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 sa->last_update_time = 0;
721 * sched_avg's period_contrib should be strictly less then 1024, so
722 * we give it 1023 to make sure it is almost a period (1024us), and
723 * will definitely be update (after enqueue).
725 sa->period_contrib = 1023;
727 * Tasks are intialized with full load to be seen as heavy tasks until
728 * they get a chance to stabilize to their real load level.
729 * Group entities are intialized with zero load to reflect the fact that
730 * nothing has been attached to the task group yet.
732 if (entity_is_task(se))
733 sa->load_avg = scale_load_down(se->load.weight);
734 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
736 * At this point, util_avg won't be used in select_task_rq_fair anyway
740 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
743 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
744 static void attach_entity_cfs_rq(struct sched_entity *se);
747 * With new tasks being created, their initial util_avgs are extrapolated
748 * based on the cfs_rq's current util_avg:
750 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
752 * However, in many cases, the above util_avg does not give a desired
753 * value. Moreover, the sum of the util_avgs may be divergent, such
754 * as when the series is a harmonic series.
756 * To solve this problem, we also cap the util_avg of successive tasks to
757 * only 1/2 of the left utilization budget:
759 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
761 * where n denotes the nth task.
763 * For example, a simplest series from the beginning would be like:
765 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
766 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
768 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
769 * if util_avg > util_avg_cap.
771 void post_init_entity_util_avg(struct sched_entity *se)
773 struct cfs_rq *cfs_rq = cfs_rq_of(se);
774 struct sched_avg *sa = &se->avg;
775 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
778 if (cfs_rq->avg.util_avg != 0) {
779 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
780 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
782 if (sa->util_avg > cap)
787 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
790 if (entity_is_task(se)) {
791 struct task_struct *p = task_of(se);
792 if (p->sched_class != &fair_sched_class) {
794 * For !fair tasks do:
796 update_cfs_rq_load_avg(now, cfs_rq);
797 attach_entity_load_avg(cfs_rq, se);
798 switched_from_fair(rq, p);
800 * such that the next switched_to_fair() has the
803 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
808 attach_entity_cfs_rq(se);
811 #else /* !CONFIG_SMP */
812 void init_entity_runnable_average(struct sched_entity *se)
815 void post_init_entity_util_avg(struct sched_entity *se)
818 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
821 #endif /* CONFIG_SMP */
824 * Update the current task's runtime statistics.
826 static void update_curr(struct cfs_rq *cfs_rq)
828 struct sched_entity *curr = cfs_rq->curr;
829 u64 now = rq_clock_task(rq_of(cfs_rq));
835 delta_exec = now - curr->exec_start;
836 if (unlikely((s64)delta_exec <= 0))
839 curr->exec_start = now;
841 schedstat_set(curr->statistics.exec_max,
842 max(delta_exec, curr->statistics.exec_max));
844 curr->sum_exec_runtime += delta_exec;
845 schedstat_add(cfs_rq->exec_clock, delta_exec);
847 curr->vruntime += calc_delta_fair(delta_exec, curr);
848 update_min_vruntime(cfs_rq);
850 if (entity_is_task(curr)) {
851 struct task_struct *curtask = task_of(curr);
853 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
854 cpuacct_charge(curtask, delta_exec);
855 account_group_exec_runtime(curtask, delta_exec);
858 account_cfs_rq_runtime(cfs_rq, delta_exec);
861 static void update_curr_fair(struct rq *rq)
863 update_curr(cfs_rq_of(&rq->curr->se));
867 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
869 u64 wait_start, prev_wait_start;
871 if (!schedstat_enabled())
874 wait_start = rq_clock(rq_of(cfs_rq));
875 prev_wait_start = schedstat_val(se->statistics.wait_start);
877 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
878 likely(wait_start > prev_wait_start))
879 wait_start -= prev_wait_start;
881 schedstat_set(se->statistics.wait_start, wait_start);
885 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
887 struct task_struct *p;
890 if (!schedstat_enabled())
893 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
895 if (entity_is_task(se)) {
897 if (task_on_rq_migrating(p)) {
899 * Preserve migrating task's wait time so wait_start
900 * time stamp can be adjusted to accumulate wait time
901 * prior to migration.
903 schedstat_set(se->statistics.wait_start, delta);
906 trace_sched_stat_wait(p, delta);
909 schedstat_set(se->statistics.wait_max,
910 max(schedstat_val(se->statistics.wait_max), delta));
911 schedstat_inc(se->statistics.wait_count);
912 schedstat_add(se->statistics.wait_sum, delta);
913 schedstat_set(se->statistics.wait_start, 0);
917 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
919 struct task_struct *tsk = NULL;
920 u64 sleep_start, block_start;
922 if (!schedstat_enabled())
925 sleep_start = schedstat_val(se->statistics.sleep_start);
926 block_start = schedstat_val(se->statistics.block_start);
928 if (entity_is_task(se))
932 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
937 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
938 schedstat_set(se->statistics.sleep_max, delta);
940 schedstat_set(se->statistics.sleep_start, 0);
941 schedstat_add(se->statistics.sum_sleep_runtime, delta);
944 account_scheduler_latency(tsk, delta >> 10, 1);
945 trace_sched_stat_sleep(tsk, delta);
949 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
954 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
955 schedstat_set(se->statistics.block_max, delta);
957 schedstat_set(se->statistics.block_start, 0);
958 schedstat_add(se->statistics.sum_sleep_runtime, delta);
961 if (tsk->in_iowait) {
962 schedstat_add(se->statistics.iowait_sum, delta);
963 schedstat_inc(se->statistics.iowait_count);
964 trace_sched_stat_iowait(tsk, delta);
967 trace_sched_stat_blocked(tsk, delta);
970 * Blocking time is in units of nanosecs, so shift by
971 * 20 to get a milliseconds-range estimation of the
972 * amount of time that the task spent sleeping:
974 if (unlikely(prof_on == SLEEP_PROFILING)) {
975 profile_hits(SLEEP_PROFILING,
976 (void *)get_wchan(tsk),
979 account_scheduler_latency(tsk, delta >> 10, 0);
985 * Task is being enqueued - update stats:
988 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
990 if (!schedstat_enabled())
994 * Are we enqueueing a waiting task? (for current tasks
995 * a dequeue/enqueue event is a NOP)
997 if (se != cfs_rq->curr)
998 update_stats_wait_start(cfs_rq, se);
1000 if (flags & ENQUEUE_WAKEUP)
1001 update_stats_enqueue_sleeper(cfs_rq, se);
1005 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1008 if (!schedstat_enabled())
1012 * Mark the end of the wait period if dequeueing a
1015 if (se != cfs_rq->curr)
1016 update_stats_wait_end(cfs_rq, se);
1018 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1019 struct task_struct *tsk = task_of(se);
1021 if (tsk->state & TASK_INTERRUPTIBLE)
1022 schedstat_set(se->statistics.sleep_start,
1023 rq_clock(rq_of(cfs_rq)));
1024 if (tsk->state & TASK_UNINTERRUPTIBLE)
1025 schedstat_set(se->statistics.block_start,
1026 rq_clock(rq_of(cfs_rq)));
1031 * We are picking a new current task - update its stats:
1034 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1037 * We are starting a new run period:
1039 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1042 /**************************************************
1043 * Scheduling class queueing methods:
1046 #ifdef CONFIG_NUMA_BALANCING
1048 * Approximate time to scan a full NUMA task in ms. The task scan period is
1049 * calculated based on the tasks virtual memory size and
1050 * numa_balancing_scan_size.
1052 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1053 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1055 /* Portion of address space to scan in MB */
1056 unsigned int sysctl_numa_balancing_scan_size = 256;
1058 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1059 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1064 spinlock_t lock; /* nr_tasks, tasks */
1069 struct rcu_head rcu;
1070 unsigned long total_faults;
1071 unsigned long max_faults_cpu;
1073 * Faults_cpu is used to decide whether memory should move
1074 * towards the CPU. As a consequence, these stats are weighted
1075 * more by CPU use than by memory faults.
1077 unsigned long *faults_cpu;
1078 unsigned long faults[0];
1081 static inline unsigned long group_faults_priv(struct numa_group *ng);
1082 static inline unsigned long group_faults_shared(struct numa_group *ng);
1084 static unsigned int task_nr_scan_windows(struct task_struct *p)
1086 unsigned long rss = 0;
1087 unsigned long nr_scan_pages;
1090 * Calculations based on RSS as non-present and empty pages are skipped
1091 * by the PTE scanner and NUMA hinting faults should be trapped based
1094 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1095 rss = get_mm_rss(p->mm);
1097 rss = nr_scan_pages;
1099 rss = round_up(rss, nr_scan_pages);
1100 return rss / nr_scan_pages;
1103 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1104 #define MAX_SCAN_WINDOW 2560
1106 static unsigned int task_scan_min(struct task_struct *p)
1108 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1109 unsigned int scan, floor;
1110 unsigned int windows = 1;
1112 if (scan_size < MAX_SCAN_WINDOW)
1113 windows = MAX_SCAN_WINDOW / scan_size;
1114 floor = 1000 / windows;
1116 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1117 return max_t(unsigned int, floor, scan);
1120 static unsigned int task_scan_start(struct task_struct *p)
1122 unsigned long smin = task_scan_min(p);
1123 unsigned long period = smin;
1125 /* Scale the maximum scan period with the amount of shared memory. */
1126 if (p->numa_group) {
1127 struct numa_group *ng = p->numa_group;
1128 unsigned long shared = group_faults_shared(ng);
1129 unsigned long private = group_faults_priv(ng);
1131 period *= atomic_read(&ng->refcount);
1132 period *= shared + 1;
1133 period /= private + shared + 1;
1136 return max(smin, period);
1139 static unsigned int task_scan_max(struct task_struct *p)
1141 unsigned long smin = task_scan_min(p);
1144 /* Watch for min being lower than max due to floor calculations */
1145 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1147 /* Scale the maximum scan period with the amount of shared memory. */
1148 if (p->numa_group) {
1149 struct numa_group *ng = p->numa_group;
1150 unsigned long shared = group_faults_shared(ng);
1151 unsigned long private = group_faults_priv(ng);
1152 unsigned long period = smax;
1154 period *= atomic_read(&ng->refcount);
1155 period *= shared + 1;
1156 period /= private + shared + 1;
1158 smax = max(smax, period);
1161 return max(smin, smax);
1164 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1166 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1167 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1170 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1172 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1173 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1176 /* Shared or private faults. */
1177 #define NR_NUMA_HINT_FAULT_TYPES 2
1179 /* Memory and CPU locality */
1180 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1182 /* Averaged statistics, and temporary buffers. */
1183 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1185 pid_t task_numa_group_id(struct task_struct *p)
1187 return p->numa_group ? p->numa_group->gid : 0;
1191 * The averaged statistics, shared & private, memory & cpu,
1192 * occupy the first half of the array. The second half of the
1193 * array is for current counters, which are averaged into the
1194 * first set by task_numa_placement.
1196 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1198 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1201 static inline unsigned long task_faults(struct task_struct *p, int nid)
1203 if (!p->numa_faults)
1206 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1207 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1210 static inline unsigned long group_faults(struct task_struct *p, int nid)
1215 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1216 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1219 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1221 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1222 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1225 static inline unsigned long group_faults_priv(struct numa_group *ng)
1227 unsigned long faults = 0;
1230 for_each_online_node(node) {
1231 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1237 static inline unsigned long group_faults_shared(struct numa_group *ng)
1239 unsigned long faults = 0;
1242 for_each_online_node(node) {
1243 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1250 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1251 * considered part of a numa group's pseudo-interleaving set. Migrations
1252 * between these nodes are slowed down, to allow things to settle down.
1254 #define ACTIVE_NODE_FRACTION 3
1256 static bool numa_is_active_node(int nid, struct numa_group *ng)
1258 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1261 /* Handle placement on systems where not all nodes are directly connected. */
1262 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1263 int maxdist, bool task)
1265 unsigned long score = 0;
1269 * All nodes are directly connected, and the same distance
1270 * from each other. No need for fancy placement algorithms.
1272 if (sched_numa_topology_type == NUMA_DIRECT)
1276 * This code is called for each node, introducing N^2 complexity,
1277 * which should be ok given the number of nodes rarely exceeds 8.
1279 for_each_online_node(node) {
1280 unsigned long faults;
1281 int dist = node_distance(nid, node);
1284 * The furthest away nodes in the system are not interesting
1285 * for placement; nid was already counted.
1287 if (dist == sched_max_numa_distance || node == nid)
1291 * On systems with a backplane NUMA topology, compare groups
1292 * of nodes, and move tasks towards the group with the most
1293 * memory accesses. When comparing two nodes at distance
1294 * "hoplimit", only nodes closer by than "hoplimit" are part
1295 * of each group. Skip other nodes.
1297 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1301 /* Add up the faults from nearby nodes. */
1303 faults = task_faults(p, node);
1305 faults = group_faults(p, node);
1308 * On systems with a glueless mesh NUMA topology, there are
1309 * no fixed "groups of nodes". Instead, nodes that are not
1310 * directly connected bounce traffic through intermediate
1311 * nodes; a numa_group can occupy any set of nodes.
1312 * The further away a node is, the less the faults count.
1313 * This seems to result in good task placement.
1315 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1316 faults *= (sched_max_numa_distance - dist);
1317 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1327 * These return the fraction of accesses done by a particular task, or
1328 * task group, on a particular numa node. The group weight is given a
1329 * larger multiplier, in order to group tasks together that are almost
1330 * evenly spread out between numa nodes.
1332 static inline unsigned long task_weight(struct task_struct *p, int nid,
1335 unsigned long faults, total_faults;
1337 if (!p->numa_faults)
1340 total_faults = p->total_numa_faults;
1345 faults = task_faults(p, nid);
1346 faults += score_nearby_nodes(p, nid, dist, true);
1348 return 1000 * faults / total_faults;
1351 static inline unsigned long group_weight(struct task_struct *p, int nid,
1354 unsigned long faults, total_faults;
1359 total_faults = p->numa_group->total_faults;
1364 faults = group_faults(p, nid);
1365 faults += score_nearby_nodes(p, nid, dist, false);
1367 return 1000 * faults / total_faults;
1370 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1371 int src_nid, int dst_cpu)
1373 struct numa_group *ng = p->numa_group;
1374 int dst_nid = cpu_to_node(dst_cpu);
1375 int last_cpupid, this_cpupid;
1377 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1380 * Multi-stage node selection is used in conjunction with a periodic
1381 * migration fault to build a temporal task<->page relation. By using
1382 * a two-stage filter we remove short/unlikely relations.
1384 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1385 * a task's usage of a particular page (n_p) per total usage of this
1386 * page (n_t) (in a given time-span) to a probability.
1388 * Our periodic faults will sample this probability and getting the
1389 * same result twice in a row, given these samples are fully
1390 * independent, is then given by P(n)^2, provided our sample period
1391 * is sufficiently short compared to the usage pattern.
1393 * This quadric squishes small probabilities, making it less likely we
1394 * act on an unlikely task<->page relation.
1396 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1397 if (!cpupid_pid_unset(last_cpupid) &&
1398 cpupid_to_nid(last_cpupid) != dst_nid)
1401 /* Always allow migrate on private faults */
1402 if (cpupid_match_pid(p, last_cpupid))
1405 /* A shared fault, but p->numa_group has not been set up yet. */
1410 * Destination node is much more heavily used than the source
1411 * node? Allow migration.
1413 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1414 ACTIVE_NODE_FRACTION)
1418 * Distribute memory according to CPU & memory use on each node,
1419 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1421 * faults_cpu(dst) 3 faults_cpu(src)
1422 * --------------- * - > ---------------
1423 * faults_mem(dst) 4 faults_mem(src)
1425 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1426 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1429 static unsigned long weighted_cpuload(struct rq *rq);
1430 static unsigned long source_load(int cpu, int type);
1431 static unsigned long target_load(int cpu, int type);
1432 static unsigned long capacity_of(int cpu);
1434 /* Cached statistics for all CPUs within a node */
1436 unsigned long nr_running;
1439 /* Total compute capacity of CPUs on a node */
1440 unsigned long compute_capacity;
1442 /* Approximate capacity in terms of runnable tasks on a node */
1443 unsigned long task_capacity;
1444 int has_free_capacity;
1448 * XXX borrowed from update_sg_lb_stats
1450 static void update_numa_stats(struct numa_stats *ns, int nid)
1452 int smt, cpu, cpus = 0;
1453 unsigned long capacity;
1455 memset(ns, 0, sizeof(*ns));
1456 for_each_cpu(cpu, cpumask_of_node(nid)) {
1457 struct rq *rq = cpu_rq(cpu);
1459 ns->nr_running += rq->nr_running;
1460 ns->load += weighted_cpuload(rq);
1461 ns->compute_capacity += capacity_of(cpu);
1467 * If we raced with hotplug and there are no CPUs left in our mask
1468 * the @ns structure is NULL'ed and task_numa_compare() will
1469 * not find this node attractive.
1471 * We'll either bail at !has_free_capacity, or we'll detect a huge
1472 * imbalance and bail there.
1477 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1478 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1479 capacity = cpus / smt; /* cores */
1481 ns->task_capacity = min_t(unsigned, capacity,
1482 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1483 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1486 struct task_numa_env {
1487 struct task_struct *p;
1489 int src_cpu, src_nid;
1490 int dst_cpu, dst_nid;
1492 struct numa_stats src_stats, dst_stats;
1497 struct task_struct *best_task;
1502 static void task_numa_assign(struct task_numa_env *env,
1503 struct task_struct *p, long imp)
1506 put_task_struct(env->best_task);
1511 env->best_imp = imp;
1512 env->best_cpu = env->dst_cpu;
1515 static bool load_too_imbalanced(long src_load, long dst_load,
1516 struct task_numa_env *env)
1519 long orig_src_load, orig_dst_load;
1520 long src_capacity, dst_capacity;
1523 * The load is corrected for the CPU capacity available on each node.
1526 * ------------ vs ---------
1527 * src_capacity dst_capacity
1529 src_capacity = env->src_stats.compute_capacity;
1530 dst_capacity = env->dst_stats.compute_capacity;
1532 /* We care about the slope of the imbalance, not the direction. */
1533 if (dst_load < src_load)
1534 swap(dst_load, src_load);
1536 /* Is the difference below the threshold? */
1537 imb = dst_load * src_capacity * 100 -
1538 src_load * dst_capacity * env->imbalance_pct;
1543 * The imbalance is above the allowed threshold.
1544 * Compare it with the old imbalance.
1546 orig_src_load = env->src_stats.load;
1547 orig_dst_load = env->dst_stats.load;
1549 if (orig_dst_load < orig_src_load)
1550 swap(orig_dst_load, orig_src_load);
1552 old_imb = orig_dst_load * src_capacity * 100 -
1553 orig_src_load * dst_capacity * env->imbalance_pct;
1555 /* Would this change make things worse? */
1556 return (imb > old_imb);
1560 * This checks if the overall compute and NUMA accesses of the system would
1561 * be improved if the source tasks was migrated to the target dst_cpu taking
1562 * into account that it might be best if task running on the dst_cpu should
1563 * be exchanged with the source task
1565 static void task_numa_compare(struct task_numa_env *env,
1566 long taskimp, long groupimp)
1568 struct rq *src_rq = cpu_rq(env->src_cpu);
1569 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1570 struct task_struct *cur;
1571 long src_load, dst_load;
1573 long imp = env->p->numa_group ? groupimp : taskimp;
1575 int dist = env->dist;
1578 cur = task_rcu_dereference(&dst_rq->curr);
1579 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1583 * Because we have preemption enabled we can get migrated around and
1584 * end try selecting ourselves (current == env->p) as a swap candidate.
1590 * "imp" is the fault differential for the source task between the
1591 * source and destination node. Calculate the total differential for
1592 * the source task and potential destination task. The more negative
1593 * the value is, the more rmeote accesses that would be expected to
1594 * be incurred if the tasks were swapped.
1597 /* Skip this swap candidate if cannot move to the source cpu */
1598 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1602 * If dst and source tasks are in the same NUMA group, or not
1603 * in any group then look only at task weights.
1605 if (cur->numa_group == env->p->numa_group) {
1606 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1607 task_weight(cur, env->dst_nid, dist);
1609 * Add some hysteresis to prevent swapping the
1610 * tasks within a group over tiny differences.
1612 if (cur->numa_group)
1616 * Compare the group weights. If a task is all by
1617 * itself (not part of a group), use the task weight
1620 if (cur->numa_group)
1621 imp += group_weight(cur, env->src_nid, dist) -
1622 group_weight(cur, env->dst_nid, dist);
1624 imp += task_weight(cur, env->src_nid, dist) -
1625 task_weight(cur, env->dst_nid, dist);
1629 if (imp <= env->best_imp && moveimp <= env->best_imp)
1633 /* Is there capacity at our destination? */
1634 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1635 !env->dst_stats.has_free_capacity)
1641 /* Balance doesn't matter much if we're running a task per cpu */
1642 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1643 dst_rq->nr_running == 1)
1647 * In the overloaded case, try and keep the load balanced.
1650 load = task_h_load(env->p);
1651 dst_load = env->dst_stats.load + load;
1652 src_load = env->src_stats.load - load;
1654 if (moveimp > imp && moveimp > env->best_imp) {
1656 * If the improvement from just moving env->p direction is
1657 * better than swapping tasks around, check if a move is
1658 * possible. Store a slightly smaller score than moveimp,
1659 * so an actually idle CPU will win.
1661 if (!load_too_imbalanced(src_load, dst_load, env)) {
1668 if (imp <= env->best_imp)
1672 load = task_h_load(cur);
1677 if (load_too_imbalanced(src_load, dst_load, env))
1681 * One idle CPU per node is evaluated for a task numa move.
1682 * Call select_idle_sibling to maybe find a better one.
1686 * select_idle_siblings() uses an per-cpu cpumask that
1687 * can be used from IRQ context.
1689 local_irq_disable();
1690 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1696 task_numa_assign(env, cur, imp);
1701 static void task_numa_find_cpu(struct task_numa_env *env,
1702 long taskimp, long groupimp)
1706 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1707 /* Skip this CPU if the source task cannot migrate */
1708 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1712 task_numa_compare(env, taskimp, groupimp);
1716 /* Only move tasks to a NUMA node less busy than the current node. */
1717 static bool numa_has_capacity(struct task_numa_env *env)
1719 struct numa_stats *src = &env->src_stats;
1720 struct numa_stats *dst = &env->dst_stats;
1722 if (src->has_free_capacity && !dst->has_free_capacity)
1726 * Only consider a task move if the source has a higher load
1727 * than the destination, corrected for CPU capacity on each node.
1729 * src->load dst->load
1730 * --------------------- vs ---------------------
1731 * src->compute_capacity dst->compute_capacity
1733 if (src->load * dst->compute_capacity * env->imbalance_pct >
1735 dst->load * src->compute_capacity * 100)
1741 static int task_numa_migrate(struct task_struct *p)
1743 struct task_numa_env env = {
1746 .src_cpu = task_cpu(p),
1747 .src_nid = task_node(p),
1749 .imbalance_pct = 112,
1755 struct sched_domain *sd;
1756 unsigned long taskweight, groupweight;
1758 long taskimp, groupimp;
1761 * Pick the lowest SD_NUMA domain, as that would have the smallest
1762 * imbalance and would be the first to start moving tasks about.
1764 * And we want to avoid any moving of tasks about, as that would create
1765 * random movement of tasks -- counter the numa conditions we're trying
1769 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1771 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1775 * Cpusets can break the scheduler domain tree into smaller
1776 * balance domains, some of which do not cross NUMA boundaries.
1777 * Tasks that are "trapped" in such domains cannot be migrated
1778 * elsewhere, so there is no point in (re)trying.
1780 if (unlikely(!sd)) {
1781 p->numa_preferred_nid = task_node(p);
1785 env.dst_nid = p->numa_preferred_nid;
1786 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1787 taskweight = task_weight(p, env.src_nid, dist);
1788 groupweight = group_weight(p, env.src_nid, dist);
1789 update_numa_stats(&env.src_stats, env.src_nid);
1790 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1791 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1792 update_numa_stats(&env.dst_stats, env.dst_nid);
1794 /* Try to find a spot on the preferred nid. */
1795 if (numa_has_capacity(&env))
1796 task_numa_find_cpu(&env, taskimp, groupimp);
1799 * Look at other nodes in these cases:
1800 * - there is no space available on the preferred_nid
1801 * - the task is part of a numa_group that is interleaved across
1802 * multiple NUMA nodes; in order to better consolidate the group,
1803 * we need to check other locations.
1805 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1806 for_each_online_node(nid) {
1807 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1810 dist = node_distance(env.src_nid, env.dst_nid);
1811 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1813 taskweight = task_weight(p, env.src_nid, dist);
1814 groupweight = group_weight(p, env.src_nid, dist);
1817 /* Only consider nodes where both task and groups benefit */
1818 taskimp = task_weight(p, nid, dist) - taskweight;
1819 groupimp = group_weight(p, nid, dist) - groupweight;
1820 if (taskimp < 0 && groupimp < 0)
1825 update_numa_stats(&env.dst_stats, env.dst_nid);
1826 if (numa_has_capacity(&env))
1827 task_numa_find_cpu(&env, taskimp, groupimp);
1832 * If the task is part of a workload that spans multiple NUMA nodes,
1833 * and is migrating into one of the workload's active nodes, remember
1834 * this node as the task's preferred numa node, so the workload can
1836 * A task that migrated to a second choice node will be better off
1837 * trying for a better one later. Do not set the preferred node here.
1839 if (p->numa_group) {
1840 struct numa_group *ng = p->numa_group;
1842 if (env.best_cpu == -1)
1847 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1848 sched_setnuma(p, env.dst_nid);
1851 /* No better CPU than the current one was found. */
1852 if (env.best_cpu == -1)
1856 * Reset the scan period if the task is being rescheduled on an
1857 * alternative node to recheck if the tasks is now properly placed.
1859 p->numa_scan_period = task_scan_start(p);
1861 if (env.best_task == NULL) {
1862 ret = migrate_task_to(p, env.best_cpu);
1864 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1868 ret = migrate_swap(p, env.best_task);
1870 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1871 put_task_struct(env.best_task);
1875 /* Attempt to migrate a task to a CPU on the preferred node. */
1876 static void numa_migrate_preferred(struct task_struct *p)
1878 unsigned long interval = HZ;
1880 /* This task has no NUMA fault statistics yet */
1881 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1884 /* Periodically retry migrating the task to the preferred node */
1885 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1886 p->numa_migrate_retry = jiffies + interval;
1888 /* Success if task is already running on preferred CPU */
1889 if (task_node(p) == p->numa_preferred_nid)
1892 /* Otherwise, try migrate to a CPU on the preferred node */
1893 task_numa_migrate(p);
1897 * Find out how many nodes on the workload is actively running on. Do this by
1898 * tracking the nodes from which NUMA hinting faults are triggered. This can
1899 * be different from the set of nodes where the workload's memory is currently
1902 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1904 unsigned long faults, max_faults = 0;
1905 int nid, active_nodes = 0;
1907 for_each_online_node(nid) {
1908 faults = group_faults_cpu(numa_group, nid);
1909 if (faults > max_faults)
1910 max_faults = faults;
1913 for_each_online_node(nid) {
1914 faults = group_faults_cpu(numa_group, nid);
1915 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1919 numa_group->max_faults_cpu = max_faults;
1920 numa_group->active_nodes = active_nodes;
1924 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1925 * increments. The more local the fault statistics are, the higher the scan
1926 * period will be for the next scan window. If local/(local+remote) ratio is
1927 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1928 * the scan period will decrease. Aim for 70% local accesses.
1930 #define NUMA_PERIOD_SLOTS 10
1931 #define NUMA_PERIOD_THRESHOLD 7
1934 * Increase the scan period (slow down scanning) if the majority of
1935 * our memory is already on our local node, or if the majority of
1936 * the page accesses are shared with other processes.
1937 * Otherwise, decrease the scan period.
1939 static void update_task_scan_period(struct task_struct *p,
1940 unsigned long shared, unsigned long private)
1942 unsigned int period_slot;
1943 int lr_ratio, ps_ratio;
1946 unsigned long remote = p->numa_faults_locality[0];
1947 unsigned long local = p->numa_faults_locality[1];
1950 * If there were no record hinting faults then either the task is
1951 * completely idle or all activity is areas that are not of interest
1952 * to automatic numa balancing. Related to that, if there were failed
1953 * migration then it implies we are migrating too quickly or the local
1954 * node is overloaded. In either case, scan slower
1956 if (local + shared == 0 || p->numa_faults_locality[2]) {
1957 p->numa_scan_period = min(p->numa_scan_period_max,
1958 p->numa_scan_period << 1);
1960 p->mm->numa_next_scan = jiffies +
1961 msecs_to_jiffies(p->numa_scan_period);
1967 * Prepare to scale scan period relative to the current period.
1968 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1969 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1970 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1972 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1973 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1974 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1976 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1978 * Most memory accesses are local. There is no need to
1979 * do fast NUMA scanning, since memory is already local.
1981 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1984 diff = slot * period_slot;
1985 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1987 * Most memory accesses are shared with other tasks.
1988 * There is no point in continuing fast NUMA scanning,
1989 * since other tasks may just move the memory elsewhere.
1991 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1994 diff = slot * period_slot;
1997 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1998 * yet they are not on the local NUMA node. Speed up
1999 * NUMA scanning to get the memory moved over.
2001 int ratio = max(lr_ratio, ps_ratio);
2002 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2005 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2006 task_scan_min(p), task_scan_max(p));
2007 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2011 * Get the fraction of time the task has been running since the last
2012 * NUMA placement cycle. The scheduler keeps similar statistics, but
2013 * decays those on a 32ms period, which is orders of magnitude off
2014 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2015 * stats only if the task is so new there are no NUMA statistics yet.
2017 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2019 u64 runtime, delta, now;
2020 /* Use the start of this time slice to avoid calculations. */
2021 now = p->se.exec_start;
2022 runtime = p->se.sum_exec_runtime;
2024 if (p->last_task_numa_placement) {
2025 delta = runtime - p->last_sum_exec_runtime;
2026 *period = now - p->last_task_numa_placement;
2028 delta = p->se.avg.load_sum / p->se.load.weight;
2029 *period = LOAD_AVG_MAX;
2032 p->last_sum_exec_runtime = runtime;
2033 p->last_task_numa_placement = now;
2039 * Determine the preferred nid for a task in a numa_group. This needs to
2040 * be done in a way that produces consistent results with group_weight,
2041 * otherwise workloads might not converge.
2043 static int preferred_group_nid(struct task_struct *p, int nid)
2048 /* Direct connections between all NUMA nodes. */
2049 if (sched_numa_topology_type == NUMA_DIRECT)
2053 * On a system with glueless mesh NUMA topology, group_weight
2054 * scores nodes according to the number of NUMA hinting faults on
2055 * both the node itself, and on nearby nodes.
2057 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2058 unsigned long score, max_score = 0;
2059 int node, max_node = nid;
2061 dist = sched_max_numa_distance;
2063 for_each_online_node(node) {
2064 score = group_weight(p, node, dist);
2065 if (score > max_score) {
2074 * Finding the preferred nid in a system with NUMA backplane
2075 * interconnect topology is more involved. The goal is to locate
2076 * tasks from numa_groups near each other in the system, and
2077 * untangle workloads from different sides of the system. This requires
2078 * searching down the hierarchy of node groups, recursively searching
2079 * inside the highest scoring group of nodes. The nodemask tricks
2080 * keep the complexity of the search down.
2082 nodes = node_online_map;
2083 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2084 unsigned long max_faults = 0;
2085 nodemask_t max_group = NODE_MASK_NONE;
2088 /* Are there nodes at this distance from each other? */
2089 if (!find_numa_distance(dist))
2092 for_each_node_mask(a, nodes) {
2093 unsigned long faults = 0;
2094 nodemask_t this_group;
2095 nodes_clear(this_group);
2097 /* Sum group's NUMA faults; includes a==b case. */
2098 for_each_node_mask(b, nodes) {
2099 if (node_distance(a, b) < dist) {
2100 faults += group_faults(p, b);
2101 node_set(b, this_group);
2102 node_clear(b, nodes);
2106 /* Remember the top group. */
2107 if (faults > max_faults) {
2108 max_faults = faults;
2109 max_group = this_group;
2111 * subtle: at the smallest distance there is
2112 * just one node left in each "group", the
2113 * winner is the preferred nid.
2118 /* Next round, evaluate the nodes within max_group. */
2126 static void task_numa_placement(struct task_struct *p)
2128 int seq, nid, max_nid = -1, max_group_nid = -1;
2129 unsigned long max_faults = 0, max_group_faults = 0;
2130 unsigned long fault_types[2] = { 0, 0 };
2131 unsigned long total_faults;
2132 u64 runtime, period;
2133 spinlock_t *group_lock = NULL;
2136 * The p->mm->numa_scan_seq field gets updated without
2137 * exclusive access. Use READ_ONCE() here to ensure
2138 * that the field is read in a single access:
2140 seq = READ_ONCE(p->mm->numa_scan_seq);
2141 if (p->numa_scan_seq == seq)
2143 p->numa_scan_seq = seq;
2144 p->numa_scan_period_max = task_scan_max(p);
2146 total_faults = p->numa_faults_locality[0] +
2147 p->numa_faults_locality[1];
2148 runtime = numa_get_avg_runtime(p, &period);
2150 /* If the task is part of a group prevent parallel updates to group stats */
2151 if (p->numa_group) {
2152 group_lock = &p->numa_group->lock;
2153 spin_lock_irq(group_lock);
2156 /* Find the node with the highest number of faults */
2157 for_each_online_node(nid) {
2158 /* Keep track of the offsets in numa_faults array */
2159 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2160 unsigned long faults = 0, group_faults = 0;
2163 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2164 long diff, f_diff, f_weight;
2166 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2167 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2168 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2169 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2171 /* Decay existing window, copy faults since last scan */
2172 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2173 fault_types[priv] += p->numa_faults[membuf_idx];
2174 p->numa_faults[membuf_idx] = 0;
2177 * Normalize the faults_from, so all tasks in a group
2178 * count according to CPU use, instead of by the raw
2179 * number of faults. Tasks with little runtime have
2180 * little over-all impact on throughput, and thus their
2181 * faults are less important.
2183 f_weight = div64_u64(runtime << 16, period + 1);
2184 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2186 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2187 p->numa_faults[cpubuf_idx] = 0;
2189 p->numa_faults[mem_idx] += diff;
2190 p->numa_faults[cpu_idx] += f_diff;
2191 faults += p->numa_faults[mem_idx];
2192 p->total_numa_faults += diff;
2193 if (p->numa_group) {
2195 * safe because we can only change our own group
2197 * mem_idx represents the offset for a given
2198 * nid and priv in a specific region because it
2199 * is at the beginning of the numa_faults array.
2201 p->numa_group->faults[mem_idx] += diff;
2202 p->numa_group->faults_cpu[mem_idx] += f_diff;
2203 p->numa_group->total_faults += diff;
2204 group_faults += p->numa_group->faults[mem_idx];
2208 if (faults > max_faults) {
2209 max_faults = faults;
2213 if (group_faults > max_group_faults) {
2214 max_group_faults = group_faults;
2215 max_group_nid = nid;
2219 update_task_scan_period(p, fault_types[0], fault_types[1]);
2221 if (p->numa_group) {
2222 numa_group_count_active_nodes(p->numa_group);
2223 spin_unlock_irq(group_lock);
2224 max_nid = preferred_group_nid(p, max_group_nid);
2228 /* Set the new preferred node */
2229 if (max_nid != p->numa_preferred_nid)
2230 sched_setnuma(p, max_nid);
2232 if (task_node(p) != p->numa_preferred_nid)
2233 numa_migrate_preferred(p);
2237 static inline int get_numa_group(struct numa_group *grp)
2239 return atomic_inc_not_zero(&grp->refcount);
2242 static inline void put_numa_group(struct numa_group *grp)
2244 if (atomic_dec_and_test(&grp->refcount))
2245 kfree_rcu(grp, rcu);
2248 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2251 struct numa_group *grp, *my_grp;
2252 struct task_struct *tsk;
2254 int cpu = cpupid_to_cpu(cpupid);
2257 if (unlikely(!p->numa_group)) {
2258 unsigned int size = sizeof(struct numa_group) +
2259 4*nr_node_ids*sizeof(unsigned long);
2261 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2265 atomic_set(&grp->refcount, 1);
2266 grp->active_nodes = 1;
2267 grp->max_faults_cpu = 0;
2268 spin_lock_init(&grp->lock);
2270 /* Second half of the array tracks nids where faults happen */
2271 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2274 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2275 grp->faults[i] = p->numa_faults[i];
2277 grp->total_faults = p->total_numa_faults;
2280 rcu_assign_pointer(p->numa_group, grp);
2284 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2286 if (!cpupid_match_pid(tsk, cpupid))
2289 grp = rcu_dereference(tsk->numa_group);
2293 my_grp = p->numa_group;
2298 * Only join the other group if its bigger; if we're the bigger group,
2299 * the other task will join us.
2301 if (my_grp->nr_tasks > grp->nr_tasks)
2305 * Tie-break on the grp address.
2307 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2310 /* Always join threads in the same process. */
2311 if (tsk->mm == current->mm)
2314 /* Simple filter to avoid false positives due to PID collisions */
2315 if (flags & TNF_SHARED)
2318 /* Update priv based on whether false sharing was detected */
2321 if (join && !get_numa_group(grp))
2329 BUG_ON(irqs_disabled());
2330 double_lock_irq(&my_grp->lock, &grp->lock);
2332 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2333 my_grp->faults[i] -= p->numa_faults[i];
2334 grp->faults[i] += p->numa_faults[i];
2336 my_grp->total_faults -= p->total_numa_faults;
2337 grp->total_faults += p->total_numa_faults;
2342 spin_unlock(&my_grp->lock);
2343 spin_unlock_irq(&grp->lock);
2345 rcu_assign_pointer(p->numa_group, grp);
2347 put_numa_group(my_grp);
2355 void task_numa_free(struct task_struct *p)
2357 struct numa_group *grp = p->numa_group;
2358 void *numa_faults = p->numa_faults;
2359 unsigned long flags;
2363 spin_lock_irqsave(&grp->lock, flags);
2364 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2365 grp->faults[i] -= p->numa_faults[i];
2366 grp->total_faults -= p->total_numa_faults;
2369 spin_unlock_irqrestore(&grp->lock, flags);
2370 RCU_INIT_POINTER(p->numa_group, NULL);
2371 put_numa_group(grp);
2374 p->numa_faults = NULL;
2379 * Got a PROT_NONE fault for a page on @node.
2381 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2383 struct task_struct *p = current;
2384 bool migrated = flags & TNF_MIGRATED;
2385 int cpu_node = task_node(current);
2386 int local = !!(flags & TNF_FAULT_LOCAL);
2387 struct numa_group *ng;
2390 if (!static_branch_likely(&sched_numa_balancing))
2393 /* for example, ksmd faulting in a user's mm */
2397 /* Allocate buffer to track faults on a per-node basis */
2398 if (unlikely(!p->numa_faults)) {
2399 int size = sizeof(*p->numa_faults) *
2400 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2402 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2403 if (!p->numa_faults)
2406 p->total_numa_faults = 0;
2407 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2411 * First accesses are treated as private, otherwise consider accesses
2412 * to be private if the accessing pid has not changed
2414 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2417 priv = cpupid_match_pid(p, last_cpupid);
2418 if (!priv && !(flags & TNF_NO_GROUP))
2419 task_numa_group(p, last_cpupid, flags, &priv);
2423 * If a workload spans multiple NUMA nodes, a shared fault that
2424 * occurs wholly within the set of nodes that the workload is
2425 * actively using should be counted as local. This allows the
2426 * scan rate to slow down when a workload has settled down.
2429 if (!priv && !local && ng && ng->active_nodes > 1 &&
2430 numa_is_active_node(cpu_node, ng) &&
2431 numa_is_active_node(mem_node, ng))
2434 task_numa_placement(p);
2437 * Retry task to preferred node migration periodically, in case it
2438 * case it previously failed, or the scheduler moved us.
2440 if (time_after(jiffies, p->numa_migrate_retry))
2441 numa_migrate_preferred(p);
2444 p->numa_pages_migrated += pages;
2445 if (flags & TNF_MIGRATE_FAIL)
2446 p->numa_faults_locality[2] += pages;
2448 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2449 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2450 p->numa_faults_locality[local] += pages;
2453 static void reset_ptenuma_scan(struct task_struct *p)
2456 * We only did a read acquisition of the mmap sem, so
2457 * p->mm->numa_scan_seq is written to without exclusive access
2458 * and the update is not guaranteed to be atomic. That's not
2459 * much of an issue though, since this is just used for
2460 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2461 * expensive, to avoid any form of compiler optimizations:
2463 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2464 p->mm->numa_scan_offset = 0;
2468 * The expensive part of numa migration is done from task_work context.
2469 * Triggered from task_tick_numa().
2471 void task_numa_work(struct callback_head *work)
2473 unsigned long migrate, next_scan, now = jiffies;
2474 struct task_struct *p = current;
2475 struct mm_struct *mm = p->mm;
2476 u64 runtime = p->se.sum_exec_runtime;
2477 struct vm_area_struct *vma;
2478 unsigned long start, end;
2479 unsigned long nr_pte_updates = 0;
2480 long pages, virtpages;
2482 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2484 work->next = work; /* protect against double add */
2486 * Who cares about NUMA placement when they're dying.
2488 * NOTE: make sure not to dereference p->mm before this check,
2489 * exit_task_work() happens _after_ exit_mm() so we could be called
2490 * without p->mm even though we still had it when we enqueued this
2493 if (p->flags & PF_EXITING)
2496 if (!mm->numa_next_scan) {
2497 mm->numa_next_scan = now +
2498 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2502 * Enforce maximal scan/migration frequency..
2504 migrate = mm->numa_next_scan;
2505 if (time_before(now, migrate))
2508 if (p->numa_scan_period == 0) {
2509 p->numa_scan_period_max = task_scan_max(p);
2510 p->numa_scan_period = task_scan_start(p);
2513 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2514 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2518 * Delay this task enough that another task of this mm will likely win
2519 * the next time around.
2521 p->node_stamp += 2 * TICK_NSEC;
2523 start = mm->numa_scan_offset;
2524 pages = sysctl_numa_balancing_scan_size;
2525 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2526 virtpages = pages * 8; /* Scan up to this much virtual space */
2531 if (!down_read_trylock(&mm->mmap_sem))
2533 vma = find_vma(mm, start);
2535 reset_ptenuma_scan(p);
2539 for (; vma; vma = vma->vm_next) {
2540 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2541 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2546 * Shared library pages mapped by multiple processes are not
2547 * migrated as it is expected they are cache replicated. Avoid
2548 * hinting faults in read-only file-backed mappings or the vdso
2549 * as migrating the pages will be of marginal benefit.
2552 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2556 * Skip inaccessible VMAs to avoid any confusion between
2557 * PROT_NONE and NUMA hinting ptes
2559 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2563 start = max(start, vma->vm_start);
2564 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2565 end = min(end, vma->vm_end);
2566 nr_pte_updates = change_prot_numa(vma, start, end);
2569 * Try to scan sysctl_numa_balancing_size worth of
2570 * hpages that have at least one present PTE that
2571 * is not already pte-numa. If the VMA contains
2572 * areas that are unused or already full of prot_numa
2573 * PTEs, scan up to virtpages, to skip through those
2577 pages -= (end - start) >> PAGE_SHIFT;
2578 virtpages -= (end - start) >> PAGE_SHIFT;
2581 if (pages <= 0 || virtpages <= 0)
2585 } while (end != vma->vm_end);
2590 * It is possible to reach the end of the VMA list but the last few
2591 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2592 * would find the !migratable VMA on the next scan but not reset the
2593 * scanner to the start so check it now.
2596 mm->numa_scan_offset = start;
2598 reset_ptenuma_scan(p);
2599 up_read(&mm->mmap_sem);
2602 * Make sure tasks use at least 32x as much time to run other code
2603 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2604 * Usually update_task_scan_period slows down scanning enough; on an
2605 * overloaded system we need to limit overhead on a per task basis.
2607 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2608 u64 diff = p->se.sum_exec_runtime - runtime;
2609 p->node_stamp += 32 * diff;
2614 * Drive the periodic memory faults..
2616 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2618 struct callback_head *work = &curr->numa_work;
2622 * We don't care about NUMA placement if we don't have memory.
2624 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2628 * Using runtime rather than walltime has the dual advantage that
2629 * we (mostly) drive the selection from busy threads and that the
2630 * task needs to have done some actual work before we bother with
2633 now = curr->se.sum_exec_runtime;
2634 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2636 if (now > curr->node_stamp + period) {
2637 if (!curr->node_stamp)
2638 curr->numa_scan_period = task_scan_start(curr);
2639 curr->node_stamp += period;
2641 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2642 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2643 task_work_add(curr, work, true);
2649 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2653 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2657 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2661 #endif /* CONFIG_NUMA_BALANCING */
2664 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2666 update_load_add(&cfs_rq->load, se->load.weight);
2667 if (!parent_entity(se))
2668 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2670 if (entity_is_task(se)) {
2671 struct rq *rq = rq_of(cfs_rq);
2673 account_numa_enqueue(rq, task_of(se));
2674 list_add(&se->group_node, &rq->cfs_tasks);
2677 cfs_rq->nr_running++;
2681 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2683 update_load_sub(&cfs_rq->load, se->load.weight);
2684 if (!parent_entity(se))
2685 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2687 if (entity_is_task(se)) {
2688 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2689 list_del_init(&se->group_node);
2692 cfs_rq->nr_running--;
2695 #ifdef CONFIG_FAIR_GROUP_SCHED
2697 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2699 long tg_weight, load, shares;
2702 * This really should be: cfs_rq->avg.load_avg, but instead we use
2703 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2704 * the shares for small weight interactive tasks.
2706 load = scale_load_down(cfs_rq->load.weight);
2708 tg_weight = atomic_long_read(&tg->load_avg);
2710 /* Ensure tg_weight >= load */
2711 tg_weight -= cfs_rq->tg_load_avg_contrib;
2714 shares = (tg->shares * load);
2716 shares /= tg_weight;
2719 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2720 * of a group with small tg->shares value. It is a floor value which is
2721 * assigned as a minimum load.weight to the sched_entity representing
2722 * the group on a CPU.
2724 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2725 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2726 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2727 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2730 if (shares < MIN_SHARES)
2731 shares = MIN_SHARES;
2732 if (shares > tg->shares)
2733 shares = tg->shares;
2737 # else /* CONFIG_SMP */
2738 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2742 # endif /* CONFIG_SMP */
2744 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2745 unsigned long weight)
2748 /* commit outstanding execution time */
2749 if (cfs_rq->curr == se)
2750 update_curr(cfs_rq);
2751 account_entity_dequeue(cfs_rq, se);
2754 update_load_set(&se->load, weight);
2757 account_entity_enqueue(cfs_rq, se);
2760 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2762 static void update_cfs_shares(struct sched_entity *se)
2764 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2765 struct task_group *tg;
2771 if (throttled_hierarchy(cfs_rq))
2777 if (likely(se->load.weight == tg->shares))
2780 shares = calc_cfs_shares(cfs_rq, tg);
2782 reweight_entity(cfs_rq_of(se), se, shares);
2785 #else /* CONFIG_FAIR_GROUP_SCHED */
2786 static inline void update_cfs_shares(struct sched_entity *se)
2789 #endif /* CONFIG_FAIR_GROUP_SCHED */
2791 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2793 struct rq *rq = rq_of(cfs_rq);
2795 if (&rq->cfs == cfs_rq) {
2797 * There are a few boundary cases this might miss but it should
2798 * get called often enough that that should (hopefully) not be
2799 * a real problem -- added to that it only calls on the local
2800 * CPU, so if we enqueue remotely we'll miss an update, but
2801 * the next tick/schedule should update.
2803 * It will not get called when we go idle, because the idle
2804 * thread is a different class (!fair), nor will the utilization
2805 * number include things like RT tasks.
2807 * As is, the util number is not freq-invariant (we'd have to
2808 * implement arch_scale_freq_capacity() for that).
2812 cpufreq_update_util(rq, 0);
2819 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2821 static u64 decay_load(u64 val, u64 n)
2823 unsigned int local_n;
2825 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2828 /* after bounds checking we can collapse to 32-bit */
2832 * As y^PERIOD = 1/2, we can combine
2833 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2834 * With a look-up table which covers y^n (n<PERIOD)
2836 * To achieve constant time decay_load.
2838 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2839 val >>= local_n / LOAD_AVG_PERIOD;
2840 local_n %= LOAD_AVG_PERIOD;
2843 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2847 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2849 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2854 c1 = decay_load((u64)d1, periods);
2858 * c2 = 1024 \Sum y^n
2862 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2865 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2867 return c1 + c2 + c3;
2870 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2873 * Accumulate the three separate parts of the sum; d1 the remainder
2874 * of the last (incomplete) period, d2 the span of full periods and d3
2875 * the remainder of the (incomplete) current period.
2880 * |<->|<----------------->|<--->|
2881 * ... |---x---|------| ... |------|-----x (now)
2884 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2887 * = u y^p + (Step 1)
2890 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2893 static __always_inline u32
2894 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2895 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2897 unsigned long scale_freq, scale_cpu;
2898 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2901 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2902 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2904 delta += sa->period_contrib;
2905 periods = delta / 1024; /* A period is 1024us (~1ms) */
2908 * Step 1: decay old *_sum if we crossed period boundaries.
2911 sa->load_sum = decay_load(sa->load_sum, periods);
2913 cfs_rq->runnable_load_sum =
2914 decay_load(cfs_rq->runnable_load_sum, periods);
2916 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2922 contrib = __accumulate_pelt_segments(periods,
2923 1024 - sa->period_contrib, delta);
2925 sa->period_contrib = delta;
2927 contrib = cap_scale(contrib, scale_freq);
2929 sa->load_sum += weight * contrib;
2931 cfs_rq->runnable_load_sum += weight * contrib;
2934 sa->util_sum += contrib * scale_cpu;
2940 * We can represent the historical contribution to runnable average as the
2941 * coefficients of a geometric series. To do this we sub-divide our runnable
2942 * history into segments of approximately 1ms (1024us); label the segment that
2943 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2945 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2947 * (now) (~1ms ago) (~2ms ago)
2949 * Let u_i denote the fraction of p_i that the entity was runnable.
2951 * We then designate the fractions u_i as our co-efficients, yielding the
2952 * following representation of historical load:
2953 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2955 * We choose y based on the with of a reasonably scheduling period, fixing:
2958 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2959 * approximately half as much as the contribution to load within the last ms
2962 * When a period "rolls over" and we have new u_0`, multiplying the previous
2963 * sum again by y is sufficient to update:
2964 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2965 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2967 static __always_inline int
2968 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2969 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2973 delta = now - sa->last_update_time;
2975 * This should only happen when time goes backwards, which it
2976 * unfortunately does during sched clock init when we swap over to TSC.
2978 if ((s64)delta < 0) {
2979 sa->last_update_time = now;
2984 * Use 1024ns as the unit of measurement since it's a reasonable
2985 * approximation of 1us and fast to compute.
2991 sa->last_update_time += delta << 10;
2994 * running is a subset of runnable (weight) so running can't be set if
2995 * runnable is clear. But there are some corner cases where the current
2996 * se has been already dequeued but cfs_rq->curr still points to it.
2997 * This means that weight will be 0 but not running for a sched_entity
2998 * but also for a cfs_rq if the latter becomes idle. As an example,
2999 * this happens during idle_balance() which calls
3000 * update_blocked_averages()
3006 * Now we know we crossed measurement unit boundaries. The *_avg
3007 * accrues by two steps:
3009 * Step 1: accumulate *_sum since last_update_time. If we haven't
3010 * crossed period boundaries, finish.
3012 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
3016 * Step 2: update *_avg.
3018 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3020 cfs_rq->runnable_load_avg =
3021 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3023 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
3029 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3031 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
3035 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3037 return ___update_load_avg(now, cpu, &se->avg,
3038 se->on_rq * scale_load_down(se->load.weight),
3039 cfs_rq->curr == se, NULL);
3043 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3045 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3046 scale_load_down(cfs_rq->load.weight),
3047 cfs_rq->curr != NULL, cfs_rq);
3051 * Signed add and clamp on underflow.
3053 * Explicitly do a load-store to ensure the intermediate value never hits
3054 * memory. This allows lockless observations without ever seeing the negative
3057 #define add_positive(_ptr, _val) do { \
3058 typeof(_ptr) ptr = (_ptr); \
3059 typeof(_val) val = (_val); \
3060 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3064 if (val < 0 && res > var) \
3067 WRITE_ONCE(*ptr, res); \
3070 #ifdef CONFIG_FAIR_GROUP_SCHED
3072 * update_tg_load_avg - update the tg's load avg
3073 * @cfs_rq: the cfs_rq whose avg changed
3074 * @force: update regardless of how small the difference
3076 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3077 * However, because tg->load_avg is a global value there are performance
3080 * In order to avoid having to look at the other cfs_rq's, we use a
3081 * differential update where we store the last value we propagated. This in
3082 * turn allows skipping updates if the differential is 'small'.
3084 * Updating tg's load_avg is necessary before update_cfs_share().
3086 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3088 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3091 * No need to update load_avg for root_task_group as it is not used.
3093 if (cfs_rq->tg == &root_task_group)
3096 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3097 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3098 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3103 * Called within set_task_rq() right before setting a task's cpu. The
3104 * caller only guarantees p->pi_lock is held; no other assumptions,
3105 * including the state of rq->lock, should be made.
3107 void set_task_rq_fair(struct sched_entity *se,
3108 struct cfs_rq *prev, struct cfs_rq *next)
3110 u64 p_last_update_time;
3111 u64 n_last_update_time;
3113 if (!sched_feat(ATTACH_AGE_LOAD))
3117 * We are supposed to update the task to "current" time, then its up to
3118 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3119 * getting what current time is, so simply throw away the out-of-date
3120 * time. This will result in the wakee task is less decayed, but giving
3121 * the wakee more load sounds not bad.
3123 if (!(se->avg.last_update_time && prev))
3126 #ifndef CONFIG_64BIT
3128 u64 p_last_update_time_copy;
3129 u64 n_last_update_time_copy;
3132 p_last_update_time_copy = prev->load_last_update_time_copy;
3133 n_last_update_time_copy = next->load_last_update_time_copy;
3137 p_last_update_time = prev->avg.last_update_time;
3138 n_last_update_time = next->avg.last_update_time;
3140 } while (p_last_update_time != p_last_update_time_copy ||
3141 n_last_update_time != n_last_update_time_copy);
3144 p_last_update_time = prev->avg.last_update_time;
3145 n_last_update_time = next->avg.last_update_time;
3147 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3148 se->avg.last_update_time = n_last_update_time;
3151 /* Take into account change of utilization of a child task group */
3153 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3155 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3156 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3158 /* Nothing to update */
3162 /* Set new sched_entity's utilization */
3163 se->avg.util_avg = gcfs_rq->avg.util_avg;
3164 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3166 /* Update parent cfs_rq utilization */
3167 add_positive(&cfs_rq->avg.util_avg, delta);
3168 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3171 /* Take into account change of load of a child task group */
3173 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3175 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3176 long delta, load = gcfs_rq->avg.load_avg;
3179 * If the load of group cfs_rq is null, the load of the
3180 * sched_entity will also be null so we can skip the formula
3185 /* Get tg's load and ensure tg_load > 0 */
3186 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3188 /* Ensure tg_load >= load and updated with current load*/
3189 tg_load -= gcfs_rq->tg_load_avg_contrib;
3193 * We need to compute a correction term in the case that the
3194 * task group is consuming more CPU than a task of equal
3195 * weight. A task with a weight equals to tg->shares will have
3196 * a load less or equal to scale_load_down(tg->shares).
3197 * Similarly, the sched_entities that represent the task group
3198 * at parent level, can't have a load higher than
3199 * scale_load_down(tg->shares). And the Sum of sched_entities'
3200 * load must be <= scale_load_down(tg->shares).
3202 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3203 /* scale gcfs_rq's load into tg's shares*/
3204 load *= scale_load_down(gcfs_rq->tg->shares);
3209 delta = load - se->avg.load_avg;
3211 /* Nothing to update */
3215 /* Set new sched_entity's load */
3216 se->avg.load_avg = load;
3217 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3219 /* Update parent cfs_rq load */
3220 add_positive(&cfs_rq->avg.load_avg, delta);
3221 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3224 * If the sched_entity is already enqueued, we also have to update the
3225 * runnable load avg.
3228 /* Update parent cfs_rq runnable_load_avg */
3229 add_positive(&cfs_rq->runnable_load_avg, delta);
3230 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3234 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3236 cfs_rq->propagate_avg = 1;
3239 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3241 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3243 if (!cfs_rq->propagate_avg)
3246 cfs_rq->propagate_avg = 0;
3250 /* Update task and its cfs_rq load average */
3251 static inline int propagate_entity_load_avg(struct sched_entity *se)
3253 struct cfs_rq *cfs_rq;
3255 if (entity_is_task(se))
3258 if (!test_and_clear_tg_cfs_propagate(se))
3261 cfs_rq = cfs_rq_of(se);
3263 set_tg_cfs_propagate(cfs_rq);
3265 update_tg_cfs_util(cfs_rq, se);
3266 update_tg_cfs_load(cfs_rq, se);
3272 * Check if we need to update the load and the utilization of a blocked
3275 static inline bool skip_blocked_update(struct sched_entity *se)
3277 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3280 * If sched_entity still have not zero load or utilization, we have to
3283 if (se->avg.load_avg || se->avg.util_avg)
3287 * If there is a pending propagation, we have to update the load and
3288 * the utilization of the sched_entity:
3290 if (gcfs_rq->propagate_avg)
3294 * Otherwise, the load and the utilization of the sched_entity is
3295 * already zero and there is no pending propagation, so it will be a
3296 * waste of time to try to decay it:
3301 #else /* CONFIG_FAIR_GROUP_SCHED */
3303 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3305 static inline int propagate_entity_load_avg(struct sched_entity *se)
3310 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3312 #endif /* CONFIG_FAIR_GROUP_SCHED */
3315 * Unsigned subtract and clamp on underflow.
3317 * Explicitly do a load-store to ensure the intermediate value never hits
3318 * memory. This allows lockless observations without ever seeing the negative
3321 #define sub_positive(_ptr, _val) do { \
3322 typeof(_ptr) ptr = (_ptr); \
3323 typeof(*ptr) val = (_val); \
3324 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3328 WRITE_ONCE(*ptr, res); \
3332 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3333 * @now: current time, as per cfs_rq_clock_task()
3334 * @cfs_rq: cfs_rq to update
3336 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3337 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3338 * post_init_entity_util_avg().
3340 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3342 * Returns true if the load decayed or we removed load.
3344 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3345 * call update_tg_load_avg() when this function returns true.
3348 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3350 struct sched_avg *sa = &cfs_rq->avg;
3351 int decayed, removed_load = 0, removed_util = 0;
3353 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3354 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3355 sub_positive(&sa->load_avg, r);
3356 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3358 set_tg_cfs_propagate(cfs_rq);
3361 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3362 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3363 sub_positive(&sa->util_avg, r);
3364 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3366 set_tg_cfs_propagate(cfs_rq);
3369 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3371 #ifndef CONFIG_64BIT
3373 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3376 if (decayed || removed_util)
3377 cfs_rq_util_change(cfs_rq);
3379 return decayed || removed_load;
3383 * Optional action to be done while updating the load average
3385 #define UPDATE_TG 0x1
3386 #define SKIP_AGE_LOAD 0x2
3388 /* Update task and its cfs_rq load average */
3389 static inline void update_load_avg(struct sched_entity *se, int flags)
3391 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3392 u64 now = cfs_rq_clock_task(cfs_rq);
3393 struct rq *rq = rq_of(cfs_rq);
3394 int cpu = cpu_of(rq);
3398 * Track task load average for carrying it to new CPU after migrated, and
3399 * track group sched_entity load average for task_h_load calc in migration
3401 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3402 __update_load_avg_se(now, cpu, cfs_rq, se);
3404 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3405 decayed |= propagate_entity_load_avg(se);
3407 if (decayed && (flags & UPDATE_TG))
3408 update_tg_load_avg(cfs_rq, 0);
3412 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3413 * @cfs_rq: cfs_rq to attach to
3414 * @se: sched_entity to attach
3416 * Must call update_cfs_rq_load_avg() before this, since we rely on
3417 * cfs_rq->avg.last_update_time being current.
3419 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3421 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3422 cfs_rq->avg.load_avg += se->avg.load_avg;
3423 cfs_rq->avg.load_sum += se->avg.load_sum;
3424 cfs_rq->avg.util_avg += se->avg.util_avg;
3425 cfs_rq->avg.util_sum += se->avg.util_sum;
3426 set_tg_cfs_propagate(cfs_rq);
3428 cfs_rq_util_change(cfs_rq);
3432 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3433 * @cfs_rq: cfs_rq to detach from
3434 * @se: sched_entity to detach
3436 * Must call update_cfs_rq_load_avg() before this, since we rely on
3437 * cfs_rq->avg.last_update_time being current.
3439 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3442 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3443 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3444 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3445 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3446 set_tg_cfs_propagate(cfs_rq);
3448 cfs_rq_util_change(cfs_rq);
3451 /* Add the load generated by se into cfs_rq's load average */
3453 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3455 struct sched_avg *sa = &se->avg;
3457 cfs_rq->runnable_load_avg += sa->load_avg;
3458 cfs_rq->runnable_load_sum += sa->load_sum;
3460 if (!sa->last_update_time) {
3461 attach_entity_load_avg(cfs_rq, se);
3462 update_tg_load_avg(cfs_rq, 0);
3466 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3468 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3470 cfs_rq->runnable_load_avg =
3471 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3472 cfs_rq->runnable_load_sum =
3473 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3476 #ifndef CONFIG_64BIT
3477 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3479 u64 last_update_time_copy;
3480 u64 last_update_time;
3483 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3485 last_update_time = cfs_rq->avg.last_update_time;
3486 } while (last_update_time != last_update_time_copy);
3488 return last_update_time;
3491 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3493 return cfs_rq->avg.last_update_time;
3498 * Synchronize entity load avg of dequeued entity without locking
3501 void sync_entity_load_avg(struct sched_entity *se)
3503 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3504 u64 last_update_time;
3506 last_update_time = cfs_rq_last_update_time(cfs_rq);
3507 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3511 * Task first catches up with cfs_rq, and then subtract
3512 * itself from the cfs_rq (task must be off the queue now).
3514 void remove_entity_load_avg(struct sched_entity *se)
3516 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3519 * tasks cannot exit without having gone through wake_up_new_task() ->
3520 * post_init_entity_util_avg() which will have added things to the
3521 * cfs_rq, so we can remove unconditionally.
3523 * Similarly for groups, they will have passed through
3524 * post_init_entity_util_avg() before unregister_sched_fair_group()
3528 sync_entity_load_avg(se);
3529 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3530 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3533 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3535 return cfs_rq->runnable_load_avg;
3538 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3540 return cfs_rq->avg.load_avg;
3543 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3545 #else /* CONFIG_SMP */
3548 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3553 #define UPDATE_TG 0x0
3554 #define SKIP_AGE_LOAD 0x0
3556 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3558 cfs_rq_util_change(cfs_rq_of(se));
3562 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3564 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3565 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3568 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3570 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3572 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3577 #endif /* CONFIG_SMP */
3579 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3581 #ifdef CONFIG_SCHED_DEBUG
3582 s64 d = se->vruntime - cfs_rq->min_vruntime;
3587 if (d > 3*sysctl_sched_latency)
3588 schedstat_inc(cfs_rq->nr_spread_over);
3593 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3595 u64 vruntime = cfs_rq->min_vruntime;
3598 * The 'current' period is already promised to the current tasks,
3599 * however the extra weight of the new task will slow them down a
3600 * little, place the new task so that it fits in the slot that
3601 * stays open at the end.
3603 if (initial && sched_feat(START_DEBIT))
3604 vruntime += sched_vslice(cfs_rq, se);
3606 /* sleeps up to a single latency don't count. */
3608 unsigned long thresh = sysctl_sched_latency;
3611 * Halve their sleep time's effect, to allow
3612 * for a gentler effect of sleepers:
3614 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3620 /* ensure we never gain time by being placed backwards. */
3621 se->vruntime = max_vruntime(se->vruntime, vruntime);
3624 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3626 static inline void check_schedstat_required(void)
3628 #ifdef CONFIG_SCHEDSTATS
3629 if (schedstat_enabled())
3632 /* Force schedstat enabled if a dependent tracepoint is active */
3633 if (trace_sched_stat_wait_enabled() ||
3634 trace_sched_stat_sleep_enabled() ||
3635 trace_sched_stat_iowait_enabled() ||
3636 trace_sched_stat_blocked_enabled() ||
3637 trace_sched_stat_runtime_enabled()) {
3638 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3639 "stat_blocked and stat_runtime require the "
3640 "kernel parameter schedstats=enable or "
3641 "kernel.sched_schedstats=1\n");
3652 * update_min_vruntime()
3653 * vruntime -= min_vruntime
3657 * update_min_vruntime()
3658 * vruntime += min_vruntime
3660 * this way the vruntime transition between RQs is done when both
3661 * min_vruntime are up-to-date.
3665 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3666 * vruntime -= min_vruntime
3670 * update_min_vruntime()
3671 * vruntime += min_vruntime
3673 * this way we don't have the most up-to-date min_vruntime on the originating
3674 * CPU and an up-to-date min_vruntime on the destination CPU.
3678 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3680 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3681 bool curr = cfs_rq->curr == se;
3684 * If we're the current task, we must renormalise before calling
3688 se->vruntime += cfs_rq->min_vruntime;
3690 update_curr(cfs_rq);
3693 * Otherwise, renormalise after, such that we're placed at the current
3694 * moment in time, instead of some random moment in the past. Being
3695 * placed in the past could significantly boost this task to the
3696 * fairness detriment of existing tasks.
3698 if (renorm && !curr)
3699 se->vruntime += cfs_rq->min_vruntime;
3702 * When enqueuing a sched_entity, we must:
3703 * - Update loads to have both entity and cfs_rq synced with now.
3704 * - Add its load to cfs_rq->runnable_avg
3705 * - For group_entity, update its weight to reflect the new share of
3707 * - Add its new weight to cfs_rq->load.weight
3709 update_load_avg(se, UPDATE_TG);
3710 enqueue_entity_load_avg(cfs_rq, se);
3711 update_cfs_shares(se);
3712 account_entity_enqueue(cfs_rq, se);
3714 if (flags & ENQUEUE_WAKEUP)
3715 place_entity(cfs_rq, se, 0);
3717 check_schedstat_required();
3718 update_stats_enqueue(cfs_rq, se, flags);
3719 check_spread(cfs_rq, se);
3721 __enqueue_entity(cfs_rq, se);
3724 if (cfs_rq->nr_running == 1) {
3725 list_add_leaf_cfs_rq(cfs_rq);
3726 check_enqueue_throttle(cfs_rq);
3730 static void __clear_buddies_last(struct sched_entity *se)
3732 for_each_sched_entity(se) {
3733 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3734 if (cfs_rq->last != se)
3737 cfs_rq->last = NULL;
3741 static void __clear_buddies_next(struct sched_entity *se)
3743 for_each_sched_entity(se) {
3744 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3745 if (cfs_rq->next != se)
3748 cfs_rq->next = NULL;
3752 static void __clear_buddies_skip(struct sched_entity *se)
3754 for_each_sched_entity(se) {
3755 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3756 if (cfs_rq->skip != se)
3759 cfs_rq->skip = NULL;
3763 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3765 if (cfs_rq->last == se)
3766 __clear_buddies_last(se);
3768 if (cfs_rq->next == se)
3769 __clear_buddies_next(se);
3771 if (cfs_rq->skip == se)
3772 __clear_buddies_skip(se);
3775 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3778 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3781 * Update run-time statistics of the 'current'.
3783 update_curr(cfs_rq);
3786 * When dequeuing a sched_entity, we must:
3787 * - Update loads to have both entity and cfs_rq synced with now.
3788 * - Substract its load from the cfs_rq->runnable_avg.
3789 * - Substract its previous weight from cfs_rq->load.weight.
3790 * - For group entity, update its weight to reflect the new share
3791 * of its group cfs_rq.
3793 update_load_avg(se, UPDATE_TG);
3794 dequeue_entity_load_avg(cfs_rq, se);
3796 update_stats_dequeue(cfs_rq, se, flags);
3798 clear_buddies(cfs_rq, se);
3800 if (se != cfs_rq->curr)
3801 __dequeue_entity(cfs_rq, se);
3803 account_entity_dequeue(cfs_rq, se);
3806 * Normalize after update_curr(); which will also have moved
3807 * min_vruntime if @se is the one holding it back. But before doing
3808 * update_min_vruntime() again, which will discount @se's position and
3809 * can move min_vruntime forward still more.
3811 if (!(flags & DEQUEUE_SLEEP))
3812 se->vruntime -= cfs_rq->min_vruntime;
3814 /* return excess runtime on last dequeue */
3815 return_cfs_rq_runtime(cfs_rq);
3817 update_cfs_shares(se);
3820 * Now advance min_vruntime if @se was the entity holding it back,
3821 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3822 * put back on, and if we advance min_vruntime, we'll be placed back
3823 * further than we started -- ie. we'll be penalized.
3825 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3826 update_min_vruntime(cfs_rq);
3830 * Preempt the current task with a newly woken task if needed:
3833 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3835 unsigned long ideal_runtime, delta_exec;
3836 struct sched_entity *se;
3839 ideal_runtime = sched_slice(cfs_rq, curr);
3840 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3841 if (delta_exec > ideal_runtime) {
3842 resched_curr(rq_of(cfs_rq));
3844 * The current task ran long enough, ensure it doesn't get
3845 * re-elected due to buddy favours.
3847 clear_buddies(cfs_rq, curr);
3852 * Ensure that a task that missed wakeup preemption by a
3853 * narrow margin doesn't have to wait for a full slice.
3854 * This also mitigates buddy induced latencies under load.
3856 if (delta_exec < sysctl_sched_min_granularity)
3859 se = __pick_first_entity(cfs_rq);
3860 delta = curr->vruntime - se->vruntime;
3865 if (delta > ideal_runtime)
3866 resched_curr(rq_of(cfs_rq));
3870 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3872 /* 'current' is not kept within the tree. */
3875 * Any task has to be enqueued before it get to execute on
3876 * a CPU. So account for the time it spent waiting on the
3879 update_stats_wait_end(cfs_rq, se);
3880 __dequeue_entity(cfs_rq, se);
3881 update_load_avg(se, UPDATE_TG);
3884 update_stats_curr_start(cfs_rq, se);
3888 * Track our maximum slice length, if the CPU's load is at
3889 * least twice that of our own weight (i.e. dont track it
3890 * when there are only lesser-weight tasks around):
3892 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3893 schedstat_set(se->statistics.slice_max,
3894 max((u64)schedstat_val(se->statistics.slice_max),
3895 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3898 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3902 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3905 * Pick the next process, keeping these things in mind, in this order:
3906 * 1) keep things fair between processes/task groups
3907 * 2) pick the "next" process, since someone really wants that to run
3908 * 3) pick the "last" process, for cache locality
3909 * 4) do not run the "skip" process, if something else is available
3911 static struct sched_entity *
3912 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3914 struct sched_entity *left = __pick_first_entity(cfs_rq);
3915 struct sched_entity *se;
3918 * If curr is set we have to see if its left of the leftmost entity
3919 * still in the tree, provided there was anything in the tree at all.
3921 if (!left || (curr && entity_before(curr, left)))
3924 se = left; /* ideally we run the leftmost entity */
3927 * Avoid running the skip buddy, if running something else can
3928 * be done without getting too unfair.
3930 if (cfs_rq->skip == se) {
3931 struct sched_entity *second;
3934 second = __pick_first_entity(cfs_rq);
3936 second = __pick_next_entity(se);
3937 if (!second || (curr && entity_before(curr, second)))
3941 if (second && wakeup_preempt_entity(second, left) < 1)
3946 * Prefer last buddy, try to return the CPU to a preempted task.
3948 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3952 * Someone really wants this to run. If it's not unfair, run it.
3954 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3957 clear_buddies(cfs_rq, se);
3962 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3964 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3967 * If still on the runqueue then deactivate_task()
3968 * was not called and update_curr() has to be done:
3971 update_curr(cfs_rq);
3973 /* throttle cfs_rqs exceeding runtime */
3974 check_cfs_rq_runtime(cfs_rq);
3976 check_spread(cfs_rq, prev);
3979 update_stats_wait_start(cfs_rq, prev);
3980 /* Put 'current' back into the tree. */
3981 __enqueue_entity(cfs_rq, prev);
3982 /* in !on_rq case, update occurred at dequeue */
3983 update_load_avg(prev, 0);
3985 cfs_rq->curr = NULL;
3989 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3992 * Update run-time statistics of the 'current'.
3994 update_curr(cfs_rq);
3997 * Ensure that runnable average is periodically updated.
3999 update_load_avg(curr, UPDATE_TG);
4000 update_cfs_shares(curr);
4002 #ifdef CONFIG_SCHED_HRTICK
4004 * queued ticks are scheduled to match the slice, so don't bother
4005 * validating it and just reschedule.
4008 resched_curr(rq_of(cfs_rq));
4012 * don't let the period tick interfere with the hrtick preemption
4014 if (!sched_feat(DOUBLE_TICK) &&
4015 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4019 if (cfs_rq->nr_running > 1)
4020 check_preempt_tick(cfs_rq, curr);
4024 /**************************************************
4025 * CFS bandwidth control machinery
4028 #ifdef CONFIG_CFS_BANDWIDTH
4030 #ifdef HAVE_JUMP_LABEL
4031 static struct static_key __cfs_bandwidth_used;
4033 static inline bool cfs_bandwidth_used(void)
4035 return static_key_false(&__cfs_bandwidth_used);
4038 void cfs_bandwidth_usage_inc(void)
4040 static_key_slow_inc(&__cfs_bandwidth_used);
4043 void cfs_bandwidth_usage_dec(void)
4045 static_key_slow_dec(&__cfs_bandwidth_used);
4047 #else /* HAVE_JUMP_LABEL */
4048 static bool cfs_bandwidth_used(void)
4053 void cfs_bandwidth_usage_inc(void) {}
4054 void cfs_bandwidth_usage_dec(void) {}
4055 #endif /* HAVE_JUMP_LABEL */
4058 * default period for cfs group bandwidth.
4059 * default: 0.1s, units: nanoseconds
4061 static inline u64 default_cfs_period(void)
4063 return 100000000ULL;
4066 static inline u64 sched_cfs_bandwidth_slice(void)
4068 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4072 * Replenish runtime according to assigned quota and update expiration time.
4073 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4074 * additional synchronization around rq->lock.
4076 * requires cfs_b->lock
4078 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4082 if (cfs_b->quota == RUNTIME_INF)
4085 now = sched_clock_cpu(smp_processor_id());
4086 cfs_b->runtime = cfs_b->quota;
4087 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4090 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4092 return &tg->cfs_bandwidth;
4095 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4096 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4098 if (unlikely(cfs_rq->throttle_count))
4099 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4101 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4104 /* returns 0 on failure to allocate runtime */
4105 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4107 struct task_group *tg = cfs_rq->tg;
4108 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4109 u64 amount = 0, min_amount, expires;
4111 /* note: this is a positive sum as runtime_remaining <= 0 */
4112 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4114 raw_spin_lock(&cfs_b->lock);
4115 if (cfs_b->quota == RUNTIME_INF)
4116 amount = min_amount;
4118 start_cfs_bandwidth(cfs_b);
4120 if (cfs_b->runtime > 0) {
4121 amount = min(cfs_b->runtime, min_amount);
4122 cfs_b->runtime -= amount;
4126 expires = cfs_b->runtime_expires;
4127 raw_spin_unlock(&cfs_b->lock);
4129 cfs_rq->runtime_remaining += amount;
4131 * we may have advanced our local expiration to account for allowed
4132 * spread between our sched_clock and the one on which runtime was
4135 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4136 cfs_rq->runtime_expires = expires;
4138 return cfs_rq->runtime_remaining > 0;
4142 * Note: This depends on the synchronization provided by sched_clock and the
4143 * fact that rq->clock snapshots this value.
4145 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4147 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4149 /* if the deadline is ahead of our clock, nothing to do */
4150 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4153 if (cfs_rq->runtime_remaining < 0)
4157 * If the local deadline has passed we have to consider the
4158 * possibility that our sched_clock is 'fast' and the global deadline
4159 * has not truly expired.
4161 * Fortunately we can check determine whether this the case by checking
4162 * whether the global deadline has advanced. It is valid to compare
4163 * cfs_b->runtime_expires without any locks since we only care about
4164 * exact equality, so a partial write will still work.
4167 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4168 /* extend local deadline, drift is bounded above by 2 ticks */
4169 cfs_rq->runtime_expires += TICK_NSEC;
4171 /* global deadline is ahead, expiration has passed */
4172 cfs_rq->runtime_remaining = 0;
4176 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4178 /* dock delta_exec before expiring quota (as it could span periods) */
4179 cfs_rq->runtime_remaining -= delta_exec;
4180 expire_cfs_rq_runtime(cfs_rq);
4182 if (likely(cfs_rq->runtime_remaining > 0))
4186 * if we're unable to extend our runtime we resched so that the active
4187 * hierarchy can be throttled
4189 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4190 resched_curr(rq_of(cfs_rq));
4193 static __always_inline
4194 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4196 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4199 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4202 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4204 return cfs_bandwidth_used() && cfs_rq->throttled;
4207 /* check whether cfs_rq, or any parent, is throttled */
4208 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4210 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4214 * Ensure that neither of the group entities corresponding to src_cpu or
4215 * dest_cpu are members of a throttled hierarchy when performing group
4216 * load-balance operations.
4218 static inline int throttled_lb_pair(struct task_group *tg,
4219 int src_cpu, int dest_cpu)
4221 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4223 src_cfs_rq = tg->cfs_rq[src_cpu];
4224 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4226 return throttled_hierarchy(src_cfs_rq) ||
4227 throttled_hierarchy(dest_cfs_rq);
4230 /* updated child weight may affect parent so we have to do this bottom up */
4231 static int tg_unthrottle_up(struct task_group *tg, void *data)
4233 struct rq *rq = data;
4234 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4236 cfs_rq->throttle_count--;
4237 if (!cfs_rq->throttle_count) {
4238 /* adjust cfs_rq_clock_task() */
4239 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4240 cfs_rq->throttled_clock_task;
4246 static int tg_throttle_down(struct task_group *tg, void *data)
4248 struct rq *rq = data;
4249 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4251 /* group is entering throttled state, stop time */
4252 if (!cfs_rq->throttle_count)
4253 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4254 cfs_rq->throttle_count++;
4259 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4261 struct rq *rq = rq_of(cfs_rq);
4262 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4263 struct sched_entity *se;
4264 long task_delta, dequeue = 1;
4267 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4269 /* freeze hierarchy runnable averages while throttled */
4271 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4274 task_delta = cfs_rq->h_nr_running;
4275 for_each_sched_entity(se) {
4276 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4277 /* throttled entity or throttle-on-deactivate */
4282 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4283 qcfs_rq->h_nr_running -= task_delta;
4285 if (qcfs_rq->load.weight)
4290 sub_nr_running(rq, task_delta);
4292 cfs_rq->throttled = 1;
4293 cfs_rq->throttled_clock = rq_clock(rq);
4294 raw_spin_lock(&cfs_b->lock);
4295 empty = list_empty(&cfs_b->throttled_cfs_rq);
4298 * Add to the _head_ of the list, so that an already-started
4299 * distribute_cfs_runtime will not see us
4301 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4304 * If we're the first throttled task, make sure the bandwidth
4308 start_cfs_bandwidth(cfs_b);
4310 raw_spin_unlock(&cfs_b->lock);
4313 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4315 struct rq *rq = rq_of(cfs_rq);
4316 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4317 struct sched_entity *se;
4321 se = cfs_rq->tg->se[cpu_of(rq)];
4323 cfs_rq->throttled = 0;
4325 update_rq_clock(rq);
4327 raw_spin_lock(&cfs_b->lock);
4328 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4329 list_del_rcu(&cfs_rq->throttled_list);
4330 raw_spin_unlock(&cfs_b->lock);
4332 /* update hierarchical throttle state */
4333 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4335 if (!cfs_rq->load.weight)
4338 task_delta = cfs_rq->h_nr_running;
4339 for_each_sched_entity(se) {
4343 cfs_rq = cfs_rq_of(se);
4345 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4346 cfs_rq->h_nr_running += task_delta;
4348 if (cfs_rq_throttled(cfs_rq))
4353 add_nr_running(rq, task_delta);
4355 /* determine whether we need to wake up potentially idle cpu */
4356 if (rq->curr == rq->idle && rq->cfs.nr_running)
4360 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4361 u64 remaining, u64 expires)
4363 struct cfs_rq *cfs_rq;
4365 u64 starting_runtime = remaining;
4368 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4370 struct rq *rq = rq_of(cfs_rq);
4374 if (!cfs_rq_throttled(cfs_rq))
4377 runtime = -cfs_rq->runtime_remaining + 1;
4378 if (runtime > remaining)
4379 runtime = remaining;
4380 remaining -= runtime;
4382 cfs_rq->runtime_remaining += runtime;
4383 cfs_rq->runtime_expires = expires;
4385 /* we check whether we're throttled above */
4386 if (cfs_rq->runtime_remaining > 0)
4387 unthrottle_cfs_rq(cfs_rq);
4397 return starting_runtime - remaining;
4401 * Responsible for refilling a task_group's bandwidth and unthrottling its
4402 * cfs_rqs as appropriate. If there has been no activity within the last
4403 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4404 * used to track this state.
4406 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4408 u64 runtime, runtime_expires;
4411 /* no need to continue the timer with no bandwidth constraint */
4412 if (cfs_b->quota == RUNTIME_INF)
4413 goto out_deactivate;
4415 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4416 cfs_b->nr_periods += overrun;
4419 * idle depends on !throttled (for the case of a large deficit), and if
4420 * we're going inactive then everything else can be deferred
4422 if (cfs_b->idle && !throttled)
4423 goto out_deactivate;
4425 __refill_cfs_bandwidth_runtime(cfs_b);
4428 /* mark as potentially idle for the upcoming period */
4433 /* account preceding periods in which throttling occurred */
4434 cfs_b->nr_throttled += overrun;
4436 runtime_expires = cfs_b->runtime_expires;
4439 * This check is repeated as we are holding onto the new bandwidth while
4440 * we unthrottle. This can potentially race with an unthrottled group
4441 * trying to acquire new bandwidth from the global pool. This can result
4442 * in us over-using our runtime if it is all used during this loop, but
4443 * only by limited amounts in that extreme case.
4445 while (throttled && cfs_b->runtime > 0) {
4446 runtime = cfs_b->runtime;
4447 raw_spin_unlock(&cfs_b->lock);
4448 /* we can't nest cfs_b->lock while distributing bandwidth */
4449 runtime = distribute_cfs_runtime(cfs_b, runtime,
4451 raw_spin_lock(&cfs_b->lock);
4453 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4455 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4459 * While we are ensured activity in the period following an
4460 * unthrottle, this also covers the case in which the new bandwidth is
4461 * insufficient to cover the existing bandwidth deficit. (Forcing the
4462 * timer to remain active while there are any throttled entities.)
4472 /* a cfs_rq won't donate quota below this amount */
4473 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4474 /* minimum remaining period time to redistribute slack quota */
4475 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4476 /* how long we wait to gather additional slack before distributing */
4477 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4480 * Are we near the end of the current quota period?
4482 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4483 * hrtimer base being cleared by hrtimer_start. In the case of
4484 * migrate_hrtimers, base is never cleared, so we are fine.
4486 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4488 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4491 /* if the call-back is running a quota refresh is already occurring */
4492 if (hrtimer_callback_running(refresh_timer))
4495 /* is a quota refresh about to occur? */
4496 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4497 if (remaining < min_expire)
4503 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4505 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4507 /* if there's a quota refresh soon don't bother with slack */
4508 if (runtime_refresh_within(cfs_b, min_left))
4511 hrtimer_start(&cfs_b->slack_timer,
4512 ns_to_ktime(cfs_bandwidth_slack_period),
4516 /* we know any runtime found here is valid as update_curr() precedes return */
4517 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4519 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4520 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4522 if (slack_runtime <= 0)
4525 raw_spin_lock(&cfs_b->lock);
4526 if (cfs_b->quota != RUNTIME_INF &&
4527 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4528 cfs_b->runtime += slack_runtime;
4530 /* we are under rq->lock, defer unthrottling using a timer */
4531 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4532 !list_empty(&cfs_b->throttled_cfs_rq))
4533 start_cfs_slack_bandwidth(cfs_b);
4535 raw_spin_unlock(&cfs_b->lock);
4537 /* even if it's not valid for return we don't want to try again */
4538 cfs_rq->runtime_remaining -= slack_runtime;
4541 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4543 if (!cfs_bandwidth_used())
4546 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4549 __return_cfs_rq_runtime(cfs_rq);
4553 * This is done with a timer (instead of inline with bandwidth return) since
4554 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4556 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4558 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4561 /* confirm we're still not at a refresh boundary */
4562 raw_spin_lock(&cfs_b->lock);
4563 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4564 raw_spin_unlock(&cfs_b->lock);
4568 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4569 runtime = cfs_b->runtime;
4571 expires = cfs_b->runtime_expires;
4572 raw_spin_unlock(&cfs_b->lock);
4577 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4579 raw_spin_lock(&cfs_b->lock);
4580 if (expires == cfs_b->runtime_expires)
4581 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4582 raw_spin_unlock(&cfs_b->lock);
4586 * When a group wakes up we want to make sure that its quota is not already
4587 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4588 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4590 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4592 if (!cfs_bandwidth_used())
4595 /* an active group must be handled by the update_curr()->put() path */
4596 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4599 /* ensure the group is not already throttled */
4600 if (cfs_rq_throttled(cfs_rq))
4603 /* update runtime allocation */
4604 account_cfs_rq_runtime(cfs_rq, 0);
4605 if (cfs_rq->runtime_remaining <= 0)
4606 throttle_cfs_rq(cfs_rq);
4609 static void sync_throttle(struct task_group *tg, int cpu)
4611 struct cfs_rq *pcfs_rq, *cfs_rq;
4613 if (!cfs_bandwidth_used())
4619 cfs_rq = tg->cfs_rq[cpu];
4620 pcfs_rq = tg->parent->cfs_rq[cpu];
4622 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4623 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4626 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4627 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4629 if (!cfs_bandwidth_used())
4632 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4636 * it's possible for a throttled entity to be forced into a running
4637 * state (e.g. set_curr_task), in this case we're finished.
4639 if (cfs_rq_throttled(cfs_rq))
4642 throttle_cfs_rq(cfs_rq);
4646 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4648 struct cfs_bandwidth *cfs_b =
4649 container_of(timer, struct cfs_bandwidth, slack_timer);
4651 do_sched_cfs_slack_timer(cfs_b);
4653 return HRTIMER_NORESTART;
4656 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4658 struct cfs_bandwidth *cfs_b =
4659 container_of(timer, struct cfs_bandwidth, period_timer);
4663 raw_spin_lock(&cfs_b->lock);
4665 overrun = hrtimer_forward_now(timer, cfs_b->period);
4669 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4672 cfs_b->period_active = 0;
4673 raw_spin_unlock(&cfs_b->lock);
4675 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4678 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4680 raw_spin_lock_init(&cfs_b->lock);
4682 cfs_b->quota = RUNTIME_INF;
4683 cfs_b->period = ns_to_ktime(default_cfs_period());
4685 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4686 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4687 cfs_b->period_timer.function = sched_cfs_period_timer;
4688 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4689 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4692 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4694 cfs_rq->runtime_enabled = 0;
4695 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4698 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4700 lockdep_assert_held(&cfs_b->lock);
4702 if (!cfs_b->period_active) {
4703 cfs_b->period_active = 1;
4704 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4705 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4709 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4711 /* init_cfs_bandwidth() was not called */
4712 if (!cfs_b->throttled_cfs_rq.next)
4715 hrtimer_cancel(&cfs_b->period_timer);
4716 hrtimer_cancel(&cfs_b->slack_timer);
4720 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4722 * The race is harmless, since modifying bandwidth settings of unhooked group
4723 * bits doesn't do much.
4726 /* cpu online calback */
4727 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4729 struct task_group *tg;
4731 lockdep_assert_held(&rq->lock);
4734 list_for_each_entry_rcu(tg, &task_groups, list) {
4735 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4736 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4738 raw_spin_lock(&cfs_b->lock);
4739 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4740 raw_spin_unlock(&cfs_b->lock);
4745 /* cpu offline callback */
4746 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4748 struct task_group *tg;
4750 lockdep_assert_held(&rq->lock);
4753 list_for_each_entry_rcu(tg, &task_groups, list) {
4754 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4756 if (!cfs_rq->runtime_enabled)
4760 * clock_task is not advancing so we just need to make sure
4761 * there's some valid quota amount
4763 cfs_rq->runtime_remaining = 1;
4765 * Offline rq is schedulable till cpu is completely disabled
4766 * in take_cpu_down(), so we prevent new cfs throttling here.
4768 cfs_rq->runtime_enabled = 0;
4770 if (cfs_rq_throttled(cfs_rq))
4771 unthrottle_cfs_rq(cfs_rq);
4776 #else /* CONFIG_CFS_BANDWIDTH */
4777 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4779 return rq_clock_task(rq_of(cfs_rq));
4782 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4783 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4784 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4785 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4786 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4788 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4793 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4798 static inline int throttled_lb_pair(struct task_group *tg,
4799 int src_cpu, int dest_cpu)
4804 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4806 #ifdef CONFIG_FAIR_GROUP_SCHED
4807 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4810 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4814 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4815 static inline void update_runtime_enabled(struct rq *rq) {}
4816 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4818 #endif /* CONFIG_CFS_BANDWIDTH */
4820 /**************************************************
4821 * CFS operations on tasks:
4824 #ifdef CONFIG_SCHED_HRTICK
4825 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4827 struct sched_entity *se = &p->se;
4828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4830 SCHED_WARN_ON(task_rq(p) != rq);
4832 if (rq->cfs.h_nr_running > 1) {
4833 u64 slice = sched_slice(cfs_rq, se);
4834 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4835 s64 delta = slice - ran;
4842 hrtick_start(rq, delta);
4847 * called from enqueue/dequeue and updates the hrtick when the
4848 * current task is from our class and nr_running is low enough
4851 static void hrtick_update(struct rq *rq)
4853 struct task_struct *curr = rq->curr;
4855 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4858 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4859 hrtick_start_fair(rq, curr);
4861 #else /* !CONFIG_SCHED_HRTICK */
4863 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4867 static inline void hrtick_update(struct rq *rq)
4873 * The enqueue_task method is called before nr_running is
4874 * increased. Here we update the fair scheduling stats and
4875 * then put the task into the rbtree:
4878 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4880 struct cfs_rq *cfs_rq;
4881 struct sched_entity *se = &p->se;
4884 * If in_iowait is set, the code below may not trigger any cpufreq
4885 * utilization updates, so do it here explicitly with the IOWAIT flag
4889 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
4891 for_each_sched_entity(se) {
4894 cfs_rq = cfs_rq_of(se);
4895 enqueue_entity(cfs_rq, se, flags);
4898 * end evaluation on encountering a throttled cfs_rq
4900 * note: in the case of encountering a throttled cfs_rq we will
4901 * post the final h_nr_running increment below.
4903 if (cfs_rq_throttled(cfs_rq))
4905 cfs_rq->h_nr_running++;
4907 flags = ENQUEUE_WAKEUP;
4910 for_each_sched_entity(se) {
4911 cfs_rq = cfs_rq_of(se);
4912 cfs_rq->h_nr_running++;
4914 if (cfs_rq_throttled(cfs_rq))
4917 update_load_avg(se, UPDATE_TG);
4918 update_cfs_shares(se);
4922 add_nr_running(rq, 1);
4927 static void set_next_buddy(struct sched_entity *se);
4930 * The dequeue_task method is called before nr_running is
4931 * decreased. We remove the task from the rbtree and
4932 * update the fair scheduling stats:
4934 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4936 struct cfs_rq *cfs_rq;
4937 struct sched_entity *se = &p->se;
4938 int task_sleep = flags & DEQUEUE_SLEEP;
4940 for_each_sched_entity(se) {
4941 cfs_rq = cfs_rq_of(se);
4942 dequeue_entity(cfs_rq, se, flags);
4945 * end evaluation on encountering a throttled cfs_rq
4947 * note: in the case of encountering a throttled cfs_rq we will
4948 * post the final h_nr_running decrement below.
4950 if (cfs_rq_throttled(cfs_rq))
4952 cfs_rq->h_nr_running--;
4954 /* Don't dequeue parent if it has other entities besides us */
4955 if (cfs_rq->load.weight) {
4956 /* Avoid re-evaluating load for this entity: */
4957 se = parent_entity(se);
4959 * Bias pick_next to pick a task from this cfs_rq, as
4960 * p is sleeping when it is within its sched_slice.
4962 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4966 flags |= DEQUEUE_SLEEP;
4969 for_each_sched_entity(se) {
4970 cfs_rq = cfs_rq_of(se);
4971 cfs_rq->h_nr_running--;
4973 if (cfs_rq_throttled(cfs_rq))
4976 update_load_avg(se, UPDATE_TG);
4977 update_cfs_shares(se);
4981 sub_nr_running(rq, 1);
4988 /* Working cpumask for: load_balance, load_balance_newidle. */
4989 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4990 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4992 #ifdef CONFIG_NO_HZ_COMMON
4994 * per rq 'load' arrray crap; XXX kill this.
4998 * The exact cpuload calculated at every tick would be:
5000 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5002 * If a cpu misses updates for n ticks (as it was idle) and update gets
5003 * called on the n+1-th tick when cpu may be busy, then we have:
5005 * load_n = (1 - 1/2^i)^n * load_0
5006 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5008 * decay_load_missed() below does efficient calculation of
5010 * load' = (1 - 1/2^i)^n * load
5012 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5013 * This allows us to precompute the above in said factors, thereby allowing the
5014 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5015 * fixed_power_int())
5017 * The calculation is approximated on a 128 point scale.
5019 #define DEGRADE_SHIFT 7
5021 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5022 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5023 { 0, 0, 0, 0, 0, 0, 0, 0 },
5024 { 64, 32, 8, 0, 0, 0, 0, 0 },
5025 { 96, 72, 40, 12, 1, 0, 0, 0 },
5026 { 112, 98, 75, 43, 15, 1, 0, 0 },
5027 { 120, 112, 98, 76, 45, 16, 2, 0 }
5031 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5032 * would be when CPU is idle and so we just decay the old load without
5033 * adding any new load.
5035 static unsigned long
5036 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5040 if (!missed_updates)
5043 if (missed_updates >= degrade_zero_ticks[idx])
5047 return load >> missed_updates;
5049 while (missed_updates) {
5050 if (missed_updates % 2)
5051 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5053 missed_updates >>= 1;
5058 #endif /* CONFIG_NO_HZ_COMMON */
5061 * __cpu_load_update - update the rq->cpu_load[] statistics
5062 * @this_rq: The rq to update statistics for
5063 * @this_load: The current load
5064 * @pending_updates: The number of missed updates
5066 * Update rq->cpu_load[] statistics. This function is usually called every
5067 * scheduler tick (TICK_NSEC).
5069 * This function computes a decaying average:
5071 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5073 * Because of NOHZ it might not get called on every tick which gives need for
5074 * the @pending_updates argument.
5076 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5077 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5078 * = A * (A * load[i]_n-2 + B) + B
5079 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5080 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5081 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5082 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5083 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5085 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5086 * any change in load would have resulted in the tick being turned back on.
5088 * For regular NOHZ, this reduces to:
5090 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5092 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5095 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5096 unsigned long pending_updates)
5098 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5101 this_rq->nr_load_updates++;
5103 /* Update our load: */
5104 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5105 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5106 unsigned long old_load, new_load;
5108 /* scale is effectively 1 << i now, and >> i divides by scale */
5110 old_load = this_rq->cpu_load[i];
5111 #ifdef CONFIG_NO_HZ_COMMON
5112 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5113 if (tickless_load) {
5114 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5116 * old_load can never be a negative value because a
5117 * decayed tickless_load cannot be greater than the
5118 * original tickless_load.
5120 old_load += tickless_load;
5123 new_load = this_load;
5125 * Round up the averaging division if load is increasing. This
5126 * prevents us from getting stuck on 9 if the load is 10, for
5129 if (new_load > old_load)
5130 new_load += scale - 1;
5132 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5135 sched_avg_update(this_rq);
5138 /* Used instead of source_load when we know the type == 0 */
5139 static unsigned long weighted_cpuload(struct rq *rq)
5141 return cfs_rq_runnable_load_avg(&rq->cfs);
5144 #ifdef CONFIG_NO_HZ_COMMON
5146 * There is no sane way to deal with nohz on smp when using jiffies because the
5147 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5148 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5150 * Therefore we need to avoid the delta approach from the regular tick when
5151 * possible since that would seriously skew the load calculation. This is why we
5152 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5153 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5154 * loop exit, nohz_idle_balance, nohz full exit...)
5156 * This means we might still be one tick off for nohz periods.
5159 static void cpu_load_update_nohz(struct rq *this_rq,
5160 unsigned long curr_jiffies,
5163 unsigned long pending_updates;
5165 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5166 if (pending_updates) {
5167 this_rq->last_load_update_tick = curr_jiffies;
5169 * In the regular NOHZ case, we were idle, this means load 0.
5170 * In the NOHZ_FULL case, we were non-idle, we should consider
5171 * its weighted load.
5173 cpu_load_update(this_rq, load, pending_updates);
5178 * Called from nohz_idle_balance() to update the load ratings before doing the
5181 static void cpu_load_update_idle(struct rq *this_rq)
5184 * bail if there's load or we're actually up-to-date.
5186 if (weighted_cpuload(this_rq))
5189 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5193 * Record CPU load on nohz entry so we know the tickless load to account
5194 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5195 * than other cpu_load[idx] but it should be fine as cpu_load readers
5196 * shouldn't rely into synchronized cpu_load[*] updates.
5198 void cpu_load_update_nohz_start(void)
5200 struct rq *this_rq = this_rq();
5203 * This is all lockless but should be fine. If weighted_cpuload changes
5204 * concurrently we'll exit nohz. And cpu_load write can race with
5205 * cpu_load_update_idle() but both updater would be writing the same.
5207 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5211 * Account the tickless load in the end of a nohz frame.
5213 void cpu_load_update_nohz_stop(void)
5215 unsigned long curr_jiffies = READ_ONCE(jiffies);
5216 struct rq *this_rq = this_rq();
5220 if (curr_jiffies == this_rq->last_load_update_tick)
5223 load = weighted_cpuload(this_rq);
5224 rq_lock(this_rq, &rf);
5225 update_rq_clock(this_rq);
5226 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5227 rq_unlock(this_rq, &rf);
5229 #else /* !CONFIG_NO_HZ_COMMON */
5230 static inline void cpu_load_update_nohz(struct rq *this_rq,
5231 unsigned long curr_jiffies,
5232 unsigned long load) { }
5233 #endif /* CONFIG_NO_HZ_COMMON */
5235 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5237 #ifdef CONFIG_NO_HZ_COMMON
5238 /* See the mess around cpu_load_update_nohz(). */
5239 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5241 cpu_load_update(this_rq, load, 1);
5245 * Called from scheduler_tick()
5247 void cpu_load_update_active(struct rq *this_rq)
5249 unsigned long load = weighted_cpuload(this_rq);
5251 if (tick_nohz_tick_stopped())
5252 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5254 cpu_load_update_periodic(this_rq, load);
5258 * Return a low guess at the load of a migration-source cpu weighted
5259 * according to the scheduling class and "nice" value.
5261 * We want to under-estimate the load of migration sources, to
5262 * balance conservatively.
5264 static unsigned long source_load(int cpu, int type)
5266 struct rq *rq = cpu_rq(cpu);
5267 unsigned long total = weighted_cpuload(rq);
5269 if (type == 0 || !sched_feat(LB_BIAS))
5272 return min(rq->cpu_load[type-1], total);
5276 * Return a high guess at the load of a migration-target cpu weighted
5277 * according to the scheduling class and "nice" value.
5279 static unsigned long target_load(int cpu, int type)
5281 struct rq *rq = cpu_rq(cpu);
5282 unsigned long total = weighted_cpuload(rq);
5284 if (type == 0 || !sched_feat(LB_BIAS))
5287 return max(rq->cpu_load[type-1], total);
5290 static unsigned long capacity_of(int cpu)
5292 return cpu_rq(cpu)->cpu_capacity;
5295 static unsigned long capacity_orig_of(int cpu)
5297 return cpu_rq(cpu)->cpu_capacity_orig;
5300 static unsigned long cpu_avg_load_per_task(int cpu)
5302 struct rq *rq = cpu_rq(cpu);
5303 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5304 unsigned long load_avg = weighted_cpuload(rq);
5307 return load_avg / nr_running;
5312 static void record_wakee(struct task_struct *p)
5315 * Only decay a single time; tasks that have less then 1 wakeup per
5316 * jiffy will not have built up many flips.
5318 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5319 current->wakee_flips >>= 1;
5320 current->wakee_flip_decay_ts = jiffies;
5323 if (current->last_wakee != p) {
5324 current->last_wakee = p;
5325 current->wakee_flips++;
5330 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5332 * A waker of many should wake a different task than the one last awakened
5333 * at a frequency roughly N times higher than one of its wakees.
5335 * In order to determine whether we should let the load spread vs consolidating
5336 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5337 * partner, and a factor of lls_size higher frequency in the other.
5339 * With both conditions met, we can be relatively sure that the relationship is
5340 * non-monogamous, with partner count exceeding socket size.
5342 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5343 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5346 static int wake_wide(struct task_struct *p)
5348 unsigned int master = current->wakee_flips;
5349 unsigned int slave = p->wakee_flips;
5350 int factor = this_cpu_read(sd_llc_size);
5353 swap(master, slave);
5354 if (slave < factor || master < slave * factor)
5360 unsigned long nr_running;
5362 unsigned long capacity;
5366 static bool get_llc_stats(struct llc_stats *stats, int cpu)
5368 struct sched_domain_shared *sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5373 stats->nr_running = READ_ONCE(sds->nr_running);
5374 stats->load = READ_ONCE(sds->load);
5375 stats->capacity = READ_ONCE(sds->capacity);
5376 stats->has_capacity = stats->nr_running < per_cpu(sd_llc_size, cpu);
5382 * Can a task be moved from prev_cpu to this_cpu without causing a load
5383 * imbalance that would trigger the load balancer?
5385 * Since we're running on 'stale' values, we might in fact create an imbalance
5386 * but recomputing these values is expensive, as that'd mean iteration 2 cache
5387 * domains worth of CPUs.
5390 wake_affine_llc(struct sched_domain *sd, struct task_struct *p,
5391 int this_cpu, int prev_cpu, int sync)
5393 struct llc_stats prev_stats, this_stats;
5394 s64 this_eff_load, prev_eff_load;
5395 unsigned long task_load;
5397 if (!get_llc_stats(&prev_stats, prev_cpu) ||
5398 !get_llc_stats(&this_stats, this_cpu))
5402 * If sync wakeup then subtract the (maximum possible)
5403 * effect of the currently running task from the load
5404 * of the current LLC.
5407 unsigned long current_load = task_h_load(current);
5409 /* in this case load hits 0 and this LLC is considered 'idle' */
5410 if (current_load > this_stats.load)
5413 this_stats.load -= current_load;
5417 * The has_capacity stuff is not SMT aware, but by trying to balance
5418 * the nr_running on both ends we try and fill the domain at equal
5419 * rates, thereby first consuming cores before siblings.
5422 /* if the old cache has capacity, stay there */
5423 if (prev_stats.has_capacity && prev_stats.nr_running < this_stats.nr_running+1)
5426 /* if this cache has capacity, come here */
5427 if (this_stats.has_capacity && this_stats.nr_running+1 < prev_stats.nr_running)
5431 * Check to see if we can move the load without causing too much
5434 task_load = task_h_load(p);
5436 this_eff_load = 100;
5437 this_eff_load *= prev_stats.capacity;
5439 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5440 prev_eff_load *= this_stats.capacity;
5442 this_eff_load *= this_stats.load + task_load;
5443 prev_eff_load *= prev_stats.load - task_load;
5445 return this_eff_load <= prev_eff_load;
5448 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5449 int prev_cpu, int sync)
5451 int this_cpu = smp_processor_id();
5455 * Default to no affine wakeups; wake_affine() should not effect a task
5456 * placement the load-balancer feels inclined to undo. The conservative
5457 * option is therefore to not move tasks when they wake up.
5462 * If the wakeup is across cache domains, try to evaluate if movement
5463 * makes sense, otherwise rely on select_idle_siblings() to do
5464 * placement inside the cache domain.
5466 if (!cpus_share_cache(prev_cpu, this_cpu))
5467 affine = wake_affine_llc(sd, p, this_cpu, prev_cpu, sync);
5469 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5471 schedstat_inc(sd->ttwu_move_affine);
5472 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5478 static inline int task_util(struct task_struct *p);
5479 static int cpu_util_wake(int cpu, struct task_struct *p);
5481 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5483 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5487 * find_idlest_group finds and returns the least busy CPU group within the
5490 static struct sched_group *
5491 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5492 int this_cpu, int sd_flag)
5494 struct sched_group *idlest = NULL, *group = sd->groups;
5495 struct sched_group *most_spare_sg = NULL;
5496 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5497 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5498 unsigned long most_spare = 0, this_spare = 0;
5499 int load_idx = sd->forkexec_idx;
5500 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5501 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5502 (sd->imbalance_pct-100) / 100;
5504 if (sd_flag & SD_BALANCE_WAKE)
5505 load_idx = sd->wake_idx;
5508 unsigned long load, avg_load, runnable_load;
5509 unsigned long spare_cap, max_spare_cap;
5513 /* Skip over this group if it has no CPUs allowed */
5514 if (!cpumask_intersects(sched_group_span(group),
5518 local_group = cpumask_test_cpu(this_cpu,
5519 sched_group_span(group));
5522 * Tally up the load of all CPUs in the group and find
5523 * the group containing the CPU with most spare capacity.
5529 for_each_cpu(i, sched_group_span(group)) {
5530 /* Bias balancing toward cpus of our domain */
5532 load = source_load(i, load_idx);
5534 load = target_load(i, load_idx);
5536 runnable_load += load;
5538 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5540 spare_cap = capacity_spare_wake(i, p);
5542 if (spare_cap > max_spare_cap)
5543 max_spare_cap = spare_cap;
5546 /* Adjust by relative CPU capacity of the group */
5547 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5548 group->sgc->capacity;
5549 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5550 group->sgc->capacity;
5553 this_runnable_load = runnable_load;
5554 this_avg_load = avg_load;
5555 this_spare = max_spare_cap;
5557 if (min_runnable_load > (runnable_load + imbalance)) {
5559 * The runnable load is significantly smaller
5560 * so we can pick this new cpu
5562 min_runnable_load = runnable_load;
5563 min_avg_load = avg_load;
5565 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5566 (100*min_avg_load > imbalance_scale*avg_load)) {
5568 * The runnable loads are close so take the
5569 * blocked load into account through avg_load.
5571 min_avg_load = avg_load;
5575 if (most_spare < max_spare_cap) {
5576 most_spare = max_spare_cap;
5577 most_spare_sg = group;
5580 } while (group = group->next, group != sd->groups);
5583 * The cross-over point between using spare capacity or least load
5584 * is too conservative for high utilization tasks on partially
5585 * utilized systems if we require spare_capacity > task_util(p),
5586 * so we allow for some task stuffing by using
5587 * spare_capacity > task_util(p)/2.
5589 * Spare capacity can't be used for fork because the utilization has
5590 * not been set yet, we must first select a rq to compute the initial
5593 if (sd_flag & SD_BALANCE_FORK)
5596 if (this_spare > task_util(p) / 2 &&
5597 imbalance_scale*this_spare > 100*most_spare)
5600 if (most_spare > task_util(p) / 2)
5601 return most_spare_sg;
5607 if (min_runnable_load > (this_runnable_load + imbalance))
5610 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5611 (100*this_avg_load < imbalance_scale*min_avg_load))
5618 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5621 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5623 unsigned long load, min_load = ULONG_MAX;
5624 unsigned int min_exit_latency = UINT_MAX;
5625 u64 latest_idle_timestamp = 0;
5626 int least_loaded_cpu = this_cpu;
5627 int shallowest_idle_cpu = -1;
5630 /* Check if we have any choice: */
5631 if (group->group_weight == 1)
5632 return cpumask_first(sched_group_span(group));
5634 /* Traverse only the allowed CPUs */
5635 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5637 struct rq *rq = cpu_rq(i);
5638 struct cpuidle_state *idle = idle_get_state(rq);
5639 if (idle && idle->exit_latency < min_exit_latency) {
5641 * We give priority to a CPU whose idle state
5642 * has the smallest exit latency irrespective
5643 * of any idle timestamp.
5645 min_exit_latency = idle->exit_latency;
5646 latest_idle_timestamp = rq->idle_stamp;
5647 shallowest_idle_cpu = i;
5648 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5649 rq->idle_stamp > latest_idle_timestamp) {
5651 * If equal or no active idle state, then
5652 * the most recently idled CPU might have
5655 latest_idle_timestamp = rq->idle_stamp;
5656 shallowest_idle_cpu = i;
5658 } else if (shallowest_idle_cpu == -1) {
5659 load = weighted_cpuload(cpu_rq(i));
5660 if (load < min_load || (load == min_load && i == this_cpu)) {
5662 least_loaded_cpu = i;
5667 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5670 #ifdef CONFIG_SCHED_SMT
5672 static inline void set_idle_cores(int cpu, int val)
5674 struct sched_domain_shared *sds;
5676 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5678 WRITE_ONCE(sds->has_idle_cores, val);
5681 static inline bool test_idle_cores(int cpu, bool def)
5683 struct sched_domain_shared *sds;
5685 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5687 return READ_ONCE(sds->has_idle_cores);
5693 * Scans the local SMT mask to see if the entire core is idle, and records this
5694 * information in sd_llc_shared->has_idle_cores.
5696 * Since SMT siblings share all cache levels, inspecting this limited remote
5697 * state should be fairly cheap.
5699 void __update_idle_core(struct rq *rq)
5701 int core = cpu_of(rq);
5705 if (test_idle_cores(core, true))
5708 for_each_cpu(cpu, cpu_smt_mask(core)) {
5716 set_idle_cores(core, 1);
5722 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5723 * there are no idle cores left in the system; tracked through
5724 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5726 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5728 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5731 if (!static_branch_likely(&sched_smt_present))
5734 if (!test_idle_cores(target, false))
5737 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5739 for_each_cpu_wrap(core, cpus, target) {
5742 for_each_cpu(cpu, cpu_smt_mask(core)) {
5743 cpumask_clear_cpu(cpu, cpus);
5753 * Failed to find an idle core; stop looking for one.
5755 set_idle_cores(target, 0);
5761 * Scan the local SMT mask for idle CPUs.
5763 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5767 if (!static_branch_likely(&sched_smt_present))
5770 for_each_cpu(cpu, cpu_smt_mask(target)) {
5771 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5780 #else /* CONFIG_SCHED_SMT */
5782 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5787 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5792 #endif /* CONFIG_SCHED_SMT */
5795 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5796 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5797 * average idle time for this rq (as found in rq->avg_idle).
5799 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5801 struct sched_domain *this_sd;
5802 u64 avg_cost, avg_idle;
5805 int cpu, nr = INT_MAX;
5807 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5812 * Due to large variance we need a large fuzz factor; hackbench in
5813 * particularly is sensitive here.
5815 avg_idle = this_rq()->avg_idle / 512;
5816 avg_cost = this_sd->avg_scan_cost + 1;
5818 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5821 if (sched_feat(SIS_PROP)) {
5822 u64 span_avg = sd->span_weight * avg_idle;
5823 if (span_avg > 4*avg_cost)
5824 nr = div_u64(span_avg, avg_cost);
5829 time = local_clock();
5831 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5834 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5840 time = local_clock() - time;
5841 cost = this_sd->avg_scan_cost;
5842 delta = (s64)(time - cost) / 8;
5843 this_sd->avg_scan_cost += delta;
5849 * Try and locate an idle core/thread in the LLC cache domain.
5851 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5853 struct sched_domain *sd;
5856 if (idle_cpu(target))
5860 * If the previous cpu is cache affine and idle, don't be stupid.
5862 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5865 sd = rcu_dereference(per_cpu(sd_llc, target));
5869 i = select_idle_core(p, sd, target);
5870 if ((unsigned)i < nr_cpumask_bits)
5873 i = select_idle_cpu(p, sd, target);
5874 if ((unsigned)i < nr_cpumask_bits)
5877 i = select_idle_smt(p, sd, target);
5878 if ((unsigned)i < nr_cpumask_bits)
5885 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5886 * tasks. The unit of the return value must be the one of capacity so we can
5887 * compare the utilization with the capacity of the CPU that is available for
5888 * CFS task (ie cpu_capacity).
5890 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5891 * recent utilization of currently non-runnable tasks on a CPU. It represents
5892 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5893 * capacity_orig is the cpu_capacity available at the highest frequency
5894 * (arch_scale_freq_capacity()).
5895 * The utilization of a CPU converges towards a sum equal to or less than the
5896 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5897 * the running time on this CPU scaled by capacity_curr.
5899 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5900 * higher than capacity_orig because of unfortunate rounding in
5901 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5902 * the average stabilizes with the new running time. We need to check that the
5903 * utilization stays within the range of [0..capacity_orig] and cap it if
5904 * necessary. Without utilization capping, a group could be seen as overloaded
5905 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5906 * available capacity. We allow utilization to overshoot capacity_curr (but not
5907 * capacity_orig) as it useful for predicting the capacity required after task
5908 * migrations (scheduler-driven DVFS).
5910 static int cpu_util(int cpu)
5912 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5913 unsigned long capacity = capacity_orig_of(cpu);
5915 return (util >= capacity) ? capacity : util;
5918 static inline int task_util(struct task_struct *p)
5920 return p->se.avg.util_avg;
5924 * cpu_util_wake: Compute cpu utilization with any contributions from
5925 * the waking task p removed.
5927 static int cpu_util_wake(int cpu, struct task_struct *p)
5929 unsigned long util, capacity;
5931 /* Task has no contribution or is new */
5932 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5933 return cpu_util(cpu);
5935 capacity = capacity_orig_of(cpu);
5936 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5938 return (util >= capacity) ? capacity : util;
5942 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5943 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5945 * In that case WAKE_AFFINE doesn't make sense and we'll let
5946 * BALANCE_WAKE sort things out.
5948 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5950 long min_cap, max_cap;
5952 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5953 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5955 /* Minimum capacity is close to max, no need to abort wake_affine */
5956 if (max_cap - min_cap < max_cap >> 3)
5959 /* Bring task utilization in sync with prev_cpu */
5960 sync_entity_load_avg(&p->se);
5962 return min_cap * 1024 < task_util(p) * capacity_margin;
5966 * select_task_rq_fair: Select target runqueue for the waking task in domains
5967 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5968 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5970 * Balances load by selecting the idlest cpu in the idlest group, or under
5971 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5973 * Returns the target cpu number.
5975 * preempt must be disabled.
5978 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5980 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5981 int cpu = smp_processor_id();
5982 int new_cpu = prev_cpu;
5983 int want_affine = 0;
5984 int sync = wake_flags & WF_SYNC;
5986 if (sd_flag & SD_BALANCE_WAKE) {
5988 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5989 && cpumask_test_cpu(cpu, &p->cpus_allowed);
5993 for_each_domain(cpu, tmp) {
5994 if (!(tmp->flags & SD_LOAD_BALANCE))
5998 * If both cpu and prev_cpu are part of this domain,
5999 * cpu is a valid SD_WAKE_AFFINE target.
6001 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6002 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6007 if (tmp->flags & sd_flag)
6009 else if (!want_affine)
6014 sd = NULL; /* Prefer wake_affine over balance flags */
6015 if (cpu == prev_cpu)
6018 if (wake_affine(affine_sd, p, prev_cpu, sync))
6024 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6025 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6028 struct sched_group *group;
6031 if (!(sd->flags & sd_flag)) {
6036 group = find_idlest_group(sd, p, cpu, sd_flag);
6042 new_cpu = find_idlest_cpu(group, p, cpu);
6043 if (new_cpu == -1 || new_cpu == cpu) {
6044 /* Now try balancing at a lower domain level of cpu */
6049 /* Now try balancing at a lower domain level of new_cpu */
6051 weight = sd->span_weight;
6053 for_each_domain(cpu, tmp) {
6054 if (weight <= tmp->span_weight)
6056 if (tmp->flags & sd_flag)
6059 /* while loop will break here if sd == NULL */
6067 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6068 * cfs_rq_of(p) references at time of call are still valid and identify the
6069 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6071 static void migrate_task_rq_fair(struct task_struct *p)
6074 * As blocked tasks retain absolute vruntime the migration needs to
6075 * deal with this by subtracting the old and adding the new
6076 * min_vruntime -- the latter is done by enqueue_entity() when placing
6077 * the task on the new runqueue.
6079 if (p->state == TASK_WAKING) {
6080 struct sched_entity *se = &p->se;
6081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6084 #ifndef CONFIG_64BIT
6085 u64 min_vruntime_copy;
6088 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6090 min_vruntime = cfs_rq->min_vruntime;
6091 } while (min_vruntime != min_vruntime_copy);
6093 min_vruntime = cfs_rq->min_vruntime;
6096 se->vruntime -= min_vruntime;
6100 * We are supposed to update the task to "current" time, then its up to date
6101 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6102 * what current time is, so simply throw away the out-of-date time. This
6103 * will result in the wakee task is less decayed, but giving the wakee more
6104 * load sounds not bad.
6106 remove_entity_load_avg(&p->se);
6108 /* Tell new CPU we are migrated */
6109 p->se.avg.last_update_time = 0;
6111 /* We have migrated, no longer consider this task hot */
6112 p->se.exec_start = 0;
6115 static void task_dead_fair(struct task_struct *p)
6117 remove_entity_load_avg(&p->se);
6119 #endif /* CONFIG_SMP */
6121 static unsigned long
6122 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6124 unsigned long gran = sysctl_sched_wakeup_granularity;
6127 * Since its curr running now, convert the gran from real-time
6128 * to virtual-time in his units.
6130 * By using 'se' instead of 'curr' we penalize light tasks, so
6131 * they get preempted easier. That is, if 'se' < 'curr' then
6132 * the resulting gran will be larger, therefore penalizing the
6133 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6134 * be smaller, again penalizing the lighter task.
6136 * This is especially important for buddies when the leftmost
6137 * task is higher priority than the buddy.
6139 return calc_delta_fair(gran, se);
6143 * Should 'se' preempt 'curr'.
6157 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6159 s64 gran, vdiff = curr->vruntime - se->vruntime;
6164 gran = wakeup_gran(curr, se);
6171 static void set_last_buddy(struct sched_entity *se)
6173 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6176 for_each_sched_entity(se) {
6177 if (SCHED_WARN_ON(!se->on_rq))
6179 cfs_rq_of(se)->last = se;
6183 static void set_next_buddy(struct sched_entity *se)
6185 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6188 for_each_sched_entity(se) {
6189 if (SCHED_WARN_ON(!se->on_rq))
6191 cfs_rq_of(se)->next = se;
6195 static void set_skip_buddy(struct sched_entity *se)
6197 for_each_sched_entity(se)
6198 cfs_rq_of(se)->skip = se;
6202 * Preempt the current task with a newly woken task if needed:
6204 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6206 struct task_struct *curr = rq->curr;
6207 struct sched_entity *se = &curr->se, *pse = &p->se;
6208 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6209 int scale = cfs_rq->nr_running >= sched_nr_latency;
6210 int next_buddy_marked = 0;
6212 if (unlikely(se == pse))
6216 * This is possible from callers such as attach_tasks(), in which we
6217 * unconditionally check_prempt_curr() after an enqueue (which may have
6218 * lead to a throttle). This both saves work and prevents false
6219 * next-buddy nomination below.
6221 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6224 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6225 set_next_buddy(pse);
6226 next_buddy_marked = 1;
6230 * We can come here with TIF_NEED_RESCHED already set from new task
6233 * Note: this also catches the edge-case of curr being in a throttled
6234 * group (e.g. via set_curr_task), since update_curr() (in the
6235 * enqueue of curr) will have resulted in resched being set. This
6236 * prevents us from potentially nominating it as a false LAST_BUDDY
6239 if (test_tsk_need_resched(curr))
6242 /* Idle tasks are by definition preempted by non-idle tasks. */
6243 if (unlikely(curr->policy == SCHED_IDLE) &&
6244 likely(p->policy != SCHED_IDLE))
6248 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6249 * is driven by the tick):
6251 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6254 find_matching_se(&se, &pse);
6255 update_curr(cfs_rq_of(se));
6257 if (wakeup_preempt_entity(se, pse) == 1) {
6259 * Bias pick_next to pick the sched entity that is
6260 * triggering this preemption.
6262 if (!next_buddy_marked)
6263 set_next_buddy(pse);
6272 * Only set the backward buddy when the current task is still
6273 * on the rq. This can happen when a wakeup gets interleaved
6274 * with schedule on the ->pre_schedule() or idle_balance()
6275 * point, either of which can * drop the rq lock.
6277 * Also, during early boot the idle thread is in the fair class,
6278 * for obvious reasons its a bad idea to schedule back to it.
6280 if (unlikely(!se->on_rq || curr == rq->idle))
6283 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6287 static struct task_struct *
6288 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6290 struct cfs_rq *cfs_rq = &rq->cfs;
6291 struct sched_entity *se;
6292 struct task_struct *p;
6296 if (!cfs_rq->nr_running)
6299 #ifdef CONFIG_FAIR_GROUP_SCHED
6300 if (prev->sched_class != &fair_sched_class)
6304 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6305 * likely that a next task is from the same cgroup as the current.
6307 * Therefore attempt to avoid putting and setting the entire cgroup
6308 * hierarchy, only change the part that actually changes.
6312 struct sched_entity *curr = cfs_rq->curr;
6315 * Since we got here without doing put_prev_entity() we also
6316 * have to consider cfs_rq->curr. If it is still a runnable
6317 * entity, update_curr() will update its vruntime, otherwise
6318 * forget we've ever seen it.
6322 update_curr(cfs_rq);
6327 * This call to check_cfs_rq_runtime() will do the
6328 * throttle and dequeue its entity in the parent(s).
6329 * Therefore the nr_running test will indeed
6332 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6335 if (!cfs_rq->nr_running)
6342 se = pick_next_entity(cfs_rq, curr);
6343 cfs_rq = group_cfs_rq(se);
6349 * Since we haven't yet done put_prev_entity and if the selected task
6350 * is a different task than we started out with, try and touch the
6351 * least amount of cfs_rqs.
6354 struct sched_entity *pse = &prev->se;
6356 while (!(cfs_rq = is_same_group(se, pse))) {
6357 int se_depth = se->depth;
6358 int pse_depth = pse->depth;
6360 if (se_depth <= pse_depth) {
6361 put_prev_entity(cfs_rq_of(pse), pse);
6362 pse = parent_entity(pse);
6364 if (se_depth >= pse_depth) {
6365 set_next_entity(cfs_rq_of(se), se);
6366 se = parent_entity(se);
6370 put_prev_entity(cfs_rq, pse);
6371 set_next_entity(cfs_rq, se);
6374 if (hrtick_enabled(rq))
6375 hrtick_start_fair(rq, p);
6381 put_prev_task(rq, prev);
6384 se = pick_next_entity(cfs_rq, NULL);
6385 set_next_entity(cfs_rq, se);
6386 cfs_rq = group_cfs_rq(se);
6391 if (hrtick_enabled(rq))
6392 hrtick_start_fair(rq, p);
6397 new_tasks = idle_balance(rq, rf);
6400 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6401 * possible for any higher priority task to appear. In that case we
6402 * must re-start the pick_next_entity() loop.
6414 * Account for a descheduled task:
6416 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6418 struct sched_entity *se = &prev->se;
6419 struct cfs_rq *cfs_rq;
6421 for_each_sched_entity(se) {
6422 cfs_rq = cfs_rq_of(se);
6423 put_prev_entity(cfs_rq, se);
6428 * sched_yield() is very simple
6430 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6432 static void yield_task_fair(struct rq *rq)
6434 struct task_struct *curr = rq->curr;
6435 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6436 struct sched_entity *se = &curr->se;
6439 * Are we the only task in the tree?
6441 if (unlikely(rq->nr_running == 1))
6444 clear_buddies(cfs_rq, se);
6446 if (curr->policy != SCHED_BATCH) {
6447 update_rq_clock(rq);
6449 * Update run-time statistics of the 'current'.
6451 update_curr(cfs_rq);
6453 * Tell update_rq_clock() that we've just updated,
6454 * so we don't do microscopic update in schedule()
6455 * and double the fastpath cost.
6457 rq_clock_skip_update(rq, true);
6463 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6465 struct sched_entity *se = &p->se;
6467 /* throttled hierarchies are not runnable */
6468 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6471 /* Tell the scheduler that we'd really like pse to run next. */
6474 yield_task_fair(rq);
6480 /**************************************************
6481 * Fair scheduling class load-balancing methods.
6485 * The purpose of load-balancing is to achieve the same basic fairness the
6486 * per-cpu scheduler provides, namely provide a proportional amount of compute
6487 * time to each task. This is expressed in the following equation:
6489 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6491 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6492 * W_i,0 is defined as:
6494 * W_i,0 = \Sum_j w_i,j (2)
6496 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6497 * is derived from the nice value as per sched_prio_to_weight[].
6499 * The weight average is an exponential decay average of the instantaneous
6502 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6504 * C_i is the compute capacity of cpu i, typically it is the
6505 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6506 * can also include other factors [XXX].
6508 * To achieve this balance we define a measure of imbalance which follows
6509 * directly from (1):
6511 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6513 * We them move tasks around to minimize the imbalance. In the continuous
6514 * function space it is obvious this converges, in the discrete case we get
6515 * a few fun cases generally called infeasible weight scenarios.
6518 * - infeasible weights;
6519 * - local vs global optima in the discrete case. ]
6524 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6525 * for all i,j solution, we create a tree of cpus that follows the hardware
6526 * topology where each level pairs two lower groups (or better). This results
6527 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6528 * tree to only the first of the previous level and we decrease the frequency
6529 * of load-balance at each level inv. proportional to the number of cpus in
6535 * \Sum { --- * --- * 2^i } = O(n) (5)
6537 * `- size of each group
6538 * | | `- number of cpus doing load-balance
6540 * `- sum over all levels
6542 * Coupled with a limit on how many tasks we can migrate every balance pass,
6543 * this makes (5) the runtime complexity of the balancer.
6545 * An important property here is that each CPU is still (indirectly) connected
6546 * to every other cpu in at most O(log n) steps:
6548 * The adjacency matrix of the resulting graph is given by:
6551 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6554 * And you'll find that:
6556 * A^(log_2 n)_i,j != 0 for all i,j (7)
6558 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6559 * The task movement gives a factor of O(m), giving a convergence complexity
6562 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6567 * In order to avoid CPUs going idle while there's still work to do, new idle
6568 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6569 * tree itself instead of relying on other CPUs to bring it work.
6571 * This adds some complexity to both (5) and (8) but it reduces the total idle
6579 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6582 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6587 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6589 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6591 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6594 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6595 * rewrite all of this once again.]
6598 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6600 enum fbq_type { regular, remote, all };
6602 #define LBF_ALL_PINNED 0x01
6603 #define LBF_NEED_BREAK 0x02
6604 #define LBF_DST_PINNED 0x04
6605 #define LBF_SOME_PINNED 0x08
6608 struct sched_domain *sd;
6616 struct cpumask *dst_grpmask;
6618 enum cpu_idle_type idle;
6620 /* The set of CPUs under consideration for load-balancing */
6621 struct cpumask *cpus;
6626 unsigned int loop_break;
6627 unsigned int loop_max;
6629 enum fbq_type fbq_type;
6630 struct list_head tasks;
6634 * Is this task likely cache-hot:
6636 static int task_hot(struct task_struct *p, struct lb_env *env)
6640 lockdep_assert_held(&env->src_rq->lock);
6642 if (p->sched_class != &fair_sched_class)
6645 if (unlikely(p->policy == SCHED_IDLE))
6649 * Buddy candidates are cache hot:
6651 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6652 (&p->se == cfs_rq_of(&p->se)->next ||
6653 &p->se == cfs_rq_of(&p->se)->last))
6656 if (sysctl_sched_migration_cost == -1)
6658 if (sysctl_sched_migration_cost == 0)
6661 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6663 return delta < (s64)sysctl_sched_migration_cost;
6666 #ifdef CONFIG_NUMA_BALANCING
6668 * Returns 1, if task migration degrades locality
6669 * Returns 0, if task migration improves locality i.e migration preferred.
6670 * Returns -1, if task migration is not affected by locality.
6672 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6674 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6675 unsigned long src_faults, dst_faults;
6676 int src_nid, dst_nid;
6678 if (!static_branch_likely(&sched_numa_balancing))
6681 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6684 src_nid = cpu_to_node(env->src_cpu);
6685 dst_nid = cpu_to_node(env->dst_cpu);
6687 if (src_nid == dst_nid)
6690 /* Migrating away from the preferred node is always bad. */
6691 if (src_nid == p->numa_preferred_nid) {
6692 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6698 /* Encourage migration to the preferred node. */
6699 if (dst_nid == p->numa_preferred_nid)
6702 /* Leaving a core idle is often worse than degrading locality. */
6703 if (env->idle != CPU_NOT_IDLE)
6707 src_faults = group_faults(p, src_nid);
6708 dst_faults = group_faults(p, dst_nid);
6710 src_faults = task_faults(p, src_nid);
6711 dst_faults = task_faults(p, dst_nid);
6714 return dst_faults < src_faults;
6718 static inline int migrate_degrades_locality(struct task_struct *p,
6726 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6729 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6733 lockdep_assert_held(&env->src_rq->lock);
6736 * We do not migrate tasks that are:
6737 * 1) throttled_lb_pair, or
6738 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6739 * 3) running (obviously), or
6740 * 4) are cache-hot on their current CPU.
6742 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6745 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6748 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6750 env->flags |= LBF_SOME_PINNED;
6753 * Remember if this task can be migrated to any other cpu in
6754 * our sched_group. We may want to revisit it if we couldn't
6755 * meet load balance goals by pulling other tasks on src_cpu.
6757 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6758 * already computed one in current iteration.
6760 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6763 /* Prevent to re-select dst_cpu via env's cpus */
6764 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6765 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6766 env->flags |= LBF_DST_PINNED;
6767 env->new_dst_cpu = cpu;
6775 /* Record that we found atleast one task that could run on dst_cpu */
6776 env->flags &= ~LBF_ALL_PINNED;
6778 if (task_running(env->src_rq, p)) {
6779 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6784 * Aggressive migration if:
6785 * 1) destination numa is preferred
6786 * 2) task is cache cold, or
6787 * 3) too many balance attempts have failed.
6789 tsk_cache_hot = migrate_degrades_locality(p, env);
6790 if (tsk_cache_hot == -1)
6791 tsk_cache_hot = task_hot(p, env);
6793 if (tsk_cache_hot <= 0 ||
6794 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6795 if (tsk_cache_hot == 1) {
6796 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6797 schedstat_inc(p->se.statistics.nr_forced_migrations);
6802 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6807 * detach_task() -- detach the task for the migration specified in env
6809 static void detach_task(struct task_struct *p, struct lb_env *env)
6811 lockdep_assert_held(&env->src_rq->lock);
6813 p->on_rq = TASK_ON_RQ_MIGRATING;
6814 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6815 set_task_cpu(p, env->dst_cpu);
6819 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6820 * part of active balancing operations within "domain".
6822 * Returns a task if successful and NULL otherwise.
6824 static struct task_struct *detach_one_task(struct lb_env *env)
6826 struct task_struct *p, *n;
6828 lockdep_assert_held(&env->src_rq->lock);
6830 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6831 if (!can_migrate_task(p, env))
6834 detach_task(p, env);
6837 * Right now, this is only the second place where
6838 * lb_gained[env->idle] is updated (other is detach_tasks)
6839 * so we can safely collect stats here rather than
6840 * inside detach_tasks().
6842 schedstat_inc(env->sd->lb_gained[env->idle]);
6848 static const unsigned int sched_nr_migrate_break = 32;
6851 * detach_tasks() -- tries to detach up to imbalance weighted load from
6852 * busiest_rq, as part of a balancing operation within domain "sd".
6854 * Returns number of detached tasks if successful and 0 otherwise.
6856 static int detach_tasks(struct lb_env *env)
6858 struct list_head *tasks = &env->src_rq->cfs_tasks;
6859 struct task_struct *p;
6863 lockdep_assert_held(&env->src_rq->lock);
6865 if (env->imbalance <= 0)
6868 while (!list_empty(tasks)) {
6870 * We don't want to steal all, otherwise we may be treated likewise,
6871 * which could at worst lead to a livelock crash.
6873 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6876 p = list_first_entry(tasks, struct task_struct, se.group_node);
6879 /* We've more or less seen every task there is, call it quits */
6880 if (env->loop > env->loop_max)
6883 /* take a breather every nr_migrate tasks */
6884 if (env->loop > env->loop_break) {
6885 env->loop_break += sched_nr_migrate_break;
6886 env->flags |= LBF_NEED_BREAK;
6890 if (!can_migrate_task(p, env))
6893 load = task_h_load(p);
6895 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6898 if ((load / 2) > env->imbalance)
6901 detach_task(p, env);
6902 list_add(&p->se.group_node, &env->tasks);
6905 env->imbalance -= load;
6907 #ifdef CONFIG_PREEMPT
6909 * NEWIDLE balancing is a source of latency, so preemptible
6910 * kernels will stop after the first task is detached to minimize
6911 * the critical section.
6913 if (env->idle == CPU_NEWLY_IDLE)
6918 * We only want to steal up to the prescribed amount of
6921 if (env->imbalance <= 0)
6926 list_move_tail(&p->se.group_node, tasks);
6930 * Right now, this is one of only two places we collect this stat
6931 * so we can safely collect detach_one_task() stats here rather
6932 * than inside detach_one_task().
6934 schedstat_add(env->sd->lb_gained[env->idle], detached);
6940 * attach_task() -- attach the task detached by detach_task() to its new rq.
6942 static void attach_task(struct rq *rq, struct task_struct *p)
6944 lockdep_assert_held(&rq->lock);
6946 BUG_ON(task_rq(p) != rq);
6947 activate_task(rq, p, ENQUEUE_NOCLOCK);
6948 p->on_rq = TASK_ON_RQ_QUEUED;
6949 check_preempt_curr(rq, p, 0);
6953 * attach_one_task() -- attaches the task returned from detach_one_task() to
6956 static void attach_one_task(struct rq *rq, struct task_struct *p)
6961 update_rq_clock(rq);
6967 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6970 static void attach_tasks(struct lb_env *env)
6972 struct list_head *tasks = &env->tasks;
6973 struct task_struct *p;
6976 rq_lock(env->dst_rq, &rf);
6977 update_rq_clock(env->dst_rq);
6979 while (!list_empty(tasks)) {
6980 p = list_first_entry(tasks, struct task_struct, se.group_node);
6981 list_del_init(&p->se.group_node);
6983 attach_task(env->dst_rq, p);
6986 rq_unlock(env->dst_rq, &rf);
6989 #ifdef CONFIG_FAIR_GROUP_SCHED
6991 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
6993 if (cfs_rq->load.weight)
6996 if (cfs_rq->avg.load_sum)
6999 if (cfs_rq->avg.util_sum)
7002 if (cfs_rq->runnable_load_sum)
7008 static void update_blocked_averages(int cpu)
7010 struct rq *rq = cpu_rq(cpu);
7011 struct cfs_rq *cfs_rq, *pos;
7014 rq_lock_irqsave(rq, &rf);
7015 update_rq_clock(rq);
7018 * Iterates the task_group tree in a bottom up fashion, see
7019 * list_add_leaf_cfs_rq() for details.
7021 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7022 struct sched_entity *se;
7024 /* throttled entities do not contribute to load */
7025 if (throttled_hierarchy(cfs_rq))
7028 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7029 update_tg_load_avg(cfs_rq, 0);
7031 /* Propagate pending load changes to the parent, if any: */
7032 se = cfs_rq->tg->se[cpu];
7033 if (se && !skip_blocked_update(se))
7034 update_load_avg(se, 0);
7037 * There can be a lot of idle CPU cgroups. Don't let fully
7038 * decayed cfs_rqs linger on the list.
7040 if (cfs_rq_is_decayed(cfs_rq))
7041 list_del_leaf_cfs_rq(cfs_rq);
7043 rq_unlock_irqrestore(rq, &rf);
7047 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7048 * This needs to be done in a top-down fashion because the load of a child
7049 * group is a fraction of its parents load.
7051 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7053 struct rq *rq = rq_of(cfs_rq);
7054 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7055 unsigned long now = jiffies;
7058 if (cfs_rq->last_h_load_update == now)
7061 cfs_rq->h_load_next = NULL;
7062 for_each_sched_entity(se) {
7063 cfs_rq = cfs_rq_of(se);
7064 cfs_rq->h_load_next = se;
7065 if (cfs_rq->last_h_load_update == now)
7070 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7071 cfs_rq->last_h_load_update = now;
7074 while ((se = cfs_rq->h_load_next) != NULL) {
7075 load = cfs_rq->h_load;
7076 load = div64_ul(load * se->avg.load_avg,
7077 cfs_rq_load_avg(cfs_rq) + 1);
7078 cfs_rq = group_cfs_rq(se);
7079 cfs_rq->h_load = load;
7080 cfs_rq->last_h_load_update = now;
7084 static unsigned long task_h_load(struct task_struct *p)
7086 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7088 update_cfs_rq_h_load(cfs_rq);
7089 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7090 cfs_rq_load_avg(cfs_rq) + 1);
7093 static inline void update_blocked_averages(int cpu)
7095 struct rq *rq = cpu_rq(cpu);
7096 struct cfs_rq *cfs_rq = &rq->cfs;
7099 rq_lock_irqsave(rq, &rf);
7100 update_rq_clock(rq);
7101 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7102 rq_unlock_irqrestore(rq, &rf);
7105 static unsigned long task_h_load(struct task_struct *p)
7107 return p->se.avg.load_avg;
7111 /********** Helpers for find_busiest_group ************************/
7120 * sg_lb_stats - stats of a sched_group required for load_balancing
7122 struct sg_lb_stats {
7123 unsigned long avg_load; /*Avg load across the CPUs of the group */
7124 unsigned long group_load; /* Total load over the CPUs of the group */
7125 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7126 unsigned long load_per_task;
7127 unsigned long group_capacity;
7128 unsigned long group_util; /* Total utilization of the group */
7129 unsigned int sum_nr_running; /* Nr tasks running in the group */
7130 unsigned int idle_cpus;
7131 unsigned int group_weight;
7132 enum group_type group_type;
7133 int group_no_capacity;
7134 #ifdef CONFIG_NUMA_BALANCING
7135 unsigned int nr_numa_running;
7136 unsigned int nr_preferred_running;
7141 * sd_lb_stats - Structure to store the statistics of a sched_domain
7142 * during load balancing.
7144 struct sd_lb_stats {
7145 struct sched_group *busiest; /* Busiest group in this sd */
7146 struct sched_group *local; /* Local group in this sd */
7147 unsigned long total_running;
7148 unsigned long total_load; /* Total load of all groups in sd */
7149 unsigned long total_capacity; /* Total capacity of all groups in sd */
7150 unsigned long avg_load; /* Average load across all groups in sd */
7152 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7153 struct sg_lb_stats local_stat; /* Statistics of the local group */
7156 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7159 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7160 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7161 * We must however clear busiest_stat::avg_load because
7162 * update_sd_pick_busiest() reads this before assignment.
7164 *sds = (struct sd_lb_stats){
7167 .total_running = 0UL,
7169 .total_capacity = 0UL,
7172 .sum_nr_running = 0,
7173 .group_type = group_other,
7179 * get_sd_load_idx - Obtain the load index for a given sched domain.
7180 * @sd: The sched_domain whose load_idx is to be obtained.
7181 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7183 * Return: The load index.
7185 static inline int get_sd_load_idx(struct sched_domain *sd,
7186 enum cpu_idle_type idle)
7192 load_idx = sd->busy_idx;
7195 case CPU_NEWLY_IDLE:
7196 load_idx = sd->newidle_idx;
7199 load_idx = sd->idle_idx;
7206 static unsigned long scale_rt_capacity(int cpu)
7208 struct rq *rq = cpu_rq(cpu);
7209 u64 total, used, age_stamp, avg;
7213 * Since we're reading these variables without serialization make sure
7214 * we read them once before doing sanity checks on them.
7216 age_stamp = READ_ONCE(rq->age_stamp);
7217 avg = READ_ONCE(rq->rt_avg);
7218 delta = __rq_clock_broken(rq) - age_stamp;
7220 if (unlikely(delta < 0))
7223 total = sched_avg_period() + delta;
7225 used = div_u64(avg, total);
7227 if (likely(used < SCHED_CAPACITY_SCALE))
7228 return SCHED_CAPACITY_SCALE - used;
7233 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7235 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7236 struct sched_group *sdg = sd->groups;
7238 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7240 capacity *= scale_rt_capacity(cpu);
7241 capacity >>= SCHED_CAPACITY_SHIFT;
7246 cpu_rq(cpu)->cpu_capacity = capacity;
7247 sdg->sgc->capacity = capacity;
7248 sdg->sgc->min_capacity = capacity;
7251 void update_group_capacity(struct sched_domain *sd, int cpu)
7253 struct sched_domain *child = sd->child;
7254 struct sched_group *group, *sdg = sd->groups;
7255 unsigned long capacity, min_capacity;
7256 unsigned long interval;
7258 interval = msecs_to_jiffies(sd->balance_interval);
7259 interval = clamp(interval, 1UL, max_load_balance_interval);
7260 sdg->sgc->next_update = jiffies + interval;
7263 update_cpu_capacity(sd, cpu);
7268 min_capacity = ULONG_MAX;
7270 if (child->flags & SD_OVERLAP) {
7272 * SD_OVERLAP domains cannot assume that child groups
7273 * span the current group.
7276 for_each_cpu(cpu, sched_group_span(sdg)) {
7277 struct sched_group_capacity *sgc;
7278 struct rq *rq = cpu_rq(cpu);
7281 * build_sched_domains() -> init_sched_groups_capacity()
7282 * gets here before we've attached the domains to the
7285 * Use capacity_of(), which is set irrespective of domains
7286 * in update_cpu_capacity().
7288 * This avoids capacity from being 0 and
7289 * causing divide-by-zero issues on boot.
7291 if (unlikely(!rq->sd)) {
7292 capacity += capacity_of(cpu);
7294 sgc = rq->sd->groups->sgc;
7295 capacity += sgc->capacity;
7298 min_capacity = min(capacity, min_capacity);
7302 * !SD_OVERLAP domains can assume that child groups
7303 * span the current group.
7306 group = child->groups;
7308 struct sched_group_capacity *sgc = group->sgc;
7310 capacity += sgc->capacity;
7311 min_capacity = min(sgc->min_capacity, min_capacity);
7312 group = group->next;
7313 } while (group != child->groups);
7316 sdg->sgc->capacity = capacity;
7317 sdg->sgc->min_capacity = min_capacity;
7321 * Check whether the capacity of the rq has been noticeably reduced by side
7322 * activity. The imbalance_pct is used for the threshold.
7323 * Return true is the capacity is reduced
7326 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7328 return ((rq->cpu_capacity * sd->imbalance_pct) <
7329 (rq->cpu_capacity_orig * 100));
7333 * Group imbalance indicates (and tries to solve) the problem where balancing
7334 * groups is inadequate due to ->cpus_allowed constraints.
7336 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7337 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7340 * { 0 1 2 3 } { 4 5 6 7 }
7343 * If we were to balance group-wise we'd place two tasks in the first group and
7344 * two tasks in the second group. Clearly this is undesired as it will overload
7345 * cpu 3 and leave one of the cpus in the second group unused.
7347 * The current solution to this issue is detecting the skew in the first group
7348 * by noticing the lower domain failed to reach balance and had difficulty
7349 * moving tasks due to affinity constraints.
7351 * When this is so detected; this group becomes a candidate for busiest; see
7352 * update_sd_pick_busiest(). And calculate_imbalance() and
7353 * find_busiest_group() avoid some of the usual balance conditions to allow it
7354 * to create an effective group imbalance.
7356 * This is a somewhat tricky proposition since the next run might not find the
7357 * group imbalance and decide the groups need to be balanced again. A most
7358 * subtle and fragile situation.
7361 static inline int sg_imbalanced(struct sched_group *group)
7363 return group->sgc->imbalance;
7367 * group_has_capacity returns true if the group has spare capacity that could
7368 * be used by some tasks.
7369 * We consider that a group has spare capacity if the * number of task is
7370 * smaller than the number of CPUs or if the utilization is lower than the
7371 * available capacity for CFS tasks.
7372 * For the latter, we use a threshold to stabilize the state, to take into
7373 * account the variance of the tasks' load and to return true if the available
7374 * capacity in meaningful for the load balancer.
7375 * As an example, an available capacity of 1% can appear but it doesn't make
7376 * any benefit for the load balance.
7379 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7381 if (sgs->sum_nr_running < sgs->group_weight)
7384 if ((sgs->group_capacity * 100) >
7385 (sgs->group_util * env->sd->imbalance_pct))
7392 * group_is_overloaded returns true if the group has more tasks than it can
7394 * group_is_overloaded is not equals to !group_has_capacity because a group
7395 * with the exact right number of tasks, has no more spare capacity but is not
7396 * overloaded so both group_has_capacity and group_is_overloaded return
7400 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7402 if (sgs->sum_nr_running <= sgs->group_weight)
7405 if ((sgs->group_capacity * 100) <
7406 (sgs->group_util * env->sd->imbalance_pct))
7413 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7414 * per-CPU capacity than sched_group ref.
7417 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7419 return sg->sgc->min_capacity * capacity_margin <
7420 ref->sgc->min_capacity * 1024;
7424 group_type group_classify(struct sched_group *group,
7425 struct sg_lb_stats *sgs)
7427 if (sgs->group_no_capacity)
7428 return group_overloaded;
7430 if (sg_imbalanced(group))
7431 return group_imbalanced;
7437 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7438 * @env: The load balancing environment.
7439 * @group: sched_group whose statistics are to be updated.
7440 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7441 * @local_group: Does group contain this_cpu.
7442 * @sgs: variable to hold the statistics for this group.
7443 * @overload: Indicate more than one runnable task for any CPU.
7445 static inline void update_sg_lb_stats(struct lb_env *env,
7446 struct sched_group *group, int load_idx,
7447 int local_group, struct sg_lb_stats *sgs,
7453 memset(sgs, 0, sizeof(*sgs));
7455 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7456 struct rq *rq = cpu_rq(i);
7458 /* Bias balancing toward cpus of our domain */
7460 load = target_load(i, load_idx);
7462 load = source_load(i, load_idx);
7464 sgs->group_load += load;
7465 sgs->group_util += cpu_util(i);
7466 sgs->sum_nr_running += rq->cfs.h_nr_running;
7468 nr_running = rq->nr_running;
7472 #ifdef CONFIG_NUMA_BALANCING
7473 sgs->nr_numa_running += rq->nr_numa_running;
7474 sgs->nr_preferred_running += rq->nr_preferred_running;
7476 sgs->sum_weighted_load += weighted_cpuload(rq);
7478 * No need to call idle_cpu() if nr_running is not 0
7480 if (!nr_running && idle_cpu(i))
7484 /* Adjust by relative CPU capacity of the group */
7485 sgs->group_capacity = group->sgc->capacity;
7486 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7488 if (sgs->sum_nr_running)
7489 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7491 sgs->group_weight = group->group_weight;
7493 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7494 sgs->group_type = group_classify(group, sgs);
7498 * update_sd_pick_busiest - return 1 on busiest group
7499 * @env: The load balancing environment.
7500 * @sds: sched_domain statistics
7501 * @sg: sched_group candidate to be checked for being the busiest
7502 * @sgs: sched_group statistics
7504 * Determine if @sg is a busier group than the previously selected
7507 * Return: %true if @sg is a busier group than the previously selected
7508 * busiest group. %false otherwise.
7510 static bool update_sd_pick_busiest(struct lb_env *env,
7511 struct sd_lb_stats *sds,
7512 struct sched_group *sg,
7513 struct sg_lb_stats *sgs)
7515 struct sg_lb_stats *busiest = &sds->busiest_stat;
7517 if (sgs->group_type > busiest->group_type)
7520 if (sgs->group_type < busiest->group_type)
7523 if (sgs->avg_load <= busiest->avg_load)
7526 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7530 * Candidate sg has no more than one task per CPU and
7531 * has higher per-CPU capacity. Migrating tasks to less
7532 * capable CPUs may harm throughput. Maximize throughput,
7533 * power/energy consequences are not considered.
7535 if (sgs->sum_nr_running <= sgs->group_weight &&
7536 group_smaller_cpu_capacity(sds->local, sg))
7540 /* This is the busiest node in its class. */
7541 if (!(env->sd->flags & SD_ASYM_PACKING))
7544 /* No ASYM_PACKING if target cpu is already busy */
7545 if (env->idle == CPU_NOT_IDLE)
7548 * ASYM_PACKING needs to move all the work to the highest
7549 * prority CPUs in the group, therefore mark all groups
7550 * of lower priority than ourself as busy.
7552 if (sgs->sum_nr_running &&
7553 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7557 /* Prefer to move from lowest priority cpu's work */
7558 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7559 sg->asym_prefer_cpu))
7566 #ifdef CONFIG_NUMA_BALANCING
7567 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7569 if (sgs->sum_nr_running > sgs->nr_numa_running)
7571 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7576 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7578 if (rq->nr_running > rq->nr_numa_running)
7580 if (rq->nr_running > rq->nr_preferred_running)
7585 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7590 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7594 #endif /* CONFIG_NUMA_BALANCING */
7597 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7598 * @env: The load balancing environment.
7599 * @sds: variable to hold the statistics for this sched_domain.
7601 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7603 struct sched_domain_shared *shared = env->sd->shared;
7604 struct sched_domain *child = env->sd->child;
7605 struct sched_group *sg = env->sd->groups;
7606 struct sg_lb_stats *local = &sds->local_stat;
7607 struct sg_lb_stats tmp_sgs;
7608 int load_idx, prefer_sibling = 0;
7609 bool overload = false;
7611 if (child && child->flags & SD_PREFER_SIBLING)
7614 load_idx = get_sd_load_idx(env->sd, env->idle);
7617 struct sg_lb_stats *sgs = &tmp_sgs;
7620 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7625 if (env->idle != CPU_NEWLY_IDLE ||
7626 time_after_eq(jiffies, sg->sgc->next_update))
7627 update_group_capacity(env->sd, env->dst_cpu);
7630 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7637 * In case the child domain prefers tasks go to siblings
7638 * first, lower the sg capacity so that we'll try
7639 * and move all the excess tasks away. We lower the capacity
7640 * of a group only if the local group has the capacity to fit
7641 * these excess tasks. The extra check prevents the case where
7642 * you always pull from the heaviest group when it is already
7643 * under-utilized (possible with a large weight task outweighs
7644 * the tasks on the system).
7646 if (prefer_sibling && sds->local &&
7647 group_has_capacity(env, local) &&
7648 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7649 sgs->group_no_capacity = 1;
7650 sgs->group_type = group_classify(sg, sgs);
7653 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7655 sds->busiest_stat = *sgs;
7659 /* Now, start updating sd_lb_stats */
7660 sds->total_running += sgs->sum_nr_running;
7661 sds->total_load += sgs->group_load;
7662 sds->total_capacity += sgs->group_capacity;
7665 } while (sg != env->sd->groups);
7667 if (env->sd->flags & SD_NUMA)
7668 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7670 if (!env->sd->parent) {
7671 /* update overload indicator if we are at root domain */
7672 if (env->dst_rq->rd->overload != overload)
7673 env->dst_rq->rd->overload = overload;
7680 * Since these are sums over groups they can contain some CPUs
7681 * multiple times for the NUMA domains.
7683 * Currently only wake_affine_llc() and find_busiest_group()
7684 * uses these numbers, only the last is affected by this problem.
7688 WRITE_ONCE(shared->nr_running, sds->total_running);
7689 WRITE_ONCE(shared->load, sds->total_load);
7690 WRITE_ONCE(shared->capacity, sds->total_capacity);
7694 * check_asym_packing - Check to see if the group is packed into the
7697 * This is primarily intended to used at the sibling level. Some
7698 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7699 * case of POWER7, it can move to lower SMT modes only when higher
7700 * threads are idle. When in lower SMT modes, the threads will
7701 * perform better since they share less core resources. Hence when we
7702 * have idle threads, we want them to be the higher ones.
7704 * This packing function is run on idle threads. It checks to see if
7705 * the busiest CPU in this domain (core in the P7 case) has a higher
7706 * CPU number than the packing function is being run on. Here we are
7707 * assuming lower CPU number will be equivalent to lower a SMT thread
7710 * Return: 1 when packing is required and a task should be moved to
7711 * this CPU. The amount of the imbalance is returned in env->imbalance.
7713 * @env: The load balancing environment.
7714 * @sds: Statistics of the sched_domain which is to be packed
7716 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7720 if (!(env->sd->flags & SD_ASYM_PACKING))
7723 if (env->idle == CPU_NOT_IDLE)
7729 busiest_cpu = sds->busiest->asym_prefer_cpu;
7730 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7733 env->imbalance = DIV_ROUND_CLOSEST(
7734 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7735 SCHED_CAPACITY_SCALE);
7741 * fix_small_imbalance - Calculate the minor imbalance that exists
7742 * amongst the groups of a sched_domain, during
7744 * @env: The load balancing environment.
7745 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7748 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7750 unsigned long tmp, capa_now = 0, capa_move = 0;
7751 unsigned int imbn = 2;
7752 unsigned long scaled_busy_load_per_task;
7753 struct sg_lb_stats *local, *busiest;
7755 local = &sds->local_stat;
7756 busiest = &sds->busiest_stat;
7758 if (!local->sum_nr_running)
7759 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7760 else if (busiest->load_per_task > local->load_per_task)
7763 scaled_busy_load_per_task =
7764 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7765 busiest->group_capacity;
7767 if (busiest->avg_load + scaled_busy_load_per_task >=
7768 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7769 env->imbalance = busiest->load_per_task;
7774 * OK, we don't have enough imbalance to justify moving tasks,
7775 * however we may be able to increase total CPU capacity used by
7779 capa_now += busiest->group_capacity *
7780 min(busiest->load_per_task, busiest->avg_load);
7781 capa_now += local->group_capacity *
7782 min(local->load_per_task, local->avg_load);
7783 capa_now /= SCHED_CAPACITY_SCALE;
7785 /* Amount of load we'd subtract */
7786 if (busiest->avg_load > scaled_busy_load_per_task) {
7787 capa_move += busiest->group_capacity *
7788 min(busiest->load_per_task,
7789 busiest->avg_load - scaled_busy_load_per_task);
7792 /* Amount of load we'd add */
7793 if (busiest->avg_load * busiest->group_capacity <
7794 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7795 tmp = (busiest->avg_load * busiest->group_capacity) /
7796 local->group_capacity;
7798 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7799 local->group_capacity;
7801 capa_move += local->group_capacity *
7802 min(local->load_per_task, local->avg_load + tmp);
7803 capa_move /= SCHED_CAPACITY_SCALE;
7805 /* Move if we gain throughput */
7806 if (capa_move > capa_now)
7807 env->imbalance = busiest->load_per_task;
7811 * calculate_imbalance - Calculate the amount of imbalance present within the
7812 * groups of a given sched_domain during load balance.
7813 * @env: load balance environment
7814 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7816 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7818 unsigned long max_pull, load_above_capacity = ~0UL;
7819 struct sg_lb_stats *local, *busiest;
7821 local = &sds->local_stat;
7822 busiest = &sds->busiest_stat;
7824 if (busiest->group_type == group_imbalanced) {
7826 * In the group_imb case we cannot rely on group-wide averages
7827 * to ensure cpu-load equilibrium, look at wider averages. XXX
7829 busiest->load_per_task =
7830 min(busiest->load_per_task, sds->avg_load);
7834 * Avg load of busiest sg can be less and avg load of local sg can
7835 * be greater than avg load across all sgs of sd because avg load
7836 * factors in sg capacity and sgs with smaller group_type are
7837 * skipped when updating the busiest sg:
7839 if (busiest->avg_load <= sds->avg_load ||
7840 local->avg_load >= sds->avg_load) {
7842 return fix_small_imbalance(env, sds);
7846 * If there aren't any idle cpus, avoid creating some.
7848 if (busiest->group_type == group_overloaded &&
7849 local->group_type == group_overloaded) {
7850 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7851 if (load_above_capacity > busiest->group_capacity) {
7852 load_above_capacity -= busiest->group_capacity;
7853 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7854 load_above_capacity /= busiest->group_capacity;
7856 load_above_capacity = ~0UL;
7860 * We're trying to get all the cpus to the average_load, so we don't
7861 * want to push ourselves above the average load, nor do we wish to
7862 * reduce the max loaded cpu below the average load. At the same time,
7863 * we also don't want to reduce the group load below the group
7864 * capacity. Thus we look for the minimum possible imbalance.
7866 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7868 /* How much load to actually move to equalise the imbalance */
7869 env->imbalance = min(
7870 max_pull * busiest->group_capacity,
7871 (sds->avg_load - local->avg_load) * local->group_capacity
7872 ) / SCHED_CAPACITY_SCALE;
7875 * if *imbalance is less than the average load per runnable task
7876 * there is no guarantee that any tasks will be moved so we'll have
7877 * a think about bumping its value to force at least one task to be
7880 if (env->imbalance < busiest->load_per_task)
7881 return fix_small_imbalance(env, sds);
7884 /******* find_busiest_group() helpers end here *********************/
7887 * find_busiest_group - Returns the busiest group within the sched_domain
7888 * if there is an imbalance.
7890 * Also calculates the amount of weighted load which should be moved
7891 * to restore balance.
7893 * @env: The load balancing environment.
7895 * Return: - The busiest group if imbalance exists.
7897 static struct sched_group *find_busiest_group(struct lb_env *env)
7899 struct sg_lb_stats *local, *busiest;
7900 struct sd_lb_stats sds;
7902 init_sd_lb_stats(&sds);
7905 * Compute the various statistics relavent for load balancing at
7908 update_sd_lb_stats(env, &sds);
7909 local = &sds.local_stat;
7910 busiest = &sds.busiest_stat;
7912 /* ASYM feature bypasses nice load balance check */
7913 if (check_asym_packing(env, &sds))
7916 /* There is no busy sibling group to pull tasks from */
7917 if (!sds.busiest || busiest->sum_nr_running == 0)
7920 /* XXX broken for overlapping NUMA groups */
7921 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7922 / sds.total_capacity;
7925 * If the busiest group is imbalanced the below checks don't
7926 * work because they assume all things are equal, which typically
7927 * isn't true due to cpus_allowed constraints and the like.
7929 if (busiest->group_type == group_imbalanced)
7932 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7933 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7934 busiest->group_no_capacity)
7938 * If the local group is busier than the selected busiest group
7939 * don't try and pull any tasks.
7941 if (local->avg_load >= busiest->avg_load)
7945 * Don't pull any tasks if this group is already above the domain
7948 if (local->avg_load >= sds.avg_load)
7951 if (env->idle == CPU_IDLE) {
7953 * This cpu is idle. If the busiest group is not overloaded
7954 * and there is no imbalance between this and busiest group
7955 * wrt idle cpus, it is balanced. The imbalance becomes
7956 * significant if the diff is greater than 1 otherwise we
7957 * might end up to just move the imbalance on another group
7959 if ((busiest->group_type != group_overloaded) &&
7960 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7964 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7965 * imbalance_pct to be conservative.
7967 if (100 * busiest->avg_load <=
7968 env->sd->imbalance_pct * local->avg_load)
7973 /* Looks like there is an imbalance. Compute it */
7974 calculate_imbalance(env, &sds);
7983 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7985 static struct rq *find_busiest_queue(struct lb_env *env,
7986 struct sched_group *group)
7988 struct rq *busiest = NULL, *rq;
7989 unsigned long busiest_load = 0, busiest_capacity = 1;
7992 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7993 unsigned long capacity, wl;
7997 rt = fbq_classify_rq(rq);
8000 * We classify groups/runqueues into three groups:
8001 * - regular: there are !numa tasks
8002 * - remote: there are numa tasks that run on the 'wrong' node
8003 * - all: there is no distinction
8005 * In order to avoid migrating ideally placed numa tasks,
8006 * ignore those when there's better options.
8008 * If we ignore the actual busiest queue to migrate another
8009 * task, the next balance pass can still reduce the busiest
8010 * queue by moving tasks around inside the node.
8012 * If we cannot move enough load due to this classification
8013 * the next pass will adjust the group classification and
8014 * allow migration of more tasks.
8016 * Both cases only affect the total convergence complexity.
8018 if (rt > env->fbq_type)
8021 capacity = capacity_of(i);
8023 wl = weighted_cpuload(rq);
8026 * When comparing with imbalance, use weighted_cpuload()
8027 * which is not scaled with the cpu capacity.
8030 if (rq->nr_running == 1 && wl > env->imbalance &&
8031 !check_cpu_capacity(rq, env->sd))
8035 * For the load comparisons with the other cpu's, consider
8036 * the weighted_cpuload() scaled with the cpu capacity, so
8037 * that the load can be moved away from the cpu that is
8038 * potentially running at a lower capacity.
8040 * Thus we're looking for max(wl_i / capacity_i), crosswise
8041 * multiplication to rid ourselves of the division works out
8042 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8043 * our previous maximum.
8045 if (wl * busiest_capacity > busiest_load * capacity) {
8047 busiest_capacity = capacity;
8056 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8057 * so long as it is large enough.
8059 #define MAX_PINNED_INTERVAL 512
8061 static int need_active_balance(struct lb_env *env)
8063 struct sched_domain *sd = env->sd;
8065 if (env->idle == CPU_NEWLY_IDLE) {
8068 * ASYM_PACKING needs to force migrate tasks from busy but
8069 * lower priority CPUs in order to pack all tasks in the
8070 * highest priority CPUs.
8072 if ((sd->flags & SD_ASYM_PACKING) &&
8073 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8078 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8079 * It's worth migrating the task if the src_cpu's capacity is reduced
8080 * because of other sched_class or IRQs if more capacity stays
8081 * available on dst_cpu.
8083 if ((env->idle != CPU_NOT_IDLE) &&
8084 (env->src_rq->cfs.h_nr_running == 1)) {
8085 if ((check_cpu_capacity(env->src_rq, sd)) &&
8086 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8090 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8093 static int active_load_balance_cpu_stop(void *data);
8095 static int should_we_balance(struct lb_env *env)
8097 struct sched_group *sg = env->sd->groups;
8098 int cpu, balance_cpu = -1;
8101 * In the newly idle case, we will allow all the cpu's
8102 * to do the newly idle load balance.
8104 if (env->idle == CPU_NEWLY_IDLE)
8107 /* Try to find first idle cpu */
8108 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8116 if (balance_cpu == -1)
8117 balance_cpu = group_balance_cpu(sg);
8120 * First idle cpu or the first cpu(busiest) in this sched group
8121 * is eligible for doing load balancing at this and above domains.
8123 return balance_cpu == env->dst_cpu;
8127 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8128 * tasks if there is an imbalance.
8130 static int load_balance(int this_cpu, struct rq *this_rq,
8131 struct sched_domain *sd, enum cpu_idle_type idle,
8132 int *continue_balancing)
8134 int ld_moved, cur_ld_moved, active_balance = 0;
8135 struct sched_domain *sd_parent = sd->parent;
8136 struct sched_group *group;
8139 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8141 struct lb_env env = {
8143 .dst_cpu = this_cpu,
8145 .dst_grpmask = sched_group_span(sd->groups),
8147 .loop_break = sched_nr_migrate_break,
8150 .tasks = LIST_HEAD_INIT(env.tasks),
8153 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8155 schedstat_inc(sd->lb_count[idle]);
8158 if (!should_we_balance(&env)) {
8159 *continue_balancing = 0;
8163 group = find_busiest_group(&env);
8165 schedstat_inc(sd->lb_nobusyg[idle]);
8169 busiest = find_busiest_queue(&env, group);
8171 schedstat_inc(sd->lb_nobusyq[idle]);
8175 BUG_ON(busiest == env.dst_rq);
8177 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8179 env.src_cpu = busiest->cpu;
8180 env.src_rq = busiest;
8183 if (busiest->nr_running > 1) {
8185 * Attempt to move tasks. If find_busiest_group has found
8186 * an imbalance but busiest->nr_running <= 1, the group is
8187 * still unbalanced. ld_moved simply stays zero, so it is
8188 * correctly treated as an imbalance.
8190 env.flags |= LBF_ALL_PINNED;
8191 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8194 rq_lock_irqsave(busiest, &rf);
8195 update_rq_clock(busiest);
8198 * cur_ld_moved - load moved in current iteration
8199 * ld_moved - cumulative load moved across iterations
8201 cur_ld_moved = detach_tasks(&env);
8204 * We've detached some tasks from busiest_rq. Every
8205 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8206 * unlock busiest->lock, and we are able to be sure
8207 * that nobody can manipulate the tasks in parallel.
8208 * See task_rq_lock() family for the details.
8211 rq_unlock(busiest, &rf);
8215 ld_moved += cur_ld_moved;
8218 local_irq_restore(rf.flags);
8220 if (env.flags & LBF_NEED_BREAK) {
8221 env.flags &= ~LBF_NEED_BREAK;
8226 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8227 * us and move them to an alternate dst_cpu in our sched_group
8228 * where they can run. The upper limit on how many times we
8229 * iterate on same src_cpu is dependent on number of cpus in our
8232 * This changes load balance semantics a bit on who can move
8233 * load to a given_cpu. In addition to the given_cpu itself
8234 * (or a ilb_cpu acting on its behalf where given_cpu is
8235 * nohz-idle), we now have balance_cpu in a position to move
8236 * load to given_cpu. In rare situations, this may cause
8237 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8238 * _independently_ and at _same_ time to move some load to
8239 * given_cpu) causing exceess load to be moved to given_cpu.
8240 * This however should not happen so much in practice and
8241 * moreover subsequent load balance cycles should correct the
8242 * excess load moved.
8244 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8246 /* Prevent to re-select dst_cpu via env's cpus */
8247 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8249 env.dst_rq = cpu_rq(env.new_dst_cpu);
8250 env.dst_cpu = env.new_dst_cpu;
8251 env.flags &= ~LBF_DST_PINNED;
8253 env.loop_break = sched_nr_migrate_break;
8256 * Go back to "more_balance" rather than "redo" since we
8257 * need to continue with same src_cpu.
8263 * We failed to reach balance because of affinity.
8266 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8268 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8269 *group_imbalance = 1;
8272 /* All tasks on this runqueue were pinned by CPU affinity */
8273 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8274 cpumask_clear_cpu(cpu_of(busiest), cpus);
8276 * Attempting to continue load balancing at the current
8277 * sched_domain level only makes sense if there are
8278 * active CPUs remaining as possible busiest CPUs to
8279 * pull load from which are not contained within the
8280 * destination group that is receiving any migrated
8283 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8285 env.loop_break = sched_nr_migrate_break;
8288 goto out_all_pinned;
8293 schedstat_inc(sd->lb_failed[idle]);
8295 * Increment the failure counter only on periodic balance.
8296 * We do not want newidle balance, which can be very
8297 * frequent, pollute the failure counter causing
8298 * excessive cache_hot migrations and active balances.
8300 if (idle != CPU_NEWLY_IDLE)
8301 sd->nr_balance_failed++;
8303 if (need_active_balance(&env)) {
8304 unsigned long flags;
8306 raw_spin_lock_irqsave(&busiest->lock, flags);
8308 /* don't kick the active_load_balance_cpu_stop,
8309 * if the curr task on busiest cpu can't be
8312 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8313 raw_spin_unlock_irqrestore(&busiest->lock,
8315 env.flags |= LBF_ALL_PINNED;
8316 goto out_one_pinned;
8320 * ->active_balance synchronizes accesses to
8321 * ->active_balance_work. Once set, it's cleared
8322 * only after active load balance is finished.
8324 if (!busiest->active_balance) {
8325 busiest->active_balance = 1;
8326 busiest->push_cpu = this_cpu;
8329 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8331 if (active_balance) {
8332 stop_one_cpu_nowait(cpu_of(busiest),
8333 active_load_balance_cpu_stop, busiest,
8334 &busiest->active_balance_work);
8337 /* We've kicked active balancing, force task migration. */
8338 sd->nr_balance_failed = sd->cache_nice_tries+1;
8341 sd->nr_balance_failed = 0;
8343 if (likely(!active_balance)) {
8344 /* We were unbalanced, so reset the balancing interval */
8345 sd->balance_interval = sd->min_interval;
8348 * If we've begun active balancing, start to back off. This
8349 * case may not be covered by the all_pinned logic if there
8350 * is only 1 task on the busy runqueue (because we don't call
8353 if (sd->balance_interval < sd->max_interval)
8354 sd->balance_interval *= 2;
8361 * We reach balance although we may have faced some affinity
8362 * constraints. Clear the imbalance flag if it was set.
8365 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8367 if (*group_imbalance)
8368 *group_imbalance = 0;
8373 * We reach balance because all tasks are pinned at this level so
8374 * we can't migrate them. Let the imbalance flag set so parent level
8375 * can try to migrate them.
8377 schedstat_inc(sd->lb_balanced[idle]);
8379 sd->nr_balance_failed = 0;
8382 /* tune up the balancing interval */
8383 if (((env.flags & LBF_ALL_PINNED) &&
8384 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8385 (sd->balance_interval < sd->max_interval))
8386 sd->balance_interval *= 2;
8393 static inline unsigned long
8394 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8396 unsigned long interval = sd->balance_interval;
8399 interval *= sd->busy_factor;
8401 /* scale ms to jiffies */
8402 interval = msecs_to_jiffies(interval);
8403 interval = clamp(interval, 1UL, max_load_balance_interval);
8409 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8411 unsigned long interval, next;
8413 /* used by idle balance, so cpu_busy = 0 */
8414 interval = get_sd_balance_interval(sd, 0);
8415 next = sd->last_balance + interval;
8417 if (time_after(*next_balance, next))
8418 *next_balance = next;
8422 * idle_balance is called by schedule() if this_cpu is about to become
8423 * idle. Attempts to pull tasks from other CPUs.
8425 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8427 unsigned long next_balance = jiffies + HZ;
8428 int this_cpu = this_rq->cpu;
8429 struct sched_domain *sd;
8430 int pulled_task = 0;
8434 * We must set idle_stamp _before_ calling idle_balance(), such that we
8435 * measure the duration of idle_balance() as idle time.
8437 this_rq->idle_stamp = rq_clock(this_rq);
8440 * Do not pull tasks towards !active CPUs...
8442 if (!cpu_active(this_cpu))
8446 * This is OK, because current is on_cpu, which avoids it being picked
8447 * for load-balance and preemption/IRQs are still disabled avoiding
8448 * further scheduler activity on it and we're being very careful to
8449 * re-start the picking loop.
8451 rq_unpin_lock(this_rq, rf);
8453 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8454 !this_rq->rd->overload) {
8456 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8458 update_next_balance(sd, &next_balance);
8464 raw_spin_unlock(&this_rq->lock);
8466 update_blocked_averages(this_cpu);
8468 for_each_domain(this_cpu, sd) {
8469 int continue_balancing = 1;
8470 u64 t0, domain_cost;
8472 if (!(sd->flags & SD_LOAD_BALANCE))
8475 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8476 update_next_balance(sd, &next_balance);
8480 if (sd->flags & SD_BALANCE_NEWIDLE) {
8481 t0 = sched_clock_cpu(this_cpu);
8483 pulled_task = load_balance(this_cpu, this_rq,
8485 &continue_balancing);
8487 domain_cost = sched_clock_cpu(this_cpu) - t0;
8488 if (domain_cost > sd->max_newidle_lb_cost)
8489 sd->max_newidle_lb_cost = domain_cost;
8491 curr_cost += domain_cost;
8494 update_next_balance(sd, &next_balance);
8497 * Stop searching for tasks to pull if there are
8498 * now runnable tasks on this rq.
8500 if (pulled_task || this_rq->nr_running > 0)
8505 raw_spin_lock(&this_rq->lock);
8507 if (curr_cost > this_rq->max_idle_balance_cost)
8508 this_rq->max_idle_balance_cost = curr_cost;
8511 * While browsing the domains, we released the rq lock, a task could
8512 * have been enqueued in the meantime. Since we're not going idle,
8513 * pretend we pulled a task.
8515 if (this_rq->cfs.h_nr_running && !pulled_task)
8519 /* Move the next balance forward */
8520 if (time_after(this_rq->next_balance, next_balance))
8521 this_rq->next_balance = next_balance;
8523 /* Is there a task of a high priority class? */
8524 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8528 this_rq->idle_stamp = 0;
8530 rq_repin_lock(this_rq, rf);
8536 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8537 * running tasks off the busiest CPU onto idle CPUs. It requires at
8538 * least 1 task to be running on each physical CPU where possible, and
8539 * avoids physical / logical imbalances.
8541 static int active_load_balance_cpu_stop(void *data)
8543 struct rq *busiest_rq = data;
8544 int busiest_cpu = cpu_of(busiest_rq);
8545 int target_cpu = busiest_rq->push_cpu;
8546 struct rq *target_rq = cpu_rq(target_cpu);
8547 struct sched_domain *sd;
8548 struct task_struct *p = NULL;
8551 rq_lock_irq(busiest_rq, &rf);
8553 * Between queueing the stop-work and running it is a hole in which
8554 * CPUs can become inactive. We should not move tasks from or to
8557 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8560 /* make sure the requested cpu hasn't gone down in the meantime */
8561 if (unlikely(busiest_cpu != smp_processor_id() ||
8562 !busiest_rq->active_balance))
8565 /* Is there any task to move? */
8566 if (busiest_rq->nr_running <= 1)
8570 * This condition is "impossible", if it occurs
8571 * we need to fix it. Originally reported by
8572 * Bjorn Helgaas on a 128-cpu setup.
8574 BUG_ON(busiest_rq == target_rq);
8576 /* Search for an sd spanning us and the target CPU. */
8578 for_each_domain(target_cpu, sd) {
8579 if ((sd->flags & SD_LOAD_BALANCE) &&
8580 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8585 struct lb_env env = {
8587 .dst_cpu = target_cpu,
8588 .dst_rq = target_rq,
8589 .src_cpu = busiest_rq->cpu,
8590 .src_rq = busiest_rq,
8593 * can_migrate_task() doesn't need to compute new_dst_cpu
8594 * for active balancing. Since we have CPU_IDLE, but no
8595 * @dst_grpmask we need to make that test go away with lying
8598 .flags = LBF_DST_PINNED,
8601 schedstat_inc(sd->alb_count);
8602 update_rq_clock(busiest_rq);
8604 p = detach_one_task(&env);
8606 schedstat_inc(sd->alb_pushed);
8607 /* Active balancing done, reset the failure counter. */
8608 sd->nr_balance_failed = 0;
8610 schedstat_inc(sd->alb_failed);
8615 busiest_rq->active_balance = 0;
8616 rq_unlock(busiest_rq, &rf);
8619 attach_one_task(target_rq, p);
8626 static inline int on_null_domain(struct rq *rq)
8628 return unlikely(!rcu_dereference_sched(rq->sd));
8631 #ifdef CONFIG_NO_HZ_COMMON
8633 * idle load balancing details
8634 * - When one of the busy CPUs notice that there may be an idle rebalancing
8635 * needed, they will kick the idle load balancer, which then does idle
8636 * load balancing for all the idle CPUs.
8639 cpumask_var_t idle_cpus_mask;
8641 unsigned long next_balance; /* in jiffy units */
8642 } nohz ____cacheline_aligned;
8644 static inline int find_new_ilb(void)
8646 int ilb = cpumask_first(nohz.idle_cpus_mask);
8648 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8655 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8656 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8657 * CPU (if there is one).
8659 static void nohz_balancer_kick(void)
8663 nohz.next_balance++;
8665 ilb_cpu = find_new_ilb();
8667 if (ilb_cpu >= nr_cpu_ids)
8670 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8673 * Use smp_send_reschedule() instead of resched_cpu().
8674 * This way we generate a sched IPI on the target cpu which
8675 * is idle. And the softirq performing nohz idle load balance
8676 * will be run before returning from the IPI.
8678 smp_send_reschedule(ilb_cpu);
8682 void nohz_balance_exit_idle(unsigned int cpu)
8684 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8686 * Completely isolated CPUs don't ever set, so we must test.
8688 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8689 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8690 atomic_dec(&nohz.nr_cpus);
8692 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8696 static inline void set_cpu_sd_state_busy(void)
8698 struct sched_domain *sd;
8699 int cpu = smp_processor_id();
8702 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8704 if (!sd || !sd->nohz_idle)
8708 atomic_inc(&sd->shared->nr_busy_cpus);
8713 void set_cpu_sd_state_idle(void)
8715 struct sched_domain *sd;
8716 int cpu = smp_processor_id();
8719 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8721 if (!sd || sd->nohz_idle)
8725 atomic_dec(&sd->shared->nr_busy_cpus);
8731 * This routine will record that the cpu is going idle with tick stopped.
8732 * This info will be used in performing idle load balancing in the future.
8734 void nohz_balance_enter_idle(int cpu)
8737 * If this cpu is going down, then nothing needs to be done.
8739 if (!cpu_active(cpu))
8742 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8743 if (!is_housekeeping_cpu(cpu))
8746 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8750 * If we're a completely isolated CPU, we don't play.
8752 if (on_null_domain(cpu_rq(cpu)))
8755 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8756 atomic_inc(&nohz.nr_cpus);
8757 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8761 static DEFINE_SPINLOCK(balancing);
8764 * Scale the max load_balance interval with the number of CPUs in the system.
8765 * This trades load-balance latency on larger machines for less cross talk.
8767 void update_max_interval(void)
8769 max_load_balance_interval = HZ*num_online_cpus()/10;
8773 * It checks each scheduling domain to see if it is due to be balanced,
8774 * and initiates a balancing operation if so.
8776 * Balancing parameters are set up in init_sched_domains.
8778 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8780 int continue_balancing = 1;
8782 unsigned long interval;
8783 struct sched_domain *sd;
8784 /* Earliest time when we have to do rebalance again */
8785 unsigned long next_balance = jiffies + 60*HZ;
8786 int update_next_balance = 0;
8787 int need_serialize, need_decay = 0;
8790 update_blocked_averages(cpu);
8793 for_each_domain(cpu, sd) {
8795 * Decay the newidle max times here because this is a regular
8796 * visit to all the domains. Decay ~1% per second.
8798 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8799 sd->max_newidle_lb_cost =
8800 (sd->max_newidle_lb_cost * 253) / 256;
8801 sd->next_decay_max_lb_cost = jiffies + HZ;
8804 max_cost += sd->max_newidle_lb_cost;
8806 if (!(sd->flags & SD_LOAD_BALANCE))
8810 * Stop the load balance at this level. There is another
8811 * CPU in our sched group which is doing load balancing more
8814 if (!continue_balancing) {
8820 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8822 need_serialize = sd->flags & SD_SERIALIZE;
8823 if (need_serialize) {
8824 if (!spin_trylock(&balancing))
8828 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8829 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8831 * The LBF_DST_PINNED logic could have changed
8832 * env->dst_cpu, so we can't know our idle
8833 * state even if we migrated tasks. Update it.
8835 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8837 sd->last_balance = jiffies;
8838 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8841 spin_unlock(&balancing);
8843 if (time_after(next_balance, sd->last_balance + interval)) {
8844 next_balance = sd->last_balance + interval;
8845 update_next_balance = 1;
8850 * Ensure the rq-wide value also decays but keep it at a
8851 * reasonable floor to avoid funnies with rq->avg_idle.
8853 rq->max_idle_balance_cost =
8854 max((u64)sysctl_sched_migration_cost, max_cost);
8859 * next_balance will be updated only when there is a need.
8860 * When the cpu is attached to null domain for ex, it will not be
8863 if (likely(update_next_balance)) {
8864 rq->next_balance = next_balance;
8866 #ifdef CONFIG_NO_HZ_COMMON
8868 * If this CPU has been elected to perform the nohz idle
8869 * balance. Other idle CPUs have already rebalanced with
8870 * nohz_idle_balance() and nohz.next_balance has been
8871 * updated accordingly. This CPU is now running the idle load
8872 * balance for itself and we need to update the
8873 * nohz.next_balance accordingly.
8875 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8876 nohz.next_balance = rq->next_balance;
8881 #ifdef CONFIG_NO_HZ_COMMON
8883 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8884 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8886 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8888 int this_cpu = this_rq->cpu;
8891 /* Earliest time when we have to do rebalance again */
8892 unsigned long next_balance = jiffies + 60*HZ;
8893 int update_next_balance = 0;
8895 if (idle != CPU_IDLE ||
8896 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8899 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8900 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8904 * If this cpu gets work to do, stop the load balancing
8905 * work being done for other cpus. Next load
8906 * balancing owner will pick it up.
8911 rq = cpu_rq(balance_cpu);
8914 * If time for next balance is due,
8917 if (time_after_eq(jiffies, rq->next_balance)) {
8920 rq_lock_irq(rq, &rf);
8921 update_rq_clock(rq);
8922 cpu_load_update_idle(rq);
8923 rq_unlock_irq(rq, &rf);
8925 rebalance_domains(rq, CPU_IDLE);
8928 if (time_after(next_balance, rq->next_balance)) {
8929 next_balance = rq->next_balance;
8930 update_next_balance = 1;
8935 * next_balance will be updated only when there is a need.
8936 * When the CPU is attached to null domain for ex, it will not be
8939 if (likely(update_next_balance))
8940 nohz.next_balance = next_balance;
8942 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8946 * Current heuristic for kicking the idle load balancer in the presence
8947 * of an idle cpu in the system.
8948 * - This rq has more than one task.
8949 * - This rq has at least one CFS task and the capacity of the CPU is
8950 * significantly reduced because of RT tasks or IRQs.
8951 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8952 * multiple busy cpu.
8953 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8954 * domain span are idle.
8956 static inline bool nohz_kick_needed(struct rq *rq)
8958 unsigned long now = jiffies;
8959 struct sched_domain_shared *sds;
8960 struct sched_domain *sd;
8961 int nr_busy, i, cpu = rq->cpu;
8964 if (unlikely(rq->idle_balance))
8968 * We may be recently in ticked or tickless idle mode. At the first
8969 * busy tick after returning from idle, we will update the busy stats.
8971 set_cpu_sd_state_busy();
8972 nohz_balance_exit_idle(cpu);
8975 * None are in tickless mode and hence no need for NOHZ idle load
8978 if (likely(!atomic_read(&nohz.nr_cpus)))
8981 if (time_before(now, nohz.next_balance))
8984 if (rq->nr_running >= 2)
8988 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8991 * XXX: write a coherent comment on why we do this.
8992 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8994 nr_busy = atomic_read(&sds->nr_busy_cpus);
9002 sd = rcu_dereference(rq->sd);
9004 if ((rq->cfs.h_nr_running >= 1) &&
9005 check_cpu_capacity(rq, sd)) {
9011 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9013 for_each_cpu(i, sched_domain_span(sd)) {
9015 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9018 if (sched_asym_prefer(i, cpu)) {
9029 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9033 * run_rebalance_domains is triggered when needed from the scheduler tick.
9034 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9036 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9038 struct rq *this_rq = this_rq();
9039 enum cpu_idle_type idle = this_rq->idle_balance ?
9040 CPU_IDLE : CPU_NOT_IDLE;
9043 * If this cpu has a pending nohz_balance_kick, then do the
9044 * balancing on behalf of the other idle cpus whose ticks are
9045 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9046 * give the idle cpus a chance to load balance. Else we may
9047 * load balance only within the local sched_domain hierarchy
9048 * and abort nohz_idle_balance altogether if we pull some load.
9050 nohz_idle_balance(this_rq, idle);
9051 rebalance_domains(this_rq, idle);
9055 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9057 void trigger_load_balance(struct rq *rq)
9059 /* Don't need to rebalance while attached to NULL domain */
9060 if (unlikely(on_null_domain(rq)))
9063 if (time_after_eq(jiffies, rq->next_balance))
9064 raise_softirq(SCHED_SOFTIRQ);
9065 #ifdef CONFIG_NO_HZ_COMMON
9066 if (nohz_kick_needed(rq))
9067 nohz_balancer_kick();
9071 static void rq_online_fair(struct rq *rq)
9075 update_runtime_enabled(rq);
9078 static void rq_offline_fair(struct rq *rq)
9082 /* Ensure any throttled groups are reachable by pick_next_task */
9083 unthrottle_offline_cfs_rqs(rq);
9086 #endif /* CONFIG_SMP */
9089 * scheduler tick hitting a task of our scheduling class:
9091 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9093 struct cfs_rq *cfs_rq;
9094 struct sched_entity *se = &curr->se;
9096 for_each_sched_entity(se) {
9097 cfs_rq = cfs_rq_of(se);
9098 entity_tick(cfs_rq, se, queued);
9101 if (static_branch_unlikely(&sched_numa_balancing))
9102 task_tick_numa(rq, curr);
9106 * called on fork with the child task as argument from the parent's context
9107 * - child not yet on the tasklist
9108 * - preemption disabled
9110 static void task_fork_fair(struct task_struct *p)
9112 struct cfs_rq *cfs_rq;
9113 struct sched_entity *se = &p->se, *curr;
9114 struct rq *rq = this_rq();
9118 update_rq_clock(rq);
9120 cfs_rq = task_cfs_rq(current);
9121 curr = cfs_rq->curr;
9123 update_curr(cfs_rq);
9124 se->vruntime = curr->vruntime;
9126 place_entity(cfs_rq, se, 1);
9128 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9130 * Upon rescheduling, sched_class::put_prev_task() will place
9131 * 'current' within the tree based on its new key value.
9133 swap(curr->vruntime, se->vruntime);
9137 se->vruntime -= cfs_rq->min_vruntime;
9142 * Priority of the task has changed. Check to see if we preempt
9146 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9148 if (!task_on_rq_queued(p))
9152 * Reschedule if we are currently running on this runqueue and
9153 * our priority decreased, or if we are not currently running on
9154 * this runqueue and our priority is higher than the current's
9156 if (rq->curr == p) {
9157 if (p->prio > oldprio)
9160 check_preempt_curr(rq, p, 0);
9163 static inline bool vruntime_normalized(struct task_struct *p)
9165 struct sched_entity *se = &p->se;
9168 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9169 * the dequeue_entity(.flags=0) will already have normalized the
9176 * When !on_rq, vruntime of the task has usually NOT been normalized.
9177 * But there are some cases where it has already been normalized:
9179 * - A forked child which is waiting for being woken up by
9180 * wake_up_new_task().
9181 * - A task which has been woken up by try_to_wake_up() and
9182 * waiting for actually being woken up by sched_ttwu_pending().
9184 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9192 * Propagate the changes of the sched_entity across the tg tree to make it
9193 * visible to the root
9195 static void propagate_entity_cfs_rq(struct sched_entity *se)
9197 struct cfs_rq *cfs_rq;
9199 /* Start to propagate at parent */
9202 for_each_sched_entity(se) {
9203 cfs_rq = cfs_rq_of(se);
9205 if (cfs_rq_throttled(cfs_rq))
9208 update_load_avg(se, UPDATE_TG);
9212 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9215 static void detach_entity_cfs_rq(struct sched_entity *se)
9217 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9219 /* Catch up with the cfs_rq and remove our load when we leave */
9220 update_load_avg(se, 0);
9221 detach_entity_load_avg(cfs_rq, se);
9222 update_tg_load_avg(cfs_rq, false);
9223 propagate_entity_cfs_rq(se);
9226 static void attach_entity_cfs_rq(struct sched_entity *se)
9228 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9230 #ifdef CONFIG_FAIR_GROUP_SCHED
9232 * Since the real-depth could have been changed (only FAIR
9233 * class maintain depth value), reset depth properly.
9235 se->depth = se->parent ? se->parent->depth + 1 : 0;
9238 /* Synchronize entity with its cfs_rq */
9239 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9240 attach_entity_load_avg(cfs_rq, se);
9241 update_tg_load_avg(cfs_rq, false);
9242 propagate_entity_cfs_rq(se);
9245 static void detach_task_cfs_rq(struct task_struct *p)
9247 struct sched_entity *se = &p->se;
9248 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9250 if (!vruntime_normalized(p)) {
9252 * Fix up our vruntime so that the current sleep doesn't
9253 * cause 'unlimited' sleep bonus.
9255 place_entity(cfs_rq, se, 0);
9256 se->vruntime -= cfs_rq->min_vruntime;
9259 detach_entity_cfs_rq(se);
9262 static void attach_task_cfs_rq(struct task_struct *p)
9264 struct sched_entity *se = &p->se;
9265 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9267 attach_entity_cfs_rq(se);
9269 if (!vruntime_normalized(p))
9270 se->vruntime += cfs_rq->min_vruntime;
9273 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9275 detach_task_cfs_rq(p);
9278 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9280 attach_task_cfs_rq(p);
9282 if (task_on_rq_queued(p)) {
9284 * We were most likely switched from sched_rt, so
9285 * kick off the schedule if running, otherwise just see
9286 * if we can still preempt the current task.
9291 check_preempt_curr(rq, p, 0);
9295 /* Account for a task changing its policy or group.
9297 * This routine is mostly called to set cfs_rq->curr field when a task
9298 * migrates between groups/classes.
9300 static void set_curr_task_fair(struct rq *rq)
9302 struct sched_entity *se = &rq->curr->se;
9304 for_each_sched_entity(se) {
9305 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9307 set_next_entity(cfs_rq, se);
9308 /* ensure bandwidth has been allocated on our new cfs_rq */
9309 account_cfs_rq_runtime(cfs_rq, 0);
9313 void init_cfs_rq(struct cfs_rq *cfs_rq)
9315 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9316 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9317 #ifndef CONFIG_64BIT
9318 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9322 cfs_rq->propagate_avg = 0;
9324 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9325 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9329 #ifdef CONFIG_FAIR_GROUP_SCHED
9330 static void task_set_group_fair(struct task_struct *p)
9332 struct sched_entity *se = &p->se;
9334 set_task_rq(p, task_cpu(p));
9335 se->depth = se->parent ? se->parent->depth + 1 : 0;
9338 static void task_move_group_fair(struct task_struct *p)
9340 detach_task_cfs_rq(p);
9341 set_task_rq(p, task_cpu(p));
9344 /* Tell se's cfs_rq has been changed -- migrated */
9345 p->se.avg.last_update_time = 0;
9347 attach_task_cfs_rq(p);
9350 static void task_change_group_fair(struct task_struct *p, int type)
9353 case TASK_SET_GROUP:
9354 task_set_group_fair(p);
9357 case TASK_MOVE_GROUP:
9358 task_move_group_fair(p);
9363 void free_fair_sched_group(struct task_group *tg)
9367 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9369 for_each_possible_cpu(i) {
9371 kfree(tg->cfs_rq[i]);
9380 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9382 struct sched_entity *se;
9383 struct cfs_rq *cfs_rq;
9386 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9389 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9393 tg->shares = NICE_0_LOAD;
9395 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9397 for_each_possible_cpu(i) {
9398 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9399 GFP_KERNEL, cpu_to_node(i));
9403 se = kzalloc_node(sizeof(struct sched_entity),
9404 GFP_KERNEL, cpu_to_node(i));
9408 init_cfs_rq(cfs_rq);
9409 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9410 init_entity_runnable_average(se);
9421 void online_fair_sched_group(struct task_group *tg)
9423 struct sched_entity *se;
9427 for_each_possible_cpu(i) {
9431 raw_spin_lock_irq(&rq->lock);
9432 update_rq_clock(rq);
9433 attach_entity_cfs_rq(se);
9434 sync_throttle(tg, i);
9435 raw_spin_unlock_irq(&rq->lock);
9439 void unregister_fair_sched_group(struct task_group *tg)
9441 unsigned long flags;
9445 for_each_possible_cpu(cpu) {
9447 remove_entity_load_avg(tg->se[cpu]);
9450 * Only empty task groups can be destroyed; so we can speculatively
9451 * check on_list without danger of it being re-added.
9453 if (!tg->cfs_rq[cpu]->on_list)
9458 raw_spin_lock_irqsave(&rq->lock, flags);
9459 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9460 raw_spin_unlock_irqrestore(&rq->lock, flags);
9464 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9465 struct sched_entity *se, int cpu,
9466 struct sched_entity *parent)
9468 struct rq *rq = cpu_rq(cpu);
9472 init_cfs_rq_runtime(cfs_rq);
9474 tg->cfs_rq[cpu] = cfs_rq;
9477 /* se could be NULL for root_task_group */
9482 se->cfs_rq = &rq->cfs;
9485 se->cfs_rq = parent->my_q;
9486 se->depth = parent->depth + 1;
9490 /* guarantee group entities always have weight */
9491 update_load_set(&se->load, NICE_0_LOAD);
9492 se->parent = parent;
9495 static DEFINE_MUTEX(shares_mutex);
9497 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9502 * We can't change the weight of the root cgroup.
9507 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9509 mutex_lock(&shares_mutex);
9510 if (tg->shares == shares)
9513 tg->shares = shares;
9514 for_each_possible_cpu(i) {
9515 struct rq *rq = cpu_rq(i);
9516 struct sched_entity *se = tg->se[i];
9519 /* Propagate contribution to hierarchy */
9520 rq_lock_irqsave(rq, &rf);
9521 update_rq_clock(rq);
9522 for_each_sched_entity(se) {
9523 update_load_avg(se, UPDATE_TG);
9524 update_cfs_shares(se);
9526 rq_unlock_irqrestore(rq, &rf);
9530 mutex_unlock(&shares_mutex);
9533 #else /* CONFIG_FAIR_GROUP_SCHED */
9535 void free_fair_sched_group(struct task_group *tg) { }
9537 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9542 void online_fair_sched_group(struct task_group *tg) { }
9544 void unregister_fair_sched_group(struct task_group *tg) { }
9546 #endif /* CONFIG_FAIR_GROUP_SCHED */
9549 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9551 struct sched_entity *se = &task->se;
9552 unsigned int rr_interval = 0;
9555 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9558 if (rq->cfs.load.weight)
9559 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9565 * All the scheduling class methods:
9567 const struct sched_class fair_sched_class = {
9568 .next = &idle_sched_class,
9569 .enqueue_task = enqueue_task_fair,
9570 .dequeue_task = dequeue_task_fair,
9571 .yield_task = yield_task_fair,
9572 .yield_to_task = yield_to_task_fair,
9574 .check_preempt_curr = check_preempt_wakeup,
9576 .pick_next_task = pick_next_task_fair,
9577 .put_prev_task = put_prev_task_fair,
9580 .select_task_rq = select_task_rq_fair,
9581 .migrate_task_rq = migrate_task_rq_fair,
9583 .rq_online = rq_online_fair,
9584 .rq_offline = rq_offline_fair,
9586 .task_dead = task_dead_fair,
9587 .set_cpus_allowed = set_cpus_allowed_common,
9590 .set_curr_task = set_curr_task_fair,
9591 .task_tick = task_tick_fair,
9592 .task_fork = task_fork_fair,
9594 .prio_changed = prio_changed_fair,
9595 .switched_from = switched_from_fair,
9596 .switched_to = switched_to_fair,
9598 .get_rr_interval = get_rr_interval_fair,
9600 .update_curr = update_curr_fair,
9602 #ifdef CONFIG_FAIR_GROUP_SCHED
9603 .task_change_group = task_change_group_fair,
9607 #ifdef CONFIG_SCHED_DEBUG
9608 void print_cfs_stats(struct seq_file *m, int cpu)
9610 struct cfs_rq *cfs_rq, *pos;
9613 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9614 print_cfs_rq(m, cpu, cfs_rq);
9618 #ifdef CONFIG_NUMA_BALANCING
9619 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9622 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9624 for_each_online_node(node) {
9625 if (p->numa_faults) {
9626 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9627 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9629 if (p->numa_group) {
9630 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9631 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9633 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9636 #endif /* CONFIG_NUMA_BALANCING */
9637 #endif /* CONFIG_SCHED_DEBUG */
9639 __init void init_sched_fair_class(void)
9642 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9644 #ifdef CONFIG_NO_HZ_COMMON
9645 nohz.next_balance = jiffies;
9646 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);