1547 return (imb > old_imb);
1551 * This checks if the overall compute and NUMA accesses of the system would
1552 * be improved if the source tasks was migrated to the target dst_cpu taking
1553 * into account that it might be best if task running on the dst_cpu should
1554 * be exchanged with the source task
1556 static void task_numa_compare(struct task_numa_env *env,
1557 long taskimp, long groupimp)
1559 struct rq *src_rq = cpu_rq(env->src_cpu);
1560 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1561 struct task_struct *cur;
1562 long src_load, dst_load;
1564 long imp = env->p->numa_group ? groupimp : taskimp;
1566 int dist = env->dist;
1569 cur = task_rcu_dereference(&dst_rq->curr);
1570 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1574 * Because we have preemption enabled we can get migrated around and
1575 * end try selecting ourselves (current == env->p) as a swap candidate.
1581 * "imp" is the fault differential for the source task between the
1582 * source and destination node. Calculate the total differential for
1583 * the source task and potential destination task. The more negative
1584 * the value is, the more rmeote accesses that would be expected to
1585 * be incurred if the tasks were swapped.
1588 /* Skip this swap candidate if cannot move to the source cpu */
1589 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1593 * If dst and source tasks are in the same NUMA group, or not
1594 * in any group then look only at task weights.
1596 if (cur->numa_group == env->p->numa_group) {
1597 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1598 task_weight(cur, env->dst_nid, dist);
1600 * Add some hysteresis to prevent swapping the
1601 * tasks within a group over tiny differences.
1603 if (cur->numa_group)
1607 * Compare the group weights. If a task is all by
1608 * itself (not part of a group), use the task weight
1611 if (cur->numa_group)
1612 imp += group_weight(cur, env->src_nid, dist) -
1613 group_weight(cur, env->dst_nid, dist);
1615 imp += task_weight(cur, env->src_nid, dist) -
1616 task_weight(cur, env->dst_nid, dist);
1620 if (imp <= env->best_imp && moveimp <= env->best_imp)
1624 /* Is there capacity at our destination? */
1625 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1626 !env->dst_stats.has_free_capacity)
1632 /* Balance doesn't matter much if we're running a task per cpu */
1633 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1634 dst_rq->nr_running == 1)
1638 * In the overloaded case, try and keep the load balanced.
1641 load = task_h_load(env->p);
1642 dst_load = env->dst_stats.load + load;
1643 src_load = env->src_stats.load - load;
1645 if (moveimp > imp && moveimp > env->best_imp) {
1647 * If the improvement from just moving env->p direction is
1648 * better than swapping tasks around, check if a move is
1649 * possible. Store a slightly smaller score than moveimp,
1650 * so an actually idle CPU will win.
1652 if (!load_too_imbalanced(src_load, dst_load, env)) {
1659 if (imp <= env->best_imp)
1663 load = task_h_load(cur);
1668 if (load_too_imbalanced(src_load, dst_load, env))
1672 * One idle CPU per node is evaluated for a task numa move.
1673 * Call select_idle_sibling to maybe find a better one.
1677 * select_idle_siblings() uses an per-cpu cpumask that
1678 * can be used from IRQ context.
1680 local_irq_disable();
1681 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1687 task_numa_assign(env, cur, imp);
1692 static void task_numa_find_cpu(struct task_numa_env *env,
1693 long taskimp, long groupimp)
1697 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1698 /* Skip this CPU if the source task cannot migrate */
1699 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1703 task_numa_compare(env, taskimp, groupimp);
1707 /* Only move tasks to a NUMA node less busy than the current node. */
1708 static bool numa_has_capacity(struct task_numa_env *env)
1710 struct numa_stats *src = &env->src_stats;
1711 struct numa_stats *dst = &env->dst_stats;
1713 if (src->has_free_capacity && !dst->has_free_capacity)
1717 * Only consider a task move if the source has a higher load
1718 * than the destination, corrected for CPU capacity on each node.
1720 * src->load dst->load
1721 * --------------------- vs ---------------------
1722 * src->compute_capacity dst->compute_capacity
1724 if (src->load * dst->compute_capacity * env->imbalance_pct >
1726 dst->load * src->compute_capacity * 100)
1732 static int task_numa_migrate(struct task_struct *p)
1734 struct task_numa_env env = {
1737 .src_cpu = task_cpu(p),
1738 .src_nid = task_node(p),
1740 .imbalance_pct = 112,
1746 struct sched_domain *sd;
1747 unsigned long taskweight, groupweight;
1749 long taskimp, groupimp;
1752 * Pick the lowest SD_NUMA domain, as that would have the smallest
1753 * imbalance and would be the first to start moving tasks about.
1755 * And we want to avoid any moving of tasks about, as that would create
1756 * random movement of tasks -- counter the numa conditions we're trying
1760 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1762 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1766 * Cpusets can break the scheduler domain tree into smaller
1767 * balance domains, some of which do not cross NUMA boundaries.
1768 * Tasks that are "trapped" in such domains cannot be migrated
1769 * elsewhere, so there is no point in (re)trying.
1771 if (unlikely(!sd)) {
1772 p->numa_preferred_nid = task_node(p);
1776 env.dst_nid = p->numa_preferred_nid;
1777 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1778 taskweight = task_weight(p, env.src_nid, dist);
1779 groupweight = group_weight(p, env.src_nid, dist);
1780 update_numa_stats(&env.src_stats, env.src_nid);
1781 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1782 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1783 update_numa_stats(&env.dst_stats, env.dst_nid);
1785 /* Try to find a spot on the preferred nid. */
1786 if (numa_has_capacity(&env))
1787 task_numa_find_cpu(&env, taskimp, groupimp);
1790 * Look at other nodes in these cases:
1791 * - there is no space available on the preferred_nid
1792 * - the task is part of a numa_group that is interleaved across
1793 * multiple NUMA nodes; in order to better consolidate the group,
1794 * we need to check other locations.
1796 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1797 for_each_online_node(nid) {
1798 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1801 dist = node_distance(env.src_nid, env.dst_nid);
1802 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1804 taskweight = task_weight(p, env.src_nid, dist);
1805 groupweight = group_weight(p, env.src_nid, dist);
1808 /* Only consider nodes where both task and groups benefit */
1809 taskimp = task_weight(p, nid, dist) - taskweight;
1810 groupimp = group_weight(p, nid, dist) - groupweight;
1811 if (taskimp < 0 && groupimp < 0)
1816 update_numa_stats(&env.dst_stats, env.dst_nid);
1817 if (numa_has_capacity(&env))
1818 task_numa_find_cpu(&env, taskimp, groupimp);
1823 * If the task is part of a workload that spans multiple NUMA nodes,
1824 * and is migrating into one of the workload's active nodes, remember
1825 * this node as the task's preferred numa node, so the workload can
1827 * A task that migrated to a second choice node will be better off
1828 * trying for a better one later. Do not set the preferred node here.
1830 if (p->numa_group) {
1831 struct numa_group *ng = p->numa_group;
1833 if (env.best_cpu == -1)
1838 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1839 sched_setnuma(p, env.dst_nid);
1842 /* No better CPU than the current one was found. */
1843 if (env.best_cpu == -1)
1847 * Reset the scan period if the task is being rescheduled on an
1848 * alternative node to recheck if the tasks is now properly placed.
1850 p->numa_scan_period = task_scan_start(p);
1852 if (env.best_task == NULL) {
1853 ret = migrate_task_to(p, env.best_cpu);
1855 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1859 ret = migrate_swap(p, env.best_task);
1861 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1862 put_task_struct(env.best_task);
1866 /* Attempt to migrate a task to a CPU on the preferred node. */
1867 static void numa_migrate_preferred(struct task_struct *p)
1869 unsigned long interval = HZ;
1871 /* This task has no NUMA fault statistics yet */
1872 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1875 /* Periodically retry migrating the task to the preferred node */
1876 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1877 p->numa_migrate_retry = jiffies + interval;
1879 /* Success if task is already running on preferred CPU */
1880 if (task_node(p) == p->numa_preferred_nid)
1883 /* Otherwise, try migrate to a CPU on the preferred node */
1884 task_numa_migrate(p);
1888 * Find out how many nodes on the workload is actively running on. Do this by
1889 * tracking the nodes from which NUMA hinting faults are triggered. This can
1890 * be different from the set of nodes where the workload's memory is currently
1893 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1895 unsigned long faults, max_faults = 0;
1896 int nid, active_nodes = 0;
1898 for_each_online_node(nid) {
1899 faults = group_faults_cpu(numa_group, nid);
1900 if (faults > max_faults)
1901 max_faults = faults;
1904 for_each_online_node(nid) {
1905 faults = group_faults_cpu(numa_group, nid);
1906 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1910 numa_group->max_faults_cpu = max_faults;
1911 numa_group->active_nodes = active_nodes;
1915 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1916 * increments. The more local the fault statistics are, the higher the scan
1917 * period will be for the next scan window. If local/(local+remote) ratio is
1918 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1919 * the scan period will decrease. Aim for 70% local accesses.
1921 #define NUMA_PERIOD_SLOTS 10
1922 #define NUMA_PERIOD_THRESHOLD 7
1925 * Increase the scan period (slow down scanning) if the majority of
1926 * our memory is already on our local node, or if the majority of
1927 * the page accesses are shared with other processes.
1928 * Otherwise, decrease the scan period.
1930 static void update_task_scan_period(struct task_struct *p,
1931 unsigned long shared, unsigned long private)
1933 unsigned int period_slot;
1934 int lr_ratio, ps_ratio;
1937 unsigned long remote = p->numa_faults_locality[0];
1938 unsigned long local = p->numa_faults_locality[1];
1941 * If there were no record hinting faults then either the task is
1942 * completely idle or all activity is areas that are not of interest
1943 * to automatic numa balancing. Related to that, if there were failed
1944 * migration then it implies we are migrating too quickly or the local
1945 * node is overloaded. In either case, scan slower
1947 if (local + shared == 0 || p->numa_faults_locality[2]) {
1948 p->numa_scan_period = min(p->numa_scan_period_max,
1949 p->numa_scan_period << 1);
1951 p->mm->numa_next_scan = jiffies +
1952 msecs_to_jiffies(p->numa_scan_period);
1958 * Prepare to scale scan period relative to the current period.
1959 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1960 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1961 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1963 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1964 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1965 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1967 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1969 * Most memory accesses are local. There is no need to
1970 * do fast NUMA scanning, since memory is already local.
1972 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1975 diff = slot * period_slot;
1976 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1978 * Most memory accesses are shared with other tasks.
1979 * There is no point in continuing fast NUMA scanning,
1980 * since other tasks may just move the memory elsewhere.
1982 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1985 diff = slot * period_slot;
1988 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1989 * yet they are not on the local NUMA node. Speed up
1990 * NUMA scanning to get the memory moved over.
1992 int ratio = max(lr_ratio, ps_ratio);
1993 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1996 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1997 task_scan_min(p), task_scan_max(p));
1998 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2002 * Get the fraction of time the task has been running since the last
2003 * NUMA placement cycle. The scheduler keeps similar statistics, but
2004 * decays those on a 32ms period, which is orders of magnitude off
2005 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2006 * stats only if the task is so new there are no NUMA statistics yet.
2008 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2010 u64 runtime, delta, now;
2011 /* Use the start of this time slice to avoid calculations. */
2012 now = p->se.exec_start;
2013 runtime = p->se.sum_exec_runtime;
2015 if (p->last_task_numa_placement) {
2016 delta = runtime - p->last_sum_exec_runtime;
2017 *period = now - p->last_task_numa_placement;
2019 delta = p->se.avg.load_sum;
2020 *period = LOAD_AVG_MAX;
2023 p->last_sum_exec_runtime = runtime;
2024 p->last_task_numa_placement = now;
2030 * Determine the preferred nid for a task in a numa_group. This needs to
2031 * be done in a way that produces consistent results with group_weight,
2032 * otherwise workloads might not converge.
2034 static int preferred_group_nid(struct task_struct *p, int nid)
2039 /* Direct connections between all NUMA nodes. */
2040 if (sched_numa_topology_type == NUMA_DIRECT)
2044 * On a system with glueless mesh NUMA topology, group_weight
2045 * scores nodes according to the number of NUMA hinting faults on
2046 * both the node itself, and on nearby nodes.
2048 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2049 unsigned long score, max_score = 0;
2050 int node, max_node = nid;
2052 dist = sched_max_numa_distance;
2054 for_each_online_node(node) {
2055 score = group_weight(p, node, dist);
2056 if (score > max_score) {
2065 * Finding the preferred nid in a system with NUMA backplane
2066 * interconnect topology is more involved. The goal is to locate
2067 * tasks from numa_groups near each other in the system, and
2068 * untangle workloads from different sides of the system. This requires
2069 * searching down the hierarchy of node groups, recursively searching
2070 * inside the highest scoring group of nodes. The nodemask tricks
2071 * keep the complexity of the search down.
2073 nodes = node_online_map;
2074 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2075 unsigned long max_faults = 0;
2076 nodemask_t max_group = NODE_MASK_NONE;
2079 /* Are there nodes at this distance from each other? */
2080 if (!find_numa_distance(dist))
2083 for_each_node_mask(a, nodes) {
2084 unsigned long faults = 0;
2085 nodemask_t this_group;
2086 nodes_clear(this_group);
2088 /* Sum group's NUMA faults; includes a==b case. */
2089 for_each_node_mask(b, nodes) {
2090 if (node_distance(a, b) < dist) {
2091 faults += group_faults(p, b);
2092 node_set(b, this_group);
2093 node_clear(b, nodes);
2097 /* Remember the top group. */
2098 if (faults > max_faults) {
2099 max_faults = faults;
2100 max_group = this_group;
2102 * subtle: at the smallest distance there is
2103 * just one node left in each "group", the
2104 * winner is the preferred nid.
2109 /* Next round, evaluate the nodes within max_group. */
2117 static void task_numa_placement(struct task_struct *p)
2119 int seq, nid, max_nid = -1, max_group_nid = -1;
2120 unsigned long max_faults = 0, max_group_faults = 0;
2121 unsigned long fault_types[2] = { 0, 0 };
2122 unsigned long total_faults;
2123 u64 runtime, period;
2124 spinlock_t *group_lock = NULL;
2127 * The p->mm->numa_scan_seq field gets updated without
2128 * exclusive access. Use READ_ONCE() here to ensure
2129 * that the field is read in a single access:
2131 seq = READ_ONCE(p->mm->numa_scan_seq);
2132 if (p->numa_scan_seq == seq)
2134 p->numa_scan_seq = seq;
2135 p->numa_scan_period_max = task_scan_max(p);
2137 total_faults = p->numa_faults_locality[0] +
2138 p->numa_faults_locality[1];
2139 runtime = numa_get_avg_runtime(p, &period);
2141 /* If the task is part of a group prevent parallel updates to group stats */
2142 if (p->numa_group) {
2143 group_lock = &p->numa_group->lock;
2144 spin_lock_irq(group_lock);
2147 /* Find the node with the highest number of faults */
2148 for_each_online_node(nid) {
2149 /* Keep track of the offsets in numa_faults array */
2150 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2151 unsigned long faults = 0, group_faults = 0;
2154 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2155 long diff, f_diff, f_weight;
2157 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2158 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2159 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2160 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2162 /* Decay existing window, copy faults since last scan */
2163 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2164 fault_types[priv] += p->numa_faults[membuf_idx];
2165 p->numa_faults[membuf_idx] = 0;
2168 * Normalize the faults_from, so all tasks in a group
2169 * count according to CPU use, instead of by the raw
2170 * number of faults. Tasks with little runtime have
2171 * little over-all impact on throughput, and thus their
2172 * faults are less important.
2174 f_weight = div64_u64(runtime << 16, period + 1);
2175 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2177 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2178 p->numa_faults[cpubuf_idx] = 0;
2180 p->numa_faults[mem_idx] += diff;
2181 p->numa_faults[cpu_idx] += f_diff;
2182 faults += p->numa_faults[mem_idx];
2183 p->total_numa_faults += diff;
2184 if (p->numa_group) {
2186 * safe because we can only change our own group
2188 * mem_idx represents the offset for a given
2189 * nid and priv in a specific region because it
2190 * is at the beginning of the numa_faults array.
2192 p->numa_group->faults[mem_idx] += diff;
2193 p->numa_group->faults_cpu[mem_idx] += f_diff;
2194 p->numa_group->total_faults += diff;
2195 group_faults += p->numa_group->faults[mem_idx];
2199 if (faults > max_faults) {
2200 max_faults = faults;
2204 if (group_faults > max_group_faults) {
2205 max_group_faults = group_faults;
2206 max_group_nid = nid;
2210 update_task_scan_period(p, fault_types[0], fault_types[1]);
2212 if (p->numa_group) {
2213 numa_group_count_active_nodes(p->numa_group);
2214 spin_unlock_irq(group_lock);
2215 max_nid = preferred_group_nid(p, max_group_nid);
2219 /* Set the new preferred node */
2220 if (max_nid != p->numa_preferred_nid)
2221 sched_setnuma(p, max_nid);
2223 if (task_node(p) != p->numa_preferred_nid)
2224 numa_migrate_preferred(p);
2228 static inline int get_numa_group(struct numa_group *grp)
2230 return atomic_inc_not_zero(&grp->refcount);
2233 static inline void put_numa_group(struct numa_group *grp)
2235 if (atomic_dec_and_test(&grp->refcount))
2236 kfree_rcu(grp, rcu);
2239 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2242 struct numa_group *grp, *my_grp;
2243 struct task_struct *tsk;
2245 int cpu = cpupid_to_cpu(cpupid);
2248 if (unlikely(!p->numa_group)) {
2249 unsigned int size = sizeof(struct numa_group) +
2250 4*nr_node_ids*sizeof(unsigned long);
2252 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2256 atomic_set(&grp->refcount, 1);
2257 grp->active_nodes = 1;
2258 grp->max_faults_cpu = 0;
2259 spin_lock_init(&grp->lock);
2261 /* Second half of the array tracks nids where faults happen */
2262 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2265 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2266 grp->faults[i] = p->numa_faults[i];
2268 grp->total_faults = p->total_numa_faults;
2271 rcu_assign_pointer(p->numa_group, grp);
2275 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2277 if (!cpupid_match_pid(tsk, cpupid))
2280 grp = rcu_dereference(tsk->numa_group);
2284 my_grp = p->numa_group;
2289 * Only join the other group if its bigger; if we're the bigger group,
2290 * the other task will join us.
2292 if (my_grp->nr_tasks > grp->nr_tasks)
2296 * Tie-break on the grp address.
2298 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2301 /* Always join threads in the same process. */
2302 if (tsk->mm == current->mm)
2305 /* Simple filter to avoid false positives due to PID collisions */
2306 if (flags & TNF_SHARED)
2309 /* Update priv based on whether false sharing was detected */
2312 if (join && !get_numa_group(grp))
2320 BUG_ON(irqs_disabled());
2321 double_lock_irq(&my_grp->lock, &grp->lock);
2323 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2324 my_grp->faults[i] -= p->numa_faults[i];
2325 grp->faults[i] += p->numa_faults[i];
2327 my_grp->total_faults -= p->total_numa_faults;
2328 grp->total_faults += p->total_numa_faults;
2333 spin_unlock(&my_grp->lock);
2334 spin_unlock_irq(&grp->lock);
2336 rcu_assign_pointer(p->numa_group, grp);
2338 put_numa_group(my_grp);
2346 void task_numa_free(struct task_struct *p)
2348 struct numa_group *grp = p->numa_group;
2349 void *numa_faults = p->numa_faults;
2350 unsigned long flags;
2354 spin_lock_irqsave(&grp->lock, flags);
2355 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2356 grp->faults[i] -= p->numa_faults[i];
2357 grp->total_faults -= p->total_numa_faults;
2360 spin_unlock_irqrestore(&grp->lock, flags);
2361 RCU_INIT_POINTER(p->numa_group, NULL);
2362 put_numa_group(grp);
2365 p->numa_faults = NULL;
2370 * Got a PROT_NONE fault for a page on @node.
2372 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2374 struct task_struct *p = current;
2375 bool migrated = flags & TNF_MIGRATED;
2376 int cpu_node = task_node(current);
2377 int local = !!(flags & TNF_FAULT_LOCAL);
2378 struct numa_group *ng;
2381 if (!static_branch_likely(&sched_numa_balancing))
2384 /* for example, ksmd faulting in a user's mm */
2388 /* Allocate buffer to track faults on a per-node basis */
2389 if (unlikely(!p->numa_faults)) {
2390 int size = sizeof(*p->numa_faults) *
2391 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2393 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2394 if (!p->numa_faults)
2397 p->total_numa_faults = 0;
2398 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2402 * First accesses are treated as private, otherwise consider accesses
2403 * to be private if the accessing pid has not changed
2405 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2408 priv = cpupid_match_pid(p, last_cpupid);
2409 if (!priv && !(flags & TNF_NO_GROUP))
2410 task_numa_group(p, last_cpupid, flags, &priv);
2414 * If a workload spans multiple NUMA nodes, a shared fault that
2415 * occurs wholly within the set of nodes that the workload is
2416 * actively using should be counted as local. This allows the
2417 * scan rate to slow down when a workload has settled down.
2420 if (!priv && !local && ng && ng->active_nodes > 1 &&
2421 numa_is_active_node(cpu_node, ng) &&
2422 numa_is_active_node(mem_node, ng))
2425 task_numa_placement(p);
2428 * Retry task to preferred node migration periodically, in case it
2429 * case it previously failed, or the scheduler moved us.
2431 if (time_after(jiffies, p->numa_migrate_retry))
2432 numa_migrate_preferred(p);
2435 p->numa_pages_migrated += pages;
2436 if (flags & TNF_MIGRATE_FAIL)
2437 p->numa_faults_locality[2] += pages;
2439 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2440 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2441 p->numa_faults_locality[local] += pages;
2444 static void reset_ptenuma_scan(struct task_struct *p)
2447 * We only did a read acquisition of the mmap sem, so
2448 * p->mm->numa_scan_seq is written to without exclusive access
2449 * and the update is not guaranteed to be atomic. That's not
2450 * much of an issue though, since this is just used for
2451 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2452 * expensive, to avoid any form of compiler optimizations:
2454 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2455 p->mm->numa_scan_offset = 0;
2459 * The expensive part of numa migration is done from task_work context.
2460 * Triggered from task_tick_numa().
2462 void task_numa_work(struct callback_head *work)
2464 unsigned long migrate, next_scan, now = jiffies;
2465 struct task_struct *p = current;
2466 struct mm_struct *mm = p->mm;
2467 u64 runtime = p->se.sum_exec_runtime;
2468 struct vm_area_struct *vma;
2469 unsigned long start, end;
2470 unsigned long nr_pte_updates = 0;
2471 long pages, virtpages;
2473 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2475 work->next = work; /* protect against double add */
2477 * Who cares about NUMA placement when they're dying.
2479 * NOTE: make sure not to dereference p->mm before this check,
2480 * exit_task_work() happens _after_ exit_mm() so we could be called
2481 * without p->mm even though we still had it when we enqueued this
2484 if (p->flags & PF_EXITING)
2487 if (!mm->numa_next_scan) {
2488 mm->numa_next_scan = now +
2489 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2493 * Enforce maximal scan/migration frequency..
2495 migrate = mm->numa_next_scan;
2496 if (time_before(now, migrate))
2499 if (p->numa_scan_period == 0) {
2500 p->numa_scan_period_max = task_scan_max(p);
2501 p->numa_scan_period = task_scan_start(p);
2504 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2505 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2509 * Delay this task enough that another task of this mm will likely win
2510 * the next time around.
2512 p->node_stamp += 2 * TICK_NSEC;
2514 start = mm->numa_scan_offset;
2515 pages = sysctl_numa_balancing_scan_size;
2516 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2517 virtpages = pages * 8; /* Scan up to this much virtual space */
2522 if (!down_read_trylock(&mm->mmap_sem))
2524 vma = find_vma(mm, start);
2526 reset_ptenuma_scan(p);
2530 for (; vma; vma = vma->vm_next) {
2531 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2532 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2537 * Shared library pages mapped by multiple processes are not
2538 * migrated as it is expected they are cache replicated. Avoid
2539 * hinting faults in read-only file-backed mappings or the vdso
2540 * as migrating the pages will be of marginal benefit.
2543 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2547 * Skip inaccessible VMAs to avoid any confusion between
2548 * PROT_NONE and NUMA hinting ptes
2550 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2554 start = max(start, vma->vm_start);
2555 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2556 end = min(end, vma->vm_end);
2557 nr_pte_updates = change_prot_numa(vma, start, end);
2560 * Try to scan sysctl_numa_balancing_size worth of
2561 * hpages that have at least one present PTE that
2562 * is not already pte-numa. If the VMA contains
2563 * areas that are unused or already full of prot_numa
2564 * PTEs, scan up to virtpages, to skip through those
2568 pages -= (end - start) >> PAGE_SHIFT;
2569 virtpages -= (end - start) >> PAGE_SHIFT;
2572 if (pages <= 0 || virtpages <= 0)
2576 } while (end != vma->vm_end);
2581 * It is possible to reach the end of the VMA list but the last few
2582 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2583 * would find the !migratable VMA on the next scan but not reset the
2584 * scanner to the start so check it now.
2587 mm->numa_scan_offset = start;
2589 reset_ptenuma_scan(p);
2590 up_read(&mm->mmap_sem);
2593 * Make sure tasks use at least 32x as much time to run other code
2594 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2595 * Usually update_task_scan_period slows down scanning enough; on an
2596 * overloaded system we need to limit overhead on a per task basis.
2598 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2599 u64 diff = p->se.sum_exec_runtime - runtime;
2600 p->node_stamp += 32 * diff;
2605 * Drive the periodic memory faults..
2607 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2609 struct callback_head *work = &curr->numa_work;
2613 * We don't care about NUMA placement if we don't have memory.
2615 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2619 * Using runtime rather than walltime has the dual advantage that
2620 * we (mostly) drive the selection from busy threads and that the
2621 * task needs to have done some actual work before we bother with
2624 now = curr->se.sum_exec_runtime;
2625 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2627 if (now > curr->node_stamp + period) {
2628 if (!curr->node_stamp)
2629 curr->numa_scan_period = task_scan_start(curr);
2630 curr->node_stamp += period;
2632 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2633 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2634 task_work_add(curr, work, true);
2640 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2644 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2648 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2652 #endif /* CONFIG_NUMA_BALANCING */
2655 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2657 update_load_add(&cfs_rq->load, se->load.weight);
2658 if (!parent_entity(se))
2659 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2661 if (entity_is_task(se)) {
2662 struct rq *rq = rq_of(cfs_rq);
2664 account_numa_enqueue(rq, task_of(se));
2665 list_add(&se->group_node, &rq->cfs_tasks);
2668 cfs_rq->nr_running++;
2672 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2674 update_load_sub(&cfs_rq->load, se->load.weight);
2675 if (!parent_entity(se))
2676 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2678 if (entity_is_task(se)) {
2679 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2680 list_del_init(&se->group_node);
2683 cfs_rq->nr_running--;
2687 * Signed add and clamp on underflow.
2689 * Explicitly do a load-store to ensure the intermediate value never hits
2690 * memory. This allows lockless observations without ever seeing the negative
2693 #define add_positive(_ptr, _val) do { \
2694 typeof(_ptr) ptr = (_ptr); \
2695 typeof(_val) val = (_val); \
2696 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2700 if (val < 0 && res > var) \
2703 WRITE_ONCE(*ptr, res); \
2707 * Unsigned subtract and clamp on underflow.
2709 * Explicitly do a load-store to ensure the intermediate value never hits
2710 * memory. This allows lockless observations without ever seeing the negative
2713 #define sub_positive(_ptr, _val) do { \
2714 typeof(_ptr) ptr = (_ptr); \
2715 typeof(*ptr) val = (_val); \
2716 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2720 WRITE_ONCE(*ptr, res); \
2725 * XXX we want to get rid of these helpers and use the full load resolution.
2727 static inline long se_weight(struct sched_entity *se)
2729 return scale_load_down(se->load.weight);
2732 static inline long se_runnable(struct sched_entity *se)
2734 return scale_load_down(se->runnable_weight);
2738 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2740 cfs_rq->runnable_weight += se->runnable_weight;
2742 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2743 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2747 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2749 cfs_rq->runnable_weight -= se->runnable_weight;
2751 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2752 sub_positive(&cfs_rq->avg.runnable_load_sum,
2753 se_runnable(se) * se->avg.runnable_load_sum);
2757 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2759 cfs_rq->avg.load_avg += se->avg.load_avg;
2760 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2764 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2766 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2767 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2771 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2773 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2775 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2777 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2780 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2781 unsigned long weight, unsigned long runnable)
2784 /* commit outstanding execution time */
2785 if (cfs_rq->curr == se)
2786 update_curr(cfs_rq);
2787 account_entity_dequeue(cfs_rq, se);
2788 dequeue_runnable_load_avg(cfs_rq, se);
2790 dequeue_load_avg(cfs_rq, se);
2792 se->runnable_weight = runnable;
2793 update_load_set(&se->load, weight);
2797 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2799 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2800 se->avg.runnable_load_avg =
2801 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2805 enqueue_load_avg(cfs_rq, se);
2807 account_entity_enqueue(cfs_rq, se);
2808 enqueue_runnable_load_avg(cfs_rq, se);
2812 void reweight_task(struct task_struct *p, int prio)
2814 struct sched_entity *se = &p->se;
2815 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2816 struct load_weight *load = &se->load;
2817 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2819 reweight_entity(cfs_rq, se, weight, weight);
2820 load->inv_weight = sched_prio_to_wmult[prio];
2823 #ifdef CONFIG_FAIR_GROUP_SCHED
2826 * All this does is approximate the hierarchical proportion which includes that
2827 * global sum we all love to hate.
2829 * That is, the weight of a group entity, is the proportional share of the
2830 * group weight based on the group runqueue weights. That is:
2832 * tg->weight * grq->load.weight
2833 * ge->load.weight = ----------------------------- (1)
2834 * \Sum grq->load.weight
2836 * Now, because computing that sum is prohibitively expensive to compute (been
2837 * there, done that) we approximate it with this average stuff. The average
2838 * moves slower and therefore the approximation is cheaper and more stable.
2840 * So instead of the above, we substitute:
2842 * grq->load.weight -> grq->avg.load_avg (2)
2844 * which yields the following:
2846 * tg->weight * grq->avg.load_avg
2847 * ge->load.weight = ------------------------------ (3)
2850 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2852 * That is shares_avg, and it is right (given the approximation (2)).
2854 * The problem with it is that because the average is slow -- it was designed
2855 * to be exactly that of course -- this leads to transients in boundary
2856 * conditions. In specific, the case where the group was idle and we start the
2857 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2858 * yielding bad latency etc..
2860 * Now, in that special case (1) reduces to:
2862 * tg->weight * grq->load.weight
2863 * ge->load.weight = ----------------------------- = tg->weight (4)
2866 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2868 * So what we do is modify our approximation (3) to approach (4) in the (near)
2873 * tg->weight * grq->load.weight
2874 * --------------------------------------------------- (5)
2875 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2877 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2878 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2881 * tg->weight * grq->load.weight
2882 * ge->load.weight = ----------------------------- (6)
2887 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2888 * max(grq->load.weight, grq->avg.load_avg)
2890 * And that is shares_weight and is icky. In the (near) UP case it approaches
2891 * (4) while in the normal case it approaches (3). It consistently
2892 * overestimates the ge->load.weight and therefore:
2894 * \Sum ge->load.weight >= tg->weight
2898 static long calc_group_shares(struct cfs_rq *cfs_rq)
2900 long tg_weight, tg_shares, load, shares;
2901 struct task_group *tg = cfs_rq->tg;
2903 tg_shares = READ_ONCE(tg->shares);
2905 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2907 tg_weight = atomic_long_read(&tg->load_avg);
2909 /* Ensure tg_weight >= load */
2910 tg_weight -= cfs_rq->tg_load_avg_contrib;
2913 shares = (tg_shares * load);
2915 shares /= tg_weight;
2918 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2919 * of a group with small tg->shares value. It is a floor value which is
2920 * assigned as a minimum load.weight to the sched_entity representing
2921 * the group on a CPU.
2923 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2924 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2925 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2926 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2929 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2933 * This calculates the effective runnable weight for a group entity based on
2934 * the group entity weight calculated above.
2936 * Because of the above approximation (2), our group entity weight is
2937 * an load_avg based ratio (3). This means that it includes blocked load and
2938 * does not represent the runnable weight.
2940 * Approximate the group entity's runnable weight per ratio from the group
2943 * grq->avg.runnable_load_avg
2944 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2947 * However, analogous to above, since the avg numbers are slow, this leads to
2948 * transients in the from-idle case. Instead we use:
2950 * ge->runnable_weight = ge->load.weight *
2952 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2953 * ----------------------------------------------------- (8)
2954 * max(grq->avg.load_avg, grq->load.weight)
2956 * Where these max() serve both to use the 'instant' values to fix the slow
2957 * from-idle and avoid the /0 on to-idle, similar to (6).
2959 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2961 long runnable, load_avg;
2963 load_avg = max(cfs_rq->avg.load_avg,
2964 scale_load_down(cfs_rq->load.weight));
2966 runnable = max(cfs_rq->avg.runnable_load_avg,
2967 scale_load_down(cfs_rq->runnable_weight));
2971 runnable /= load_avg;
2973 return clamp_t(long, runnable, MIN_SHARES, shares);
2975 # endif /* CONFIG_SMP */
2977 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2980 * Recomputes the group entity based on the current state of its group
2983 static void update_cfs_group(struct sched_entity *se)
2985 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2986 long shares, runnable;
2991 if (throttled_hierarchy(gcfs_rq))
2995 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2997 if (likely(se->load.weight == shares))
3000 shares = calc_group_shares(gcfs_rq);
3001 runnable = calc_group_runnable(gcfs_rq, shares);
3004 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3007 #else /* CONFIG_FAIR_GROUP_SCHED */
3008 static inline void update_cfs_group(struct sched_entity *se)
3011 #endif /* CONFIG_FAIR_GROUP_SCHED */
3013 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3015 struct rq *rq = rq_of(cfs_rq);
3017 if (&rq->cfs == cfs_rq) {
3019 * There are a few boundary cases this might miss but it should
3020 * get called often enough that that should (hopefully) not be
3021 * a real problem -- added to that it only calls on the local
3022 * CPU, so if we enqueue remotely we'll miss an update, but
3023 * the next tick/schedule should update.
3025 * It will not get called when we go idle, because the idle
3026 * thread is a different class (!fair), nor will the utilization
3027 * number include things like RT tasks.
3029 * As is, the util number is not freq-invariant (we'd have to
3030 * implement arch_scale_freq_capacity() for that).
3034 cpufreq_update_util(rq, 0);
3041 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3043 static u64 decay_load(u64 val, u64 n)
3045 unsigned int local_n;
3047 if (unlikely(n > LOAD_AVG_PERIOD * 63))
3050 /* after bounds checking we can collapse to 32-bit */
3054 * As y^PERIOD = 1/2, we can combine
3055 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3056 * With a look-up table which covers y^n (n<PERIOD)
3058 * To achieve constant time decay_load.
3060 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3061 val >>= local_n / LOAD_AVG_PERIOD;
3062 local_n %= LOAD_AVG_PERIOD;
3065 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3069 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3071 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3076 c1 = decay_load((u64)d1, periods);
3080 * c2 = 1024 \Sum y^n
3084 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3087 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3089 return c1 + c2 + c3;
3092 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
3095 * Accumulate the three separate parts of the sum; d1 the remainder
3096 * of the last (incomplete) period, d2 the span of full periods and d3
3097 * the remainder of the (incomplete) current period.
3102 * |<->|<----------------->|<--->|
3103 * ... |---x---|------| ... |------|-----x (now)
3106 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3109 * = u y^p + (Step 1)
3112 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3115 static __always_inline u32
3116 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3117 unsigned long load, unsigned long runnable, int running)
3119 unsigned long scale_freq, scale_cpu;
3120 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3123 scale_freq = arch_scale_freq_capacity(NULL, cpu);
3124 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3126 delta += sa->period_contrib;
3127 periods = delta / 1024; /* A period is 1024us (~1ms) */
3130 * Step 1: decay old *_sum if we crossed period boundaries.
3133 sa->load_sum = decay_load(sa->load_sum, periods);
3134 sa->runnable_load_sum =
3135 decay_load(sa->runnable_load_sum, periods);
3136 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3142 contrib = __accumulate_pelt_segments(periods,
3143 1024 - sa->period_contrib, delta);
3145 sa->period_contrib = delta;
3147 contrib = cap_scale(contrib, scale_freq);
3149 sa->load_sum += load * contrib;
3151 sa->runnable_load_sum += runnable * contrib;
3153 sa->util_sum += contrib * scale_cpu;
3159 * We can represent the historical contribution to runnable average as the
3160 * coefficients of a geometric series. To do this we sub-divide our runnable
3161 * history into segments of approximately 1ms (1024us); label the segment that
3162 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3164 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3166 * (now) (~1ms ago) (~2ms ago)
3168 * Let u_i denote the fraction of p_i that the entity was runnable.
3170 * We then designate the fractions u_i as our co-efficients, yielding the
3171 * following representation of historical load:
3172 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3174 * We choose y based on the with of a reasonably scheduling period, fixing:
3177 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3178 * approximately half as much as the contribution to load within the last ms
3181 * When a period "rolls over" and we have new u_0`, multiplying the previous
3182 * sum again by y is sufficient to update:
3183 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3184 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3186 static __always_inline int
3187 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3188 unsigned long load, unsigned long runnable, int running)
3192 delta = now - sa->last_update_time;
3194 * This should only happen when time goes backwards, which it
3195 * unfortunately does during sched clock init when we swap over to TSC.
3197 if ((s64)delta < 0) {
3198 sa->last_update_time = now;
3203 * Use 1024ns as the unit of measurement since it's a reasonable
3204 * approximation of 1us and fast to compute.
3210 sa->last_update_time += delta << 10;
3213 * running is a subset of runnable (weight) so running can't be set if
3214 * runnable is clear. But there are some corner cases where the current
3215 * se has been already dequeued but cfs_rq->curr still points to it.
3216 * This means that weight will be 0 but not running for a sched_entity
3217 * but also for a cfs_rq if the latter becomes idle. As an example,
3218 * this happens during idle_balance() which calls
3219 * update_blocked_averages()
3222 runnable = running = 0;
3225 * Now we know we crossed measurement unit boundaries. The *_avg
3226 * accrues by two steps:
3228 * Step 1: accumulate *_sum since last_update_time. If we haven't
3229 * crossed period boundaries, finish.
3231 if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3237 static __always_inline void
3238 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3240 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3243 * Step 2: update *_avg.
3245 sa->load_avg = div_u64(load * sa->load_sum, divider);
3246 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3247 sa->util_avg = sa->util_sum / divider;
3254 * se_runnable() == se_weight()
3256 * group: [ see update_cfs_group() ]
3257 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3258 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3260 * load_sum := runnable_sum
3261 * load_avg = se_weight(se) * runnable_avg
3263 * runnable_load_sum := runnable_sum
3264 * runnable_load_avg = se_runnable(se) * runnable_avg
3266 * XXX collapse load_sum and runnable_load_sum
3270 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3271 * load_avg = \Sum se->avg.load_avg
3273 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3274 * runnable_load_avg = \Sum se->avg.runable_load_avg
3278 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3280 if (entity_is_task(se))
3281 se->runnable_weight = se->load.weight;
3283 if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3284 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3292 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3294 if (entity_is_task(se))
3295 se->runnable_weight = se->load.weight;
3297 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3298 cfs_rq->curr == se)) {
3300 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3308 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3310 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3311 scale_load_down(cfs_rq->load.weight),
3312 scale_load_down(cfs_rq->runnable_weight),
3313 cfs_rq->curr != NULL)) {
3315 ___update_load_avg(&cfs_rq->avg, 1, 1);
3322 #ifdef CONFIG_FAIR_GROUP_SCHED
3324 * update_tg_load_avg - update the tg's load avg
3325 * @cfs_rq: the cfs_rq whose avg changed
3326 * @force: update regardless of how small the difference
3328 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3329 * However, because tg->load_avg is a global value there are performance
3332 * In order to avoid having to look at the other cfs_rq's, we use a
3333 * differential update where we store the last value we propagated. This in
3334 * turn allows skipping updates if the differential is 'small'.
3336 * Updating tg's load_avg is necessary before update_cfs_share().
3338 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3340 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3343 * No need to update load_avg for root_task_group as it is not used.
3345 if (cfs_rq->tg == &root_task_group)
3348 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3349 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3350 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3355 * Called within set_task_rq() right before setting a task's cpu. The
3356 * caller only guarantees p->pi_lock is held; no other assumptions,
3357 * including the state of rq->lock, should be made.
3359 void set_task_rq_fair(struct sched_entity *se,
3360 struct cfs_rq *prev, struct cfs_rq *next)
3362 u64 p_last_update_time;
3363 u64 n_last_update_time;
3365 if (!sched_feat(ATTACH_AGE_LOAD))
3369 * We are supposed to update the task to "current" time, then its up to
3370 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3371 * getting what current time is, so simply throw away the out-of-date
3372 * time. This will result in the wakee task is less decayed, but giving
3373 * the wakee more load sounds not bad.
3375 if (!(se->avg.last_update_time && prev))
3378 #ifndef CONFIG_64BIT
3380 u64 p_last_update_time_copy;
3381 u64 n_last_update_time_copy;
3384 p_last_update_time_copy = prev->load_last_update_time_copy;
3385 n_last_update_time_copy = next->load_last_update_time_copy;
3389 p_last_update_time = prev->avg.last_update_time;
3390 n_last_update_time = next->avg.last_update_time;
3392 } while (p_last_update_time != p_last_update_time_copy ||
3393 n_last_update_time != n_last_update_time_copy);
3396 p_last_update_time = prev->avg.last_update_time;
3397 n_last_update_time = next->avg.last_update_time;
3399 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3400 se->avg.last_update_time = n_last_update_time;
3405 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3406 * propagate its contribution. The key to this propagation is the invariant
3407 * that for each group:
3409 * ge->avg == grq->avg (1)
3411 * _IFF_ we look at the pure running and runnable sums. Because they
3412 * represent the very same entity, just at different points in the hierarchy.
3415 * Per the above update_tg_cfs_util() is trivial (and still 'wrong') and
3416 * simply copies the running sum over.
3418 * However, update_tg_cfs_runnable() is more complex. So we have:
3420 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3422 * And since, like util, the runnable part should be directly transferable,
3423 * the following would _appear_ to be the straight forward approach:
3425 * grq->avg.load_avg = grq->load.weight * grq->avg.running_avg (3)
3427 * And per (1) we have:
3429 * ge->avg.running_avg == grq->avg.running_avg
3433 * ge->load.weight * grq->avg.load_avg
3434 * ge->avg.load_avg = ----------------------------------- (4)
3437 * Except that is wrong!
3439 * Because while for entities historical weight is not important and we
3440 * really only care about our future and therefore can consider a pure
3441 * runnable sum, runqueues can NOT do this.
3443 * We specifically want runqueues to have a load_avg that includes
3444 * historical weights. Those represent the blocked load, the load we expect
3445 * to (shortly) return to us. This only works by keeping the weights as
3446 * integral part of the sum. We therefore cannot decompose as per (3).
3451 * Another way to look at things is:
3453 * grq->avg.load_avg = \Sum se->avg.load_avg
3455 * Therefore, per (2):
3457 * grq->avg.load_avg = \Sum se->load.weight * se->avg.runnable_avg
3459 * And the very thing we're propagating is a change in that sum (someone
3460 * joined/left). So we can easily know the runnable change, which would be, per
3461 * (2) the already tracked se->load_avg divided by the corresponding
3464 * Basically (4) but in differential form:
3466 * d(runnable_avg) += se->avg.load_avg / se->load.weight
3468 * ge->avg.load_avg += ge->load.weight * d(runnable_avg)
3472 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3474 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3476 /* Nothing to update */
3480 /* Set new sched_entity's utilization */
3481 se->avg.util_avg = gcfs_rq->avg.util_avg;
3482 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3484 /* Update parent cfs_rq utilization */
3485 add_positive(&cfs_rq->avg.util_avg, delta);
3486 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3490 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3492 long runnable_sum = gcfs_rq->prop_runnable_sum;
3493 long runnable_load_avg, load_avg;
3494 s64 runnable_load_sum, load_sum;
3499 gcfs_rq->prop_runnable_sum = 0;
3501 load_sum = (s64)se_weight(se) * runnable_sum;
3502 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3504 add_positive(&se->avg.load_sum, runnable_sum);
3505 add_positive(&se->avg.load_avg, load_avg);
3507 add_positive(&cfs_rq->avg.load_avg, load_avg);
3508 add_positive(&cfs_rq->avg.load_sum, load_sum);
3510 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3511 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3513 add_positive(&se->avg.runnable_load_sum, runnable_sum);
3514 add_positive(&se->avg.runnable_load_avg, runnable_load_avg);
3517 add_positive(&cfs_rq->avg.runnable_load_avg, runnable_load_avg);
3518 add_positive(&cfs_rq->avg.runnable_load_sum, runnable_load_sum);
3522 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3524 cfs_rq->propagate = 1;
3525 cfs_rq->prop_runnable_sum += runnable_sum;
3528 /* Update task and its cfs_rq load average */
3529 static inline int propagate_entity_load_avg(struct sched_entity *se)
3531 struct cfs_rq *cfs_rq, *gcfs_rq;
3533 if (entity_is_task(se))
3536 gcfs_rq = group_cfs_rq(se);
3537 if (!gcfs_rq->propagate)
3540 gcfs_rq->propagate = 0;
3542 cfs_rq = cfs_rq_of(se);
3544 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3546 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3547 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3553 * Check if we need to update the load and the utilization of a blocked
3556 static inline bool skip_blocked_update(struct sched_entity *se)
3558 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3561 * If sched_entity still have not zero load or utilization, we have to
3564 if (se->avg.load_avg || se->avg.util_avg)
3568 * If there is a pending propagation, we have to update the load and
3569 * the utilization of the sched_entity:
3571 if (gcfs_rq->propagate)
3575 * Otherwise, the load and the utilization of the sched_entity is
3576 * already zero and there is no pending propagation, so it will be a
3577 * waste of time to try to decay it:
3582 #else /* CONFIG_FAIR_GROUP_SCHED */
3584 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3586 static inline int propagate_entity_load_avg(struct sched_entity *se)
3591 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3593 #endif /* CONFIG_FAIR_GROUP_SCHED */
3596 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3597 * @now: current time, as per cfs_rq_clock_task()
3598 * @cfs_rq: cfs_rq to update
3600 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3601 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3602 * post_init_entity_util_avg().
3604 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3606 * Returns true if the load decayed or we removed load.
3608 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3609 * call update_tg_load_avg() when this function returns true.
3612 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3614 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3615 struct sched_avg *sa = &cfs_rq->avg;
3618 if (cfs_rq->removed.nr) {
3620 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3622 raw_spin_lock(&cfs_rq->removed.lock);
3623 swap(cfs_rq->removed.util_avg, removed_util);
3624 swap(cfs_rq->removed.load_avg, removed_load);
3625 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3626 cfs_rq->removed.nr = 0;
3627 raw_spin_unlock(&cfs_rq->removed.lock);
3630 sub_positive(&sa->load_avg, r);
3631 sub_positive(&sa->load_sum, r * divider);
3634 sub_positive(&sa->util_avg, r);
3635 sub_positive(&sa->util_sum, r * divider);
3637 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3642 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3644 #ifndef CONFIG_64BIT
3646 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3650 cfs_rq_util_change(cfs_rq);
3656 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3657 * @cfs_rq: cfs_rq to attach to
3658 * @se: sched_entity to attach
3660 * Must call update_cfs_rq_load_avg() before this, since we rely on
3661 * cfs_rq->avg.last_update_time being current.
3663 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3665 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3668 * When we attach the @se to the @cfs_rq, we must align the decay
3669 * window because without that, really weird and wonderful things can
3674 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3675 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3678 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3679 * period_contrib. This isn't strictly correct, but since we're
3680 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3683 se->avg.util_sum = se->avg.util_avg * divider;
3685 se->avg.load_sum = divider;
3686 if (se_weight(se)) {
3688 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3691 se->avg.runnable_load_sum = se->avg.load_sum;
3693 enqueue_load_avg(cfs_rq, se);
3694 cfs_rq->avg.util_avg += se->avg.util_avg;
3695 cfs_rq->avg.util_sum += se->avg.util_sum;
3697 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3699 cfs_rq_util_change(cfs_rq);
3703 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3704 * @cfs_rq: cfs_rq to detach from
3705 * @se: sched_entity to detach
3707 * Must call update_cfs_rq_load_avg() before this, since we rely on
3708 * cfs_rq->avg.last_update_time being current.
3710 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3712 dequeue_load_avg(cfs_rq, se);
3713 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3714 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3716 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3718 cfs_rq_util_change(cfs_rq);
3722 * Optional action to be done while updating the load average
3724 #define UPDATE_TG 0x1
3725 #define SKIP_AGE_LOAD 0x2
3726 #define DO_ATTACH 0x4
3728 /* Update task and its cfs_rq load average */
3729 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3731 u64 now = cfs_rq_clock_task(cfs_rq);
3732 struct rq *rq = rq_of(cfs_rq);
3733 int cpu = cpu_of(rq);
3737 * Track task load average for carrying it to new CPU after migrated, and
3738 * track group sched_entity load average for task_h_load calc in migration
3740 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3741 __update_load_avg_se(now, cpu, cfs_rq, se);
3743 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3744 decayed |= propagate_entity_load_avg(se);
3746 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3748 attach_entity_load_avg(cfs_rq, se);
3749 update_tg_load_avg(cfs_rq, 0);
3751 } else if (decayed && (flags & UPDATE_TG))
3752 update_tg_load_avg(cfs_rq, 0);
3755 #ifndef CONFIG_64BIT
3756 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3758 u64 last_update_time_copy;
3759 u64 last_update_time;
3762 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3764 last_update_time = cfs_rq->avg.last_update_time;
3765 } while (last_update_time != last_update_time_copy);
3767 return last_update_time;
3770 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3772 return cfs_rq->avg.last_update_time;
3777 * Synchronize entity load avg of dequeued entity without locking
3780 void sync_entity_load_avg(struct sched_entity *se)
3782 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3783 u64 last_update_time;
3785 last_update_time = cfs_rq_last_update_time(cfs_rq);
3786 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3790 * Task first catches up with cfs_rq, and then subtract
3791 * itself from the cfs_rq (task must be off the queue now).
3793 void remove_entity_load_avg(struct sched_entity *se)
3795 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3796 unsigned long flags;
3799 * tasks cannot exit without having gone through wake_up_new_task() ->
3800 * post_init_entity_util_avg() which will have added things to the
3801 * cfs_rq, so we can remove unconditionally.
3803 * Similarly for groups, they will have passed through
3804 * post_init_entity_util_avg() before unregister_sched_fair_group()
3808 sync_entity_load_avg(se);
3810 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3811 ++cfs_rq->removed.nr;
3812 cfs_rq->removed.util_avg += se->avg.util_avg;
3813 cfs_rq->removed.load_avg += se->avg.load_avg;
3814 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3815 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3818 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3820 return cfs_rq->avg.runnable_load_avg;
3823 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3825 return cfs_rq->avg.load_avg;
3828 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3830 #else /* CONFIG_SMP */
3833 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3838 #define UPDATE_TG 0x0
3839 #define SKIP_AGE_LOAD 0x0
3840 #define DO_ATTACH 0x0
3842 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3844 cfs_rq_util_change(cfs_rq);
3847 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3850 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3852 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3854 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3859 #endif /* CONFIG_SMP */
3861 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3863 #ifdef CONFIG_SCHED_DEBUG
3864 s64 d = se->vruntime - cfs_rq->min_vruntime;
3869 if (d > 3*sysctl_sched_latency)
3870 schedstat_inc(cfs_rq->nr_spread_over);
3875 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3877 u64 vruntime = cfs_rq->min_vruntime;
3880 * The 'current' period is already promised to the current tasks,
3881 * however the extra weight of the new task will slow them down a
3882 * little, place the new task so that it fits in the slot that
3883 * stays open at the end.
3885 if (initial && sched_feat(START_DEBIT))
3886 vruntime += sched_vslice(cfs_rq, se);
3888 /* sleeps up to a single latency don't count. */
3890 unsigned long thresh = sysctl_sched_latency;
3893 * Halve their sleep time's effect, to allow
3894 * for a gentler effect of sleepers:
3896 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3902 /* ensure we never gain time by being placed backwards. */
3903 se->vruntime = max_vruntime(se->vruntime, vruntime);
3906 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3908 static inline void check_schedstat_required(void)
3910 #ifdef CONFIG_SCHEDSTATS
3911 if (schedstat_enabled())
3914 /* Force schedstat enabled if a dependent tracepoint is active */
3915 if (trace_sched_stat_wait_enabled() ||
3916 trace_sched_stat_sleep_enabled() ||
3917 trace_sched_stat_iowait_enabled() ||
3918 trace_sched_stat_blocked_enabled() ||
3919 trace_sched_stat_runtime_enabled()) {
3920 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3921 "stat_blocked and stat_runtime require the "
3922 "kernel parameter schedstats=enable or "
3923 "kernel.sched_schedstats=1\n");
3934 * update_min_vruntime()
3935 * vruntime -= min_vruntime
3939 * update_min_vruntime()
3940 * vruntime += min_vruntime
3942 * this way the vruntime transition between RQs is done when both
3943 * min_vruntime are up-to-date.
3947 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3948 * vruntime -= min_vruntime
3952 * update_min_vruntime()
3953 * vruntime += min_vruntime
3955 * this way we don't have the most up-to-date min_vruntime on the originating
3956 * CPU and an up-to-date min_vruntime on the destination CPU.
3960 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3962 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3963 bool curr = cfs_rq->curr == se;
3966 * If we're the current task, we must renormalise before calling
3970 se->vruntime += cfs_rq->min_vruntime;
3972 update_curr(cfs_rq);
3975 * Otherwise, renormalise after, such that we're placed at the current
3976 * moment in time, instead of some random moment in the past. Being
3977 * placed in the past could significantly boost this task to the
3978 * fairness detriment of existing tasks.
3980 if (renorm && !curr)
3981 se->vruntime += cfs_rq->min_vruntime;
3984 * When enqueuing a sched_entity, we must:
3985 * - Update loads to have both entity and cfs_rq synced with now.
3986 * - Add its load to cfs_rq->runnable_avg
3987 * - For group_entity, update its weight to reflect the new share of
3989 * - Add its new weight to cfs_rq->load.weight
3991 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3992 update_cfs_group(se);
3993 enqueue_runnable_load_avg(cfs_rq, se);
3994 account_entity_enqueue(cfs_rq, se);
3996 if (flags & ENQUEUE_WAKEUP)
3997 place_entity(cfs_rq, se, 0);
3999 check_schedstat_required();
4000 update_stats_enqueue(cfs_rq, se, flags);
4001 check_spread(cfs_rq, se);
4003 __enqueue_entity(cfs_rq, se);
4006 if (cfs_rq->nr_running == 1) {
4007 list_add_leaf_cfs_rq(cfs_rq);
4008 check_enqueue_throttle(cfs_rq);
4012 static void __clear_buddies_last(struct sched_entity *se)
4014 for_each_sched_entity(se) {
4015 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4016 if (cfs_rq->last != se)
4019 cfs_rq->last = NULL;
4023 static void __clear_buddies_next(struct sched_entity *se)
4025 for_each_sched_entity(se) {
4026 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4027 if (cfs_rq->next != se)
4030 cfs_rq->next = NULL;
4034 static void __clear_buddies_skip(struct sched_entity *se)
4036 for_each_sched_entity(se) {
4037 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4038 if (cfs_rq->skip != se)
4041 cfs_rq->skip = NULL;
4045 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4047 if (cfs_rq->last == se)
4048 __clear_buddies_last(se);
4050 if (cfs_rq->next == se)
4051 __clear_buddies_next(se);
4053 if (cfs_rq->skip == se)
4054 __clear_buddies_skip(se);
4057 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4060 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4063 * Update run-time statistics of the 'current'.
4065 update_curr(cfs_rq);
4068 * When dequeuing a sched_entity, we must:
4069 * - Update loads to have both entity and cfs_rq synced with now.
4070 * - Substract its load from the cfs_rq->runnable_avg.
4071 * - Substract its previous weight from cfs_rq->load.weight.
4072 * - For group entity, update its weight to reflect the new share
4073 * of its group cfs_rq.
4075 update_load_avg(cfs_rq, se, UPDATE_TG);
4076 dequeue_runnable_load_avg(cfs_rq, se);
4078 update_stats_dequeue(cfs_rq, se, flags);
4080 clear_buddies(cfs_rq, se);
4082 if (se != cfs_rq->curr)
4083 __dequeue_entity(cfs_rq, se);
4085 account_entity_dequeue(cfs_rq, se);
4088 * Normalize after update_curr(); which will also have moved
4089 * min_vruntime if @se is the one holding it back. But before doing
4090 * update_min_vruntime() again, which will discount @se's position and
4091 * can move min_vruntime forward still more.
4093 if (!(flags & DEQUEUE_SLEEP))
4094 se->vruntime -= cfs_rq->min_vruntime;
4096 /* return excess runtime on last dequeue */
4097 return_cfs_rq_runtime(cfs_rq);
4099 update_cfs_group(se);
4102 * Now advance min_vruntime if @se was the entity holding it back,
4103 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4104 * put back on, and if we advance min_vruntime, we'll be placed back
4105 * further than we started -- ie. we'll be penalized.
4107 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4108 update_min_vruntime(cfs_rq);
4112 * Preempt the current task with a newly woken task if needed:
4115 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4117 unsigned long ideal_runtime, delta_exec;
4118 struct sched_entity *se;
4121 ideal_runtime = sched_slice(cfs_rq, curr);
4122 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4123 if (delta_exec > ideal_runtime) {
4124 resched_curr(rq_of(cfs_rq));
4126 * The current task ran long enough, ensure it doesn't get
4127 * re-elected due to buddy favours.
4129 clear_buddies(cfs_rq, curr);
4134 * Ensure that a task that missed wakeup preemption by a
4135 * narrow margin doesn't have to wait for a full slice.
4136 * This also mitigates buddy induced latencies under load.
4138 if (delta_exec < sysctl_sched_min_granularity)
4141 se = __pick_first_entity(cfs_rq);
4142 delta = curr->vruntime - se->vruntime;
4147 if (delta > ideal_runtime)
4148 resched_curr(rq_of(cfs_rq));
4152 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4154 /* 'current' is not kept within the tree. */
4157 * Any task has to be enqueued before it get to execute on
4158 * a CPU. So account for the time it spent waiting on the
4161 update_stats_wait_end(cfs_rq, se);
4162 __dequeue_entity(cfs_rq, se);
4163 update_load_avg(cfs_rq, se, UPDATE_TG);
4166 update_stats_curr_start(cfs_rq, se);
4170 * Track our maximum slice length, if the CPU's load is at
4171 * least twice that of our own weight (i.e. dont track it
4172 * when there are only lesser-weight tasks around):
4174 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4175 schedstat_set(se->statistics.slice_max,
4176 max((u64)schedstat_val(se->statistics.slice_max),
4177 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4180 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4184 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4187 * Pick the next process, keeping these things in mind, in this order:
4188 * 1) keep things fair between processes/task groups
4189 * 2) pick the "next" process, since someone really wants that to run
4190 * 3) pick the "last" process, for cache locality
4191 * 4) do not run the "skip" process, if something else is available
4193 static struct sched_entity *
4194 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4196 struct sched_entity *left = __pick_first_entity(cfs_rq);
4197 struct sched_entity *se;
4200 * If curr is set we have to see if its left of the leftmost entity
4201 * still in the tree, provided there was anything in the tree at all.
4203 if (!left || (curr && entity_before(curr, left)))
4206 se = left; /* ideally we run the leftmost entity */
4209 * Avoid running the skip buddy, if running something else can
4210 * be done without getting too unfair.
4212 if (cfs_rq->skip == se) {
4213 struct sched_entity *second;
4216 second = __pick_first_entity(cfs_rq);
4218 second = __pick_next_entity(se);
4219 if (!second || (curr && entity_before(curr, second)))
4223 if (second && wakeup_preempt_entity(second, left) < 1)
4228 * Prefer last buddy, try to return the CPU to a preempted task.
4230 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4234 * Someone really wants this to run. If it's not unfair, run it.
4236 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4239 clear_buddies(cfs_rq, se);
4244 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4246 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4249 * If still on the runqueue then deactivate_task()
4250 * was not called and update_curr() has to be done:
4253 update_curr(cfs_rq);
4255 /* throttle cfs_rqs exceeding runtime */
4256 check_cfs_rq_runtime(cfs_rq);
4258 check_spread(cfs_rq, prev);
4261 update_stats_wait_start(cfs_rq, prev);
4262 /* Put 'current' back into the tree. */
4263 __enqueue_entity(cfs_rq, prev);
4264 /* in !on_rq case, update occurred at dequeue */
4265 update_load_avg(cfs_rq, prev, 0);
4267 cfs_rq->curr = NULL;
4271 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4274 * Update run-time statistics of the 'current'.
4276 update_curr(cfs_rq);
4279 * Ensure that runnable average is periodically updated.
4281 update_load_avg(cfs_rq, curr, UPDATE_TG);
4282 update_cfs_group(curr);
4284 #ifdef CONFIG_SCHED_HRTICK
4286 * queued ticks are scheduled to match the slice, so don't bother
4287 * validating it and just reschedule.
4290 resched_curr(rq_of(cfs_rq));
4294 * don't let the period tick interfere with the hrtick preemption
4296 if (!sched_feat(DOUBLE_TICK) &&
4297 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4301 if (cfs_rq->nr_running > 1)
4302 check_preempt_tick(cfs_rq, curr);
4306 /**************************************************
4307 * CFS bandwidth control machinery
4310 #ifdef CONFIG_CFS_BANDWIDTH
4312 #ifdef HAVE_JUMP_LABEL
4313 static struct static_key __cfs_bandwidth_used;
4315 static inline bool cfs_bandwidth_used(void)
4317 return static_key_false(&__cfs_bandwidth_used);
4320 void cfs_bandwidth_usage_inc(void)
4322 static_key_slow_inc(&__cfs_bandwidth_used);
4325 void cfs_bandwidth_usage_dec(void)
4327 static_key_slow_dec(&__cfs_bandwidth_used);
4329 #else /* HAVE_JUMP_LABEL */
4330 static bool cfs_bandwidth_used(void)
4335 void cfs_bandwidth_usage_inc(void) {}
4336 void cfs_bandwidth_usage_dec(void) {}
4337 #endif /* HAVE_JUMP_LABEL */
4340 * default period for cfs group bandwidth.
4341 * default: 0.1s, units: nanoseconds
4343 static inline u64 default_cfs_period(void)
4345 return 100000000ULL;
4348 static inline u64 sched_cfs_bandwidth_slice(void)
4350 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4354 * Replenish runtime according to assigned quota and update expiration time.
4355 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4356 * additional synchronization around rq->lock.
4358 * requires cfs_b->lock
4360 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4364 if (cfs_b->quota == RUNTIME_INF)
4367 now = sched_clock_cpu(smp_processor_id());
4368 cfs_b->runtime = cfs_b->quota;
4369 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4372 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4374 return &tg->cfs_bandwidth;
4377 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4378 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4380 if (unlikely(cfs_rq->throttle_count))
4381 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4383 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4386 /* returns 0 on failure to allocate runtime */
4387 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4389 struct task_group *tg = cfs_rq->tg;
4390 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4391 u64 amount = 0, min_amount, expires;
4393 /* note: this is a positive sum as runtime_remaining <= 0 */
4394 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4396 raw_spin_lock(&cfs_b->lock);
4397 if (cfs_b->quota == RUNTIME_INF)
4398 amount = min_amount;
4400 start_cfs_bandwidth(cfs_b);
4402 if (cfs_b->runtime > 0) {
4403 amount = min(cfs_b->runtime, min_amount);
4404 cfs_b->runtime -= amount;
4408 expires = cfs_b->runtime_expires;
4409 raw_spin_unlock(&cfs_b->lock);
4411 cfs_rq->runtime_remaining += amount;
4413 * we may have advanced our local expiration to account for allowed
4414 * spread between our sched_clock and the one on which runtime was
4417 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4418 cfs_rq->runtime_expires = expires;
4420 return cfs_rq->runtime_remaining > 0;
4424 * Note: This depends on the synchronization provided by sched_clock and the
4425 * fact that rq->clock snapshots this value.
4427 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4429 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4431 /* if the deadline is ahead of our clock, nothing to do */
4432 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4435 if (cfs_rq->runtime_remaining < 0)
4439 * If the local deadline has passed we have to consider the
4440 * possibility that our sched_clock is 'fast' and the global deadline
4441 * has not truly expired.
4443 * Fortunately we can check determine whether this the case by checking
4444 * whether the global deadline has advanced. It is valid to compare
4445 * cfs_b->runtime_expires without any locks since we only care about
4446 * exact equality, so a partial write will still work.
4449 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4450 /* extend local deadline, drift is bounded above by 2 ticks */
4451 cfs_rq->runtime_expires += TICK_NSEC;
4453 /* global deadline is ahead, expiration has passed */
4454 cfs_rq->runtime_remaining = 0;
4458 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4460 /* dock delta_exec before expiring quota (as it could span periods) */
4461 cfs_rq->runtime_remaining -= delta_exec;
4462 expire_cfs_rq_runtime(cfs_rq);
4464 if (likely(cfs_rq->runtime_remaining > 0))
4468 * if we're unable to extend our runtime we resched so that the active
4469 * hierarchy can be throttled
4471 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4472 resched_curr(rq_of(cfs_rq));
4475 static __always_inline
4476 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4478 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4481 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4484 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4486 return cfs_bandwidth_used() && cfs_rq->throttled;
4489 /* check whether cfs_rq, or any parent, is throttled */
4490 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4492 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4496 * Ensure that neither of the group entities corresponding to src_cpu or
4497 * dest_cpu are members of a throttled hierarchy when performing group
4498 * load-balance operations.
4500 static inline int throttled_lb_pair(struct task_group *tg,
4501 int src_cpu, int dest_cpu)
4503 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4505 src_cfs_rq = tg->cfs_rq[src_cpu];
4506 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4508 return throttled_hierarchy(src_cfs_rq) ||
4509 throttled_hierarchy(dest_cfs_rq);
4512 /* updated child weight may affect parent so we have to do this bottom up */
4513 static int tg_unthrottle_up(struct task_group *tg, void *data)
4515 struct rq *rq = data;
4516 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4518 cfs_rq->throttle_count--;
4519 if (!cfs_rq->throttle_count) {
4520 /* adjust cfs_rq_clock_task() */
4521 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4522 cfs_rq->throttled_clock_task;
4528 static int tg_throttle_down(struct task_group *tg, void *data)
4530 struct rq *rq = data;
4531 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4533 /* group is entering throttled state, stop time */
4534 if (!cfs_rq->throttle_count)
4535 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4536 cfs_rq->throttle_count++;
4541 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4543 struct rq *rq = rq_of(cfs_rq);
4544 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4545 struct sched_entity *se;
4546 long task_delta, dequeue = 1;
4549 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4551 /* freeze hierarchy runnable averages while throttled */
4553 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4556 task_delta = cfs_rq->h_nr_running;
4557 for_each_sched_entity(se) {
4558 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4559 /* throttled entity or throttle-on-deactivate */
4564 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4565 qcfs_rq->h_nr_running -= task_delta;
4567 if (qcfs_rq->load.weight)
4572 sub_nr_running(rq, task_delta);
4574 cfs_rq->throttled = 1;
4575 cfs_rq->throttled_clock = rq_clock(rq);
4576 raw_spin_lock(&cfs_b->lock);
4577 empty = list_empty(&cfs_b->throttled_cfs_rq);
4580 * Add to the _head_ of the list, so that an already-started
4581 * distribute_cfs_runtime will not see us
4583 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4586 * If we're the first throttled task, make sure the bandwidth
4590 start_cfs_bandwidth(cfs_b);
4592 raw_spin_unlock(&cfs_b->lock);
4595 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4597 struct rq *rq = rq_of(cfs_rq);
4598 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4599 struct sched_entity *se;
4603 se = cfs_rq->tg->se[cpu_of(rq)];
4605 cfs_rq->throttled = 0;
4607 update_rq_clock(rq);
4609 raw_spin_lock(&cfs_b->lock);
4610 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4611 list_del_rcu(&cfs_rq->throttled_list);
4612 raw_spin_unlock(&cfs_b->lock);
4614 /* update hierarchical throttle state */
4615 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4617 if (!cfs_rq->load.weight)
4620 task_delta = cfs_rq->h_nr_running;
4621 for_each_sched_entity(se) {
4625 cfs_rq = cfs_rq_of(se);
4627 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4628 cfs_rq->h_nr_running += task_delta;
4630 if (cfs_rq_throttled(cfs_rq))
4635 add_nr_running(rq, task_delta);
4637 /* determine whether we need to wake up potentially idle cpu */
4638 if (rq->curr == rq->idle && rq->cfs.nr_running)
4642 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4643 u64 remaining, u64 expires)
4645 struct cfs_rq *cfs_rq;
4647 u64 starting_runtime = remaining;
4650 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4652 struct rq *rq = rq_of(cfs_rq);
4656 if (!cfs_rq_throttled(cfs_rq))
4659 runtime = -cfs_rq->runtime_remaining + 1;
4660 if (runtime > remaining)
4661 runtime = remaining;
4662 remaining -= runtime;
4664 cfs_rq->runtime_remaining += runtime;
4665 cfs_rq->runtime_expires = expires;
4667 /* we check whether we're throttled above */
4668 if (cfs_rq->runtime_remaining > 0)
4669 unthrottle_cfs_rq(cfs_rq);
4679 return starting_runtime - remaining;
4683 * Responsible for refilling a task_group's bandwidth and unthrottling its
4684 * cfs_rqs as appropriate. If there has been no activity within the last
4685 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4686 * used to track this state.
4688 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4690 u64 runtime, runtime_expires;
4693 /* no need to continue the timer with no bandwidth constraint */
4694 if (cfs_b->quota == RUNTIME_INF)
4695 goto out_deactivate;
4697 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4698 cfs_b->nr_periods += overrun;
4701 * idle depends on !throttled (for the case of a large deficit), and if
4702 * we're going inactive then everything else can be deferred
4704 if (cfs_b->idle && !throttled)
4705 goto out_deactivate;
4707 __refill_cfs_bandwidth_runtime(cfs_b);
4710 /* mark as potentially idle for the upcoming period */
4715 /* account preceding periods in which throttling occurred */
4716 cfs_b->nr_throttled += overrun;
4718 runtime_expires = cfs_b->runtime_expires;
4721 * This check is repeated as we are holding onto the new bandwidth while
4722 * we unthrottle. This can potentially race with an unthrottled group
4723 * trying to acquire new bandwidth from the global pool. This can result
4724 * in us over-using our runtime if it is all used during this loop, but
4725 * only by limited amounts in that extreme case.
4727 while (throttled && cfs_b->runtime > 0) {
4728 runtime = cfs_b->runtime;
4729 raw_spin_unlock(&cfs_b->lock);
4730 /* we can't nest cfs_b->lock while distributing bandwidth */
4731 runtime = distribute_cfs_runtime(cfs_b, runtime,
4733 raw_spin_lock(&cfs_b->lock);
4735 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4737 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4741 * While we are ensured activity in the period following an
4742 * unthrottle, this also covers the case in which the new bandwidth is
4743 * insufficient to cover the existing bandwidth deficit. (Forcing the
4744 * timer to remain active while there are any throttled entities.)
4754 /* a cfs_rq won't donate quota below this amount */
4755 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4756 /* minimum remaining period time to redistribute slack quota */
4757 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4758 /* how long we wait to gather additional slack before distributing */
4759 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4762 * Are we near the end of the current quota period?
4764 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4765 * hrtimer base being cleared by hrtimer_start. In the case of
4766 * migrate_hrtimers, base is never cleared, so we are fine.
4768 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4770 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4773 /* if the call-back is running a quota refresh is already occurring */
4774 if (hrtimer_callback_running(refresh_timer))
4777 /* is a quota refresh about to occur? */
4778 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4779 if (remaining < min_expire)
4785 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4787 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4789 /* if there's a quota refresh soon don't bother with slack */
4790 if (runtime_refresh_within(cfs_b, min_left))
4793 hrtimer_start(&cfs_b->slack_timer,
4794 ns_to_ktime(cfs_bandwidth_slack_period),
4798 /* we know any runtime found here is valid as update_curr() precedes return */
4799 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4801 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4802 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4804 if (slack_runtime <= 0)
4807 raw_spin_lock(&cfs_b->lock);
4808 if (cfs_b->quota != RUNTIME_INF &&
4809 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4810 cfs_b->runtime += slack_runtime;
4812 /* we are under rq->lock, defer unthrottling using a timer */
4813 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4814 !list_empty(&cfs_b->throttled_cfs_rq))
4815 start_cfs_slack_bandwidth(cfs_b);
4817 raw_spin_unlock(&cfs_b->lock);
4819 /* even if it's not valid for return we don't want to try again */
4820 cfs_rq->runtime_remaining -= slack_runtime;
4823 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4825 if (!cfs_bandwidth_used())
4828 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4831 __return_cfs_rq_runtime(cfs_rq);
4835 * This is done with a timer (instead of inline with bandwidth return) since
4836 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4838 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4840 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4843 /* confirm we're still not at a refresh boundary */
4844 raw_spin_lock(&cfs_b->lock);
4845 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4846 raw_spin_unlock(&cfs_b->lock);
4850 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4851 runtime = cfs_b->runtime;
4853 expires = cfs_b->runtime_expires;
4854 raw_spin_unlock(&cfs_b->lock);
4859 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4861 raw_spin_lock(&cfs_b->lock);
4862 if (expires == cfs_b->runtime_expires)
4863 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4864 raw_spin_unlock(&cfs_b->lock);
4868 * When a group wakes up we want to make sure that its quota is not already
4869 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4870 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4872 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4874 if (!cfs_bandwidth_used())
4877 /* an active group must be handled by the update_curr()->put() path */
4878 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4881 /* ensure the group is not already throttled */
4882 if (cfs_rq_throttled(cfs_rq))
4885 /* update runtime allocation */
4886 account_cfs_rq_runtime(cfs_rq, 0);
4887 if (cfs_rq->runtime_remaining <= 0)
4888 throttle_cfs_rq(cfs_rq);
4891 static void sync_throttle(struct task_group *tg, int cpu)
4893 struct cfs_rq *pcfs_rq, *cfs_rq;
4895 if (!cfs_bandwidth_used())
4901 cfs_rq = tg->cfs_rq[cpu];
4902 pcfs_rq = tg->parent->cfs_rq[cpu];
4904 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4905 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4908 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4909 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4911 if (!cfs_bandwidth_used())
4914 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4918 * it's possible for a throttled entity to be forced into a running
4919 * state (e.g. set_curr_task), in this case we're finished.
4921 if (cfs_rq_throttled(cfs_rq))
4924 throttle_cfs_rq(cfs_rq);
4928 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4930 struct cfs_bandwidth *cfs_b =
4931 container_of(timer, struct cfs_bandwidth, slack_timer);
4933 do_sched_cfs_slack_timer(cfs_b);
4935 return HRTIMER_NORESTART;
4938 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4940 struct cfs_bandwidth *cfs_b =
4941 container_of(timer, struct cfs_bandwidth, period_timer);
4945 raw_spin_lock(&cfs_b->lock);
4947 overrun = hrtimer_forward_now(timer, cfs_b->period);
4951 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4954 cfs_b->period_active = 0;
4955 raw_spin_unlock(&cfs_b->lock);
4957 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4960 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4962 raw_spin_lock_init(&cfs_b->lock);
4964 cfs_b->quota = RUNTIME_INF;
4965 cfs_b->period = ns_to_ktime(default_cfs_period());
4967 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4968 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4969 cfs_b->period_timer.function = sched_cfs_period_timer;
4970 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4971 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4974 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4976 cfs_rq->runtime_enabled = 0;
4977 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4980 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4982 lockdep_assert_held(&cfs_b->lock);
4984 if (!cfs_b->period_active) {
4985 cfs_b->period_active = 1;
4986 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4987 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4991 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4993 /* init_cfs_bandwidth() was not called */
4994 if (!cfs_b->throttled_cfs_rq.next)
4997 hrtimer_cancel(&cfs_b->period_timer);
4998 hrtimer_cancel(&cfs_b->slack_timer);
5002 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
5004 * The race is harmless, since modifying bandwidth settings of unhooked group
5005 * bits doesn't do much.
5008 /* cpu online calback */
5009 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5011 struct task_group *tg;
5013 lockdep_assert_held(&rq->lock);
5016 list_for_each_entry_rcu(tg, &task_groups, list) {
5017 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5018 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5020 raw_spin_lock(&cfs_b->lock);
5021 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5022 raw_spin_unlock(&cfs_b->lock);
5027 /* cpu offline callback */
5028 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5030 struct task_group *tg;
5032 lockdep_assert_held(&rq->lock);
5035 list_for_each_entry_rcu(tg, &task_groups, list) {
5036 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5038 if (!cfs_rq->runtime_enabled)
5042 * clock_task is not advancing so we just need to make sure
5043 * there's some valid quota amount
5045 cfs_rq->runtime_remaining = 1;
5047 * Offline rq is schedulable till cpu is completely disabled
5048 * in take_cpu_down(), so we prevent new cfs throttling here.
5050 cfs_rq->runtime_enabled = 0;
5052 if (cfs_rq_throttled(cfs_rq))
5053 unthrottle_cfs_rq(cfs_rq);
5058 #else /* CONFIG_CFS_BANDWIDTH */
5059 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5061 return rq_clock_task(rq_of(cfs_rq));
5064 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5065 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5066 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5067 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5068 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5070 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5075 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5080 static inline int throttled_lb_pair(struct task_group *tg,
5081 int src_cpu, int dest_cpu)
5086 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5088 #ifdef CONFIG_FAIR_GROUP_SCHED
5089 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5092 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5096 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5097 static inline void update_runtime_enabled(struct rq *rq) {}
5098 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5100 #endif /* CONFIG_CFS_BANDWIDTH */
5102 /**************************************************
5103 * CFS operations on tasks:
5106 #ifdef CONFIG_SCHED_HRTICK
5107 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5109 struct sched_entity *se = &p->se;
5110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5112 SCHED_WARN_ON(task_rq(p) != rq);
5114 if (rq->cfs.h_nr_running > 1) {
5115 u64 slice = sched_slice(cfs_rq, se);
5116 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5117 s64 delta = slice - ran;
5124 hrtick_start(rq, delta);
5129 * called from enqueue/dequeue and updates the hrtick when the
5130 * current task is from our class and nr_running is low enough
5133 static void hrtick_update(struct rq *rq)
5135 struct task_struct *curr = rq->curr;
5137 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5140 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5141 hrtick_start_fair(rq, curr);
5143 #else /* !CONFIG_SCHED_HRTICK */
5145 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5149 static inline void hrtick_update(struct rq *rq)
5155 * The enqueue_task method is called before nr_running is
5156 * increased. Here we update the fair scheduling stats and
5157 * then put the task into the rbtree:
5160 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5162 struct cfs_rq *cfs_rq;
5163 struct sched_entity *se = &p->se;
5166 * If in_iowait is set, the code below may not trigger any cpufreq
5167 * utilization updates, so do it here explicitly with the IOWAIT flag
5171 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5173 for_each_sched_entity(se) {
5176 cfs_rq = cfs_rq_of(se);
5177 enqueue_entity(cfs_rq, se, flags);
5180 * end evaluation on encountering a throttled cfs_rq
5182 * note: in the case of encountering a throttled cfs_rq we will
5183 * post the final h_nr_running increment below.
5185 if (cfs_rq_throttled(cfs_rq))
5187 cfs_rq->h_nr_running++;
5189 flags = ENQUEUE_WAKEUP;
5192 for_each_sched_entity(se) {
5193 cfs_rq = cfs_rq_of(se);
5194 cfs_rq->h_nr_running++;
5196 if (cfs_rq_throttled(cfs_rq))
5199 update_load_avg(cfs_rq, se, UPDATE_TG);
5200 update_cfs_group(se);
5204 add_nr_running(rq, 1);
5209 static void set_next_buddy(struct sched_entity *se);
5212 * The dequeue_task method is called before nr_running is
5213 * decreased. We remove the task from the rbtree and
5214 * update the fair scheduling stats:
5216 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5218 struct cfs_rq *cfs_rq;
5219 struct sched_entity *se = &p->se;
5220 int task_sleep = flags & DEQUEUE_SLEEP;
5222 for_each_sched_entity(se) {
5223 cfs_rq = cfs_rq_of(se);
5224 dequeue_entity(cfs_rq, se, flags);
5227 * end evaluation on encountering a throttled cfs_rq
5229 * note: in the case of encountering a throttled cfs_rq we will
5230 * post the final h_nr_running decrement below.
5232 if (cfs_rq_throttled(cfs_rq))
5234 cfs_rq->h_nr_running--;
5236 /* Don't dequeue parent if it has other entities besides us */
5237 if (cfs_rq->load.weight) {
5238 /* Avoid re-evaluating load for this entity: */
5239 se = parent_entity(se);
5241 * Bias pick_next to pick a task from this cfs_rq, as
5242 * p is sleeping when it is within its sched_slice.
5244 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5248 flags |= DEQUEUE_SLEEP;
5251 for_each_sched_entity(se) {
5252 cfs_rq = cfs_rq_of(se);
5253 cfs_rq->h_nr_running--;
5255 if (cfs_rq_throttled(cfs_rq))
5258 update_load_avg(cfs_rq, se, UPDATE_TG);
5259 update_cfs_group(se);
5263 sub_nr_running(rq, 1);
5270 /* Working cpumask for: load_balance, load_balance_newidle. */
5271 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5272 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5274 #ifdef CONFIG_NO_HZ_COMMON
5276 * per rq 'load' arrray crap; XXX kill this.
5280 * The exact cpuload calculated at every tick would be:
5282 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5284 * If a cpu misses updates for n ticks (as it was idle) and update gets
5285 * called on the n+1-th tick when cpu may be busy, then we have:
5287 * load_n = (1 - 1/2^i)^n * load_0
5288 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5290 * decay_load_missed() below does efficient calculation of
5292 * load' = (1 - 1/2^i)^n * load
5294 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5295 * This allows us to precompute the above in said factors, thereby allowing the
5296 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5297 * fixed_power_int())
5299 * The calculation is approximated on a 128 point scale.
5301 #define DEGRADE_SHIFT 7
5303 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5304 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5305 { 0, 0, 0, 0, 0, 0, 0, 0 },
5306 { 64, 32, 8, 0, 0, 0, 0, 0 },
5307 { 96, 72, 40, 12, 1, 0, 0, 0 },
5308 { 112, 98, 75, 43, 15, 1, 0, 0 },
5309 { 120, 112, 98, 76, 45, 16, 2, 0 }
5313 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5314 * would be when CPU is idle and so we just decay the old load without
5315 * adding any new load.
5317 static unsigned long
5318 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5322 if (!missed_updates)
5325 if (missed_updates >= degrade_zero_ticks[idx])
5329 return load >> missed_updates;
5331 while (missed_updates) {
5332 if (missed_updates % 2)
5333 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5335 missed_updates >>= 1;
5340 #endif /* CONFIG_NO_HZ_COMMON */
5343 * __cpu_load_update - update the rq->cpu_load[] statistics
5344 * @this_rq: The rq to update statistics for
5345 * @this_load: The current load
5346 * @pending_updates: The number of missed updates
5348 * Update rq->cpu_load[] statistics. This function is usually called every
5349 * scheduler tick (TICK_NSEC).
5351 * This function computes a decaying average:
5353 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5355 * Because of NOHZ it might not get called on every tick which gives need for
5356 * the @pending_updates argument.
5358 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5359 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5360 * = A * (A * load[i]_n-2 + B) + B
5361 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5362 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5363 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5364 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5365 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5367 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5368 * any change in load would have resulted in the tick being turned back on.
5370 * For regular NOHZ, this reduces to:
5372 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5374 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5377 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5378 unsigned long pending_updates)
5380 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5383 this_rq->nr_load_updates++;
5385 /* Update our load: */
5386 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5387 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5388 unsigned long old_load, new_load;
5390 /* scale is effectively 1 << i now, and >> i divides by scale */
5392 old_load = this_rq->cpu_load[i];
5393 #ifdef CONFIG_NO_HZ_COMMON
5394 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5395 if (tickless_load) {
5396 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5398 * old_load can never be a negative value because a
5399 * decayed tickless_load cannot be greater than the
5400 * original tickless_load.
5402 old_load += tickless_load;
5405 new_load = this_load;
5407 * Round up the averaging division if load is increasing. This
5408 * prevents us from getting stuck on 9 if the load is 10, for
5411 if (new_load > old_load)
5412 new_load += scale - 1;
5414 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5417 sched_avg_update(this_rq);
5420 /* Used instead of source_load when we know the type == 0 */
5421 static unsigned long weighted_cpuload(struct rq *rq)
5423 return cfs_rq_runnable_load_avg(&rq->cfs);
5426 #ifdef CONFIG_NO_HZ_COMMON
5428 * There is no sane way to deal with nohz on smp when using jiffies because the
5429 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5430 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5432 * Therefore we need to avoid the delta approach from the regular tick when
5433 * possible since that would seriously skew the load calculation. This is why we
5434 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5435 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5436 * loop exit, nohz_idle_balance, nohz full exit...)
5438 * This means we might still be one tick off for nohz periods.
5441 static void cpu_load_update_nohz(struct rq *this_rq,
5442 unsigned long curr_jiffies,
5445 unsigned long pending_updates;
5447 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5448 if (pending_updates) {
5449 this_rq->last_load_update_tick = curr_jiffies;
5451 * In the regular NOHZ case, we were idle, this means load 0.
5452 * In the NOHZ_FULL case, we were non-idle, we should consider
5453 * its weighted load.
5455 cpu_load_update(this_rq, load, pending_updates);
5460 * Called from nohz_idle_balance() to update the load ratings before doing the
5463 static void cpu_load_update_idle(struct rq *this_rq)
5466 * bail if there's load or we're actually up-to-date.
5468 if (weighted_cpuload(this_rq))
5471 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5475 * Record CPU load on nohz entry so we know the tickless load to account
5476 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5477 * than other cpu_load[idx] but it should be fine as cpu_load readers
5478 * shouldn't rely into synchronized cpu_load[*] updates.
5480 void cpu_load_update_nohz_start(void)
5482 struct rq *this_rq = this_rq();
5485 * This is all lockless but should be fine. If weighted_cpuload changes
5486 * concurrently we'll exit nohz. And cpu_load write can race with
5487 * cpu_load_update_idle() but both updater would be writing the same.
5489 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5493 * Account the tickless load in the end of a nohz frame.
5495 void cpu_load_update_nohz_stop(void)
5497 unsigned long curr_jiffies = READ_ONCE(jiffies);
5498 struct rq *this_rq = this_rq();
5502 if (curr_jiffies == this_rq->last_load_update_tick)
5505 load = weighted_cpuload(this_rq);
5506 rq_lock(this_rq, &rf);
5507 update_rq_clock(this_rq);
5508 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5509 rq_unlock(this_rq, &rf);
5511 #else /* !CONFIG_NO_HZ_COMMON */
5512 static inline void cpu_load_update_nohz(struct rq *this_rq,
5513 unsigned long curr_jiffies,
5514 unsigned long load) { }
5515 #endif /* CONFIG_NO_HZ_COMMON */
5517 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5519 #ifdef CONFIG_NO_HZ_COMMON
5520 /* See the mess around cpu_load_update_nohz(). */
5521 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5523 cpu_load_update(this_rq, load, 1);
5527 * Called from scheduler_tick()
5529 void cpu_load_update_active(struct rq *this_rq)
5531 unsigned long load = weighted_cpuload(this_rq);
5533 if (tick_nohz_tick_stopped())
5534 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5536 cpu_load_update_periodic(this_rq, load);
5540 * Return a low guess at the load of a migration-source cpu weighted
5541 * according to the scheduling class and "nice" value.
5543 * We want to under-estimate the load of migration sources, to
5544 * balance conservatively.
5546 static unsigned long source_load(int cpu, int type)
5548 struct rq *rq = cpu_rq(cpu);
5549 unsigned long total = weighted_cpuload(rq);
5551 if (type == 0 || !sched_feat(LB_BIAS))
5554 return min(rq->cpu_load[type-1], total);
5558 * Return a high guess at the load of a migration-target cpu weighted
5559 * according to the scheduling class and "nice" value.
5561 static unsigned long target_load(int cpu, int type)
5563 struct rq *rq = cpu_rq(cpu);
5564 unsigned long total = weighted_cpuload(rq);
5566 if (type == 0 || !sched_feat(LB_BIAS))
5569 return max(rq->cpu_load[type-1], total);
5572 static unsigned long capacity_of(int cpu)
5574 return cpu_rq(cpu)->cpu_capacity;
5577 static unsigned long capacity_orig_of(int cpu)
5579 return cpu_rq(cpu)->cpu_capacity_orig;
5582 static unsigned long cpu_avg_load_per_task(int cpu)
5584 struct rq *rq = cpu_rq(cpu);
5585 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5586 unsigned long load_avg = weighted_cpuload(rq);
5589 return load_avg / nr_running;
5594 static void record_wakee(struct task_struct *p)
5597 * Only decay a single time; tasks that have less then 1 wakeup per
5598 * jiffy will not have built up many flips.
5600 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5601 current->wakee_flips >>= 1;
5602 current->wakee_flip_decay_ts = jiffies;
5605 if (current->last_wakee != p) {
5606 current->last_wakee = p;
5607 current->wakee_flips++;
5612 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5614 * A waker of many should wake a different task than the one last awakened
5615 * at a frequency roughly N times higher than one of its wakees.
5617 * In order to determine whether we should let the load spread vs consolidating
5618 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5619 * partner, and a factor of lls_size higher frequency in the other.
5621 * With both conditions met, we can be relatively sure that the relationship is
5622 * non-monogamous, with partner count exceeding socket size.
5624 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5625 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5628 static int wake_wide(struct task_struct *p)
5630 unsigned int master = current->wakee_flips;
5631 unsigned int slave = p->wakee_flips;
5632 int factor = this_cpu_read(sd_llc_size);
5635 swap(master, slave);
5636 if (slave < factor || master < slave * factor)
5642 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5643 * soonest. For the purpose of speed we only consider the waking and previous
5646 * wake_affine_idle() - only considers 'now', it check if the waking CPU is (or
5649 * wake_affine_weight() - considers the weight to reflect the average
5650 * scheduling latency of the CPUs. This seems to work
5651 * for the overloaded case.
5655 wake_affine_idle(struct sched_domain *sd, struct task_struct *p,
5656 int this_cpu, int prev_cpu, int sync)
5658 if (idle_cpu(this_cpu))
5661 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5668 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5669 int this_cpu, int prev_cpu, int sync)
5671 s64 this_eff_load, prev_eff_load;
5672 unsigned long task_load;
5674 this_eff_load = target_load(this_cpu, sd->wake_idx);
5675 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5678 unsigned long current_load = task_h_load(current);
5680 if (current_load > this_eff_load)
5683 this_eff_load -= current_load;
5686 task_load = task_h_load(p);
5688 this_eff_load += task_load;
5689 if (sched_feat(WA_BIAS))
5690 this_eff_load *= 100;
5691 this_eff_load *= capacity_of(prev_cpu);
5693 prev_eff_load -= task_load;
5694 if (sched_feat(WA_BIAS))
5695 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5696 prev_eff_load *= capacity_of(this_cpu);
5698 return this_eff_load <= prev_eff_load;
5701 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5702 int prev_cpu, int sync)
5704 int this_cpu = smp_processor_id();
5705 bool affine = false;
5707 if (sched_feat(WA_IDLE) && !affine)
5708 affine = wake_affine_idle(sd, p, this_cpu, prev_cpu, sync);
5710 if (sched_feat(WA_WEIGHT) && !affine)
5711 affine = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5713 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5715 schedstat_inc(sd->ttwu_move_affine);
5716 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5722 static inline int task_util(struct task_struct *p);
5723 static int cpu_util_wake(int cpu, struct task_struct *p);
5725 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5727 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5731 * find_idlest_group finds and returns the least busy CPU group within the
5734 static struct sched_group *
5735 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5736 int this_cpu, int sd_flag)
5738 struct sched_group *idlest = NULL, *group = sd->groups;
5739 struct sched_group *most_spare_sg = NULL;
5740 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5741 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5742 unsigned long most_spare = 0, this_spare = 0;
5743 int load_idx = sd->forkexec_idx;
5744 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5745 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5746 (sd->imbalance_pct-100) / 100;
5748 if (sd_flag & SD_BALANCE_WAKE)
5749 load_idx = sd->wake_idx;
5752 unsigned long load, avg_load, runnable_load;
5753 unsigned long spare_cap, max_spare_cap;
5757 /* Skip over this group if it has no CPUs allowed */
5758 if (!cpumask_intersects(sched_group_span(group),
5762 local_group = cpumask_test_cpu(this_cpu,
5763 sched_group_span(group));
5766 * Tally up the load of all CPUs in the group and find
5767 * the group containing the CPU with most spare capacity.
5773 for_each_cpu(i, sched_group_span(group)) {
5774 /* Bias balancing toward cpus of our domain */
5776 load = source_load(i, load_idx);
5778 load = target_load(i, load_idx);
5780 runnable_load += load;
5782 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5784 spare_cap = capacity_spare_wake(i, p);
5786 if (spare_cap > max_spare_cap)
5787 max_spare_cap = spare_cap;
5790 /* Adjust by relative CPU capacity of the group */
5791 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5792 group->sgc->capacity;
5793 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5794 group->sgc->capacity;
5797 this_runnable_load = runnable_load;
5798 this_avg_load = avg_load;
5799 this_spare = max_spare_cap;
5801 if (min_runnable_load > (runnable_load + imbalance)) {
5803 * The runnable load is significantly smaller
5804 * so we can pick this new cpu
5806 min_runnable_load = runnable_load;
5807 min_avg_load = avg_load;
5809 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5810 (100*min_avg_load > imbalance_scale*avg_load)) {
5812 * The runnable loads are close so take the
5813 * blocked load into account through avg_load.
5815 min_avg_load = avg_load;
5819 if (most_spare < max_spare_cap) {
5820 most_spare = max_spare_cap;
5821 most_spare_sg = group;
5824 } while (group = group->next, group != sd->groups);
5827 * The cross-over point between using spare capacity or least load
5828 * is too conservative for high utilization tasks on partially
5829 * utilized systems if we require spare_capacity > task_util(p),
5830 * so we allow for some task stuffing by using
5831 * spare_capacity > task_util(p)/2.
5833 * Spare capacity can't be used for fork because the utilization has
5834 * not been set yet, we must first select a rq to compute the initial
5837 if (sd_flag & SD_BALANCE_FORK)
5840 if (this_spare > task_util(p) / 2 &&
5841 imbalance_scale*this_spare > 100*most_spare)
5844 if (most_spare > task_util(p) / 2)
5845 return most_spare_sg;
5851 if (min_runnable_load > (this_runnable_load + imbalance))
5854 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5855 (100*this_avg_load < imbalance_scale*min_avg_load))
5862 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5865 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5867 unsigned long load, min_load = ULONG_MAX;
5868 unsigned int min_exit_latency = UINT_MAX;
5869 u64 latest_idle_timestamp = 0;
5870 int least_loaded_cpu = this_cpu;
5871 int shallowest_idle_cpu = -1;
5874 /* Check if we have any choice: */
5875 if (group->group_weight == 1)
5876 return cpumask_first(sched_group_span(group));
5878 /* Traverse only the allowed CPUs */
5879 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5881 struct rq *rq = cpu_rq(i);
5882 struct cpuidle_state *idle = idle_get_state(rq);
5883 if (idle && idle->exit_latency < min_exit_latency) {
5885 * We give priority to a CPU whose idle state
5886 * has the smallest exit latency irrespective
5887 * of any idle timestamp.
5889 min_exit_latency = idle->exit_latency;
5890 latest_idle_timestamp = rq->idle_stamp;
5891 shallowest_idle_cpu = i;
5892 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5893 rq->idle_stamp > latest_idle_timestamp) {
5895 * If equal or no active idle state, then
5896 * the most recently idled CPU might have
5899 latest_idle_timestamp = rq->idle_stamp;
5900 shallowest_idle_cpu = i;
5902 } else if (shallowest_idle_cpu == -1) {
5903 load = weighted_cpuload(cpu_rq(i));
5904 if (load < min_load || (load == min_load && i == this_cpu)) {
5906 least_loaded_cpu = i;
5911 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5914 #ifdef CONFIG_SCHED_SMT
5916 static inline void set_idle_cores(int cpu, int val)
5918 struct sched_domain_shared *sds;
5920 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5922 WRITE_ONCE(sds->has_idle_cores, val);
5925 static inline bool test_idle_cores(int cpu, bool def)
5927 struct sched_domain_shared *sds;
5929 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5931 return READ_ONCE(sds->has_idle_cores);
5937 * Scans the local SMT mask to see if the entire core is idle, and records this
5938 * information in sd_llc_shared->has_idle_cores.
5940 * Since SMT siblings share all cache levels, inspecting this limited remote
5941 * state should be fairly cheap.
5943 void __update_idle_core(struct rq *rq)
5945 int core = cpu_of(rq);
5949 if (test_idle_cores(core, true))
5952 for_each_cpu(cpu, cpu_smt_mask(core)) {
5960 set_idle_cores(core, 1);
5966 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5967 * there are no idle cores left in the system; tracked through
5968 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5970 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5972 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5975 if (!static_branch_likely(&sched_smt_present))
5978 if (!test_idle_cores(target, false))
5981 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5983 for_each_cpu_wrap(core, cpus, target) {
5986 for_each_cpu(cpu, cpu_smt_mask(core)) {
5987 cpumask_clear_cpu(cpu, cpus);
5997 * Failed to find an idle core; stop looking for one.
5999 set_idle_cores(target, 0);
6005 * Scan the local SMT mask for idle CPUs.
6007 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6011 if (!static_branch_likely(&sched_smt_present))
6014 for_each_cpu(cpu, cpu_smt_mask(target)) {
6015 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6024 #else /* CONFIG_SCHED_SMT */
6026 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6031 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6036 #endif /* CONFIG_SCHED_SMT */
6039 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6040 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6041 * average idle time for this rq (as found in rq->avg_idle).
6043 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6045 struct sched_domain *this_sd;
6046 u64 avg_cost, avg_idle;
6049 int cpu, nr = INT_MAX;
6051 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6056 * Due to large variance we need a large fuzz factor; hackbench in
6057 * particularly is sensitive here.
6059 avg_idle = this_rq()->avg_idle / 512;
6060 avg_cost = this_sd->avg_scan_cost + 1;
6062 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6065 if (sched_feat(SIS_PROP)) {
6066 u64 span_avg = sd->span_weight * avg_idle;
6067 if (span_avg > 4*avg_cost)
6068 nr = div_u64(span_avg, avg_cost);
6073 time = local_clock();
6075 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6078 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6084 time = local_clock() - time;
6085 cost = this_sd->avg_scan_cost;
6086 delta = (s64)(time - cost) / 8;
6087 this_sd->avg_scan_cost += delta;
6093 * Try and locate an idle core/thread in the LLC cache domain.
6095 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6097 struct sched_domain *sd;
6100 if (idle_cpu(target))
6104 * If the previous cpu is cache affine and idle, don't be stupid.
6106 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6109 sd = rcu_dereference(per_cpu(sd_llc, target));
6113 i = select_idle_core(p, sd, target);
6114 if ((unsigned)i < nr_cpumask_bits)
6117 i = select_idle_cpu(p, sd, target);
6118 if ((unsigned)i < nr_cpumask_bits)
6121 i = select_idle_smt(p, sd, target);
6122 if ((unsigned)i < nr_cpumask_bits)
6129 * cpu_util returns the amount of capacity of a CPU that is used by CFS
6130 * tasks. The unit of the return value must be the one of capacity so we can
6131 * compare the utilization with the capacity of the CPU that is available for
6132 * CFS task (ie cpu_capacity).
6134 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6135 * recent utilization of currently non-runnable tasks on a CPU. It represents
6136 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6137 * capacity_orig is the cpu_capacity available at the highest frequency
6138 * (arch_scale_freq_capacity()).
6139 * The utilization of a CPU converges towards a sum equal to or less than the
6140 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6141 * the running time on this CPU scaled by capacity_curr.
6143 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6144 * higher than capacity_orig because of unfortunate rounding in
6145 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6146 * the average stabilizes with the new running time. We need to check that the
6147 * utilization stays within the range of [0..capacity_orig] and cap it if
6148 * necessary. Without utilization capping, a group could be seen as overloaded
6149 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6150 * available capacity. We allow utilization to overshoot capacity_curr (but not
6151 * capacity_orig) as it useful for predicting the capacity required after task
6152 * migrations (scheduler-driven DVFS).
6154 static int cpu_util(int cpu)
6156 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
6157 unsigned long capacity = capacity_orig_of(cpu);
6159 return (util >= capacity) ? capacity : util;
6162 static inline int task_util(struct task_struct *p)
6164 return p->se.avg.util_avg;
6168 * cpu_util_wake: Compute cpu utilization with any contributions from
6169 * the waking task p removed.
6171 static int cpu_util_wake(int cpu, struct task_struct *p)
6173 unsigned long util, capacity;
6175 /* Task has no contribution or is new */
6176 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6177 return cpu_util(cpu);
6179 capacity = capacity_orig_of(cpu);
6180 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6182 return (util >= capacity) ? capacity : util;
6186 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6187 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6189 * In that case WAKE_AFFINE doesn't make sense and we'll let
6190 * BALANCE_WAKE sort things out.
6192 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6194 long min_cap, max_cap;
6196 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6197 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6199 /* Minimum capacity is close to max, no need to abort wake_affine */
6200 if (max_cap - min_cap < max_cap >> 3)
6203 /* Bring task utilization in sync with prev_cpu */
6204 sync_entity_load_avg(&p->se);
6206 return min_cap * 1024 < task_util(p) * capacity_margin;
6210 * select_task_rq_fair: Select target runqueue for the waking task in domains
6211 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6212 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6214 * Balances load by selecting the idlest cpu in the idlest group, or under
6215 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6217 * Returns the target cpu number.
6219 * preempt must be disabled.
6222 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6224 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6225 int cpu = smp_processor_id();
6226 int new_cpu = prev_cpu;
6227 int want_affine = 0;
6228 int sync = wake_flags & WF_SYNC;
6230 if (sd_flag & SD_BALANCE_WAKE) {
6232 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6233 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6237 for_each_domain(cpu, tmp) {
6238 if (!(tmp->flags & SD_LOAD_BALANCE))
6242 * If both cpu and prev_cpu are part of this domain,
6243 * cpu is a valid SD_WAKE_AFFINE target.
6245 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6246 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6251 if (tmp->flags & sd_flag)
6253 else if (!want_affine)
6258 sd = NULL; /* Prefer wake_affine over balance flags */
6259 if (cpu == prev_cpu)
6262 if (wake_affine(affine_sd, p, prev_cpu, sync))
6268 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6269 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6272 struct sched_group *group;
6275 if (!(sd->flags & sd_flag)) {
6280 group = find_idlest_group(sd, p, cpu, sd_flag);
6286 new_cpu = find_idlest_cpu(group, p, cpu);
6287 if (new_cpu == -1 || new_cpu == cpu) {
6288 /* Now try balancing at a lower domain level of cpu */
6293 /* Now try balancing at a lower domain level of new_cpu */
6295 weight = sd->span_weight;
6297 for_each_domain(cpu, tmp) {
6298 if (weight <= tmp->span_weight)
6300 if (tmp->flags & sd_flag)
6303 /* while loop will break here if sd == NULL */
6310 static void detach_entity_cfs_rq(struct sched_entity *se);
6313 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6314 * cfs_rq_of(p) references at time of call are still valid and identify the
6315 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6317 static void migrate_task_rq_fair(struct task_struct *p)
6320 * As blocked tasks retain absolute vruntime the migration needs to
6321 * deal with this by subtracting the old and adding the new
6322 * min_vruntime -- the latter is done by enqueue_entity() when placing
6323 * the task on the new runqueue.
6325 if (p->state == TASK_WAKING) {
6326 struct sched_entity *se = &p->se;
6327 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6330 #ifndef CONFIG_64BIT
6331 u64 min_vruntime_copy;
6334 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6336 min_vruntime = cfs_rq->min_vruntime;
6337 } while (min_vruntime != min_vruntime_copy);
6339 min_vruntime = cfs_rq->min_vruntime;
6342 se->vruntime -= min_vruntime;
6345 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6347 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6348 * rq->lock and can modify state directly.
6350 lockdep_assert_held(&task_rq(p)->lock);
6351 detach_entity_cfs_rq(&p->se);
6355 * We are supposed to update the task to "current" time, then
6356 * its up to date and ready to go to new CPU/cfs_rq. But we
6357 * have difficulty in getting what current time is, so simply
6358 * throw away the out-of-date time. This will result in the
6359 * wakee task is less decayed, but giving the wakee more load
6362 remove_entity_load_avg(&p->se);
6365 /* Tell new CPU we are migrated */
6366 p->se.avg.last_update_time = 0;
6368 /* We have migrated, no longer consider this task hot */
6369 p->se.exec_start = 0;
6372 static void task_dead_fair(struct task_struct *p)
6374 remove_entity_load_avg(&p->se);
6376 #endif /* CONFIG_SMP */
6378 static unsigned long
6379 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6381 unsigned long gran = sysctl_sched_wakeup_granularity;
6384 * Since its curr running now, convert the gran from real-time
6385 * to virtual-time in his units.
6387 * By using 'se' instead of 'curr' we penalize light tasks, so
6388 * they get preempted easier. That is, if 'se' < 'curr' then
6389 * the resulting gran will be larger, therefore penalizing the
6390 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6391 * be smaller, again penalizing the lighter task.
6393 * This is especially important for buddies when the leftmost
6394 * task is higher priority than the buddy.
6396 return calc_delta_fair(gran, se);
6400 * Should 'se' preempt 'curr'.
6414 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6416 s64 gran, vdiff = curr->vruntime - se->vruntime;
6421 gran = wakeup_gran(curr, se);
6428 static void set_last_buddy(struct sched_entity *se)
6430 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6433 for_each_sched_entity(se) {
6434 if (SCHED_WARN_ON(!se->on_rq))
6436 cfs_rq_of(se)->last = se;
6440 static void set_next_buddy(struct sched_entity *se)
6442 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6445 for_each_sched_entity(se) {
6446 if (SCHED_WARN_ON(!se->on_rq))
6448 cfs_rq_of(se)->next = se;
6452 static void set_skip_buddy(struct sched_entity *se)
6454 for_each_sched_entity(se)
6455 cfs_rq_of(se)->skip = se;
6459 * Preempt the current task with a newly woken task if needed:
6461 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6463 struct task_struct *curr = rq->curr;
6464 struct sched_entity *se = &curr->se, *pse = &p->se;
6465 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6466 int scale = cfs_rq->nr_running >= sched_nr_latency;
6467 int next_buddy_marked = 0;
6469 if (unlikely(se == pse))
6473 * This is possible from callers such as attach_tasks(), in which we
6474 * unconditionally check_prempt_curr() after an enqueue (which may have
6475 * lead to a throttle). This both saves work and prevents false
6476 * next-buddy nomination below.
6478 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6481 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6482 set_next_buddy(pse);
6483 next_buddy_marked = 1;
6487 * We can come here with TIF_NEED_RESCHED already set from new task
6490 * Note: this also catches the edge-case of curr being in a throttled
6491 * group (e.g. via set_curr_task), since update_curr() (in the
6492 * enqueue of curr) will have resulted in resched being set. This
6493 * prevents us from potentially nominating it as a false LAST_BUDDY
6496 if (test_tsk_need_resched(curr))
6499 /* Idle tasks are by definition preempted by non-idle tasks. */
6500 if (unlikely(curr->policy == SCHED_IDLE) &&
6501 likely(p->policy != SCHED_IDLE))
6505 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6506 * is driven by the tick):
6508 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6511 find_matching_se(&se, &pse);
6512 update_curr(cfs_rq_of(se));
6514 if (wakeup_preempt_entity(se, pse) == 1) {
6516 * Bias pick_next to pick the sched entity that is
6517 * triggering this preemption.
6519 if (!next_buddy_marked)
6520 set_next_buddy(pse);
6529 * Only set the backward buddy when the current task is still
6530 * on the rq. This can happen when a wakeup gets interleaved
6531 * with schedule on the ->pre_schedule() or idle_balance()
6532 * point, either of which can * drop the rq lock.
6534 * Also, during early boot the idle thread is in the fair class,
6535 * for obvious reasons its a bad idea to schedule back to it.
6537 if (unlikely(!se->on_rq || curr == rq->idle))
6540 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6544 static struct task_struct *
6545 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6547 struct cfs_rq *cfs_rq = &rq->cfs;
6548 struct sched_entity *se;
6549 struct task_struct *p;
6553 if (!cfs_rq->nr_running)
6556 #ifdef CONFIG_FAIR_GROUP_SCHED
6557 if (prev->sched_class != &fair_sched_class)
6561 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6562 * likely that a next task is from the same cgroup as the current.
6564 * Therefore attempt to avoid putting and setting the entire cgroup
6565 * hierarchy, only change the part that actually changes.
6569 struct sched_entity *curr = cfs_rq->curr;
6572 * Since we got here without doing put_prev_entity() we also
6573 * have to consider cfs_rq->curr. If it is still a runnable
6574 * entity, update_curr() will update its vruntime, otherwise
6575 * forget we've ever seen it.
6579 update_curr(cfs_rq);
6584 * This call to check_cfs_rq_runtime() will do the
6585 * throttle and dequeue its entity in the parent(s).
6586 * Therefore the nr_running test will indeed
6589 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6592 if (!cfs_rq->nr_running)
6599 se = pick_next_entity(cfs_rq, curr);
6600 cfs_rq = group_cfs_rq(se);
6606 * Since we haven't yet done put_prev_entity and if the selected task
6607 * is a different task than we started out with, try and touch the
6608 * least amount of cfs_rqs.
6611 struct sched_entity *pse = &prev->se;
6613 while (!(cfs_rq = is_same_group(se, pse))) {
6614 int se_depth = se->depth;
6615 int pse_depth = pse->depth;
6617 if (se_depth <= pse_depth) {
6618 put_prev_entity(cfs_rq_of(pse), pse);
6619 pse = parent_entity(pse);
6621 if (se_depth >= pse_depth) {
6622 set_next_entity(cfs_rq_of(se), se);
6623 se = parent_entity(se);
6627 put_prev_entity(cfs_rq, pse);
6628 set_next_entity(cfs_rq, se);
6631 if (hrtick_enabled(rq))
6632 hrtick_start_fair(rq, p);
6638 put_prev_task(rq, prev);
6641 se = pick_next_entity(cfs_rq, NULL);
6642 set_next_entity(cfs_rq, se);
6643 cfs_rq = group_cfs_rq(se);
6648 if (hrtick_enabled(rq))
6649 hrtick_start_fair(rq, p);
6654 new_tasks = idle_balance(rq, rf);
6657 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6658 * possible for any higher priority task to appear. In that case we
6659 * must re-start the pick_next_entity() loop.
6671 * Account for a descheduled task:
6673 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6675 struct sched_entity *se = &prev->se;
6676 struct cfs_rq *cfs_rq;
6678 for_each_sched_entity(se) {
6679 cfs_rq = cfs_rq_of(se);
6680 put_prev_entity(cfs_rq, se);
6685 * sched_yield() is very simple
6687 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6689 static void yield_task_fair(struct rq *rq)
6691 struct task_struct *curr = rq->curr;
6692 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6693 struct sched_entity *se = &curr->se;
6696 * Are we the only task in the tree?
6698 if (unlikely(rq->nr_running == 1))
6701 clear_buddies(cfs_rq, se);
6703 if (curr->policy != SCHED_BATCH) {
6704 update_rq_clock(rq);
6706 * Update run-time statistics of the 'current'.
6708 update_curr(cfs_rq);
6710 * Tell update_rq_clock() that we've just updated,
6711 * so we don't do microscopic update in schedule()
6712 * and double the fastpath cost.
6714 rq_clock_skip_update(rq, true);
6720 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6722 struct sched_entity *se = &p->se;
6724 /* throttled hierarchies are not runnable */
6725 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6728 /* Tell the scheduler that we'd really like pse to run next. */
6731 yield_task_fair(rq);
6737 /**************************************************
6738 * Fair scheduling class load-balancing methods.
6742 * The purpose of load-balancing is to achieve the same basic fairness the
6743 * per-cpu scheduler provides, namely provide a proportional amount of compute
6744 * time to each task. This is expressed in the following equation:
6746 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6748 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6749 * W_i,0 is defined as:
6751 * W_i,0 = \Sum_j w_i,j (2)
6753 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6754 * is derived from the nice value as per sched_prio_to_weight[].
6756 * The weight average is an exponential decay average of the instantaneous
6759 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6761 * C_i is the compute capacity of cpu i, typically it is the
6762 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6763 * can also include other factors [XXX].
6765 * To achieve this balance we define a measure of imbalance which follows
6766 * directly from (1):
6768 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6770 * We them move tasks around to minimize the imbalance. In the continuous
6771 * function space it is obvious this converges, in the discrete case we get
6772 * a few fun cases generally called infeasible weight scenarios.
6775 * - infeasible weights;
6776 * - local vs global optima in the discrete case. ]
6781 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6782 * for all i,j solution, we create a tree of cpus that follows the hardware
6783 * topology where each level pairs two lower groups (or better). This results
6784 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6785 * tree to only the first of the previous level and we decrease the frequency
6786 * of load-balance at each level inv. proportional to the number of cpus in
6792 * \Sum { --- * --- * 2^i } = O(n) (5)
6794 * `- size of each group
6795 * | | `- number of cpus doing load-balance
6797 * `- sum over all levels
6799 * Coupled with a limit on how many tasks we can migrate every balance pass,
6800 * this makes (5) the runtime complexity of the balancer.
6802 * An important property here is that each CPU is still (indirectly) connected
6803 * to every other cpu in at most O(log n) steps:
6805 * The adjacency matrix of the resulting graph is given by:
6808 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6811 * And you'll find that:
6813 * A^(log_2 n)_i,j != 0 for all i,j (7)
6815 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6816 * The task movement gives a factor of O(m), giving a convergence complexity
6819 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6824 * In order to avoid CPUs going idle while there's still work to do, new idle
6825 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6826 * tree itself instead of relying on other CPUs to bring it work.
6828 * This adds some complexity to both (5) and (8) but it reduces the total idle
6836 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6839 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6844 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6846 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6848 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6851 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6852 * rewrite all of this once again.]
6855 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6857 enum fbq_type { regular, remote, all };
6859 #define LBF_ALL_PINNED 0x01
6860 #define LBF_NEED_BREAK 0x02
6861 #define LBF_DST_PINNED 0x04
6862 #define LBF_SOME_PINNED 0x08
6865 struct sched_domain *sd;
6873 struct cpumask *dst_grpmask;
6875 enum cpu_idle_type idle;
6877 /* The set of CPUs under consideration for load-balancing */
6878 struct cpumask *cpus;
6883 unsigned int loop_break;
6884 unsigned int loop_max;
6886 enum fbq_type fbq_type;
6887 struct list_head tasks;
6891 * Is this task likely cache-hot:
6893 static int task_hot(struct task_struct *p, struct lb_env *env)
6897 lockdep_assert_held(&env->src_rq->lock);
6899 if (p->sched_class != &fair_sched_class)
6902 if (unlikely(p->policy == SCHED_IDLE))
6906 * Buddy candidates are cache hot:
6908 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6909 (&p->se == cfs_rq_of(&p->se)->next ||
6910 &p->se == cfs_rq_of(&p->se)->last))
6913 if (sysctl_sched_migration_cost == -1)
6915 if (sysctl_sched_migration_cost == 0)
6918 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6920 return delta < (s64)sysctl_sched_migration_cost;
6923 #ifdef CONFIG_NUMA_BALANCING
6925 * Returns 1, if task migration degrades locality
6926 * Returns 0, if task migration improves locality i.e migration preferred.
6927 * Returns -1, if task migration is not affected by locality.
6929 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6931 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6932 unsigned long src_faults, dst_faults;
6933 int src_nid, dst_nid;
6935 if (!static_branch_likely(&sched_numa_balancing))
6938 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6941 src_nid = cpu_to_node(env->src_cpu);
6942 dst_nid = cpu_to_node(env->dst_cpu);
6944 if (src_nid == dst_nid)
6947 /* Migrating away from the preferred node is always bad. */
6948 if (src_nid == p->numa_preferred_nid) {
6949 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6955 /* Encourage migration to the preferred node. */
6956 if (dst_nid == p->numa_preferred_nid)
6959 /* Leaving a core idle is often worse than degrading locality. */
6960 if (env->idle != CPU_NOT_IDLE)
6964 src_faults = group_faults(p, src_nid);
6965 dst_faults = group_faults(p, dst_nid);
6967 src_faults = task_faults(p, src_nid);
6968 dst_faults = task_faults(p, dst_nid);
6971 return dst_faults < src_faults;
6975 static inline int migrate_degrades_locality(struct task_struct *p,
6983 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6986 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6990 lockdep_assert_held(&env->src_rq->lock);
6993 * We do not migrate tasks that are:
6994 * 1) throttled_lb_pair, or
6995 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6996 * 3) running (obviously), or
6997 * 4) are cache-hot on their current CPU.
6999 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7002 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7005 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7007 env->flags |= LBF_SOME_PINNED;
7010 * Remember if this task can be migrated to any other cpu in
7011 * our sched_group. We may want to revisit it if we couldn't
7012 * meet load balance goals by pulling other tasks on src_cpu.
7014 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7015 * already computed one in current iteration.
7017 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7020 /* Prevent to re-select dst_cpu via env's cpus */
7021 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7022 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7023 env->flags |= LBF_DST_PINNED;
7024 env->new_dst_cpu = cpu;
7032 /* Record that we found atleast one task that could run on dst_cpu */
7033 env->flags &= ~LBF_ALL_PINNED;
7035 if (task_running(env->src_rq, p)) {
7036 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7041 * Aggressive migration if:
7042 * 1) destination numa is preferred
7043 * 2) task is cache cold, or
7044 * 3) too many balance attempts have failed.
7046 tsk_cache_hot = migrate_degrades_locality(p, env);
7047 if (tsk_cache_hot == -1)
7048 tsk_cache_hot = task_hot(p, env);
7050 if (tsk_cache_hot <= 0 ||
7051 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7052 if (tsk_cache_hot == 1) {
7053 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7054 schedstat_inc(p->se.statistics.nr_forced_migrations);
7059 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7064 * detach_task() -- detach the task for the migration specified in env
7066 static void detach_task(struct task_struct *p, struct lb_env *env)
7068 lockdep_assert_held(&env->src_rq->lock);
7070 p->on_rq = TASK_ON_RQ_MIGRATING;
7071 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7072 set_task_cpu(p, env->dst_cpu);
7076 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7077 * part of active balancing operations within "domain".
7079 * Returns a task if successful and NULL otherwise.
7081 static struct task_struct *detach_one_task(struct lb_env *env)
7083 struct task_struct *p, *n;
7085 lockdep_assert_held(&env->src_rq->lock);
7087 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7088 if (!can_migrate_task(p, env))
7091 detach_task(p, env);
7094 * Right now, this is only the second place where
7095 * lb_gained[env->idle] is updated (other is detach_tasks)
7096 * so we can safely collect stats here rather than
7097 * inside detach_tasks().
7099 schedstat_inc(env->sd->lb_gained[env->idle]);
7105 static const unsigned int sched_nr_migrate_break = 32;
7108 * detach_tasks() -- tries to detach up to imbalance weighted load from
7109 * busiest_rq, as part of a balancing operation within domain "sd".
7111 * Returns number of detached tasks if successful and 0 otherwise.
7113 static int detach_tasks(struct lb_env *env)
7115 struct list_head *tasks = &env->src_rq->cfs_tasks;
7116 struct task_struct *p;
7120 lockdep_assert_held(&env->src_rq->lock);
7122 if (env->imbalance <= 0)
7125 while (!list_empty(tasks)) {
7127 * We don't want to steal all, otherwise we may be treated likewise,
7128 * which could at worst lead to a livelock crash.
7130 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7133 p = list_first_entry(tasks, struct task_struct, se.group_node);
7136 /* We've more or less seen every task there is, call it quits */
7137 if (env->loop > env->loop_max)
7140 /* take a breather every nr_migrate tasks */
7141 if (env->loop > env->loop_break) {
7142 env->loop_break += sched_nr_migrate_break;
7143 env->flags |= LBF_NEED_BREAK;
7147 if (!can_migrate_task(p, env))
7150 load = task_h_load(p);
7152 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7155 if ((load / 2) > env->imbalance)
7158 detach_task(p, env);
7159 list_add(&p->se.group_node, &env->tasks);
7162 env->imbalance -= load;
7164 #ifdef CONFIG_PREEMPT
7166 * NEWIDLE balancing is a source of latency, so preemptible
7167 * kernels will stop after the first task is detached to minimize
7168 * the critical section.
7170 if (env->idle == CPU_NEWLY_IDLE)
7175 * We only want to steal up to the prescribed amount of
7178 if (env->imbalance <= 0)
7183 list_move_tail(&p->se.group_node, tasks);
7187 * Right now, this is one of only two places we collect this stat
7188 * so we can safely collect detach_one_task() stats here rather
7189 * than inside detach_one_task().
7191 schedstat_add(env->sd->lb_gained[env->idle], detached);
7197 * attach_task() -- attach the task detached by detach_task() to its new rq.
7199 static void attach_task(struct rq *rq, struct task_struct *p)
7201 lockdep_assert_held(&rq->lock);
7203 BUG_ON(task_rq(p) != rq);
7204 activate_task(rq, p, ENQUEUE_NOCLOCK);
7205 p->on_rq = TASK_ON_RQ_QUEUED;
7206 check_preempt_curr(rq, p, 0);
7210 * attach_one_task() -- attaches the task returned from detach_one_task() to
7213 static void attach_one_task(struct rq *rq, struct task_struct *p)
7218 update_rq_clock(rq);
7224 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7227 static void attach_tasks(struct lb_env *env)
7229 struct list_head *tasks = &env->tasks;
7230 struct task_struct *p;
7233 rq_lock(env->dst_rq, &rf);
7234 update_rq_clock(env->dst_rq);
7236 while (!list_empty(tasks)) {
7237 p = list_first_entry(tasks, struct task_struct, se.group_node);
7238 list_del_init(&p->se.group_node);
7240 attach_task(env->dst_rq, p);
7243 rq_unlock(env->dst_rq, &rf);
7246 #ifdef CONFIG_FAIR_GROUP_SCHED
7248 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7250 if (cfs_rq->load.weight)
7253 if (cfs_rq->avg.load_sum)
7256 if (cfs_rq->avg.util_sum)
7259 if (cfs_rq->avg.runnable_load_sum)
7265 static void update_blocked_averages(int cpu)
7267 struct rq *rq = cpu_rq(cpu);
7268 struct cfs_rq *cfs_rq, *pos;
7271 rq_lock_irqsave(rq, &rf);
7272 update_rq_clock(rq);
7275 * Iterates the task_group tree in a bottom up fashion, see
7276 * list_add_leaf_cfs_rq() for details.
7278 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7279 struct sched_entity *se;
7281 /* throttled entities do not contribute to load */
7282 if (throttled_hierarchy(cfs_rq))
7285 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7286 update_tg_load_avg(cfs_rq, 0);
7288 /* Propagate pending load changes to the parent, if any: */
7289 se = cfs_rq->tg->se[cpu];
7290 if (se && !skip_blocked_update(se))
7291 update_load_avg(cfs_rq_of(se), se, 0);
7294 * There can be a lot of idle CPU cgroups. Don't let fully
7295 * decayed cfs_rqs linger on the list.
7297 if (cfs_rq_is_decayed(cfs_rq))
7298 list_del_leaf_cfs_rq(cfs_rq);
7300 rq_unlock_irqrestore(rq, &rf);
7304 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7305 * This needs to be done in a top-down fashion because the load of a child
7306 * group is a fraction of its parents load.
7308 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7310 struct rq *rq = rq_of(cfs_rq);
7311 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7312 unsigned long now = jiffies;
7315 if (cfs_rq->last_h_load_update == now)
7318 cfs_rq->h_load_next = NULL;
7319 for_each_sched_entity(se) {
7320 cfs_rq = cfs_rq_of(se);
7321 cfs_rq->h_load_next = se;
7322 if (cfs_rq->last_h_load_update == now)
7327 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7328 cfs_rq->last_h_load_update = now;
7331 while ((se = cfs_rq->h_load_next) != NULL) {
7332 load = cfs_rq->h_load;
7333 load = div64_ul(load * se->avg.load_avg,
7334 cfs_rq_load_avg(cfs_rq) + 1);
7335 cfs_rq = group_cfs_rq(se);
7336 cfs_rq->h_load = load;
7337 cfs_rq->last_h_load_update = now;
7341 static unsigned long task_h_load(struct task_struct *p)
7343 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7345 update_cfs_rq_h_load(cfs_rq);
7346 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7347 cfs_rq_load_avg(cfs_rq) + 1);
7350 static inline void update_blocked_averages(int cpu)
7352 struct rq *rq = cpu_rq(cpu);
7353 struct cfs_rq *cfs_rq = &rq->cfs;
7356 rq_lock_irqsave(rq, &rf);
7357 update_rq_clock(rq);
7358 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7359 rq_unlock_irqrestore(rq, &rf);
7362 static unsigned long task_h_load(struct task_struct *p)
7364 return p->se.avg.load_avg;
7368 /********** Helpers for find_busiest_group ************************/
7377 * sg_lb_stats - stats of a sched_group required for load_balancing
7379 struct sg_lb_stats {
7380 unsigned long avg_load; /*Avg load across the CPUs of the group */
7381 unsigned long group_load; /* Total load over the CPUs of the group */
7382 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7383 unsigned long load_per_task;
7384 unsigned long group_capacity;
7385 unsigned long group_util; /* Total utilization of the group */
7386 unsigned int sum_nr_running; /* Nr tasks running in the group */
7387 unsigned int idle_cpus;
7388 unsigned int group_weight;
7389 enum group_type group_type;
7390 int group_no_capacity;
7391 #ifdef CONFIG_NUMA_BALANCING
7392 unsigned int nr_numa_running;
7393 unsigned int nr_preferred_running;
7398 * sd_lb_stats - Structure to store the statistics of a sched_domain
7399 * during load balancing.
7401 struct sd_lb_stats {
7402 struct sched_group *busiest; /* Busiest group in this sd */
7403 struct sched_group *local; /* Local group in this sd */
7404 unsigned long total_running;
7405 unsigned long total_load; /* Total load of all groups in sd */
7406 unsigned long total_capacity; /* Total capacity of all groups in sd */
7407 unsigned long avg_load; /* Average load across all groups in sd */
7409 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7410 struct sg_lb_stats local_stat; /* Statistics of the local group */
7413 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7416 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7417 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7418 * We must however clear busiest_stat::avg_load because
7419 * update_sd_pick_busiest() reads this before assignment.
7421 *sds = (struct sd_lb_stats){
7424 .total_running = 0UL,
7426 .total_capacity = 0UL,
7429 .sum_nr_running = 0,
7430 .group_type = group_other,
7436 * get_sd_load_idx - Obtain the load index for a given sched domain.
7437 * @sd: The sched_domain whose load_idx is to be obtained.
7438 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7440 * Return: The load index.
7442 static inline int get_sd_load_idx(struct sched_domain *sd,
7443 enum cpu_idle_type idle)
7449 load_idx = sd->busy_idx;
7452 case CPU_NEWLY_IDLE:
7453 load_idx = sd->newidle_idx;
7456 load_idx = sd->idle_idx;
7463 static unsigned long scale_rt_capacity(int cpu)
7465 struct rq *rq = cpu_rq(cpu);
7466 u64 total, used, age_stamp, avg;
7470 * Since we're reading these variables without serialization make sure
7471 * we read them once before doing sanity checks on them.
7473 age_stamp = READ_ONCE(rq->age_stamp);
7474 avg = READ_ONCE(rq->rt_avg);
7475 delta = __rq_clock_broken(rq) - age_stamp;
7477 if (unlikely(delta < 0))
7480 total = sched_avg_period() + delta;
7482 used = div_u64(avg, total);
7484 if (likely(used < SCHED_CAPACITY_SCALE))
7485 return SCHED_CAPACITY_SCALE - used;
7490 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7492 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7493 struct sched_group *sdg = sd->groups;
7495 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7497 capacity *= scale_rt_capacity(cpu);
7498 capacity >>= SCHED_CAPACITY_SHIFT;
7503 cpu_rq(cpu)->cpu_capacity = capacity;
7504 sdg->sgc->capacity = capacity;
7505 sdg->sgc->min_capacity = capacity;
7508 void update_group_capacity(struct sched_domain *sd, int cpu)
7510 struct sched_domain *child = sd->child;
7511 struct sched_group *group, *sdg = sd->groups;
7512 unsigned long capacity, min_capacity;
7513 unsigned long interval;
7515 interval = msecs_to_jiffies(sd->balance_interval);
7516 interval = clamp(interval, 1UL, max_load_balance_interval);
7517 sdg->sgc->next_update = jiffies + interval;
7520 update_cpu_capacity(sd, cpu);
7525 min_capacity = ULONG_MAX;
7527 if (child->flags & SD_OVERLAP) {
7529 * SD_OVERLAP domains cannot assume that child groups
7530 * span the current group.
7533 for_each_cpu(cpu, sched_group_span(sdg)) {
7534 struct sched_group_capacity *sgc;
7535 struct rq *rq = cpu_rq(cpu);
7538 * build_sched_domains() -> init_sched_groups_capacity()
7539 * gets here before we've attached the domains to the
7542 * Use capacity_of(), which is set irrespective of domains
7543 * in update_cpu_capacity().
7545 * This avoids capacity from being 0 and
7546 * causing divide-by-zero issues on boot.
7548 if (unlikely(!rq->sd)) {
7549 capacity += capacity_of(cpu);
7551 sgc = rq->sd->groups->sgc;
7552 capacity += sgc->capacity;
7555 min_capacity = min(capacity, min_capacity);
7559 * !SD_OVERLAP domains can assume that child groups
7560 * span the current group.
7563 group = child->groups;
7565 struct sched_group_capacity *sgc = group->sgc;
7567 capacity += sgc->capacity;
7568 min_capacity = min(sgc->min_capacity, min_capacity);
7569 group = group->next;
7570 } while (group != child->groups);
7573 sdg->sgc->capacity = capacity;
7574 sdg->sgc->min_capacity = min_capacity;
7578 * Check whether the capacity of the rq has been noticeably reduced by side
7579 * activity. The imbalance_pct is used for the threshold.
7580 * Return true is the capacity is reduced
7583 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7585 return ((rq->cpu_capacity * sd->imbalance_pct) <
7586 (rq->cpu_capacity_orig * 100));
7590 * Group imbalance indicates (and tries to solve) the problem where balancing
7591 * groups is inadequate due to ->cpus_allowed constraints.
7593 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7594 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7597 * { 0 1 2 3 } { 4 5 6 7 }
7600 * If we were to balance group-wise we'd place two tasks in the first group and
7601 * two tasks in the second group. Clearly this is undesired as it will overload
7602 * cpu 3 and leave one of the cpus in the second group unused.
7604 * The current solution to this issue is detecting the skew in the first group
7605 * by noticing the lower domain failed to reach balance and had difficulty
7606 * moving tasks due to affinity constraints.
7608 * When this is so detected; this group becomes a candidate for busiest; see
7609 * update_sd_pick_busiest(). And calculate_imbalance() and
7610 * find_busiest_group() avoid some of the usual balance conditions to allow it
7611 * to create an effective group imbalance.
7613 * This is a somewhat tricky proposition since the next run might not find the
7614 * group imbalance and decide the groups need to be balanced again. A most
7615 * subtle and fragile situation.
7618 static inline int sg_imbalanced(struct sched_group *group)
7620 return group->sgc->imbalance;
7624 * group_has_capacity returns true if the group has spare capacity that could
7625 * be used by some tasks.
7626 * We consider that a group has spare capacity if the * number of task is
7627 * smaller than the number of CPUs or if the utilization is lower than the
7628 * available capacity for CFS tasks.
7629 * For the latter, we use a threshold to stabilize the state, to take into
7630 * account the variance of the tasks' load and to return true if the available
7631 * capacity in meaningful for the load balancer.
7632 * As an example, an available capacity of 1% can appear but it doesn't make
7633 * any benefit for the load balance.
7636 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7638 if (sgs->sum_nr_running < sgs->group_weight)
7641 if ((sgs->group_capacity * 100) >
7642 (sgs->group_util * env->sd->imbalance_pct))
7649 * group_is_overloaded returns true if the group has more tasks than it can
7651 * group_is_overloaded is not equals to !group_has_capacity because a group
7652 * with the exact right number of tasks, has no more spare capacity but is not
7653 * overloaded so both group_has_capacity and group_is_overloaded return
7657 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7659 if (sgs->sum_nr_running <= sgs->group_weight)
7662 if ((sgs->group_capacity * 100) <
7663 (sgs->group_util * env->sd->imbalance_pct))
7670 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7671 * per-CPU capacity than sched_group ref.
7674 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7676 return sg->sgc->min_capacity * capacity_margin <
7677 ref->sgc->min_capacity * 1024;
7681 group_type group_classify(struct sched_group *group,
7682 struct sg_lb_stats *sgs)
7684 if (sgs->group_no_capacity)
7685 return group_overloaded;
7687 if (sg_imbalanced(group))
7688 return group_imbalanced;
7694 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7695 * @env: The load balancing environment.
7696 * @group: sched_group whose statistics are to be updated.
7697 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7698 * @local_group: Does group contain this_cpu.
7699 * @sgs: variable to hold the statistics for this group.
7700 * @overload: Indicate more than one runnable task for any CPU.
7702 static inline void update_sg_lb_stats(struct lb_env *env,
7703 struct sched_group *group, int load_idx,
7704 int local_group, struct sg_lb_stats *sgs,
7710 memset(sgs, 0, sizeof(*sgs));
7712 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7713 struct rq *rq = cpu_rq(i);
7715 /* Bias balancing toward cpus of our domain */
7717 load = target_load(i, load_idx);
7719 load = source_load(i, load_idx);
7721 sgs->group_load += load;
7722 sgs->group_util += cpu_util(i);
7723 sgs->sum_nr_running += rq->cfs.h_nr_running;
7725 nr_running = rq->nr_running;
7729 #ifdef CONFIG_NUMA_BALANCING
7730 sgs->nr_numa_running += rq->nr_numa_running;
7731 sgs->nr_preferred_running += rq->nr_preferred_running;
7733 sgs->sum_weighted_load += weighted_cpuload(rq);
7735 * No need to call idle_cpu() if nr_running is not 0
7737 if (!nr_running && idle_cpu(i))
7741 /* Adjust by relative CPU capacity of the group */
7742 sgs->group_capacity = group->sgc->capacity;
7743 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7745 if (sgs->sum_nr_running)
7746 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7748 sgs->group_weight = group->group_weight;
7750 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7751 sgs->group_type = group_classify(group, sgs);
7755 * update_sd_pick_busiest - return 1 on busiest group
7756 * @env: The load balancing environment.
7757 * @sds: sched_domain statistics
7758 * @sg: sched_group candidate to be checked for being the busiest
7759 * @sgs: sched_group statistics
7761 * Determine if @sg is a busier group than the previously selected
7764 * Return: %true if @sg is a busier group than the previously selected
7765 * busiest group. %false otherwise.
7767 static bool update_sd_pick_busiest(struct lb_env *env,
7768 struct sd_lb_stats *sds,
7769 struct sched_group *sg,
7770 struct sg_lb_stats *sgs)
7772 struct sg_lb_stats *busiest = &sds->busiest_stat;
7774 if (sgs->group_type > busiest->group_type)
7777 if (sgs->group_type < busiest->group_type)
7780 if (sgs->avg_load <= busiest->avg_load)
7783 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7787 * Candidate sg has no more than one task per CPU and
7788 * has higher per-CPU capacity. Migrating tasks to less
7789 * capable CPUs may harm throughput. Maximize throughput,
7790 * power/energy consequences are not considered.
7792 if (sgs->sum_nr_running <= sgs->group_weight &&
7793 group_smaller_cpu_capacity(sds->local, sg))
7797 /* This is the busiest node in its class. */
7798 if (!(env->sd->flags & SD_ASYM_PACKING))
7801 /* No ASYM_PACKING if target cpu is already busy */
7802 if (env->idle == CPU_NOT_IDLE)
7805 * ASYM_PACKING needs to move all the work to the highest
7806 * prority CPUs in the group, therefore mark all groups
7807 * of lower priority than ourself as busy.
7809 if (sgs->sum_nr_running &&
7810 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7814 /* Prefer to move from lowest priority cpu's work */
7815 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7816 sg->asym_prefer_cpu))
7823 #ifdef CONFIG_NUMA_BALANCING
7824 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7826 if (sgs->sum_nr_running > sgs->nr_numa_running)
7828 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7833 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7835 if (rq->nr_running > rq->nr_numa_running)
7837 if (rq->nr_running > rq->nr_preferred_running)
7842 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7847 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7851 #endif /* CONFIG_NUMA_BALANCING */
7854 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7855 * @env: The load balancing environment.
7856 * @sds: variable to hold the statistics for this sched_domain.
7858 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7860 struct sched_domain *child = env->sd->child;
7861 struct sched_group *sg = env->sd->groups;
7862 struct sg_lb_stats *local = &sds->local_stat;
7863 struct sg_lb_stats tmp_sgs;
7864 int load_idx, prefer_sibling = 0;
7865 bool overload = false;
7867 if (child && child->flags & SD_PREFER_SIBLING)
7870 load_idx = get_sd_load_idx(env->sd, env->idle);
7873 struct sg_lb_stats *sgs = &tmp_sgs;
7876 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7881 if (env->idle != CPU_NEWLY_IDLE ||
7882 time_after_eq(jiffies, sg->sgc->next_update))
7883 update_group_capacity(env->sd, env->dst_cpu);
7886 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7893 * In case the child domain prefers tasks go to siblings
7894 * first, lower the sg capacity so that we'll try
7895 * and move all the excess tasks away. We lower the capacity
7896 * of a group only if the local group has the capacity to fit
7897 * these excess tasks. The extra check prevents the case where
7898 * you always pull from the heaviest group when it is already
7899 * under-utilized (possible with a large weight task outweighs
7900 * the tasks on the system).
7902 if (prefer_sibling && sds->local &&
7903 group_has_capacity(env, local) &&
7904 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7905 sgs->group_no_capacity = 1;
7906 sgs->group_type = group_classify(sg, sgs);
7909 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7911 sds->busiest_stat = *sgs;
7915 /* Now, start updating sd_lb_stats */
7916 sds->total_running += sgs->sum_nr_running;
7917 sds->total_load += sgs->group_load;
7918 sds->total_capacity += sgs->group_capacity;
7921 } while (sg != env->sd->groups);
7923 if (env->sd->flags & SD_NUMA)
7924 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7926 if (!env->sd->parent) {
7927 /* update overload indicator if we are at root domain */
7928 if (env->dst_rq->rd->overload != overload)
7929 env->dst_rq->rd->overload = overload;
7934 * check_asym_packing - Check to see if the group is packed into the
7937 * This is primarily intended to used at the sibling level. Some
7938 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7939 * case of POWER7, it can move to lower SMT modes only when higher
7940 * threads are idle. When in lower SMT modes, the threads will
7941 * perform better since they share less core resources. Hence when we
7942 * have idle threads, we want them to be the higher ones.
7944 * This packing function is run on idle threads. It checks to see if
7945 * the busiest CPU in this domain (core in the P7 case) has a higher
7946 * CPU number than the packing function is being run on. Here we are
7947 * assuming lower CPU number will be equivalent to lower a SMT thread
7950 * Return: 1 when packing is required and a task should be moved to
7951 * this CPU. The amount of the imbalance is returned in env->imbalance.
7953 * @env: The load balancing environment.
7954 * @sds: Statistics of the sched_domain which is to be packed
7956 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7960 if (!(env->sd->flags & SD_ASYM_PACKING))
7963 if (env->idle == CPU_NOT_IDLE)
7969 busiest_cpu = sds->busiest->asym_prefer_cpu;
7970 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7973 env->imbalance = DIV_ROUND_CLOSEST(
7974 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7975 SCHED_CAPACITY_SCALE);
7981 * fix_small_imbalance - Calculate the minor imbalance that exists
7982 * amongst the groups of a sched_domain, during
7984 * @env: The load balancing environment.
7985 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7988 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7990 unsigned long tmp, capa_now = 0, capa_move = 0;
7991 unsigned int imbn = 2;
7992 unsigned long scaled_busy_load_per_task;
7993 struct sg_lb_stats *local, *busiest;
7995 local = &sds->local_stat;
7996 busiest = &sds->busiest_stat;
7998 if (!local->sum_nr_running)
7999 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8000 else if (busiest->load_per_task > local->load_per_task)
8003 scaled_busy_load_per_task =
8004 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8005 busiest->group_capacity;
8007 if (busiest->avg_load + scaled_busy_load_per_task >=
8008 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8009 env->imbalance = busiest->load_per_task;
8014 * OK, we don't have enough imbalance to justify moving tasks,
8015 * however we may be able to increase total CPU capacity used by
8019 capa_now += busiest->group_capacity *
8020 min(busiest->load_per_task, busiest->avg_load);
8021 capa_now += local->group_capacity *
8022 min(local->load_per_task, local->avg_load);
8023 capa_now /= SCHED_CAPACITY_SCALE;
8025 /* Amount of load we'd subtract */
8026 if (busiest->avg_load > scaled_busy_load_per_task) {
8027 capa_move += busiest->group_capacity *
8028 min(busiest->load_per_task,
8029 busiest->avg_load - scaled_busy_load_per_task);
8032 /* Amount of load we'd add */
8033 if (busiest->avg_load * busiest->group_capacity <
8034 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8035 tmp = (busiest->avg_load * busiest->group_capacity) /
8036 local->group_capacity;
8038 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8039 local->group_capacity;
8041 capa_move += local->group_capacity *
8042 min(local->load_per_task, local->avg_load + tmp);
8043 capa_move /= SCHED_CAPACITY_SCALE;
8045 /* Move if we gain throughput */
8046 if (capa_move > capa_now)
8047 env->imbalance = busiest->load_per_task;
8051 * calculate_imbalance - Calculate the amount of imbalance present within the
8052 * groups of a given sched_domain during load balance.
8053 * @env: load balance environment
8054 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8056 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8058 unsigned long max_pull, load_above_capacity = ~0UL;
8059 struct sg_lb_stats *local, *busiest;
8061 local = &sds->local_stat;
8062 busiest = &sds->busiest_stat;
8064 if (busiest->group_type == group_imbalanced) {
8066 * In the group_imb case we cannot rely on group-wide averages
8067 * to ensure cpu-load equilibrium, look at wider averages. XXX
8069 busiest->load_per_task =
8070 min(busiest->load_per_task, sds->avg_load);
8074 * Avg load of busiest sg can be less and avg load of local sg can
8075 * be greater than avg load across all sgs of sd because avg load
8076 * factors in sg capacity and sgs with smaller group_type are
8077 * skipped when updating the busiest sg:
8079 if (busiest->avg_load <= sds->avg_load ||
8080 local->avg_load >= sds->avg_load) {
8082 return fix_small_imbalance(env, sds);
8086 * If there aren't any idle cpus, avoid creating some.
8088 if (busiest->group_type == group_overloaded &&
8089 local->group_type == group_overloaded) {
8090 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8091 if (load_above_capacity > busiest->group_capacity) {
8092 load_above_capacity -= busiest->group_capacity;
8093 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8094 load_above_capacity /= busiest->group_capacity;
8096 load_above_capacity = ~0UL;
8100 * We're trying to get all the cpus to the average_load, so we don't
8101 * want to push ourselves above the average load, nor do we wish to
8102 * reduce the max loaded cpu below the average load. At the same time,
8103 * we also don't want to reduce the group load below the group
8104 * capacity. Thus we look for the minimum possible imbalance.
8106 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8108 /* How much load to actually move to equalise the imbalance */
8109 env->imbalance = min(
8110 max_pull * busiest->group_capacity,
8111 (sds->avg_load - local->avg_load) * local->group_capacity
8112 ) / SCHED_CAPACITY_SCALE;
8115 * if *imbalance is less than the average load per runnable task
8116 * there is no guarantee that any tasks will be moved so we'll have
8117 * a think about bumping its value to force at least one task to be
8120 if (env->imbalance < busiest->load_per_task)
8121 return fix_small_imbalance(env, sds);
8124 /******* find_busiest_group() helpers end here *********************/
8127 * find_busiest_group - Returns the busiest group within the sched_domain
8128 * if there is an imbalance.
8130 * Also calculates the amount of weighted load which should be moved
8131 * to restore balance.
8133 * @env: The load balancing environment.
8135 * Return: - The busiest group if imbalance exists.
8137 static struct sched_group *find_busiest_group(struct lb_env *env)
8139 struct sg_lb_stats *local, *busiest;
8140 struct sd_lb_stats sds;
8142 init_sd_lb_stats(&sds);
8145 * Compute the various statistics relavent for load balancing at
8148 update_sd_lb_stats(env, &sds);
8149 local = &sds.local_stat;
8150 busiest = &sds.busiest_stat;
8152 /* ASYM feature bypasses nice load balance check */
8153 if (check_asym_packing(env, &sds))
8156 /* There is no busy sibling group to pull tasks from */
8157 if (!sds.busiest || busiest->sum_nr_running == 0)
8160 /* XXX broken for overlapping NUMA groups */
8161 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8162 / sds.total_capacity;
8165 * If the busiest group is imbalanced the below checks don't
8166 * work because they assume all things are equal, which typically
8167 * isn't true due to cpus_allowed constraints and the like.
8169 if (busiest->group_type == group_imbalanced)
8172 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8173 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8174 busiest->group_no_capacity)
8178 * If the local group is busier than the selected busiest group
8179 * don't try and pull any tasks.
8181 if (local->avg_load >= busiest->avg_load)
8185 * Don't pull any tasks if this group is already above the domain
8188 if (local->avg_load >= sds.avg_load)
8191 if (env->idle == CPU_IDLE) {
8193 * This cpu is idle. If the busiest group is not overloaded
8194 * and there is no imbalance between this and busiest group
8195 * wrt idle cpus, it is balanced. The imbalance becomes
8196 * significant if the diff is greater than 1 otherwise we
8197 * might end up to just move the imbalance on another group
8199 if ((busiest->group_type != group_overloaded) &&
8200 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8204 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8205 * imbalance_pct to be conservative.
8207 if (100 * busiest->avg_load <=
8208 env->sd->imbalance_pct * local->avg_load)
8213 /* Looks like there is an imbalance. Compute it */
8214 calculate_imbalance(env, &sds);
8223 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8225 static struct rq *find_busiest_queue(struct lb_env *env,
8226 struct sched_group *group)
8228 struct rq *busiest = NULL, *rq;
8229 unsigned long busiest_load = 0, busiest_capacity = 1;
8232 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8233 unsigned long capacity, wl;
8237 rt = fbq_classify_rq(rq);
8240 * We classify groups/runqueues into three groups:
8241 * - regular: there are !numa tasks
8242 * - remote: there are numa tasks that run on the 'wrong' node
8243 * - all: there is no distinction
8245 * In order to avoid migrating ideally placed numa tasks,
8246 * ignore those when there's better options.
8248 * If we ignore the actual busiest queue to migrate another
8249 * task, the next balance pass can still reduce the busiest
8250 * queue by moving tasks around inside the node.
8252 * If we cannot move enough load due to this classification
8253 * the next pass will adjust the group classification and
8254 * allow migration of more tasks.
8256 * Both cases only affect the total convergence complexity.
8258 if (rt > env->fbq_type)
8261 capacity = capacity_of(i);
8263 wl = weighted_cpuload(rq);
8266 * When comparing with imbalance, use weighted_cpuload()
8267 * which is not scaled with the cpu capacity.
8270 if (rq->nr_running == 1 && wl > env->imbalance &&
8271 !check_cpu_capacity(rq, env->sd))
8275 * For the load comparisons with the other cpu's, consider
8276 * the weighted_cpuload() scaled with the cpu capacity, so
8277 * that the load can be moved away from the cpu that is
8278 * potentially running at a lower capacity.
8280 * Thus we're looking for max(wl_i / capacity_i), crosswise
8281 * multiplication to rid ourselves of the division works out
8282 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8283 * our previous maximum.
8285 if (wl * busiest_capacity > busiest_load * capacity) {
8287 busiest_capacity = capacity;
8296 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8297 * so long as it is large enough.
8299 #define MAX_PINNED_INTERVAL 512
8301 static int need_active_balance(struct lb_env *env)
8303 struct sched_domain *sd = env->sd;
8305 if (env->idle == CPU_NEWLY_IDLE) {
8308 * ASYM_PACKING needs to force migrate tasks from busy but
8309 * lower priority CPUs in order to pack all tasks in the
8310 * highest priority CPUs.
8312 if ((sd->flags & SD_ASYM_PACKING) &&
8313 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8318 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8319 * It's worth migrating the task if the src_cpu's capacity is reduced
8320 * because of other sched_class or IRQs if more capacity stays
8321 * available on dst_cpu.
8323 if ((env->idle != CPU_NOT_IDLE) &&
8324 (env->src_rq->cfs.h_nr_running == 1)) {
8325 if ((check_cpu_capacity(env->src_rq, sd)) &&
8326 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8330 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8333 static int active_load_balance_cpu_stop(void *data);
8335 static int should_we_balance(struct lb_env *env)
8337 struct sched_group *sg = env->sd->groups;
8338 int cpu, balance_cpu = -1;
8341 * Ensure the balancing environment is consistent; can happen
8342 * when the softirq triggers 'during' hotplug.
8344 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8348 * In the newly idle case, we will allow all the cpu's
8349 * to do the newly idle load balance.
8351 if (env->idle == CPU_NEWLY_IDLE)
8354 /* Try to find first idle cpu */
8355 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8363 if (balance_cpu == -1)
8364 balance_cpu = group_balance_cpu(sg);
8367 * First idle cpu or the first cpu(busiest) in this sched group
8368 * is eligible for doing load balancing at this and above domains.
8370 return balance_cpu == env->dst_cpu;
8374 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8375 * tasks if there is an imbalance.
8377 static int load_balance(int this_cpu, struct rq *this_rq,
8378 struct sched_domain *sd, enum cpu_idle_type idle,
8379 int *continue_balancing)
8381 int ld_moved, cur_ld_moved, active_balance = 0;
8382 struct sched_domain *sd_parent = sd->parent;
8383 struct sched_group *group;
8386 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8388 struct lb_env env = {
8390 .dst_cpu = this_cpu,
8392 .dst_grpmask = sched_group_span(sd->groups),
8394 .loop_break = sched_nr_migrate_break,
8397 .tasks = LIST_HEAD_INIT(env.tasks),
8400 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8402 schedstat_inc(sd->lb_count[idle]);
8405 if (!should_we_balance(&env)) {
8406 *continue_balancing = 0;
8410 group = find_busiest_group(&env);
8412 schedstat_inc(sd->lb_nobusyg[idle]);
8416 busiest = find_busiest_queue(&env, group);
8418 schedstat_inc(sd->lb_nobusyq[idle]);
8422 BUG_ON(busiest == env.dst_rq);
8424 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8426 env.src_cpu = busiest->cpu;
8427 env.src_rq = busiest;
8430 if (busiest->nr_running > 1) {
8432 * Attempt to move tasks. If find_busiest_group has found
8433 * an imbalance but busiest->nr_running <= 1, the group is
8434 * still unbalanced. ld_moved simply stays zero, so it is
8435 * correctly treated as an imbalance.
8437 env.flags |= LBF_ALL_PINNED;
8438 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8441 rq_lock_irqsave(busiest, &rf);
8442 update_rq_clock(busiest);
8445 * cur_ld_moved - load moved in current iteration
8446 * ld_moved - cumulative load moved across iterations
8448 cur_ld_moved = detach_tasks(&env);
8451 * We've detached some tasks from busiest_rq. Every
8452 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8453 * unlock busiest->lock, and we are able to be sure
8454 * that nobody can manipulate the tasks in parallel.
8455 * See task_rq_lock() family for the details.
8458 rq_unlock(busiest, &rf);
8462 ld_moved += cur_ld_moved;
8465 local_irq_restore(rf.flags);
8467 if (env.flags & LBF_NEED_BREAK) {
8468 env.flags &= ~LBF_NEED_BREAK;
8473 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8474 * us and move them to an alternate dst_cpu in our sched_group
8475 * where they can run. The upper limit on how many times we
8476 * iterate on same src_cpu is dependent on number of cpus in our
8479 * This changes load balance semantics a bit on who can move
8480 * load to a given_cpu. In addition to the given_cpu itself
8481 * (or a ilb_cpu acting on its behalf where given_cpu is
8482 * nohz-idle), we now have balance_cpu in a position to move
8483 * load to given_cpu. In rare situations, this may cause
8484 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8485 * _independently_ and at _same_ time to move some load to
8486 * given_cpu) causing exceess load to be moved to given_cpu.
8487 * This however should not happen so much in practice and
8488 * moreover subsequent load balance cycles should correct the
8489 * excess load moved.
8491 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8493 /* Prevent to re-select dst_cpu via env's cpus */
8494 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8496 env.dst_rq = cpu_rq(env.new_dst_cpu);
8497 env.dst_cpu = env.new_dst_cpu;
8498 env.flags &= ~LBF_DST_PINNED;
8500 env.loop_break = sched_nr_migrate_break;
8503 * Go back to "more_balance" rather than "redo" since we
8504 * need to continue with same src_cpu.
8510 * We failed to reach balance because of affinity.
8513 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8515 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8516 *group_imbalance = 1;
8519 /* All tasks on this runqueue were pinned by CPU affinity */
8520 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8521 cpumask_clear_cpu(cpu_of(busiest), cpus);
8523 * Attempting to continue load balancing at the current
8524 * sched_domain level only makes sense if there are
8525 * active CPUs remaining as possible busiest CPUs to
8526 * pull load from which are not contained within the
8527 * destination group that is receiving any migrated
8530 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8532 env.loop_break = sched_nr_migrate_break;
8535 goto out_all_pinned;
8540 schedstat_inc(sd->lb_failed[idle]);
8542 * Increment the failure counter only on periodic balance.
8543 * We do not want newidle balance, which can be very
8544 * frequent, pollute the failure counter causing
8545 * excessive cache_hot migrations and active balances.
8547 if (idle != CPU_NEWLY_IDLE)
8548 sd->nr_balance_failed++;
8550 if (need_active_balance(&env)) {
8551 unsigned long flags;
8553 raw_spin_lock_irqsave(&busiest->lock, flags);
8555 /* don't kick the active_load_balance_cpu_stop,
8556 * if the curr task on busiest cpu can't be
8559 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8560 raw_spin_unlock_irqrestore(&busiest->lock,
8562 env.flags |= LBF_ALL_PINNED;
8563 goto out_one_pinned;
8567 * ->active_balance synchronizes accesses to
8568 * ->active_balance_work. Once set, it's cleared
8569 * only after active load balance is finished.
8571 if (!busiest->active_balance) {
8572 busiest->active_balance = 1;
8573 busiest->push_cpu = this_cpu;
8576 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8578 if (active_balance) {
8579 stop_one_cpu_nowait(cpu_of(busiest),
8580 active_load_balance_cpu_stop, busiest,
8581 &busiest->active_balance_work);
8584 /* We've kicked active balancing, force task migration. */
8585 sd->nr_balance_failed = sd->cache_nice_tries+1;
8588 sd->nr_balance_failed = 0;
8590 if (likely(!active_balance)) {
8591 /* We were unbalanced, so reset the balancing interval */
8592 sd->balance_interval = sd->min_interval;
8595 * If we've begun active balancing, start to back off. This
8596 * case may not be covered by the all_pinned logic if there
8597 * is only 1 task on the busy runqueue (because we don't call
8600 if (sd->balance_interval < sd->max_interval)
8601 sd->balance_interval *= 2;
8608 * We reach balance although we may have faced some affinity
8609 * constraints. Clear the imbalance flag if it was set.
8612 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8614 if (*group_imbalance)
8615 *group_imbalance = 0;
8620 * We reach balance because all tasks are pinned at this level so
8621 * we can't migrate them. Let the imbalance flag set so parent level
8622 * can try to migrate them.
8624 schedstat_inc(sd->lb_balanced[idle]);
8626 sd->nr_balance_failed = 0;
8629 /* tune up the balancing interval */
8630 if (((env.flags & LBF_ALL_PINNED) &&
8631 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8632 (sd->balance_interval < sd->max_interval))
8633 sd->balance_interval *= 2;
8640 static inline unsigned long
8641 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8643 unsigned long interval = sd->balance_interval;
8646 interval *= sd->busy_factor;
8648 /* scale ms to jiffies */
8649 interval = msecs_to_jiffies(interval);
8650 interval = clamp(interval, 1UL, max_load_balance_interval);
8656 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8658 unsigned long interval, next;
8660 /* used by idle balance, so cpu_busy = 0 */
8661 interval = get_sd_balance_interval(sd, 0);
8662 next = sd->last_balance + interval;
8664 if (time_after(*next_balance, next))
8665 *next_balance = next;
8669 * idle_balance is called by schedule() if this_cpu is about to become
8670 * idle. Attempts to pull tasks from other CPUs.
8672 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8674 unsigned long next_balance = jiffies + HZ;
8675 int this_cpu = this_rq->cpu;
8676 struct sched_domain *sd;
8677 int pulled_task = 0;
8681 * We must set idle_stamp _before_ calling idle_balance(), such that we
8682 * measure the duration of idle_balance() as idle time.
8684 this_rq->idle_stamp = rq_clock(this_rq);
8687 * Do not pull tasks towards !active CPUs...
8689 if (!cpu_active(this_cpu))
8693 * This is OK, because current is on_cpu, which avoids it being picked
8694 * for load-balance and preemption/IRQs are still disabled avoiding
8695 * further scheduler activity on it and we're being very careful to
8696 * re-start the picking loop.
8698 rq_unpin_lock(this_rq, rf);
8700 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8701 !this_rq->rd->overload) {
8703 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8705 update_next_balance(sd, &next_balance);
8711 raw_spin_unlock(&this_rq->lock);
8713 update_blocked_averages(this_cpu);
8715 for_each_domain(this_cpu, sd) {
8716 int continue_balancing = 1;
8717 u64 t0, domain_cost;
8719 if (!(sd->flags & SD_LOAD_BALANCE))
8722 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8723 update_next_balance(sd, &next_balance);
8727 if (sd->flags & SD_BALANCE_NEWIDLE) {
8728 t0 = sched_clock_cpu(this_cpu);
8730 pulled_task = load_balance(this_cpu, this_rq,
8732 &continue_balancing);
8734 domain_cost = sched_clock_cpu(this_cpu) - t0;
8735 if (domain_cost > sd->max_newidle_lb_cost)
8736 sd->max_newidle_lb_cost = domain_cost;
8738 curr_cost += domain_cost;
8741 update_next_balance(sd, &next_balance);
8744 * Stop searching for tasks to pull if there are
8745 * now runnable tasks on this rq.
8747 if (pulled_task || this_rq->nr_running > 0)
8752 raw_spin_lock(&this_rq->lock);
8754 if (curr_cost > this_rq->max_idle_balance_cost)
8755 this_rq->max_idle_balance_cost = curr_cost;
8758 * While browsing the domains, we released the rq lock, a task could
8759 * have been enqueued in the meantime. Since we're not going idle,
8760 * pretend we pulled a task.
8762 if (this_rq->cfs.h_nr_running && !pulled_task)
8766 /* Move the next balance forward */
8767 if (time_after(this_rq->next_balance, next_balance))
8768 this_rq->next_balance = next_balance;
8770 /* Is there a task of a high priority class? */
8771 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8775 this_rq->idle_stamp = 0;
8777 rq_repin_lock(this_rq, rf);
8783 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8784 * running tasks off the busiest CPU onto idle CPUs. It requires at
8785 * least 1 task to be running on each physical CPU where possible, and
8786 * avoids physical / logical imbalances.
8788 static int active_load_balance_cpu_stop(void *data)
8790 struct rq *busiest_rq = data;
8791 int busiest_cpu = cpu_of(busiest_rq);
8792 int target_cpu = busiest_rq->push_cpu;
8793 struct rq *target_rq = cpu_rq(target_cpu);
8794 struct sched_domain *sd;
8795 struct task_struct *p = NULL;
8798 rq_lock_irq(busiest_rq, &rf);
8800 * Between queueing the stop-work and running it is a hole in which
8801 * CPUs can become inactive. We should not move tasks from or to
8804 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8807 /* make sure the requested cpu hasn't gone down in the meantime */
8808 if (unlikely(busiest_cpu != smp_processor_id() ||
8809 !busiest_rq->active_balance))
8812 /* Is there any task to move? */
8813 if (busiest_rq->nr_running <= 1)
8817 * This condition is "impossible", if it occurs
8818 * we need to fix it. Originally reported by
8819 * Bjorn Helgaas on a 128-cpu setup.
8821 BUG_ON(busiest_rq == target_rq);
8823 /* Search for an sd spanning us and the target CPU. */
8825 for_each_domain(target_cpu, sd) {
8826 if ((sd->flags & SD_LOAD_BALANCE) &&
8827 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8832 struct lb_env env = {
8834 .dst_cpu = target_cpu,
8835 .dst_rq = target_rq,
8836 .src_cpu = busiest_rq->cpu,
8837 .src_rq = busiest_rq,
8840 * can_migrate_task() doesn't need to compute new_dst_cpu
8841 * for active balancing. Since we have CPU_IDLE, but no
8842 * @dst_grpmask we need to make that test go away with lying
8845 .flags = LBF_DST_PINNED,
8848 schedstat_inc(sd->alb_count);
8849 update_rq_clock(busiest_rq);
8851 p = detach_one_task(&env);
8853 schedstat_inc(sd->alb_pushed);
8854 /* Active balancing done, reset the failure counter. */
8855 sd->nr_balance_failed = 0;
8857 schedstat_inc(sd->alb_failed);
8862 busiest_rq->active_balance = 0;
8863 rq_unlock(busiest_rq, &rf);
8866 attach_one_task(target_rq, p);
8873 static inline int on_null_domain(struct rq *rq)
8875 return unlikely(!rcu_dereference_sched(rq->sd));
8878 #ifdef CONFIG_NO_HZ_COMMON
8880 * idle load balancing details
8881 * - When one of the busy CPUs notice that there may be an idle rebalancing
8882 * needed, they will kick the idle load balancer, which then does idle
8883 * load balancing for all the idle CPUs.
8886 cpumask_var_t idle_cpus_mask;
8888 unsigned long next_balance; /* in jiffy units */
8889 } nohz ____cacheline_aligned;
8891 static inline int find_new_ilb(void)
8893 int ilb = cpumask_first(nohz.idle_cpus_mask);
8895 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8902 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8903 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8904 * CPU (if there is one).
8906 static void nohz_balancer_kick(void)
8910 nohz.next_balance++;
8912 ilb_cpu = find_new_ilb();
8914 if (ilb_cpu >= nr_cpu_ids)
8917 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8920 * Use smp_send_reschedule() instead of resched_cpu().
8921 * This way we generate a sched IPI on the target cpu which
8922 * is idle. And the softirq performing nohz idle load balance
8923 * will be run before returning from the IPI.
8925 smp_send_reschedule(ilb_cpu);
8929 void nohz_balance_exit_idle(unsigned int cpu)
8931 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8933 * Completely isolated CPUs don't ever set, so we must test.
8935 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8936 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8937 atomic_dec(&nohz.nr_cpus);
8939 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8943 static inline void set_cpu_sd_state_busy(void)
8945 struct sched_domain *sd;
8946 int cpu = smp_processor_id();
8949 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8951 if (!sd || !sd->nohz_idle)
8955 atomic_inc(&sd->shared->nr_busy_cpus);
8960 void set_cpu_sd_state_idle(void)
8962 struct sched_domain *sd;
8963 int cpu = smp_processor_id();
8966 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8968 if (!sd || sd->nohz_idle)
8972 atomic_dec(&sd->shared->nr_busy_cpus);
8978 * This routine will record that the cpu is going idle with tick stopped.
8979 * This info will be used in performing idle load balancing in the future.
8981 void nohz_balance_enter_idle(int cpu)
8984 * If this cpu is going down, then nothing needs to be done.
8986 if (!cpu_active(cpu))
8989 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8990 if (!is_housekeeping_cpu(cpu))
8993 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8997 * If we're a completely isolated CPU, we don't play.
8999 if (on_null_domain(cpu_rq(cpu)))
9002 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9003 atomic_inc(&nohz.nr_cpus);
9004 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9008 static DEFINE_SPINLOCK(balancing);
9011 * Scale the max load_balance interval with the number of CPUs in the system.
9012 * This trades load-balance latency on larger machines for less cross talk.
9014 void update_max_interval(void)
9016 max_load_balance_interval = HZ*num_online_cpus()/10;
9020 * It checks each scheduling domain to see if it is due to be balanced,
9021 * and initiates a balancing operation if so.
9023 * Balancing parameters are set up in init_sched_domains.
9025 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9027 int continue_balancing = 1;
9029 unsigned long interval;
9030 struct sched_domain *sd;
9031 /* Earliest time when we have to do rebalance again */
9032 unsigned long next_balance = jiffies + 60*HZ;
9033 int update_next_balance = 0;
9034 int need_serialize, need_decay = 0;
9037 update_blocked_averages(cpu);
9040 for_each_domain(cpu, sd) {
9042 * Decay the newidle max times here because this is a regular
9043 * visit to all the domains. Decay ~1% per second.
9045 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9046 sd->max_newidle_lb_cost =
9047 (sd->max_newidle_lb_cost * 253) / 256;
9048 sd->next_decay_max_lb_cost = jiffies + HZ;
9051 max_cost += sd->max_newidle_lb_cost;
9053 if (!(sd->flags & SD_LOAD_BALANCE))
9057 * Stop the load balance at this level. There is another
9058 * CPU in our sched group which is doing load balancing more
9061 if (!continue_balancing) {
9067 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9069 need_serialize = sd->flags & SD_SERIALIZE;
9070 if (need_serialize) {
9071 if (!spin_trylock(&balancing))
9075 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9076 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9078 * The LBF_DST_PINNED logic could have changed
9079 * env->dst_cpu, so we can't know our idle
9080 * state even if we migrated tasks. Update it.
9082 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9084 sd->last_balance = jiffies;
9085 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9088 spin_unlock(&balancing);
9090 if (time_after(next_balance, sd->last_balance + interval)) {
9091 next_balance = sd->last_balance + interval;
9092 update_next_balance = 1;
9097 * Ensure the rq-wide value also decays but keep it at a
9098 * reasonable floor to avoid funnies with rq->avg_idle.
9100 rq->max_idle_balance_cost =
9101 max((u64)sysctl_sched_migration_cost, max_cost);
9106 * next_balance will be updated only when there is a need.
9107 * When the cpu is attached to null domain for ex, it will not be
9110 if (likely(update_next_balance)) {
9111 rq->next_balance = next_balance;
9113 #ifdef CONFIG_NO_HZ_COMMON
9115 * If this CPU has been elected to perform the nohz idle
9116 * balance. Other idle CPUs have already rebalanced with
9117 * nohz_idle_balance() and nohz.next_balance has been
9118 * updated accordingly. This CPU is now running the idle load
9119 * balance for itself and we need to update the
9120 * nohz.next_balance accordingly.
9122 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9123 nohz.next_balance = rq->next_balance;
9128 #ifdef CONFIG_NO_HZ_COMMON
9130 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9131 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9133 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9135 int this_cpu = this_rq->cpu;
9138 /* Earliest time when we have to do rebalance again */
9139 unsigned long next_balance = jiffies + 60*HZ;
9140 int update_next_balance = 0;
9142 if (idle != CPU_IDLE ||
9143 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9146 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9147 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9151 * If this cpu gets work to do, stop the load balancing
9152 * work being done for other cpus. Next load
9153 * balancing owner will pick it up.
9158 rq = cpu_rq(balance_cpu);
9161 * If time for next balance is due,
9164 if (time_after_eq(jiffies, rq->next_balance)) {
9167 rq_lock_irq(rq, &rf);
9168 update_rq_clock(rq);
9169 cpu_load_update_idle(rq);
9170 rq_unlock_irq(rq, &rf);
9172 rebalance_domains(rq, CPU_IDLE);
9175 if (time_after(next_balance, rq->next_balance)) {
9176 next_balance = rq->next_balance;
9177 update_next_balance = 1;
9182 * next_balance will be updated only when there is a need.
9183 * When the CPU is attached to null domain for ex, it will not be
9186 if (likely(update_next_balance))
9187 nohz.next_balance = next_balance;
9189 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9193 * Current heuristic for kicking the idle load balancer in the presence
9194 * of an idle cpu in the system.
9195 * - This rq has more than one task.
9196 * - This rq has at least one CFS task and the capacity of the CPU is
9197 * significantly reduced because of RT tasks or IRQs.
9198 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9199 * multiple busy cpu.
9200 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9201 * domain span are idle.
9203 static inline bool nohz_kick_needed(struct rq *rq)
9205 unsigned long now = jiffies;
9206 struct sched_domain_shared *sds;
9207 struct sched_domain *sd;
9208 int nr_busy, i, cpu = rq->cpu;
9211 if (unlikely(rq->idle_balance))
9215 * We may be recently in ticked or tickless idle mode. At the first
9216 * busy tick after returning from idle, we will update the busy stats.
9218 set_cpu_sd_state_busy();
9219 nohz_balance_exit_idle(cpu);
9222 * None are in tickless mode and hence no need for NOHZ idle load
9225 if (likely(!atomic_read(&nohz.nr_cpus)))
9228 if (time_before(now, nohz.next_balance))
9231 if (rq->nr_running >= 2)
9235 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9238 * XXX: write a coherent comment on why we do this.
9239 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9241 nr_busy = atomic_read(&sds->nr_busy_cpus);
9249 sd = rcu_dereference(rq->sd);
9251 if ((rq->cfs.h_nr_running >= 1) &&
9252 check_cpu_capacity(rq, sd)) {
9258 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9260 for_each_cpu(i, sched_domain_span(sd)) {
9262 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9265 if (sched_asym_prefer(i, cpu)) {
9276 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9280 * run_rebalance_domains is triggered when needed from the scheduler tick.
9281 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9283 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9285 struct rq *this_rq = this_rq();
9286 enum cpu_idle_type idle = this_rq->idle_balance ?
9287 CPU_IDLE : CPU_NOT_IDLE;
9290 * If this cpu has a pending nohz_balance_kick, then do the
9291 * balancing on behalf of the other idle cpus whose ticks are
9292 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9293 * give the idle cpus a chance to load balance. Else we may
9294 * load balance only within the local sched_domain hierarchy
9295 * and abort nohz_idle_balance altogether if we pull some load.
9297 nohz_idle_balance(this_rq, idle);
9298 rebalance_domains(this_rq, idle);
9302 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9304 void trigger_load_balance(struct rq *rq)
9306 /* Don't need to rebalance while attached to NULL domain */
9307 if (unlikely(on_null_domain(rq)))
9310 if (time_after_eq(jiffies, rq->next_balance))
9311 raise_softirq(SCHED_SOFTIRQ);
9312 #ifdef CONFIG_NO_HZ_COMMON
9313 if (nohz_kick_needed(rq))
9314 nohz_balancer_kick();
9318 static void rq_online_fair(struct rq *rq)
9322 update_runtime_enabled(rq);
9325 static void rq_offline_fair(struct rq *rq)
9329 /* Ensure any throttled groups are reachable by pick_next_task */
9330 unthrottle_offline_cfs_rqs(rq);
9333 #endif /* CONFIG_SMP */
9336 * scheduler tick hitting a task of our scheduling class:
9338 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9340 struct cfs_rq *cfs_rq;
9341 struct sched_entity *se = &curr->se;
9343 for_each_sched_entity(se) {
9344 cfs_rq = cfs_rq_of(se);
9345 entity_tick(cfs_rq, se, queued);
9348 if (static_branch_unlikely(&sched_numa_balancing))
9349 task_tick_numa(rq, curr);
9353 * called on fork with the child task as argument from the parent's context
9354 * - child not yet on the tasklist
9355 * - preemption disabled
9357 static void task_fork_fair(struct task_struct *p)
9359 struct cfs_rq *cfs_rq;
9360 struct sched_entity *se = &p->se, *curr;
9361 struct rq *rq = this_rq();
9365 update_rq_clock(rq);
9367 cfs_rq = task_cfs_rq(current);
9368 curr = cfs_rq->curr;
9370 update_curr(cfs_rq);
9371 se->vruntime = curr->vruntime;
9373 place_entity(cfs_rq, se, 1);
9375 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9377 * Upon rescheduling, sched_class::put_prev_task() will place
9378 * 'current' within the tree based on its new key value.
9380 swap(curr->vruntime, se->vruntime);
9384 se->vruntime -= cfs_rq->min_vruntime;
9389 * Priority of the task has changed. Check to see if we preempt
9393 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9395 if (!task_on_rq_queued(p))
9399 * Reschedule if we are currently running on this runqueue and
9400 * our priority decreased, or if we are not currently running on
9401 * this runqueue and our priority is higher than the current's
9403 if (rq->curr == p) {
9404 if (p->prio > oldprio)
9407 check_preempt_curr(rq, p, 0);
9410 static inline bool vruntime_normalized(struct task_struct *p)
9412 struct sched_entity *se = &p->se;
9415 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9416 * the dequeue_entity(.flags=0) will already have normalized the
9423 * When !on_rq, vruntime of the task has usually NOT been normalized.
9424 * But there are some cases where it has already been normalized:
9426 * - A forked child which is waiting for being woken up by
9427 * wake_up_new_task().
9428 * - A task which has been woken up by try_to_wake_up() and
9429 * waiting for actually being woken up by sched_ttwu_pending().
9431 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9437 #ifdef CONFIG_FAIR_GROUP_SCHED
9439 * Propagate the changes of the sched_entity across the tg tree to make it
9440 * visible to the root
9442 static void propagate_entity_cfs_rq(struct sched_entity *se)
9444 struct cfs_rq *cfs_rq;
9446 /* Start to propagate at parent */
9449 for_each_sched_entity(se) {
9450 cfs_rq = cfs_rq_of(se);
9452 if (cfs_rq_throttled(cfs_rq))
9455 update_load_avg(cfs_rq, se, UPDATE_TG);
9459 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9462 static void detach_entity_cfs_rq(struct sched_entity *se)
9464 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9466 /* Catch up with the cfs_rq and remove our load when we leave */
9467 update_load_avg(cfs_rq, se, 0);
9468 detach_entity_load_avg(cfs_rq, se);
9469 update_tg_load_avg(cfs_rq, false);
9470 propagate_entity_cfs_rq(se);
9473 static void attach_entity_cfs_rq(struct sched_entity *se)
9475 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9477 #ifdef CONFIG_FAIR_GROUP_SCHED
9479 * Since the real-depth could have been changed (only FAIR
9480 * class maintain depth value), reset depth properly.
9482 se->depth = se->parent ? se->parent->depth + 1 : 0;
9485 /* Synchronize entity with its cfs_rq */
9486 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9487 attach_entity_load_avg(cfs_rq, se);
9488 update_tg_load_avg(cfs_rq, false);
9489 propagate_entity_cfs_rq(se);
9492 static void detach_task_cfs_rq(struct task_struct *p)
9494 struct sched_entity *se = &p->se;
9495 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9497 if (!vruntime_normalized(p)) {
9499 * Fix up our vruntime so that the current sleep doesn't
9500 * cause 'unlimited' sleep bonus.
9502 place_entity(cfs_rq, se, 0);
9503 se->vruntime -= cfs_rq->min_vruntime;
9506 detach_entity_cfs_rq(se);
9509 static void attach_task_cfs_rq(struct task_struct *p)
9511 struct sched_entity *se = &p->se;
9512 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9514 attach_entity_cfs_rq(se);
9516 if (!vruntime_normalized(p))
9517 se->vruntime += cfs_rq->min_vruntime;
9520 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9522 detach_task_cfs_rq(p);
9525 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9527 attach_task_cfs_rq(p);
9529 if (task_on_rq_queued(p)) {
9531 * We were most likely switched from sched_rt, so
9532 * kick off the schedule if running, otherwise just see
9533 * if we can still preempt the current task.
9538 check_preempt_curr(rq, p, 0);
9542 /* Account for a task changing its policy or group.
9544 * This routine is mostly called to set cfs_rq->curr field when a task
9545 * migrates between groups/classes.
9547 static void set_curr_task_fair(struct rq *rq)
9549 struct sched_entity *se = &rq->curr->se;
9551 for_each_sched_entity(se) {
9552 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9554 set_next_entity(cfs_rq, se);
9555 /* ensure bandwidth has been allocated on our new cfs_rq */
9556 account_cfs_rq_runtime(cfs_rq, 0);
9560 void init_cfs_rq(struct cfs_rq *cfs_rq)
9562 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9563 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9564 #ifndef CONFIG_64BIT
9565 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9568 raw_spin_lock_init(&cfs_rq->removed.lock);
9572 #ifdef CONFIG_FAIR_GROUP_SCHED
9573 static void task_set_group_fair(struct task_struct *p)
9575 struct sched_entity *se = &p->se;
9577 set_task_rq(p, task_cpu(p));
9578 se->depth = se->parent ? se->parent->depth + 1 : 0;
9581 static void task_move_group_fair(struct task_struct *p)
9583 detach_task_cfs_rq(p);
9584 set_task_rq(p, task_cpu(p));
9587 /* Tell se's cfs_rq has been changed -- migrated */
9588 p->se.avg.last_update_time = 0;
9590 attach_task_cfs_rq(p);
9593 static void task_change_group_fair(struct task_struct *p, int type)
9596 case TASK_SET_GROUP:
9597 task_set_group_fair(p);
9600 case TASK_MOVE_GROUP:
9601 task_move_group_fair(p);
9606 void free_fair_sched_group(struct task_group *tg)
9610 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9612 for_each_possible_cpu(i) {
9614 kfree(tg->cfs_rq[i]);
9623 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9625 struct sched_entity *se;
9626 struct cfs_rq *cfs_rq;
9629 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9632 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9636 tg->shares = NICE_0_LOAD;
9638 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9640 for_each_possible_cpu(i) {
9641 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9642 GFP_KERNEL, cpu_to_node(i));
9646 se = kzalloc_node(sizeof(struct sched_entity),
9647 GFP_KERNEL, cpu_to_node(i));
9651 init_cfs_rq(cfs_rq);
9652 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9653 init_entity_runnable_average(se);
9664 void online_fair_sched_group(struct task_group *tg)
9666 struct sched_entity *se;
9670 for_each_possible_cpu(i) {
9674 raw_spin_lock_irq(&rq->lock);
9675 update_rq_clock(rq);
9676 attach_entity_cfs_rq(se);
9677 sync_throttle(tg, i);
9678 raw_spin_unlock_irq(&rq->lock);
9682 void unregister_fair_sched_group(struct task_group *tg)
9684 unsigned long flags;
9688 for_each_possible_cpu(cpu) {
9690 remove_entity_load_avg(tg->se[cpu]);
9693 * Only empty task groups can be destroyed; so we can speculatively
9694 * check on_list without danger of it being re-added.
9696 if (!tg->cfs_rq[cpu]->on_list)
9701 raw_spin_lock_irqsave(&rq->lock, flags);
9702 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9703 raw_spin_unlock_irqrestore(&rq->lock, flags);
9707 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9708 struct sched_entity *se, int cpu,
9709 struct sched_entity *parent)
9711 struct rq *rq = cpu_rq(cpu);
9715 init_cfs_rq_runtime(cfs_rq);
9717 tg->cfs_rq[cpu] = cfs_rq;
9720 /* se could be NULL for root_task_group */
9725 se->cfs_rq = &rq->cfs;
9728 se->cfs_rq = parent->my_q;
9729 se->depth = parent->depth + 1;
9733 /* guarantee group entities always have weight */
9734 update_load_set(&se->load, NICE_0_LOAD);
9735 se->parent = parent;
9738 static DEFINE_MUTEX(shares_mutex);
9740 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9745 * We can't change the weight of the root cgroup.
9750 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9752 mutex_lock(&shares_mutex);
9753 if (tg->shares == shares)
9756 tg->shares = shares;
9757 for_each_possible_cpu(i) {
9758 struct rq *rq = cpu_rq(i);
9759 struct sched_entity *se = tg->se[i];
9762 /* Propagate contribution to hierarchy */
9763 rq_lock_irqsave(rq, &rf);
9764 update_rq_clock(rq);
9765 for_each_sched_entity(se) {
9766 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9767 update_cfs_group(se);
9769 rq_unlock_irqrestore(rq, &rf);
9773 mutex_unlock(&shares_mutex);
9776 #else /* CONFIG_FAIR_GROUP_SCHED */
9778 void free_fair_sched_group(struct task_group *tg) { }
9780 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9785 void online_fair_sched_group(struct task_group *tg) { }
9787 void unregister_fair_sched_group(struct task_group *tg) { }
9789 #endif /* CONFIG_FAIR_GROUP_SCHED */
9792 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9794 struct sched_entity *se = &task->se;
9795 unsigned int rr_interval = 0;
9798 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9801 if (rq->cfs.load.weight)
9802 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9808 * All the scheduling class methods:
9810 const struct sched_class fair_sched_class = {
9811 .next = &idle_sched_class,
9812 .enqueue_task = enqueue_task_fair,
9813 .dequeue_task = dequeue_task_fair,
9814 .yield_task = yield_task_fair,
9815 .yield_to_task = yield_to_task_fair,
9817 .check_preempt_curr = check_preempt_wakeup,
9819 .pick_next_task = pick_next_task_fair,
9820 .put_prev_task = put_prev_task_fair,
9823 .select_task_rq = select_task_rq_fair,
9824 .migrate_task_rq = migrate_task_rq_fair,
9826 .rq_online = rq_online_fair,
9827 .rq_offline = rq_offline_fair,
9829 .task_dead = task_dead_fair,
9830 .set_cpus_allowed = set_cpus_allowed_common,
9833 .set_curr_task = set_curr_task_fair,
9834 .task_tick = task_tick_fair,
9835 .task_fork = task_fork_fair,
9837 .prio_changed = prio_changed_fair,
9838 .switched_from = switched_from_fair,
9839 .switched_to = switched_to_fair,
9841 .get_rr_interval = get_rr_interval_fair,
9843 .update_curr = update_curr_fair,
9845 #ifdef CONFIG_FAIR_GROUP_SCHED
9846 .task_change_group = task_change_group_fair,
9850 #ifdef CONFIG_SCHED_DEBUG
9851 void print_cfs_stats(struct seq_file *m, int cpu)
9853 struct cfs_rq *cfs_rq, *pos;
9856 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9857 print_cfs_rq(m, cpu, cfs_rq);
9861 #ifdef CONFIG_NUMA_BALANCING
9862 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9865 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9867 for_each_online_node(node) {
9868 if (p->numa_faults) {
9869 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9870 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9872 if (p->numa_group) {
9873 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9874 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9876 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9879 #endif /* CONFIG_NUMA_BALANCING */
9880 #endif /* CONFIG_SCHED_DEBUG */
9882 __init void init_sched_fair_class(void)
9885 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9887 #ifdef CONFIG_NO_HZ_COMMON
9888 nohz.next_balance = jiffies;
9889 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);