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 = 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;
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--;
2696 * Signed add and clamp on underflow.
2698 * Explicitly do a load-store to ensure the intermediate value never hits
2699 * memory. This allows lockless observations without ever seeing the negative
2702 #define add_positive(_ptr, _val) do { \
2703 typeof(_ptr) ptr = (_ptr); \
2704 typeof(_val) val = (_val); \
2705 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2709 if (val < 0 && res > var) \
2712 WRITE_ONCE(*ptr, res); \
2716 * Unsigned subtract and clamp on underflow.
2718 * Explicitly do a load-store to ensure the intermediate value never hits
2719 * memory. This allows lockless observations without ever seeing the negative
2722 #define sub_positive(_ptr, _val) do { \
2723 typeof(_ptr) ptr = (_ptr); \
2724 typeof(*ptr) val = (_val); \
2725 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2729 WRITE_ONCE(*ptr, res); \
2734 * XXX we want to get rid of this helper and use the full load resolution.
2736 static inline long se_weight(struct sched_entity *se)
2738 return scale_load_down(se->load.weight);
2742 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2744 cfs_rq->runnable_load_avg += se->avg.load_avg;
2745 cfs_rq->runnable_load_sum += se_weight(se) * se->avg.load_sum;
2749 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2751 sub_positive(&cfs_rq->runnable_load_avg, se->avg.load_avg);
2752 sub_positive(&cfs_rq->runnable_load_sum, se_weight(se) * se->avg.load_sum);
2756 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2758 cfs_rq->avg.load_avg += se->avg.load_avg;
2759 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2763 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2765 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2766 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2770 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2772 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2774 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2776 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2779 #ifdef CONFIG_FAIR_GROUP_SCHED
2782 * All this does is approximate the hierarchical proportion which includes that
2783 * global sum we all love to hate.
2785 * That is, the weight of a group entity, is the proportional share of the
2786 * group weight based on the group runqueue weights. That is:
2788 * tg->weight * grq->load.weight
2789 * ge->load.weight = ----------------------------- (1)
2790 * \Sum grq->load.weight
2792 * Now, because computing that sum is prohibitively expensive to compute (been
2793 * there, done that) we approximate it with this average stuff. The average
2794 * moves slower and therefore the approximation is cheaper and more stable.
2796 * So instead of the above, we substitute:
2798 * grq->load.weight -> grq->avg.load_avg (2)
2800 * which yields the following:
2802 * tg->weight * grq->avg.load_avg
2803 * ge->load.weight = ------------------------------ (3)
2806 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2808 * That is shares_avg, and it is right (given the approximation (2)).
2810 * The problem with it is that because the average is slow -- it was designed
2811 * to be exactly that of course -- this leads to transients in boundary
2812 * conditions. In specific, the case where the group was idle and we start the
2813 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2814 * yielding bad latency etc..
2816 * Now, in that special case (1) reduces to:
2818 * tg->weight * grq->load.weight
2819 * ge->load.weight = ----------------------------- = tg>weight (4)
2822 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2824 * So what we do is modify our approximation (3) to approach (4) in the (near)
2829 * tg->weight * grq->load.weight
2830 * --------------------------------------------------- (5)
2831 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2834 * And that is shares_weight and is icky. In the (near) UP case it approaches
2835 * (4) while in the normal case it approaches (3). It consistently
2836 * overestimates the ge->load.weight and therefore:
2838 * \Sum ge->load.weight >= tg->weight
2842 static long calc_cfs_shares(struct cfs_rq *cfs_rq)
2844 long tg_weight, tg_shares, load, shares;
2845 struct task_group *tg = cfs_rq->tg;
2847 tg_shares = READ_ONCE(tg->shares);
2850 * Because (5) drops to 0 when the cfs_rq is idle, we need to use (3)
2853 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2855 tg_weight = atomic_long_read(&tg->load_avg);
2857 /* Ensure tg_weight >= load */
2858 tg_weight -= cfs_rq->tg_load_avg_contrib;
2861 shares = (tg_shares * load);
2863 shares /= tg_weight;
2866 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2867 * of a group with small tg->shares value. It is a floor value which is
2868 * assigned as a minimum load.weight to the sched_entity representing
2869 * the group on a CPU.
2871 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2872 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2873 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2874 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2877 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2879 # endif /* CONFIG_SMP */
2881 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2882 unsigned long weight)
2885 /* commit outstanding execution time */
2886 if (cfs_rq->curr == se)
2887 update_curr(cfs_rq);
2888 account_entity_dequeue(cfs_rq, se);
2889 dequeue_runnable_load_avg(cfs_rq, se);
2891 dequeue_load_avg(cfs_rq, se);
2893 update_load_set(&se->load, weight);
2896 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum,
2897 LOAD_AVG_MAX - 1024 + se->avg.period_contrib);
2900 enqueue_load_avg(cfs_rq, se);
2902 account_entity_enqueue(cfs_rq, se);
2903 enqueue_runnable_load_avg(cfs_rq, se);
2907 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2909 static void update_cfs_shares(struct sched_entity *se)
2911 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2917 if (throttled_hierarchy(cfs_rq))
2921 shares = READ_ONCE(cfs_rq->tg->shares);
2923 if (likely(se->load.weight == shares))
2926 shares = calc_cfs_shares(cfs_rq);
2929 reweight_entity(cfs_rq_of(se), se, shares);
2932 #else /* CONFIG_FAIR_GROUP_SCHED */
2933 static inline void update_cfs_shares(struct sched_entity *se)
2936 #endif /* CONFIG_FAIR_GROUP_SCHED */
2938 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2940 struct rq *rq = rq_of(cfs_rq);
2942 if (&rq->cfs == cfs_rq) {
2944 * There are a few boundary cases this might miss but it should
2945 * get called often enough that that should (hopefully) not be
2946 * a real problem -- added to that it only calls on the local
2947 * CPU, so if we enqueue remotely we'll miss an update, but
2948 * the next tick/schedule should update.
2950 * It will not get called when we go idle, because the idle
2951 * thread is a different class (!fair), nor will the utilization
2952 * number include things like RT tasks.
2954 * As is, the util number is not freq-invariant (we'd have to
2955 * implement arch_scale_freq_capacity() for that).
2959 cpufreq_update_util(rq, 0);
2966 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2968 static u64 decay_load(u64 val, u64 n)
2970 unsigned int local_n;
2972 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2975 /* after bounds checking we can collapse to 32-bit */
2979 * As y^PERIOD = 1/2, we can combine
2980 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2981 * With a look-up table which covers y^n (n<PERIOD)
2983 * To achieve constant time decay_load.
2985 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2986 val >>= local_n / LOAD_AVG_PERIOD;
2987 local_n %= LOAD_AVG_PERIOD;
2990 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2994 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2996 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3001 c1 = decay_load((u64)d1, periods);
3005 * c2 = 1024 \Sum y^n
3009 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3012 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3014 return c1 + c2 + c3;
3017 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
3020 * Accumulate the three separate parts of the sum; d1 the remainder
3021 * of the last (incomplete) period, d2 the span of full periods and d3
3022 * the remainder of the (incomplete) current period.
3027 * |<->|<----------------->|<--->|
3028 * ... |---x---|------| ... |------|-----x (now)
3031 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3034 * = u y^p + (Step 1)
3037 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3040 static __always_inline u32
3041 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3042 unsigned long weight, int running, struct cfs_rq *cfs_rq)
3044 unsigned long scale_freq, scale_cpu;
3045 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3048 scale_freq = arch_scale_freq_capacity(NULL, cpu);
3049 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3051 delta += sa->period_contrib;
3052 periods = delta / 1024; /* A period is 1024us (~1ms) */
3055 * Step 1: decay old *_sum if we crossed period boundaries.
3058 sa->load_sum = decay_load(sa->load_sum, periods);
3060 cfs_rq->runnable_load_sum =
3061 decay_load(cfs_rq->runnable_load_sum, periods);
3063 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3069 contrib = __accumulate_pelt_segments(periods,
3070 1024 - sa->period_contrib, delta);
3072 sa->period_contrib = delta;
3074 contrib = cap_scale(contrib, scale_freq);
3076 sa->load_sum += weight * contrib;
3078 cfs_rq->runnable_load_sum += weight * contrib;
3081 sa->util_sum += contrib * scale_cpu;
3087 * We can represent the historical contribution to runnable average as the
3088 * coefficients of a geometric series. To do this we sub-divide our runnable
3089 * history into segments of approximately 1ms (1024us); label the segment that
3090 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3092 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3094 * (now) (~1ms ago) (~2ms ago)
3096 * Let u_i denote the fraction of p_i that the entity was runnable.
3098 * We then designate the fractions u_i as our co-efficients, yielding the
3099 * following representation of historical load:
3100 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3102 * We choose y based on the with of a reasonably scheduling period, fixing:
3105 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3106 * approximately half as much as the contribution to load within the last ms
3109 * When a period "rolls over" and we have new u_0`, multiplying the previous
3110 * sum again by y is sufficient to update:
3111 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3112 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3114 static __always_inline int
3115 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3116 unsigned long weight, int running, struct cfs_rq *cfs_rq)
3120 delta = now - sa->last_update_time;
3122 * This should only happen when time goes backwards, which it
3123 * unfortunately does during sched clock init when we swap over to TSC.
3125 if ((s64)delta < 0) {
3126 sa->last_update_time = now;
3131 * Use 1024ns as the unit of measurement since it's a reasonable
3132 * approximation of 1us and fast to compute.
3138 sa->last_update_time += delta << 10;
3141 * running is a subset of runnable (weight) so running can't be set if
3142 * runnable is clear. But there are some corner cases where the current
3143 * se has been already dequeued but cfs_rq->curr still points to it.
3144 * This means that weight will be 0 but not running for a sched_entity
3145 * but also for a cfs_rq if the latter becomes idle. As an example,
3146 * this happens during idle_balance() which calls
3147 * update_blocked_averages()
3153 * Now we know we crossed measurement unit boundaries. The *_avg
3154 * accrues by two steps:
3156 * Step 1: accumulate *_sum since last_update_time. If we haven't
3157 * crossed period boundaries, finish.
3159 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
3165 static __always_inline void
3166 ___update_load_avg(struct sched_avg *sa, unsigned long weight, struct cfs_rq *cfs_rq)
3168 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3171 * Step 2: update *_avg.
3173 sa->load_avg = div_u64(weight * sa->load_sum, divider);
3175 cfs_rq->runnable_load_avg =
3176 div_u64(cfs_rq->runnable_load_sum, divider);
3178 sa->util_avg = sa->util_sum / divider;
3184 * load_sum := runnable_sum
3185 * load_avg = se_weight(se) * runnable_avg
3189 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3190 * load_avg = \Sum se->avg.load_avg
3194 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3196 if (___update_load_sum(now, cpu, &se->avg, 0, 0, NULL)) {
3197 ___update_load_avg(&se->avg, se_weight(se), NULL);
3205 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3207 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq,
3208 cfs_rq->curr == se, NULL)) {
3210 ___update_load_avg(&se->avg, se_weight(se), NULL);
3218 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3220 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3221 scale_load_down(cfs_rq->load.weight),
3222 cfs_rq->curr != NULL, cfs_rq)) {
3223 ___update_load_avg(&cfs_rq->avg, 1, cfs_rq);
3230 #ifdef CONFIG_FAIR_GROUP_SCHED
3232 * update_tg_load_avg - update the tg's load avg
3233 * @cfs_rq: the cfs_rq whose avg changed
3234 * @force: update regardless of how small the difference
3236 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3237 * However, because tg->load_avg is a global value there are performance
3240 * In order to avoid having to look at the other cfs_rq's, we use a
3241 * differential update where we store the last value we propagated. This in
3242 * turn allows skipping updates if the differential is 'small'.
3244 * Updating tg's load_avg is necessary before update_cfs_share().
3246 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3248 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3251 * No need to update load_avg for root_task_group as it is not used.
3253 if (cfs_rq->tg == &root_task_group)
3256 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3257 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3258 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3263 * Called within set_task_rq() right before setting a task's cpu. The
3264 * caller only guarantees p->pi_lock is held; no other assumptions,
3265 * including the state of rq->lock, should be made.
3267 void set_task_rq_fair(struct sched_entity *se,
3268 struct cfs_rq *prev, struct cfs_rq *next)
3270 u64 p_last_update_time;
3271 u64 n_last_update_time;
3273 if (!sched_feat(ATTACH_AGE_LOAD))
3277 * We are supposed to update the task to "current" time, then its up to
3278 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3279 * getting what current time is, so simply throw away the out-of-date
3280 * time. This will result in the wakee task is less decayed, but giving
3281 * the wakee more load sounds not bad.
3283 if (!(se->avg.last_update_time && prev))
3286 #ifndef CONFIG_64BIT
3288 u64 p_last_update_time_copy;
3289 u64 n_last_update_time_copy;
3292 p_last_update_time_copy = prev->load_last_update_time_copy;
3293 n_last_update_time_copy = next->load_last_update_time_copy;
3297 p_last_update_time = prev->avg.last_update_time;
3298 n_last_update_time = next->avg.last_update_time;
3300 } while (p_last_update_time != p_last_update_time_copy ||
3301 n_last_update_time != n_last_update_time_copy);
3304 p_last_update_time = prev->avg.last_update_time;
3305 n_last_update_time = next->avg.last_update_time;
3307 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3308 se->avg.last_update_time = n_last_update_time;
3311 /* Take into account change of utilization of a child task group */
3313 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3315 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3316 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3318 /* Nothing to update */
3322 /* Set new sched_entity's utilization */
3323 se->avg.util_avg = gcfs_rq->avg.util_avg;
3324 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3326 /* Update parent cfs_rq utilization */
3327 add_positive(&cfs_rq->avg.util_avg, delta);
3328 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3331 /* Take into account change of load of a child task group */
3333 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3335 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3336 long delta, load = gcfs_rq->avg.load_avg;
3339 * If the load of group cfs_rq is null, the load of the
3340 * sched_entity will also be null so we can skip the formula
3345 /* Get tg's load and ensure tg_load > 0 */
3346 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3348 /* Ensure tg_load >= load and updated with current load*/
3349 tg_load -= gcfs_rq->tg_load_avg_contrib;
3353 * We need to compute a correction term in the case that the
3354 * task group is consuming more CPU than a task of equal
3355 * weight. A task with a weight equals to tg->shares will have
3356 * a load less or equal to scale_load_down(tg->shares).
3357 * Similarly, the sched_entities that represent the task group
3358 * at parent level, can't have a load higher than
3359 * scale_load_down(tg->shares). And the Sum of sched_entities'
3360 * load must be <= scale_load_down(tg->shares).
3362 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3363 /* scale gcfs_rq's load into tg's shares*/
3364 load *= scale_load_down(gcfs_rq->tg->shares);
3369 delta = load - se->avg.load_avg;
3371 /* Nothing to update */
3375 /* Set new sched_entity's load */
3376 se->avg.load_avg = load;
3377 se->avg.load_sum = LOAD_AVG_MAX;
3379 /* Update parent cfs_rq load */
3380 add_positive(&cfs_rq->avg.load_avg, delta);
3381 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3384 * If the sched_entity is already enqueued, we also have to update the
3385 * runnable load avg.
3388 /* Update parent cfs_rq runnable_load_avg */
3389 add_positive(&cfs_rq->runnable_load_avg, delta);
3390 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3394 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3396 cfs_rq->propagate_avg = 1;
3399 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3401 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3403 if (!cfs_rq->propagate_avg)
3406 cfs_rq->propagate_avg = 0;
3410 /* Update task and its cfs_rq load average */
3411 static inline int propagate_entity_load_avg(struct sched_entity *se)
3413 struct cfs_rq *cfs_rq;
3415 if (entity_is_task(se))
3418 if (!test_and_clear_tg_cfs_propagate(se))
3421 cfs_rq = cfs_rq_of(se);
3423 set_tg_cfs_propagate(cfs_rq);
3425 update_tg_cfs_util(cfs_rq, se);
3426 update_tg_cfs_load(cfs_rq, se);
3432 * Check if we need to update the load and the utilization of a blocked
3435 static inline bool skip_blocked_update(struct sched_entity *se)
3437 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3440 * If sched_entity still have not zero load or utilization, we have to
3443 if (se->avg.load_avg || se->avg.util_avg)
3447 * If there is a pending propagation, we have to update the load and
3448 * the utilization of the sched_entity:
3450 if (gcfs_rq->propagate_avg)
3454 * Otherwise, the load and the utilization of the sched_entity is
3455 * already zero and there is no pending propagation, so it will be a
3456 * waste of time to try to decay it:
3461 #else /* CONFIG_FAIR_GROUP_SCHED */
3463 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3465 static inline int propagate_entity_load_avg(struct sched_entity *se)
3470 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3472 #endif /* CONFIG_FAIR_GROUP_SCHED */
3475 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3476 * @now: current time, as per cfs_rq_clock_task()
3477 * @cfs_rq: cfs_rq to update
3479 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3480 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3481 * post_init_entity_util_avg().
3483 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3485 * Returns true if the load decayed or we removed load.
3487 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3488 * call update_tg_load_avg() when this function returns true.
3491 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3493 struct sched_avg *sa = &cfs_rq->avg;
3494 int decayed, removed_load = 0, removed_util = 0;
3496 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3497 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3498 sub_positive(&sa->load_avg, r);
3499 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3501 set_tg_cfs_propagate(cfs_rq);
3504 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3505 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3506 sub_positive(&sa->util_avg, r);
3507 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3509 set_tg_cfs_propagate(cfs_rq);
3512 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3514 #ifndef CONFIG_64BIT
3516 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3519 if (decayed || removed_util)
3520 cfs_rq_util_change(cfs_rq);
3522 return decayed || removed_load;
3526 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3527 * @cfs_rq: cfs_rq to attach to
3528 * @se: sched_entity to attach
3530 * Must call update_cfs_rq_load_avg() before this, since we rely on
3531 * cfs_rq->avg.last_update_time being current.
3533 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3535 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3536 enqueue_load_avg(cfs_rq, se);
3537 cfs_rq->avg.util_avg += se->avg.util_avg;
3538 cfs_rq->avg.util_sum += se->avg.util_sum;
3539 set_tg_cfs_propagate(cfs_rq);
3541 cfs_rq_util_change(cfs_rq);
3545 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3546 * @cfs_rq: cfs_rq to detach from
3547 * @se: sched_entity to detach
3549 * Must call update_cfs_rq_load_avg() before this, since we rely on
3550 * cfs_rq->avg.last_update_time being current.
3552 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3554 dequeue_load_avg(cfs_rq, se);
3555 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3556 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3557 set_tg_cfs_propagate(cfs_rq);
3559 cfs_rq_util_change(cfs_rq);
3563 * Optional action to be done while updating the load average
3565 #define UPDATE_TG 0x1
3566 #define SKIP_AGE_LOAD 0x2
3567 #define DO_ATTACH 0x4
3569 /* Update task and its cfs_rq load average */
3570 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3572 u64 now = cfs_rq_clock_task(cfs_rq);
3573 struct rq *rq = rq_of(cfs_rq);
3574 int cpu = cpu_of(rq);
3578 * Track task load average for carrying it to new CPU after migrated, and
3579 * track group sched_entity load average for task_h_load calc in migration
3581 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3582 __update_load_avg_se(now, cpu, cfs_rq, se);
3584 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3585 decayed |= propagate_entity_load_avg(se);
3587 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3589 attach_entity_load_avg(cfs_rq, se);
3590 update_tg_load_avg(cfs_rq, 0);
3592 } else if (decayed && (flags & UPDATE_TG))
3593 update_tg_load_avg(cfs_rq, 0);
3596 #ifndef CONFIG_64BIT
3597 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3599 u64 last_update_time_copy;
3600 u64 last_update_time;
3603 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3605 last_update_time = cfs_rq->avg.last_update_time;
3606 } while (last_update_time != last_update_time_copy);
3608 return last_update_time;
3611 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3613 return cfs_rq->avg.last_update_time;
3618 * Synchronize entity load avg of dequeued entity without locking
3621 void sync_entity_load_avg(struct sched_entity *se)
3623 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3624 u64 last_update_time;
3626 last_update_time = cfs_rq_last_update_time(cfs_rq);
3627 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3631 * Task first catches up with cfs_rq, and then subtract
3632 * itself from the cfs_rq (task must be off the queue now).
3634 void remove_entity_load_avg(struct sched_entity *se)
3636 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3639 * tasks cannot exit without having gone through wake_up_new_task() ->
3640 * post_init_entity_util_avg() which will have added things to the
3641 * cfs_rq, so we can remove unconditionally.
3643 * Similarly for groups, they will have passed through
3644 * post_init_entity_util_avg() before unregister_sched_fair_group()
3648 sync_entity_load_avg(se);
3649 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3650 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3653 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3655 return cfs_rq->runnable_load_avg;
3658 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3660 return cfs_rq->avg.load_avg;
3663 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3665 #else /* CONFIG_SMP */
3668 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3673 #define UPDATE_TG 0x0
3674 #define SKIP_AGE_LOAD 0x0
3675 #define DO_ATTACH 0x0
3677 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3679 cfs_rq_util_change(cfs_rq);
3682 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3685 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3687 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3689 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3694 #endif /* CONFIG_SMP */
3696 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3698 #ifdef CONFIG_SCHED_DEBUG
3699 s64 d = se->vruntime - cfs_rq->min_vruntime;
3704 if (d > 3*sysctl_sched_latency)
3705 schedstat_inc(cfs_rq->nr_spread_over);
3710 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3712 u64 vruntime = cfs_rq->min_vruntime;
3715 * The 'current' period is already promised to the current tasks,
3716 * however the extra weight of the new task will slow them down a
3717 * little, place the new task so that it fits in the slot that
3718 * stays open at the end.
3720 if (initial && sched_feat(START_DEBIT))
3721 vruntime += sched_vslice(cfs_rq, se);
3723 /* sleeps up to a single latency don't count. */
3725 unsigned long thresh = sysctl_sched_latency;
3728 * Halve their sleep time's effect, to allow
3729 * for a gentler effect of sleepers:
3731 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3737 /* ensure we never gain time by being placed backwards. */
3738 se->vruntime = max_vruntime(se->vruntime, vruntime);
3741 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3743 static inline void check_schedstat_required(void)
3745 #ifdef CONFIG_SCHEDSTATS
3746 if (schedstat_enabled())
3749 /* Force schedstat enabled if a dependent tracepoint is active */
3750 if (trace_sched_stat_wait_enabled() ||
3751 trace_sched_stat_sleep_enabled() ||
3752 trace_sched_stat_iowait_enabled() ||
3753 trace_sched_stat_blocked_enabled() ||
3754 trace_sched_stat_runtime_enabled()) {
3755 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3756 "stat_blocked and stat_runtime require the "
3757 "kernel parameter schedstats=enable or "
3758 "kernel.sched_schedstats=1\n");
3769 * update_min_vruntime()
3770 * vruntime -= min_vruntime
3774 * update_min_vruntime()
3775 * vruntime += min_vruntime
3777 * this way the vruntime transition between RQs is done when both
3778 * min_vruntime are up-to-date.
3782 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3783 * vruntime -= min_vruntime
3787 * update_min_vruntime()
3788 * vruntime += min_vruntime
3790 * this way we don't have the most up-to-date min_vruntime on the originating
3791 * CPU and an up-to-date min_vruntime on the destination CPU.
3795 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3797 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3798 bool curr = cfs_rq->curr == se;
3801 * If we're the current task, we must renormalise before calling
3805 se->vruntime += cfs_rq->min_vruntime;
3807 update_curr(cfs_rq);
3810 * Otherwise, renormalise after, such that we're placed at the current
3811 * moment in time, instead of some random moment in the past. Being
3812 * placed in the past could significantly boost this task to the
3813 * fairness detriment of existing tasks.
3815 if (renorm && !curr)
3816 se->vruntime += cfs_rq->min_vruntime;
3819 * When enqueuing a sched_entity, we must:
3820 * - Update loads to have both entity and cfs_rq synced with now.
3821 * - Add its load to cfs_rq->runnable_avg
3822 * - For group_entity, update its weight to reflect the new share of
3824 * - Add its new weight to cfs_rq->load.weight
3826 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3827 enqueue_runnable_load_avg(cfs_rq, se);
3828 update_cfs_shares(se);
3829 account_entity_enqueue(cfs_rq, se);
3831 if (flags & ENQUEUE_WAKEUP)
3832 place_entity(cfs_rq, se, 0);
3834 check_schedstat_required();
3835 update_stats_enqueue(cfs_rq, se, flags);
3836 check_spread(cfs_rq, se);
3838 __enqueue_entity(cfs_rq, se);
3841 if (cfs_rq->nr_running == 1) {
3842 list_add_leaf_cfs_rq(cfs_rq);
3843 check_enqueue_throttle(cfs_rq);
3847 static void __clear_buddies_last(struct sched_entity *se)
3849 for_each_sched_entity(se) {
3850 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3851 if (cfs_rq->last != se)
3854 cfs_rq->last = NULL;
3858 static void __clear_buddies_next(struct sched_entity *se)
3860 for_each_sched_entity(se) {
3861 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3862 if (cfs_rq->next != se)
3865 cfs_rq->next = NULL;
3869 static void __clear_buddies_skip(struct sched_entity *se)
3871 for_each_sched_entity(se) {
3872 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3873 if (cfs_rq->skip != se)
3876 cfs_rq->skip = NULL;
3880 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3882 if (cfs_rq->last == se)
3883 __clear_buddies_last(se);
3885 if (cfs_rq->next == se)
3886 __clear_buddies_next(se);
3888 if (cfs_rq->skip == se)
3889 __clear_buddies_skip(se);
3892 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3895 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3898 * Update run-time statistics of the 'current'.
3900 update_curr(cfs_rq);
3903 * When dequeuing a sched_entity, we must:
3904 * - Update loads to have both entity and cfs_rq synced with now.
3905 * - Substract its load from the cfs_rq->runnable_avg.
3906 * - Substract its previous weight from cfs_rq->load.weight.
3907 * - For group entity, update its weight to reflect the new share
3908 * of its group cfs_rq.
3910 update_load_avg(cfs_rq, se, UPDATE_TG);
3911 dequeue_runnable_load_avg(cfs_rq, se);
3913 update_stats_dequeue(cfs_rq, se, flags);
3915 clear_buddies(cfs_rq, se);
3917 if (se != cfs_rq->curr)
3918 __dequeue_entity(cfs_rq, se);
3920 account_entity_dequeue(cfs_rq, se);
3923 * Normalize after update_curr(); which will also have moved
3924 * min_vruntime if @se is the one holding it back. But before doing
3925 * update_min_vruntime() again, which will discount @se's position and
3926 * can move min_vruntime forward still more.
3928 if (!(flags & DEQUEUE_SLEEP))
3929 se->vruntime -= cfs_rq->min_vruntime;
3931 /* return excess runtime on last dequeue */
3932 return_cfs_rq_runtime(cfs_rq);
3934 update_cfs_shares(se);
3937 * Now advance min_vruntime if @se was the entity holding it back,
3938 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3939 * put back on, and if we advance min_vruntime, we'll be placed back
3940 * further than we started -- ie. we'll be penalized.
3942 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3943 update_min_vruntime(cfs_rq);
3947 * Preempt the current task with a newly woken task if needed:
3950 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3952 unsigned long ideal_runtime, delta_exec;
3953 struct sched_entity *se;
3956 ideal_runtime = sched_slice(cfs_rq, curr);
3957 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3958 if (delta_exec > ideal_runtime) {
3959 resched_curr(rq_of(cfs_rq));
3961 * The current task ran long enough, ensure it doesn't get
3962 * re-elected due to buddy favours.
3964 clear_buddies(cfs_rq, curr);
3969 * Ensure that a task that missed wakeup preemption by a
3970 * narrow margin doesn't have to wait for a full slice.
3971 * This also mitigates buddy induced latencies under load.
3973 if (delta_exec < sysctl_sched_min_granularity)
3976 se = __pick_first_entity(cfs_rq);
3977 delta = curr->vruntime - se->vruntime;
3982 if (delta > ideal_runtime)
3983 resched_curr(rq_of(cfs_rq));
3987 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3989 /* 'current' is not kept within the tree. */
3992 * Any task has to be enqueued before it get to execute on
3993 * a CPU. So account for the time it spent waiting on the
3996 update_stats_wait_end(cfs_rq, se);
3997 __dequeue_entity(cfs_rq, se);
3998 update_load_avg(cfs_rq, se, UPDATE_TG);
4001 update_stats_curr_start(cfs_rq, se);
4005 * Track our maximum slice length, if the CPU's load is at
4006 * least twice that of our own weight (i.e. dont track it
4007 * when there are only lesser-weight tasks around):
4009 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4010 schedstat_set(se->statistics.slice_max,
4011 max((u64)schedstat_val(se->statistics.slice_max),
4012 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4015 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4019 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4022 * Pick the next process, keeping these things in mind, in this order:
4023 * 1) keep things fair between processes/task groups
4024 * 2) pick the "next" process, since someone really wants that to run
4025 * 3) pick the "last" process, for cache locality
4026 * 4) do not run the "skip" process, if something else is available
4028 static struct sched_entity *
4029 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4031 struct sched_entity *left = __pick_first_entity(cfs_rq);
4032 struct sched_entity *se;
4035 * If curr is set we have to see if its left of the leftmost entity
4036 * still in the tree, provided there was anything in the tree at all.
4038 if (!left || (curr && entity_before(curr, left)))
4041 se = left; /* ideally we run the leftmost entity */
4044 * Avoid running the skip buddy, if running something else can
4045 * be done without getting too unfair.
4047 if (cfs_rq->skip == se) {
4048 struct sched_entity *second;
4051 second = __pick_first_entity(cfs_rq);
4053 second = __pick_next_entity(se);
4054 if (!second || (curr && entity_before(curr, second)))
4058 if (second && wakeup_preempt_entity(second, left) < 1)
4063 * Prefer last buddy, try to return the CPU to a preempted task.
4065 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4069 * Someone really wants this to run. If it's not unfair, run it.
4071 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4074 clear_buddies(cfs_rq, se);
4079 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4081 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4084 * If still on the runqueue then deactivate_task()
4085 * was not called and update_curr() has to be done:
4088 update_curr(cfs_rq);
4090 /* throttle cfs_rqs exceeding runtime */
4091 check_cfs_rq_runtime(cfs_rq);
4093 check_spread(cfs_rq, prev);
4096 update_stats_wait_start(cfs_rq, prev);
4097 /* Put 'current' back into the tree. */
4098 __enqueue_entity(cfs_rq, prev);
4099 /* in !on_rq case, update occurred at dequeue */
4100 update_load_avg(cfs_rq, prev, 0);
4102 cfs_rq->curr = NULL;
4106 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4109 * Update run-time statistics of the 'current'.
4111 update_curr(cfs_rq);
4114 * Ensure that runnable average is periodically updated.
4116 update_load_avg(cfs_rq, curr, UPDATE_TG);
4117 update_cfs_shares(curr);
4119 #ifdef CONFIG_SCHED_HRTICK
4121 * queued ticks are scheduled to match the slice, so don't bother
4122 * validating it and just reschedule.
4125 resched_curr(rq_of(cfs_rq));
4129 * don't let the period tick interfere with the hrtick preemption
4131 if (!sched_feat(DOUBLE_TICK) &&
4132 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4136 if (cfs_rq->nr_running > 1)
4137 check_preempt_tick(cfs_rq, curr);
4141 /**************************************************
4142 * CFS bandwidth control machinery
4145 #ifdef CONFIG_CFS_BANDWIDTH
4147 #ifdef HAVE_JUMP_LABEL
4148 static struct static_key __cfs_bandwidth_used;
4150 static inline bool cfs_bandwidth_used(void)
4152 return static_key_false(&__cfs_bandwidth_used);
4155 void cfs_bandwidth_usage_inc(void)
4157 static_key_slow_inc(&__cfs_bandwidth_used);
4160 void cfs_bandwidth_usage_dec(void)
4162 static_key_slow_dec(&__cfs_bandwidth_used);
4164 #else /* HAVE_JUMP_LABEL */
4165 static bool cfs_bandwidth_used(void)
4170 void cfs_bandwidth_usage_inc(void) {}
4171 void cfs_bandwidth_usage_dec(void) {}
4172 #endif /* HAVE_JUMP_LABEL */
4175 * default period for cfs group bandwidth.
4176 * default: 0.1s, units: nanoseconds
4178 static inline u64 default_cfs_period(void)
4180 return 100000000ULL;
4183 static inline u64 sched_cfs_bandwidth_slice(void)
4185 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4189 * Replenish runtime according to assigned quota and update expiration time.
4190 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4191 * additional synchronization around rq->lock.
4193 * requires cfs_b->lock
4195 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4199 if (cfs_b->quota == RUNTIME_INF)
4202 now = sched_clock_cpu(smp_processor_id());
4203 cfs_b->runtime = cfs_b->quota;
4204 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4207 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4209 return &tg->cfs_bandwidth;
4212 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4213 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4215 if (unlikely(cfs_rq->throttle_count))
4216 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4218 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4221 /* returns 0 on failure to allocate runtime */
4222 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4224 struct task_group *tg = cfs_rq->tg;
4225 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4226 u64 amount = 0, min_amount, expires;
4228 /* note: this is a positive sum as runtime_remaining <= 0 */
4229 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4231 raw_spin_lock(&cfs_b->lock);
4232 if (cfs_b->quota == RUNTIME_INF)
4233 amount = min_amount;
4235 start_cfs_bandwidth(cfs_b);
4237 if (cfs_b->runtime > 0) {
4238 amount = min(cfs_b->runtime, min_amount);
4239 cfs_b->runtime -= amount;
4243 expires = cfs_b->runtime_expires;
4244 raw_spin_unlock(&cfs_b->lock);
4246 cfs_rq->runtime_remaining += amount;
4248 * we may have advanced our local expiration to account for allowed
4249 * spread between our sched_clock and the one on which runtime was
4252 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4253 cfs_rq->runtime_expires = expires;
4255 return cfs_rq->runtime_remaining > 0;
4259 * Note: This depends on the synchronization provided by sched_clock and the
4260 * fact that rq->clock snapshots this value.
4262 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4264 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4266 /* if the deadline is ahead of our clock, nothing to do */
4267 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4270 if (cfs_rq->runtime_remaining < 0)
4274 * If the local deadline has passed we have to consider the
4275 * possibility that our sched_clock is 'fast' and the global deadline
4276 * has not truly expired.
4278 * Fortunately we can check determine whether this the case by checking
4279 * whether the global deadline has advanced. It is valid to compare
4280 * cfs_b->runtime_expires without any locks since we only care about
4281 * exact equality, so a partial write will still work.
4284 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4285 /* extend local deadline, drift is bounded above by 2 ticks */
4286 cfs_rq->runtime_expires += TICK_NSEC;
4288 /* global deadline is ahead, expiration has passed */
4289 cfs_rq->runtime_remaining = 0;
4293 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4295 /* dock delta_exec before expiring quota (as it could span periods) */
4296 cfs_rq->runtime_remaining -= delta_exec;
4297 expire_cfs_rq_runtime(cfs_rq);
4299 if (likely(cfs_rq->runtime_remaining > 0))
4303 * if we're unable to extend our runtime we resched so that the active
4304 * hierarchy can be throttled
4306 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4307 resched_curr(rq_of(cfs_rq));
4310 static __always_inline
4311 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4313 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4316 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4319 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4321 return cfs_bandwidth_used() && cfs_rq->throttled;
4324 /* check whether cfs_rq, or any parent, is throttled */
4325 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4327 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4331 * Ensure that neither of the group entities corresponding to src_cpu or
4332 * dest_cpu are members of a throttled hierarchy when performing group
4333 * load-balance operations.
4335 static inline int throttled_lb_pair(struct task_group *tg,
4336 int src_cpu, int dest_cpu)
4338 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4340 src_cfs_rq = tg->cfs_rq[src_cpu];
4341 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4343 return throttled_hierarchy(src_cfs_rq) ||
4344 throttled_hierarchy(dest_cfs_rq);
4347 /* updated child weight may affect parent so we have to do this bottom up */
4348 static int tg_unthrottle_up(struct task_group *tg, void *data)
4350 struct rq *rq = data;
4351 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4353 cfs_rq->throttle_count--;
4354 if (!cfs_rq->throttle_count) {
4355 /* adjust cfs_rq_clock_task() */
4356 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4357 cfs_rq->throttled_clock_task;
4363 static int tg_throttle_down(struct task_group *tg, void *data)
4365 struct rq *rq = data;
4366 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4368 /* group is entering throttled state, stop time */
4369 if (!cfs_rq->throttle_count)
4370 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4371 cfs_rq->throttle_count++;
4376 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4378 struct rq *rq = rq_of(cfs_rq);
4379 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4380 struct sched_entity *se;
4381 long task_delta, dequeue = 1;
4384 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4386 /* freeze hierarchy runnable averages while throttled */
4388 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4391 task_delta = cfs_rq->h_nr_running;
4392 for_each_sched_entity(se) {
4393 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4394 /* throttled entity or throttle-on-deactivate */
4399 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4400 qcfs_rq->h_nr_running -= task_delta;
4402 if (qcfs_rq->load.weight)
4407 sub_nr_running(rq, task_delta);
4409 cfs_rq->throttled = 1;
4410 cfs_rq->throttled_clock = rq_clock(rq);
4411 raw_spin_lock(&cfs_b->lock);
4412 empty = list_empty(&cfs_b->throttled_cfs_rq);
4415 * Add to the _head_ of the list, so that an already-started
4416 * distribute_cfs_runtime will not see us
4418 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4421 * If we're the first throttled task, make sure the bandwidth
4425 start_cfs_bandwidth(cfs_b);
4427 raw_spin_unlock(&cfs_b->lock);
4430 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4432 struct rq *rq = rq_of(cfs_rq);
4433 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4434 struct sched_entity *se;
4438 se = cfs_rq->tg->se[cpu_of(rq)];
4440 cfs_rq->throttled = 0;
4442 update_rq_clock(rq);
4444 raw_spin_lock(&cfs_b->lock);
4445 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4446 list_del_rcu(&cfs_rq->throttled_list);
4447 raw_spin_unlock(&cfs_b->lock);
4449 /* update hierarchical throttle state */
4450 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4452 if (!cfs_rq->load.weight)
4455 task_delta = cfs_rq->h_nr_running;
4456 for_each_sched_entity(se) {
4460 cfs_rq = cfs_rq_of(se);
4462 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4463 cfs_rq->h_nr_running += task_delta;
4465 if (cfs_rq_throttled(cfs_rq))
4470 add_nr_running(rq, task_delta);
4472 /* determine whether we need to wake up potentially idle cpu */
4473 if (rq->curr == rq->idle && rq->cfs.nr_running)
4477 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4478 u64 remaining, u64 expires)
4480 struct cfs_rq *cfs_rq;
4482 u64 starting_runtime = remaining;
4485 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4487 struct rq *rq = rq_of(cfs_rq);
4491 if (!cfs_rq_throttled(cfs_rq))
4494 runtime = -cfs_rq->runtime_remaining + 1;
4495 if (runtime > remaining)
4496 runtime = remaining;
4497 remaining -= runtime;
4499 cfs_rq->runtime_remaining += runtime;
4500 cfs_rq->runtime_expires = expires;
4502 /* we check whether we're throttled above */
4503 if (cfs_rq->runtime_remaining > 0)
4504 unthrottle_cfs_rq(cfs_rq);
4514 return starting_runtime - remaining;
4518 * Responsible for refilling a task_group's bandwidth and unthrottling its
4519 * cfs_rqs as appropriate. If there has been no activity within the last
4520 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4521 * used to track this state.
4523 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4525 u64 runtime, runtime_expires;
4528 /* no need to continue the timer with no bandwidth constraint */
4529 if (cfs_b->quota == RUNTIME_INF)
4530 goto out_deactivate;
4532 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4533 cfs_b->nr_periods += overrun;
4536 * idle depends on !throttled (for the case of a large deficit), and if
4537 * we're going inactive then everything else can be deferred
4539 if (cfs_b->idle && !throttled)
4540 goto out_deactivate;
4542 __refill_cfs_bandwidth_runtime(cfs_b);
4545 /* mark as potentially idle for the upcoming period */
4550 /* account preceding periods in which throttling occurred */
4551 cfs_b->nr_throttled += overrun;
4553 runtime_expires = cfs_b->runtime_expires;
4556 * This check is repeated as we are holding onto the new bandwidth while
4557 * we unthrottle. This can potentially race with an unthrottled group
4558 * trying to acquire new bandwidth from the global pool. This can result
4559 * in us over-using our runtime if it is all used during this loop, but
4560 * only by limited amounts in that extreme case.
4562 while (throttled && cfs_b->runtime > 0) {
4563 runtime = cfs_b->runtime;
4564 raw_spin_unlock(&cfs_b->lock);
4565 /* we can't nest cfs_b->lock while distributing bandwidth */
4566 runtime = distribute_cfs_runtime(cfs_b, runtime,
4568 raw_spin_lock(&cfs_b->lock);
4570 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4572 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4576 * While we are ensured activity in the period following an
4577 * unthrottle, this also covers the case in which the new bandwidth is
4578 * insufficient to cover the existing bandwidth deficit. (Forcing the
4579 * timer to remain active while there are any throttled entities.)
4589 /* a cfs_rq won't donate quota below this amount */
4590 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4591 /* minimum remaining period time to redistribute slack quota */
4592 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4593 /* how long we wait to gather additional slack before distributing */
4594 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4597 * Are we near the end of the current quota period?
4599 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4600 * hrtimer base being cleared by hrtimer_start. In the case of
4601 * migrate_hrtimers, base is never cleared, so we are fine.
4603 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4605 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4608 /* if the call-back is running a quota refresh is already occurring */
4609 if (hrtimer_callback_running(refresh_timer))
4612 /* is a quota refresh about to occur? */
4613 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4614 if (remaining < min_expire)
4620 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4622 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4624 /* if there's a quota refresh soon don't bother with slack */
4625 if (runtime_refresh_within(cfs_b, min_left))
4628 hrtimer_start(&cfs_b->slack_timer,
4629 ns_to_ktime(cfs_bandwidth_slack_period),
4633 /* we know any runtime found here is valid as update_curr() precedes return */
4634 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4636 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4637 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4639 if (slack_runtime <= 0)
4642 raw_spin_lock(&cfs_b->lock);
4643 if (cfs_b->quota != RUNTIME_INF &&
4644 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4645 cfs_b->runtime += slack_runtime;
4647 /* we are under rq->lock, defer unthrottling using a timer */
4648 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4649 !list_empty(&cfs_b->throttled_cfs_rq))
4650 start_cfs_slack_bandwidth(cfs_b);
4652 raw_spin_unlock(&cfs_b->lock);
4654 /* even if it's not valid for return we don't want to try again */
4655 cfs_rq->runtime_remaining -= slack_runtime;
4658 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4660 if (!cfs_bandwidth_used())
4663 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4666 __return_cfs_rq_runtime(cfs_rq);
4670 * This is done with a timer (instead of inline with bandwidth return) since
4671 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4673 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4675 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4678 /* confirm we're still not at a refresh boundary */
4679 raw_spin_lock(&cfs_b->lock);
4680 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4681 raw_spin_unlock(&cfs_b->lock);
4685 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4686 runtime = cfs_b->runtime;
4688 expires = cfs_b->runtime_expires;
4689 raw_spin_unlock(&cfs_b->lock);
4694 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4696 raw_spin_lock(&cfs_b->lock);
4697 if (expires == cfs_b->runtime_expires)
4698 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4699 raw_spin_unlock(&cfs_b->lock);
4703 * When a group wakes up we want to make sure that its quota is not already
4704 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4705 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4707 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4709 if (!cfs_bandwidth_used())
4712 /* an active group must be handled by the update_curr()->put() path */
4713 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4716 /* ensure the group is not already throttled */
4717 if (cfs_rq_throttled(cfs_rq))
4720 /* update runtime allocation */
4721 account_cfs_rq_runtime(cfs_rq, 0);
4722 if (cfs_rq->runtime_remaining <= 0)
4723 throttle_cfs_rq(cfs_rq);
4726 static void sync_throttle(struct task_group *tg, int cpu)
4728 struct cfs_rq *pcfs_rq, *cfs_rq;
4730 if (!cfs_bandwidth_used())
4736 cfs_rq = tg->cfs_rq[cpu];
4737 pcfs_rq = tg->parent->cfs_rq[cpu];
4739 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4740 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4743 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4744 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4746 if (!cfs_bandwidth_used())
4749 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4753 * it's possible for a throttled entity to be forced into a running
4754 * state (e.g. set_curr_task), in this case we're finished.
4756 if (cfs_rq_throttled(cfs_rq))
4759 throttle_cfs_rq(cfs_rq);
4763 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4765 struct cfs_bandwidth *cfs_b =
4766 container_of(timer, struct cfs_bandwidth, slack_timer);
4768 do_sched_cfs_slack_timer(cfs_b);
4770 return HRTIMER_NORESTART;
4773 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4775 struct cfs_bandwidth *cfs_b =
4776 container_of(timer, struct cfs_bandwidth, period_timer);
4780 raw_spin_lock(&cfs_b->lock);
4782 overrun = hrtimer_forward_now(timer, cfs_b->period);
4786 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4789 cfs_b->period_active = 0;
4790 raw_spin_unlock(&cfs_b->lock);
4792 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4795 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4797 raw_spin_lock_init(&cfs_b->lock);
4799 cfs_b->quota = RUNTIME_INF;
4800 cfs_b->period = ns_to_ktime(default_cfs_period());
4802 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4803 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4804 cfs_b->period_timer.function = sched_cfs_period_timer;
4805 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4806 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4809 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4811 cfs_rq->runtime_enabled = 0;
4812 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4815 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4817 lockdep_assert_held(&cfs_b->lock);
4819 if (!cfs_b->period_active) {
4820 cfs_b->period_active = 1;
4821 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4822 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4826 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4828 /* init_cfs_bandwidth() was not called */
4829 if (!cfs_b->throttled_cfs_rq.next)
4832 hrtimer_cancel(&cfs_b->period_timer);
4833 hrtimer_cancel(&cfs_b->slack_timer);
4837 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4839 * The race is harmless, since modifying bandwidth settings of unhooked group
4840 * bits doesn't do much.
4843 /* cpu online calback */
4844 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4846 struct task_group *tg;
4848 lockdep_assert_held(&rq->lock);
4851 list_for_each_entry_rcu(tg, &task_groups, list) {
4852 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4853 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4855 raw_spin_lock(&cfs_b->lock);
4856 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4857 raw_spin_unlock(&cfs_b->lock);
4862 /* cpu offline callback */
4863 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4865 struct task_group *tg;
4867 lockdep_assert_held(&rq->lock);
4870 list_for_each_entry_rcu(tg, &task_groups, list) {
4871 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4873 if (!cfs_rq->runtime_enabled)
4877 * clock_task is not advancing so we just need to make sure
4878 * there's some valid quota amount
4880 cfs_rq->runtime_remaining = 1;
4882 * Offline rq is schedulable till cpu is completely disabled
4883 * in take_cpu_down(), so we prevent new cfs throttling here.
4885 cfs_rq->runtime_enabled = 0;
4887 if (cfs_rq_throttled(cfs_rq))
4888 unthrottle_cfs_rq(cfs_rq);
4893 #else /* CONFIG_CFS_BANDWIDTH */
4894 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4896 return rq_clock_task(rq_of(cfs_rq));
4899 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4900 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4901 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4902 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4903 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4905 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4910 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4915 static inline int throttled_lb_pair(struct task_group *tg,
4916 int src_cpu, int dest_cpu)
4921 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4923 #ifdef CONFIG_FAIR_GROUP_SCHED
4924 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4927 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4931 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4932 static inline void update_runtime_enabled(struct rq *rq) {}
4933 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4935 #endif /* CONFIG_CFS_BANDWIDTH */
4937 /**************************************************
4938 * CFS operations on tasks:
4941 #ifdef CONFIG_SCHED_HRTICK
4942 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4944 struct sched_entity *se = &p->se;
4945 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4947 SCHED_WARN_ON(task_rq(p) != rq);
4949 if (rq->cfs.h_nr_running > 1) {
4950 u64 slice = sched_slice(cfs_rq, se);
4951 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4952 s64 delta = slice - ran;
4959 hrtick_start(rq, delta);
4964 * called from enqueue/dequeue and updates the hrtick when the
4965 * current task is from our class and nr_running is low enough
4968 static void hrtick_update(struct rq *rq)
4970 struct task_struct *curr = rq->curr;
4972 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4975 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4976 hrtick_start_fair(rq, curr);
4978 #else /* !CONFIG_SCHED_HRTICK */
4980 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4984 static inline void hrtick_update(struct rq *rq)
4990 * The enqueue_task method is called before nr_running is
4991 * increased. Here we update the fair scheduling stats and
4992 * then put the task into the rbtree:
4995 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4997 struct cfs_rq *cfs_rq;
4998 struct sched_entity *se = &p->se;
5001 * If in_iowait is set, the code below may not trigger any cpufreq
5002 * utilization updates, so do it here explicitly with the IOWAIT flag
5006 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5008 for_each_sched_entity(se) {
5011 cfs_rq = cfs_rq_of(se);
5012 enqueue_entity(cfs_rq, se, flags);
5015 * end evaluation on encountering a throttled cfs_rq
5017 * note: in the case of encountering a throttled cfs_rq we will
5018 * post the final h_nr_running increment below.
5020 if (cfs_rq_throttled(cfs_rq))
5022 cfs_rq->h_nr_running++;
5024 flags = ENQUEUE_WAKEUP;
5027 for_each_sched_entity(se) {
5028 cfs_rq = cfs_rq_of(se);
5029 cfs_rq->h_nr_running++;
5031 if (cfs_rq_throttled(cfs_rq))
5034 update_load_avg(cfs_rq, se, UPDATE_TG);
5035 update_cfs_shares(se);
5039 add_nr_running(rq, 1);
5044 static void set_next_buddy(struct sched_entity *se);
5047 * The dequeue_task method is called before nr_running is
5048 * decreased. We remove the task from the rbtree and
5049 * update the fair scheduling stats:
5051 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5053 struct cfs_rq *cfs_rq;
5054 struct sched_entity *se = &p->se;
5055 int task_sleep = flags & DEQUEUE_SLEEP;
5057 for_each_sched_entity(se) {
5058 cfs_rq = cfs_rq_of(se);
5059 dequeue_entity(cfs_rq, se, flags);
5062 * end evaluation on encountering a throttled cfs_rq
5064 * note: in the case of encountering a throttled cfs_rq we will
5065 * post the final h_nr_running decrement below.
5067 if (cfs_rq_throttled(cfs_rq))
5069 cfs_rq->h_nr_running--;
5071 /* Don't dequeue parent if it has other entities besides us */
5072 if (cfs_rq->load.weight) {
5073 /* Avoid re-evaluating load for this entity: */
5074 se = parent_entity(se);
5076 * Bias pick_next to pick a task from this cfs_rq, as
5077 * p is sleeping when it is within its sched_slice.
5079 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5083 flags |= DEQUEUE_SLEEP;
5086 for_each_sched_entity(se) {
5087 cfs_rq = cfs_rq_of(se);
5088 cfs_rq->h_nr_running--;
5090 if (cfs_rq_throttled(cfs_rq))
5093 update_load_avg(cfs_rq, se, UPDATE_TG);
5094 update_cfs_shares(se);
5098 sub_nr_running(rq, 1);
5105 /* Working cpumask for: load_balance, load_balance_newidle. */
5106 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5107 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5109 #ifdef CONFIG_NO_HZ_COMMON
5111 * per rq 'load' arrray crap; XXX kill this.
5115 * The exact cpuload calculated at every tick would be:
5117 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5119 * If a cpu misses updates for n ticks (as it was idle) and update gets
5120 * called on the n+1-th tick when cpu may be busy, then we have:
5122 * load_n = (1 - 1/2^i)^n * load_0
5123 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5125 * decay_load_missed() below does efficient calculation of
5127 * load' = (1 - 1/2^i)^n * load
5129 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5130 * This allows us to precompute the above in said factors, thereby allowing the
5131 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5132 * fixed_power_int())
5134 * The calculation is approximated on a 128 point scale.
5136 #define DEGRADE_SHIFT 7
5138 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5139 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5140 { 0, 0, 0, 0, 0, 0, 0, 0 },
5141 { 64, 32, 8, 0, 0, 0, 0, 0 },
5142 { 96, 72, 40, 12, 1, 0, 0, 0 },
5143 { 112, 98, 75, 43, 15, 1, 0, 0 },
5144 { 120, 112, 98, 76, 45, 16, 2, 0 }
5148 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5149 * would be when CPU is idle and so we just decay the old load without
5150 * adding any new load.
5152 static unsigned long
5153 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5157 if (!missed_updates)
5160 if (missed_updates >= degrade_zero_ticks[idx])
5164 return load >> missed_updates;
5166 while (missed_updates) {
5167 if (missed_updates % 2)
5168 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5170 missed_updates >>= 1;
5175 #endif /* CONFIG_NO_HZ_COMMON */
5178 * __cpu_load_update - update the rq->cpu_load[] statistics
5179 * @this_rq: The rq to update statistics for
5180 * @this_load: The current load
5181 * @pending_updates: The number of missed updates
5183 * Update rq->cpu_load[] statistics. This function is usually called every
5184 * scheduler tick (TICK_NSEC).
5186 * This function computes a decaying average:
5188 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5190 * Because of NOHZ it might not get called on every tick which gives need for
5191 * the @pending_updates argument.
5193 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5194 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5195 * = A * (A * load[i]_n-2 + B) + B
5196 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5197 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5198 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5199 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5200 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5202 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5203 * any change in load would have resulted in the tick being turned back on.
5205 * For regular NOHZ, this reduces to:
5207 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5209 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5212 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5213 unsigned long pending_updates)
5215 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5218 this_rq->nr_load_updates++;
5220 /* Update our load: */
5221 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5222 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5223 unsigned long old_load, new_load;
5225 /* scale is effectively 1 << i now, and >> i divides by scale */
5227 old_load = this_rq->cpu_load[i];
5228 #ifdef CONFIG_NO_HZ_COMMON
5229 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5230 if (tickless_load) {
5231 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5233 * old_load can never be a negative value because a
5234 * decayed tickless_load cannot be greater than the
5235 * original tickless_load.
5237 old_load += tickless_load;
5240 new_load = this_load;
5242 * Round up the averaging division if load is increasing. This
5243 * prevents us from getting stuck on 9 if the load is 10, for
5246 if (new_load > old_load)
5247 new_load += scale - 1;
5249 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5252 sched_avg_update(this_rq);
5255 /* Used instead of source_load when we know the type == 0 */
5256 static unsigned long weighted_cpuload(struct rq *rq)
5258 return cfs_rq_runnable_load_avg(&rq->cfs);
5261 #ifdef CONFIG_NO_HZ_COMMON
5263 * There is no sane way to deal with nohz on smp when using jiffies because the
5264 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5265 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5267 * Therefore we need to avoid the delta approach from the regular tick when
5268 * possible since that would seriously skew the load calculation. This is why we
5269 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5270 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5271 * loop exit, nohz_idle_balance, nohz full exit...)
5273 * This means we might still be one tick off for nohz periods.
5276 static void cpu_load_update_nohz(struct rq *this_rq,
5277 unsigned long curr_jiffies,
5280 unsigned long pending_updates;
5282 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5283 if (pending_updates) {
5284 this_rq->last_load_update_tick = curr_jiffies;
5286 * In the regular NOHZ case, we were idle, this means load 0.
5287 * In the NOHZ_FULL case, we were non-idle, we should consider
5288 * its weighted load.
5290 cpu_load_update(this_rq, load, pending_updates);
5295 * Called from nohz_idle_balance() to update the load ratings before doing the
5298 static void cpu_load_update_idle(struct rq *this_rq)
5301 * bail if there's load or we're actually up-to-date.
5303 if (weighted_cpuload(this_rq))
5306 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5310 * Record CPU load on nohz entry so we know the tickless load to account
5311 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5312 * than other cpu_load[idx] but it should be fine as cpu_load readers
5313 * shouldn't rely into synchronized cpu_load[*] updates.
5315 void cpu_load_update_nohz_start(void)
5317 struct rq *this_rq = this_rq();
5320 * This is all lockless but should be fine. If weighted_cpuload changes
5321 * concurrently we'll exit nohz. And cpu_load write can race with
5322 * cpu_load_update_idle() but both updater would be writing the same.
5324 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5328 * Account the tickless load in the end of a nohz frame.
5330 void cpu_load_update_nohz_stop(void)
5332 unsigned long curr_jiffies = READ_ONCE(jiffies);
5333 struct rq *this_rq = this_rq();
5337 if (curr_jiffies == this_rq->last_load_update_tick)
5340 load = weighted_cpuload(this_rq);
5341 rq_lock(this_rq, &rf);
5342 update_rq_clock(this_rq);
5343 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5344 rq_unlock(this_rq, &rf);
5346 #else /* !CONFIG_NO_HZ_COMMON */
5347 static inline void cpu_load_update_nohz(struct rq *this_rq,
5348 unsigned long curr_jiffies,
5349 unsigned long load) { }
5350 #endif /* CONFIG_NO_HZ_COMMON */
5352 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5354 #ifdef CONFIG_NO_HZ_COMMON
5355 /* See the mess around cpu_load_update_nohz(). */
5356 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5358 cpu_load_update(this_rq, load, 1);
5362 * Called from scheduler_tick()
5364 void cpu_load_update_active(struct rq *this_rq)
5366 unsigned long load = weighted_cpuload(this_rq);
5368 if (tick_nohz_tick_stopped())
5369 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5371 cpu_load_update_periodic(this_rq, load);
5375 * Return a low guess at the load of a migration-source cpu weighted
5376 * according to the scheduling class and "nice" value.
5378 * We want to under-estimate the load of migration sources, to
5379 * balance conservatively.
5381 static unsigned long source_load(int cpu, int type)
5383 struct rq *rq = cpu_rq(cpu);
5384 unsigned long total = weighted_cpuload(rq);
5386 if (type == 0 || !sched_feat(LB_BIAS))
5389 return min(rq->cpu_load[type-1], total);
5393 * Return a high guess at the load of a migration-target cpu weighted
5394 * according to the scheduling class and "nice" value.
5396 static unsigned long target_load(int cpu, int type)
5398 struct rq *rq = cpu_rq(cpu);
5399 unsigned long total = weighted_cpuload(rq);
5401 if (type == 0 || !sched_feat(LB_BIAS))
5404 return max(rq->cpu_load[type-1], total);
5407 static unsigned long capacity_of(int cpu)
5409 return cpu_rq(cpu)->cpu_capacity;
5412 static unsigned long capacity_orig_of(int cpu)
5414 return cpu_rq(cpu)->cpu_capacity_orig;
5417 static unsigned long cpu_avg_load_per_task(int cpu)
5419 struct rq *rq = cpu_rq(cpu);
5420 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5421 unsigned long load_avg = weighted_cpuload(rq);
5424 return load_avg / nr_running;
5429 static void record_wakee(struct task_struct *p)
5432 * Only decay a single time; tasks that have less then 1 wakeup per
5433 * jiffy will not have built up many flips.
5435 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5436 current->wakee_flips >>= 1;
5437 current->wakee_flip_decay_ts = jiffies;
5440 if (current->last_wakee != p) {
5441 current->last_wakee = p;
5442 current->wakee_flips++;
5447 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5449 * A waker of many should wake a different task than the one last awakened
5450 * at a frequency roughly N times higher than one of its wakees.
5452 * In order to determine whether we should let the load spread vs consolidating
5453 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5454 * partner, and a factor of lls_size higher frequency in the other.
5456 * With both conditions met, we can be relatively sure that the relationship is
5457 * non-monogamous, with partner count exceeding socket size.
5459 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5460 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5463 static int wake_wide(struct task_struct *p)
5465 unsigned int master = current->wakee_flips;
5466 unsigned int slave = p->wakee_flips;
5467 int factor = this_cpu_read(sd_llc_size);
5470 swap(master, slave);
5471 if (slave < factor || master < slave * factor)
5477 unsigned long nr_running;
5479 unsigned long capacity;
5483 static bool get_llc_stats(struct llc_stats *stats, int cpu)
5485 struct sched_domain_shared *sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5490 stats->nr_running = READ_ONCE(sds->nr_running);
5491 stats->load = READ_ONCE(sds->load);
5492 stats->capacity = READ_ONCE(sds->capacity);
5493 stats->has_capacity = stats->nr_running < per_cpu(sd_llc_size, cpu);
5499 * Can a task be moved from prev_cpu to this_cpu without causing a load
5500 * imbalance that would trigger the load balancer?
5502 * Since we're running on 'stale' values, we might in fact create an imbalance
5503 * but recomputing these values is expensive, as that'd mean iteration 2 cache
5504 * domains worth of CPUs.
5507 wake_affine_llc(struct sched_domain *sd, struct task_struct *p,
5508 int this_cpu, int prev_cpu, int sync)
5510 struct llc_stats prev_stats, this_stats;
5511 s64 this_eff_load, prev_eff_load;
5512 unsigned long task_load;
5514 if (!get_llc_stats(&prev_stats, prev_cpu) ||
5515 !get_llc_stats(&this_stats, this_cpu))
5519 * If sync wakeup then subtract the (maximum possible)
5520 * effect of the currently running task from the load
5521 * of the current LLC.
5524 unsigned long current_load = task_h_load(current);
5526 /* in this case load hits 0 and this LLC is considered 'idle' */
5527 if (current_load > this_stats.load)
5530 this_stats.load -= current_load;
5534 * The has_capacity stuff is not SMT aware, but by trying to balance
5535 * the nr_running on both ends we try and fill the domain at equal
5536 * rates, thereby first consuming cores before siblings.
5539 /* if the old cache has capacity, stay there */
5540 if (prev_stats.has_capacity && prev_stats.nr_running < this_stats.nr_running+1)
5543 /* if this cache has capacity, come here */
5544 if (this_stats.has_capacity && this_stats.nr_running+1 < prev_stats.nr_running)
5548 * Check to see if we can move the load without causing too much
5551 task_load = task_h_load(p);
5553 this_eff_load = 100;
5554 this_eff_load *= prev_stats.capacity;
5556 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5557 prev_eff_load *= this_stats.capacity;
5559 this_eff_load *= this_stats.load + task_load;
5560 prev_eff_load *= prev_stats.load - task_load;
5562 return this_eff_load <= prev_eff_load;
5565 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5566 int prev_cpu, int sync)
5568 int this_cpu = smp_processor_id();
5572 * Default to no affine wakeups; wake_affine() should not effect a task
5573 * placement the load-balancer feels inclined to undo. The conservative
5574 * option is therefore to not move tasks when they wake up.
5579 * If the wakeup is across cache domains, try to evaluate if movement
5580 * makes sense, otherwise rely on select_idle_siblings() to do
5581 * placement inside the cache domain.
5583 if (!cpus_share_cache(prev_cpu, this_cpu))
5584 affine = wake_affine_llc(sd, p, this_cpu, prev_cpu, sync);
5586 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5588 schedstat_inc(sd->ttwu_move_affine);
5589 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5595 static inline int task_util(struct task_struct *p);
5596 static int cpu_util_wake(int cpu, struct task_struct *p);
5598 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5600 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5604 * find_idlest_group finds and returns the least busy CPU group within the
5607 static struct sched_group *
5608 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5609 int this_cpu, int sd_flag)
5611 struct sched_group *idlest = NULL, *group = sd->groups;
5612 struct sched_group *most_spare_sg = NULL;
5613 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5614 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5615 unsigned long most_spare = 0, this_spare = 0;
5616 int load_idx = sd->forkexec_idx;
5617 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5618 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5619 (sd->imbalance_pct-100) / 100;
5621 if (sd_flag & SD_BALANCE_WAKE)
5622 load_idx = sd->wake_idx;
5625 unsigned long load, avg_load, runnable_load;
5626 unsigned long spare_cap, max_spare_cap;
5630 /* Skip over this group if it has no CPUs allowed */
5631 if (!cpumask_intersects(sched_group_span(group),
5635 local_group = cpumask_test_cpu(this_cpu,
5636 sched_group_span(group));
5639 * Tally up the load of all CPUs in the group and find
5640 * the group containing the CPU with most spare capacity.
5646 for_each_cpu(i, sched_group_span(group)) {
5647 /* Bias balancing toward cpus of our domain */
5649 load = source_load(i, load_idx);
5651 load = target_load(i, load_idx);
5653 runnable_load += load;
5655 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5657 spare_cap = capacity_spare_wake(i, p);
5659 if (spare_cap > max_spare_cap)
5660 max_spare_cap = spare_cap;
5663 /* Adjust by relative CPU capacity of the group */
5664 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5665 group->sgc->capacity;
5666 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5667 group->sgc->capacity;
5670 this_runnable_load = runnable_load;
5671 this_avg_load = avg_load;
5672 this_spare = max_spare_cap;
5674 if (min_runnable_load > (runnable_load + imbalance)) {
5676 * The runnable load is significantly smaller
5677 * so we can pick this new cpu
5679 min_runnable_load = runnable_load;
5680 min_avg_load = avg_load;
5682 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5683 (100*min_avg_load > imbalance_scale*avg_load)) {
5685 * The runnable loads are close so take the
5686 * blocked load into account through avg_load.
5688 min_avg_load = avg_load;
5692 if (most_spare < max_spare_cap) {
5693 most_spare = max_spare_cap;
5694 most_spare_sg = group;
5697 } while (group = group->next, group != sd->groups);
5700 * The cross-over point between using spare capacity or least load
5701 * is too conservative for high utilization tasks on partially
5702 * utilized systems if we require spare_capacity > task_util(p),
5703 * so we allow for some task stuffing by using
5704 * spare_capacity > task_util(p)/2.
5706 * Spare capacity can't be used for fork because the utilization has
5707 * not been set yet, we must first select a rq to compute the initial
5710 if (sd_flag & SD_BALANCE_FORK)
5713 if (this_spare > task_util(p) / 2 &&
5714 imbalance_scale*this_spare > 100*most_spare)
5717 if (most_spare > task_util(p) / 2)
5718 return most_spare_sg;
5724 if (min_runnable_load > (this_runnable_load + imbalance))
5727 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5728 (100*this_avg_load < imbalance_scale*min_avg_load))
5735 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5738 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5740 unsigned long load, min_load = ULONG_MAX;
5741 unsigned int min_exit_latency = UINT_MAX;
5742 u64 latest_idle_timestamp = 0;
5743 int least_loaded_cpu = this_cpu;
5744 int shallowest_idle_cpu = -1;
5747 /* Check if we have any choice: */
5748 if (group->group_weight == 1)
5749 return cpumask_first(sched_group_span(group));
5751 /* Traverse only the allowed CPUs */
5752 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5754 struct rq *rq = cpu_rq(i);
5755 struct cpuidle_state *idle = idle_get_state(rq);
5756 if (idle && idle->exit_latency < min_exit_latency) {
5758 * We give priority to a CPU whose idle state
5759 * has the smallest exit latency irrespective
5760 * of any idle timestamp.
5762 min_exit_latency = idle->exit_latency;
5763 latest_idle_timestamp = rq->idle_stamp;
5764 shallowest_idle_cpu = i;
5765 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5766 rq->idle_stamp > latest_idle_timestamp) {
5768 * If equal or no active idle state, then
5769 * the most recently idled CPU might have
5772 latest_idle_timestamp = rq->idle_stamp;
5773 shallowest_idle_cpu = i;
5775 } else if (shallowest_idle_cpu == -1) {
5776 load = weighted_cpuload(cpu_rq(i));
5777 if (load < min_load || (load == min_load && i == this_cpu)) {
5779 least_loaded_cpu = i;
5784 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5787 #ifdef CONFIG_SCHED_SMT
5789 static inline void set_idle_cores(int cpu, int val)
5791 struct sched_domain_shared *sds;
5793 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5795 WRITE_ONCE(sds->has_idle_cores, val);
5798 static inline bool test_idle_cores(int cpu, bool def)
5800 struct sched_domain_shared *sds;
5802 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5804 return READ_ONCE(sds->has_idle_cores);
5810 * Scans the local SMT mask to see if the entire core is idle, and records this
5811 * information in sd_llc_shared->has_idle_cores.
5813 * Since SMT siblings share all cache levels, inspecting this limited remote
5814 * state should be fairly cheap.
5816 void __update_idle_core(struct rq *rq)
5818 int core = cpu_of(rq);
5822 if (test_idle_cores(core, true))
5825 for_each_cpu(cpu, cpu_smt_mask(core)) {
5833 set_idle_cores(core, 1);
5839 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5840 * there are no idle cores left in the system; tracked through
5841 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5843 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5845 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5848 if (!static_branch_likely(&sched_smt_present))
5851 if (!test_idle_cores(target, false))
5854 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5856 for_each_cpu_wrap(core, cpus, target) {
5859 for_each_cpu(cpu, cpu_smt_mask(core)) {
5860 cpumask_clear_cpu(cpu, cpus);
5870 * Failed to find an idle core; stop looking for one.
5872 set_idle_cores(target, 0);
5878 * Scan the local SMT mask for idle CPUs.
5880 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5884 if (!static_branch_likely(&sched_smt_present))
5887 for_each_cpu(cpu, cpu_smt_mask(target)) {
5888 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5897 #else /* CONFIG_SCHED_SMT */
5899 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5904 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5909 #endif /* CONFIG_SCHED_SMT */
5912 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5913 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5914 * average idle time for this rq (as found in rq->avg_idle).
5916 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5918 struct sched_domain *this_sd;
5919 u64 avg_cost, avg_idle;
5922 int cpu, nr = INT_MAX;
5924 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5929 * Due to large variance we need a large fuzz factor; hackbench in
5930 * particularly is sensitive here.
5932 avg_idle = this_rq()->avg_idle / 512;
5933 avg_cost = this_sd->avg_scan_cost + 1;
5935 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5938 if (sched_feat(SIS_PROP)) {
5939 u64 span_avg = sd->span_weight * avg_idle;
5940 if (span_avg > 4*avg_cost)
5941 nr = div_u64(span_avg, avg_cost);
5946 time = local_clock();
5948 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5951 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5957 time = local_clock() - time;
5958 cost = this_sd->avg_scan_cost;
5959 delta = (s64)(time - cost) / 8;
5960 this_sd->avg_scan_cost += delta;
5966 * Try and locate an idle core/thread in the LLC cache domain.
5968 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5970 struct sched_domain *sd;
5973 if (idle_cpu(target))
5977 * If the previous cpu is cache affine and idle, don't be stupid.
5979 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5982 sd = rcu_dereference(per_cpu(sd_llc, target));
5986 i = select_idle_core(p, sd, target);
5987 if ((unsigned)i < nr_cpumask_bits)
5990 i = select_idle_cpu(p, sd, target);
5991 if ((unsigned)i < nr_cpumask_bits)
5994 i = select_idle_smt(p, sd, target);
5995 if ((unsigned)i < nr_cpumask_bits)
6002 * cpu_util returns the amount of capacity of a CPU that is used by CFS
6003 * tasks. The unit of the return value must be the one of capacity so we can
6004 * compare the utilization with the capacity of the CPU that is available for
6005 * CFS task (ie cpu_capacity).
6007 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6008 * recent utilization of currently non-runnable tasks on a CPU. It represents
6009 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6010 * capacity_orig is the cpu_capacity available at the highest frequency
6011 * (arch_scale_freq_capacity()).
6012 * The utilization of a CPU converges towards a sum equal to or less than the
6013 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6014 * the running time on this CPU scaled by capacity_curr.
6016 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6017 * higher than capacity_orig because of unfortunate rounding in
6018 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6019 * the average stabilizes with the new running time. We need to check that the
6020 * utilization stays within the range of [0..capacity_orig] and cap it if
6021 * necessary. Without utilization capping, a group could be seen as overloaded
6022 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6023 * available capacity. We allow utilization to overshoot capacity_curr (but not
6024 * capacity_orig) as it useful for predicting the capacity required after task
6025 * migrations (scheduler-driven DVFS).
6027 static int cpu_util(int cpu)
6029 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
6030 unsigned long capacity = capacity_orig_of(cpu);
6032 return (util >= capacity) ? capacity : util;
6035 static inline int task_util(struct task_struct *p)
6037 return p->se.avg.util_avg;
6041 * cpu_util_wake: Compute cpu utilization with any contributions from
6042 * the waking task p removed.
6044 static int cpu_util_wake(int cpu, struct task_struct *p)
6046 unsigned long util, capacity;
6048 /* Task has no contribution or is new */
6049 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6050 return cpu_util(cpu);
6052 capacity = capacity_orig_of(cpu);
6053 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6055 return (util >= capacity) ? capacity : util;
6059 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6060 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6062 * In that case WAKE_AFFINE doesn't make sense and we'll let
6063 * BALANCE_WAKE sort things out.
6065 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6067 long min_cap, max_cap;
6069 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6070 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6072 /* Minimum capacity is close to max, no need to abort wake_affine */
6073 if (max_cap - min_cap < max_cap >> 3)
6076 /* Bring task utilization in sync with prev_cpu */
6077 sync_entity_load_avg(&p->se);
6079 return min_cap * 1024 < task_util(p) * capacity_margin;
6083 * select_task_rq_fair: Select target runqueue for the waking task in domains
6084 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6085 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6087 * Balances load by selecting the idlest cpu in the idlest group, or under
6088 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6090 * Returns the target cpu number.
6092 * preempt must be disabled.
6095 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6097 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6098 int cpu = smp_processor_id();
6099 int new_cpu = prev_cpu;
6100 int want_affine = 0;
6101 int sync = wake_flags & WF_SYNC;
6103 if (sd_flag & SD_BALANCE_WAKE) {
6105 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6106 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6110 for_each_domain(cpu, tmp) {
6111 if (!(tmp->flags & SD_LOAD_BALANCE))
6115 * If both cpu and prev_cpu are part of this domain,
6116 * cpu is a valid SD_WAKE_AFFINE target.
6118 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6119 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6124 if (tmp->flags & sd_flag)
6126 else if (!want_affine)
6131 sd = NULL; /* Prefer wake_affine over balance flags */
6132 if (cpu == prev_cpu)
6135 if (wake_affine(affine_sd, p, prev_cpu, sync))
6141 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6142 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6145 struct sched_group *group;
6148 if (!(sd->flags & sd_flag)) {
6153 group = find_idlest_group(sd, p, cpu, sd_flag);
6159 new_cpu = find_idlest_cpu(group, p, cpu);
6160 if (new_cpu == -1 || new_cpu == cpu) {
6161 /* Now try balancing at a lower domain level of cpu */
6166 /* Now try balancing at a lower domain level of new_cpu */
6168 weight = sd->span_weight;
6170 for_each_domain(cpu, tmp) {
6171 if (weight <= tmp->span_weight)
6173 if (tmp->flags & sd_flag)
6176 /* while loop will break here if sd == NULL */
6184 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6185 * cfs_rq_of(p) references at time of call are still valid and identify the
6186 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6188 static void migrate_task_rq_fair(struct task_struct *p)
6191 * As blocked tasks retain absolute vruntime the migration needs to
6192 * deal with this by subtracting the old and adding the new
6193 * min_vruntime -- the latter is done by enqueue_entity() when placing
6194 * the task on the new runqueue.
6196 if (p->state == TASK_WAKING) {
6197 struct sched_entity *se = &p->se;
6198 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6201 #ifndef CONFIG_64BIT
6202 u64 min_vruntime_copy;
6205 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6207 min_vruntime = cfs_rq->min_vruntime;
6208 } while (min_vruntime != min_vruntime_copy);
6210 min_vruntime = cfs_rq->min_vruntime;
6213 se->vruntime -= min_vruntime;
6217 * We are supposed to update the task to "current" time, then its up to date
6218 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6219 * what current time is, so simply throw away the out-of-date time. This
6220 * will result in the wakee task is less decayed, but giving the wakee more
6221 * load sounds not bad.
6223 remove_entity_load_avg(&p->se);
6225 /* Tell new CPU we are migrated */
6226 p->se.avg.last_update_time = 0;
6228 /* We have migrated, no longer consider this task hot */
6229 p->se.exec_start = 0;
6232 static void task_dead_fair(struct task_struct *p)
6234 remove_entity_load_avg(&p->se);
6236 #endif /* CONFIG_SMP */
6238 static unsigned long
6239 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6241 unsigned long gran = sysctl_sched_wakeup_granularity;
6244 * Since its curr running now, convert the gran from real-time
6245 * to virtual-time in his units.
6247 * By using 'se' instead of 'curr' we penalize light tasks, so
6248 * they get preempted easier. That is, if 'se' < 'curr' then
6249 * the resulting gran will be larger, therefore penalizing the
6250 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6251 * be smaller, again penalizing the lighter task.
6253 * This is especially important for buddies when the leftmost
6254 * task is higher priority than the buddy.
6256 return calc_delta_fair(gran, se);
6260 * Should 'se' preempt 'curr'.
6274 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6276 s64 gran, vdiff = curr->vruntime - se->vruntime;
6281 gran = wakeup_gran(curr, se);
6288 static void set_last_buddy(struct sched_entity *se)
6290 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6293 for_each_sched_entity(se) {
6294 if (SCHED_WARN_ON(!se->on_rq))
6296 cfs_rq_of(se)->last = se;
6300 static void set_next_buddy(struct sched_entity *se)
6302 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6305 for_each_sched_entity(se) {
6306 if (SCHED_WARN_ON(!se->on_rq))
6308 cfs_rq_of(se)->next = se;
6312 static void set_skip_buddy(struct sched_entity *se)
6314 for_each_sched_entity(se)
6315 cfs_rq_of(se)->skip = se;
6319 * Preempt the current task with a newly woken task if needed:
6321 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6323 struct task_struct *curr = rq->curr;
6324 struct sched_entity *se = &curr->se, *pse = &p->se;
6325 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6326 int scale = cfs_rq->nr_running >= sched_nr_latency;
6327 int next_buddy_marked = 0;
6329 if (unlikely(se == pse))
6333 * This is possible from callers such as attach_tasks(), in which we
6334 * unconditionally check_prempt_curr() after an enqueue (which may have
6335 * lead to a throttle). This both saves work and prevents false
6336 * next-buddy nomination below.
6338 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6341 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6342 set_next_buddy(pse);
6343 next_buddy_marked = 1;
6347 * We can come here with TIF_NEED_RESCHED already set from new task
6350 * Note: this also catches the edge-case of curr being in a throttled
6351 * group (e.g. via set_curr_task), since update_curr() (in the
6352 * enqueue of curr) will have resulted in resched being set. This
6353 * prevents us from potentially nominating it as a false LAST_BUDDY
6356 if (test_tsk_need_resched(curr))
6359 /* Idle tasks are by definition preempted by non-idle tasks. */
6360 if (unlikely(curr->policy == SCHED_IDLE) &&
6361 likely(p->policy != SCHED_IDLE))
6365 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6366 * is driven by the tick):
6368 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6371 find_matching_se(&se, &pse);
6372 update_curr(cfs_rq_of(se));
6374 if (wakeup_preempt_entity(se, pse) == 1) {
6376 * Bias pick_next to pick the sched entity that is
6377 * triggering this preemption.
6379 if (!next_buddy_marked)
6380 set_next_buddy(pse);
6389 * Only set the backward buddy when the current task is still
6390 * on the rq. This can happen when a wakeup gets interleaved
6391 * with schedule on the ->pre_schedule() or idle_balance()
6392 * point, either of which can * drop the rq lock.
6394 * Also, during early boot the idle thread is in the fair class,
6395 * for obvious reasons its a bad idea to schedule back to it.
6397 if (unlikely(!se->on_rq || curr == rq->idle))
6400 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6404 static struct task_struct *
6405 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6407 struct cfs_rq *cfs_rq = &rq->cfs;
6408 struct sched_entity *se;
6409 struct task_struct *p;
6413 if (!cfs_rq->nr_running)
6416 #ifdef CONFIG_FAIR_GROUP_SCHED
6417 if (prev->sched_class != &fair_sched_class)
6421 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6422 * likely that a next task is from the same cgroup as the current.
6424 * Therefore attempt to avoid putting and setting the entire cgroup
6425 * hierarchy, only change the part that actually changes.
6429 struct sched_entity *curr = cfs_rq->curr;
6432 * Since we got here without doing put_prev_entity() we also
6433 * have to consider cfs_rq->curr. If it is still a runnable
6434 * entity, update_curr() will update its vruntime, otherwise
6435 * forget we've ever seen it.
6439 update_curr(cfs_rq);
6444 * This call to check_cfs_rq_runtime() will do the
6445 * throttle and dequeue its entity in the parent(s).
6446 * Therefore the nr_running test will indeed
6449 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6452 if (!cfs_rq->nr_running)
6459 se = pick_next_entity(cfs_rq, curr);
6460 cfs_rq = group_cfs_rq(se);
6466 * Since we haven't yet done put_prev_entity and if the selected task
6467 * is a different task than we started out with, try and touch the
6468 * least amount of cfs_rqs.
6471 struct sched_entity *pse = &prev->se;
6473 while (!(cfs_rq = is_same_group(se, pse))) {
6474 int se_depth = se->depth;
6475 int pse_depth = pse->depth;
6477 if (se_depth <= pse_depth) {
6478 put_prev_entity(cfs_rq_of(pse), pse);
6479 pse = parent_entity(pse);
6481 if (se_depth >= pse_depth) {
6482 set_next_entity(cfs_rq_of(se), se);
6483 se = parent_entity(se);
6487 put_prev_entity(cfs_rq, pse);
6488 set_next_entity(cfs_rq, se);
6491 if (hrtick_enabled(rq))
6492 hrtick_start_fair(rq, p);
6498 put_prev_task(rq, prev);
6501 se = pick_next_entity(cfs_rq, NULL);
6502 set_next_entity(cfs_rq, se);
6503 cfs_rq = group_cfs_rq(se);
6508 if (hrtick_enabled(rq))
6509 hrtick_start_fair(rq, p);
6514 new_tasks = idle_balance(rq, rf);
6517 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6518 * possible for any higher priority task to appear. In that case we
6519 * must re-start the pick_next_entity() loop.
6531 * Account for a descheduled task:
6533 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6535 struct sched_entity *se = &prev->se;
6536 struct cfs_rq *cfs_rq;
6538 for_each_sched_entity(se) {
6539 cfs_rq = cfs_rq_of(se);
6540 put_prev_entity(cfs_rq, se);
6545 * sched_yield() is very simple
6547 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6549 static void yield_task_fair(struct rq *rq)
6551 struct task_struct *curr = rq->curr;
6552 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6553 struct sched_entity *se = &curr->se;
6556 * Are we the only task in the tree?
6558 if (unlikely(rq->nr_running == 1))
6561 clear_buddies(cfs_rq, se);
6563 if (curr->policy != SCHED_BATCH) {
6564 update_rq_clock(rq);
6566 * Update run-time statistics of the 'current'.
6568 update_curr(cfs_rq);
6570 * Tell update_rq_clock() that we've just updated,
6571 * so we don't do microscopic update in schedule()
6572 * and double the fastpath cost.
6574 rq_clock_skip_update(rq, true);
6580 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6582 struct sched_entity *se = &p->se;
6584 /* throttled hierarchies are not runnable */
6585 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6588 /* Tell the scheduler that we'd really like pse to run next. */
6591 yield_task_fair(rq);
6597 /**************************************************
6598 * Fair scheduling class load-balancing methods.
6602 * The purpose of load-balancing is to achieve the same basic fairness the
6603 * per-cpu scheduler provides, namely provide a proportional amount of compute
6604 * time to each task. This is expressed in the following equation:
6606 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6608 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6609 * W_i,0 is defined as:
6611 * W_i,0 = \Sum_j w_i,j (2)
6613 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6614 * is derived from the nice value as per sched_prio_to_weight[].
6616 * The weight average is an exponential decay average of the instantaneous
6619 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6621 * C_i is the compute capacity of cpu i, typically it is the
6622 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6623 * can also include other factors [XXX].
6625 * To achieve this balance we define a measure of imbalance which follows
6626 * directly from (1):
6628 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6630 * We them move tasks around to minimize the imbalance. In the continuous
6631 * function space it is obvious this converges, in the discrete case we get
6632 * a few fun cases generally called infeasible weight scenarios.
6635 * - infeasible weights;
6636 * - local vs global optima in the discrete case. ]
6641 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6642 * for all i,j solution, we create a tree of cpus that follows the hardware
6643 * topology where each level pairs two lower groups (or better). This results
6644 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6645 * tree to only the first of the previous level and we decrease the frequency
6646 * of load-balance at each level inv. proportional to the number of cpus in
6652 * \Sum { --- * --- * 2^i } = O(n) (5)
6654 * `- size of each group
6655 * | | `- number of cpus doing load-balance
6657 * `- sum over all levels
6659 * Coupled with a limit on how many tasks we can migrate every balance pass,
6660 * this makes (5) the runtime complexity of the balancer.
6662 * An important property here is that each CPU is still (indirectly) connected
6663 * to every other cpu in at most O(log n) steps:
6665 * The adjacency matrix of the resulting graph is given by:
6668 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6671 * And you'll find that:
6673 * A^(log_2 n)_i,j != 0 for all i,j (7)
6675 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6676 * The task movement gives a factor of O(m), giving a convergence complexity
6679 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6684 * In order to avoid CPUs going idle while there's still work to do, new idle
6685 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6686 * tree itself instead of relying on other CPUs to bring it work.
6688 * This adds some complexity to both (5) and (8) but it reduces the total idle
6696 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6699 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6704 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6706 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6708 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6711 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6712 * rewrite all of this once again.]
6715 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6717 enum fbq_type { regular, remote, all };
6719 #define LBF_ALL_PINNED 0x01
6720 #define LBF_NEED_BREAK 0x02
6721 #define LBF_DST_PINNED 0x04
6722 #define LBF_SOME_PINNED 0x08
6725 struct sched_domain *sd;
6733 struct cpumask *dst_grpmask;
6735 enum cpu_idle_type idle;
6737 /* The set of CPUs under consideration for load-balancing */
6738 struct cpumask *cpus;
6743 unsigned int loop_break;
6744 unsigned int loop_max;
6746 enum fbq_type fbq_type;
6747 struct list_head tasks;
6751 * Is this task likely cache-hot:
6753 static int task_hot(struct task_struct *p, struct lb_env *env)
6757 lockdep_assert_held(&env->src_rq->lock);
6759 if (p->sched_class != &fair_sched_class)
6762 if (unlikely(p->policy == SCHED_IDLE))
6766 * Buddy candidates are cache hot:
6768 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6769 (&p->se == cfs_rq_of(&p->se)->next ||
6770 &p->se == cfs_rq_of(&p->se)->last))
6773 if (sysctl_sched_migration_cost == -1)
6775 if (sysctl_sched_migration_cost == 0)
6778 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6780 return delta < (s64)sysctl_sched_migration_cost;
6783 #ifdef CONFIG_NUMA_BALANCING
6785 * Returns 1, if task migration degrades locality
6786 * Returns 0, if task migration improves locality i.e migration preferred.
6787 * Returns -1, if task migration is not affected by locality.
6789 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6791 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6792 unsigned long src_faults, dst_faults;
6793 int src_nid, dst_nid;
6795 if (!static_branch_likely(&sched_numa_balancing))
6798 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6801 src_nid = cpu_to_node(env->src_cpu);
6802 dst_nid = cpu_to_node(env->dst_cpu);
6804 if (src_nid == dst_nid)
6807 /* Migrating away from the preferred node is always bad. */
6808 if (src_nid == p->numa_preferred_nid) {
6809 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6815 /* Encourage migration to the preferred node. */
6816 if (dst_nid == p->numa_preferred_nid)
6819 /* Leaving a core idle is often worse than degrading locality. */
6820 if (env->idle != CPU_NOT_IDLE)
6824 src_faults = group_faults(p, src_nid);
6825 dst_faults = group_faults(p, dst_nid);
6827 src_faults = task_faults(p, src_nid);
6828 dst_faults = task_faults(p, dst_nid);
6831 return dst_faults < src_faults;
6835 static inline int migrate_degrades_locality(struct task_struct *p,
6843 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6846 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6850 lockdep_assert_held(&env->src_rq->lock);
6853 * We do not migrate tasks that are:
6854 * 1) throttled_lb_pair, or
6855 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6856 * 3) running (obviously), or
6857 * 4) are cache-hot on their current CPU.
6859 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6862 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6865 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6867 env->flags |= LBF_SOME_PINNED;
6870 * Remember if this task can be migrated to any other cpu in
6871 * our sched_group. We may want to revisit it if we couldn't
6872 * meet load balance goals by pulling other tasks on src_cpu.
6874 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6875 * already computed one in current iteration.
6877 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6880 /* Prevent to re-select dst_cpu via env's cpus */
6881 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6882 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6883 env->flags |= LBF_DST_PINNED;
6884 env->new_dst_cpu = cpu;
6892 /* Record that we found atleast one task that could run on dst_cpu */
6893 env->flags &= ~LBF_ALL_PINNED;
6895 if (task_running(env->src_rq, p)) {
6896 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6901 * Aggressive migration if:
6902 * 1) destination numa is preferred
6903 * 2) task is cache cold, or
6904 * 3) too many balance attempts have failed.
6906 tsk_cache_hot = migrate_degrades_locality(p, env);
6907 if (tsk_cache_hot == -1)
6908 tsk_cache_hot = task_hot(p, env);
6910 if (tsk_cache_hot <= 0 ||
6911 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6912 if (tsk_cache_hot == 1) {
6913 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6914 schedstat_inc(p->se.statistics.nr_forced_migrations);
6919 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6924 * detach_task() -- detach the task for the migration specified in env
6926 static void detach_task(struct task_struct *p, struct lb_env *env)
6928 lockdep_assert_held(&env->src_rq->lock);
6930 p->on_rq = TASK_ON_RQ_MIGRATING;
6931 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6932 set_task_cpu(p, env->dst_cpu);
6936 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6937 * part of active balancing operations within "domain".
6939 * Returns a task if successful and NULL otherwise.
6941 static struct task_struct *detach_one_task(struct lb_env *env)
6943 struct task_struct *p, *n;
6945 lockdep_assert_held(&env->src_rq->lock);
6947 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6948 if (!can_migrate_task(p, env))
6951 detach_task(p, env);
6954 * Right now, this is only the second place where
6955 * lb_gained[env->idle] is updated (other is detach_tasks)
6956 * so we can safely collect stats here rather than
6957 * inside detach_tasks().
6959 schedstat_inc(env->sd->lb_gained[env->idle]);
6965 static const unsigned int sched_nr_migrate_break = 32;
6968 * detach_tasks() -- tries to detach up to imbalance weighted load from
6969 * busiest_rq, as part of a balancing operation within domain "sd".
6971 * Returns number of detached tasks if successful and 0 otherwise.
6973 static int detach_tasks(struct lb_env *env)
6975 struct list_head *tasks = &env->src_rq->cfs_tasks;
6976 struct task_struct *p;
6980 lockdep_assert_held(&env->src_rq->lock);
6982 if (env->imbalance <= 0)
6985 while (!list_empty(tasks)) {
6987 * We don't want to steal all, otherwise we may be treated likewise,
6988 * which could at worst lead to a livelock crash.
6990 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6993 p = list_first_entry(tasks, struct task_struct, se.group_node);
6996 /* We've more or less seen every task there is, call it quits */
6997 if (env->loop > env->loop_max)
7000 /* take a breather every nr_migrate tasks */
7001 if (env->loop > env->loop_break) {
7002 env->loop_break += sched_nr_migrate_break;
7003 env->flags |= LBF_NEED_BREAK;
7007 if (!can_migrate_task(p, env))
7010 load = task_h_load(p);
7012 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7015 if ((load / 2) > env->imbalance)
7018 detach_task(p, env);
7019 list_add(&p->se.group_node, &env->tasks);
7022 env->imbalance -= load;
7024 #ifdef CONFIG_PREEMPT
7026 * NEWIDLE balancing is a source of latency, so preemptible
7027 * kernels will stop after the first task is detached to minimize
7028 * the critical section.
7030 if (env->idle == CPU_NEWLY_IDLE)
7035 * We only want to steal up to the prescribed amount of
7038 if (env->imbalance <= 0)
7043 list_move_tail(&p->se.group_node, tasks);
7047 * Right now, this is one of only two places we collect this stat
7048 * so we can safely collect detach_one_task() stats here rather
7049 * than inside detach_one_task().
7051 schedstat_add(env->sd->lb_gained[env->idle], detached);
7057 * attach_task() -- attach the task detached by detach_task() to its new rq.
7059 static void attach_task(struct rq *rq, struct task_struct *p)
7061 lockdep_assert_held(&rq->lock);
7063 BUG_ON(task_rq(p) != rq);
7064 activate_task(rq, p, ENQUEUE_NOCLOCK);
7065 p->on_rq = TASK_ON_RQ_QUEUED;
7066 check_preempt_curr(rq, p, 0);
7070 * attach_one_task() -- attaches the task returned from detach_one_task() to
7073 static void attach_one_task(struct rq *rq, struct task_struct *p)
7078 update_rq_clock(rq);
7084 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7087 static void attach_tasks(struct lb_env *env)
7089 struct list_head *tasks = &env->tasks;
7090 struct task_struct *p;
7093 rq_lock(env->dst_rq, &rf);
7094 update_rq_clock(env->dst_rq);
7096 while (!list_empty(tasks)) {
7097 p = list_first_entry(tasks, struct task_struct, se.group_node);
7098 list_del_init(&p->se.group_node);
7100 attach_task(env->dst_rq, p);
7103 rq_unlock(env->dst_rq, &rf);
7106 #ifdef CONFIG_FAIR_GROUP_SCHED
7108 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7110 if (cfs_rq->load.weight)
7113 if (cfs_rq->avg.load_sum)
7116 if (cfs_rq->avg.util_sum)
7119 if (cfs_rq->runnable_load_sum)
7125 static void update_blocked_averages(int cpu)
7127 struct rq *rq = cpu_rq(cpu);
7128 struct cfs_rq *cfs_rq, *pos;
7131 rq_lock_irqsave(rq, &rf);
7132 update_rq_clock(rq);
7135 * Iterates the task_group tree in a bottom up fashion, see
7136 * list_add_leaf_cfs_rq() for details.
7138 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7139 struct sched_entity *se;
7141 /* throttled entities do not contribute to load */
7142 if (throttled_hierarchy(cfs_rq))
7145 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7146 update_tg_load_avg(cfs_rq, 0);
7148 /* Propagate pending load changes to the parent, if any: */
7149 se = cfs_rq->tg->se[cpu];
7150 if (se && !skip_blocked_update(se))
7151 update_load_avg(cfs_rq_of(se), se, 0);
7154 * There can be a lot of idle CPU cgroups. Don't let fully
7155 * decayed cfs_rqs linger on the list.
7157 if (cfs_rq_is_decayed(cfs_rq))
7158 list_del_leaf_cfs_rq(cfs_rq);
7160 rq_unlock_irqrestore(rq, &rf);
7164 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7165 * This needs to be done in a top-down fashion because the load of a child
7166 * group is a fraction of its parents load.
7168 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7170 struct rq *rq = rq_of(cfs_rq);
7171 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7172 unsigned long now = jiffies;
7175 if (cfs_rq->last_h_load_update == now)
7178 cfs_rq->h_load_next = NULL;
7179 for_each_sched_entity(se) {
7180 cfs_rq = cfs_rq_of(se);
7181 cfs_rq->h_load_next = se;
7182 if (cfs_rq->last_h_load_update == now)
7187 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7188 cfs_rq->last_h_load_update = now;
7191 while ((se = cfs_rq->h_load_next) != NULL) {
7192 load = cfs_rq->h_load;
7193 load = div64_ul(load * se->avg.load_avg,
7194 cfs_rq_load_avg(cfs_rq) + 1);
7195 cfs_rq = group_cfs_rq(se);
7196 cfs_rq->h_load = load;
7197 cfs_rq->last_h_load_update = now;
7201 static unsigned long task_h_load(struct task_struct *p)
7203 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7205 update_cfs_rq_h_load(cfs_rq);
7206 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7207 cfs_rq_load_avg(cfs_rq) + 1);
7210 static inline void update_blocked_averages(int cpu)
7212 struct rq *rq = cpu_rq(cpu);
7213 struct cfs_rq *cfs_rq = &rq->cfs;
7216 rq_lock_irqsave(rq, &rf);
7217 update_rq_clock(rq);
7218 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7219 rq_unlock_irqrestore(rq, &rf);
7222 static unsigned long task_h_load(struct task_struct *p)
7224 return p->se.avg.load_avg;
7228 /********** Helpers for find_busiest_group ************************/
7237 * sg_lb_stats - stats of a sched_group required for load_balancing
7239 struct sg_lb_stats {
7240 unsigned long avg_load; /*Avg load across the CPUs of the group */
7241 unsigned long group_load; /* Total load over the CPUs of the group */
7242 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7243 unsigned long load_per_task;
7244 unsigned long group_capacity;
7245 unsigned long group_util; /* Total utilization of the group */
7246 unsigned int sum_nr_running; /* Nr tasks running in the group */
7247 unsigned int idle_cpus;
7248 unsigned int group_weight;
7249 enum group_type group_type;
7250 int group_no_capacity;
7251 #ifdef CONFIG_NUMA_BALANCING
7252 unsigned int nr_numa_running;
7253 unsigned int nr_preferred_running;
7258 * sd_lb_stats - Structure to store the statistics of a sched_domain
7259 * during load balancing.
7261 struct sd_lb_stats {
7262 struct sched_group *busiest; /* Busiest group in this sd */
7263 struct sched_group *local; /* Local group in this sd */
7264 unsigned long total_running;
7265 unsigned long total_load; /* Total load of all groups in sd */
7266 unsigned long total_capacity; /* Total capacity of all groups in sd */
7267 unsigned long avg_load; /* Average load across all groups in sd */
7269 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7270 struct sg_lb_stats local_stat; /* Statistics of the local group */
7273 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7276 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7277 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7278 * We must however clear busiest_stat::avg_load because
7279 * update_sd_pick_busiest() reads this before assignment.
7281 *sds = (struct sd_lb_stats){
7284 .total_running = 0UL,
7286 .total_capacity = 0UL,
7289 .sum_nr_running = 0,
7290 .group_type = group_other,
7296 * get_sd_load_idx - Obtain the load index for a given sched domain.
7297 * @sd: The sched_domain whose load_idx is to be obtained.
7298 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7300 * Return: The load index.
7302 static inline int get_sd_load_idx(struct sched_domain *sd,
7303 enum cpu_idle_type idle)
7309 load_idx = sd->busy_idx;
7312 case CPU_NEWLY_IDLE:
7313 load_idx = sd->newidle_idx;
7316 load_idx = sd->idle_idx;
7323 static unsigned long scale_rt_capacity(int cpu)
7325 struct rq *rq = cpu_rq(cpu);
7326 u64 total, used, age_stamp, avg;
7330 * Since we're reading these variables without serialization make sure
7331 * we read them once before doing sanity checks on them.
7333 age_stamp = READ_ONCE(rq->age_stamp);
7334 avg = READ_ONCE(rq->rt_avg);
7335 delta = __rq_clock_broken(rq) - age_stamp;
7337 if (unlikely(delta < 0))
7340 total = sched_avg_period() + delta;
7342 used = div_u64(avg, total);
7344 if (likely(used < SCHED_CAPACITY_SCALE))
7345 return SCHED_CAPACITY_SCALE - used;
7350 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7352 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7353 struct sched_group *sdg = sd->groups;
7355 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7357 capacity *= scale_rt_capacity(cpu);
7358 capacity >>= SCHED_CAPACITY_SHIFT;
7363 cpu_rq(cpu)->cpu_capacity = capacity;
7364 sdg->sgc->capacity = capacity;
7365 sdg->sgc->min_capacity = capacity;
7368 void update_group_capacity(struct sched_domain *sd, int cpu)
7370 struct sched_domain *child = sd->child;
7371 struct sched_group *group, *sdg = sd->groups;
7372 unsigned long capacity, min_capacity;
7373 unsigned long interval;
7375 interval = msecs_to_jiffies(sd->balance_interval);
7376 interval = clamp(interval, 1UL, max_load_balance_interval);
7377 sdg->sgc->next_update = jiffies + interval;
7380 update_cpu_capacity(sd, cpu);
7385 min_capacity = ULONG_MAX;
7387 if (child->flags & SD_OVERLAP) {
7389 * SD_OVERLAP domains cannot assume that child groups
7390 * span the current group.
7393 for_each_cpu(cpu, sched_group_span(sdg)) {
7394 struct sched_group_capacity *sgc;
7395 struct rq *rq = cpu_rq(cpu);
7398 * build_sched_domains() -> init_sched_groups_capacity()
7399 * gets here before we've attached the domains to the
7402 * Use capacity_of(), which is set irrespective of domains
7403 * in update_cpu_capacity().
7405 * This avoids capacity from being 0 and
7406 * causing divide-by-zero issues on boot.
7408 if (unlikely(!rq->sd)) {
7409 capacity += capacity_of(cpu);
7411 sgc = rq->sd->groups->sgc;
7412 capacity += sgc->capacity;
7415 min_capacity = min(capacity, min_capacity);
7419 * !SD_OVERLAP domains can assume that child groups
7420 * span the current group.
7423 group = child->groups;
7425 struct sched_group_capacity *sgc = group->sgc;
7427 capacity += sgc->capacity;
7428 min_capacity = min(sgc->min_capacity, min_capacity);
7429 group = group->next;
7430 } while (group != child->groups);
7433 sdg->sgc->capacity = capacity;
7434 sdg->sgc->min_capacity = min_capacity;
7438 * Check whether the capacity of the rq has been noticeably reduced by side
7439 * activity. The imbalance_pct is used for the threshold.
7440 * Return true is the capacity is reduced
7443 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7445 return ((rq->cpu_capacity * sd->imbalance_pct) <
7446 (rq->cpu_capacity_orig * 100));
7450 * Group imbalance indicates (and tries to solve) the problem where balancing
7451 * groups is inadequate due to ->cpus_allowed constraints.
7453 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7454 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7457 * { 0 1 2 3 } { 4 5 6 7 }
7460 * If we were to balance group-wise we'd place two tasks in the first group and
7461 * two tasks in the second group. Clearly this is undesired as it will overload
7462 * cpu 3 and leave one of the cpus in the second group unused.
7464 * The current solution to this issue is detecting the skew in the first group
7465 * by noticing the lower domain failed to reach balance and had difficulty
7466 * moving tasks due to affinity constraints.
7468 * When this is so detected; this group becomes a candidate for busiest; see
7469 * update_sd_pick_busiest(). And calculate_imbalance() and
7470 * find_busiest_group() avoid some of the usual balance conditions to allow it
7471 * to create an effective group imbalance.
7473 * This is a somewhat tricky proposition since the next run might not find the
7474 * group imbalance and decide the groups need to be balanced again. A most
7475 * subtle and fragile situation.
7478 static inline int sg_imbalanced(struct sched_group *group)
7480 return group->sgc->imbalance;
7484 * group_has_capacity returns true if the group has spare capacity that could
7485 * be used by some tasks.
7486 * We consider that a group has spare capacity if the * number of task is
7487 * smaller than the number of CPUs or if the utilization is lower than the
7488 * available capacity for CFS tasks.
7489 * For the latter, we use a threshold to stabilize the state, to take into
7490 * account the variance of the tasks' load and to return true if the available
7491 * capacity in meaningful for the load balancer.
7492 * As an example, an available capacity of 1% can appear but it doesn't make
7493 * any benefit for the load balance.
7496 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7498 if (sgs->sum_nr_running < sgs->group_weight)
7501 if ((sgs->group_capacity * 100) >
7502 (sgs->group_util * env->sd->imbalance_pct))
7509 * group_is_overloaded returns true if the group has more tasks than it can
7511 * group_is_overloaded is not equals to !group_has_capacity because a group
7512 * with the exact right number of tasks, has no more spare capacity but is not
7513 * overloaded so both group_has_capacity and group_is_overloaded return
7517 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7519 if (sgs->sum_nr_running <= sgs->group_weight)
7522 if ((sgs->group_capacity * 100) <
7523 (sgs->group_util * env->sd->imbalance_pct))
7530 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7531 * per-CPU capacity than sched_group ref.
7534 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7536 return sg->sgc->min_capacity * capacity_margin <
7537 ref->sgc->min_capacity * 1024;
7541 group_type group_classify(struct sched_group *group,
7542 struct sg_lb_stats *sgs)
7544 if (sgs->group_no_capacity)
7545 return group_overloaded;
7547 if (sg_imbalanced(group))
7548 return group_imbalanced;
7554 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7555 * @env: The load balancing environment.
7556 * @group: sched_group whose statistics are to be updated.
7557 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7558 * @local_group: Does group contain this_cpu.
7559 * @sgs: variable to hold the statistics for this group.
7560 * @overload: Indicate more than one runnable task for any CPU.
7562 static inline void update_sg_lb_stats(struct lb_env *env,
7563 struct sched_group *group, int load_idx,
7564 int local_group, struct sg_lb_stats *sgs,
7570 memset(sgs, 0, sizeof(*sgs));
7572 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7573 struct rq *rq = cpu_rq(i);
7575 /* Bias balancing toward cpus of our domain */
7577 load = target_load(i, load_idx);
7579 load = source_load(i, load_idx);
7581 sgs->group_load += load;
7582 sgs->group_util += cpu_util(i);
7583 sgs->sum_nr_running += rq->cfs.h_nr_running;
7585 nr_running = rq->nr_running;
7589 #ifdef CONFIG_NUMA_BALANCING
7590 sgs->nr_numa_running += rq->nr_numa_running;
7591 sgs->nr_preferred_running += rq->nr_preferred_running;
7593 sgs->sum_weighted_load += weighted_cpuload(rq);
7595 * No need to call idle_cpu() if nr_running is not 0
7597 if (!nr_running && idle_cpu(i))
7601 /* Adjust by relative CPU capacity of the group */
7602 sgs->group_capacity = group->sgc->capacity;
7603 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7605 if (sgs->sum_nr_running)
7606 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7608 sgs->group_weight = group->group_weight;
7610 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7611 sgs->group_type = group_classify(group, sgs);
7615 * update_sd_pick_busiest - return 1 on busiest group
7616 * @env: The load balancing environment.
7617 * @sds: sched_domain statistics
7618 * @sg: sched_group candidate to be checked for being the busiest
7619 * @sgs: sched_group statistics
7621 * Determine if @sg is a busier group than the previously selected
7624 * Return: %true if @sg is a busier group than the previously selected
7625 * busiest group. %false otherwise.
7627 static bool update_sd_pick_busiest(struct lb_env *env,
7628 struct sd_lb_stats *sds,
7629 struct sched_group *sg,
7630 struct sg_lb_stats *sgs)
7632 struct sg_lb_stats *busiest = &sds->busiest_stat;
7634 if (sgs->group_type > busiest->group_type)
7637 if (sgs->group_type < busiest->group_type)
7640 if (sgs->avg_load <= busiest->avg_load)
7643 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7647 * Candidate sg has no more than one task per CPU and
7648 * has higher per-CPU capacity. Migrating tasks to less
7649 * capable CPUs may harm throughput. Maximize throughput,
7650 * power/energy consequences are not considered.
7652 if (sgs->sum_nr_running <= sgs->group_weight &&
7653 group_smaller_cpu_capacity(sds->local, sg))
7657 /* This is the busiest node in its class. */
7658 if (!(env->sd->flags & SD_ASYM_PACKING))
7661 /* No ASYM_PACKING if target cpu is already busy */
7662 if (env->idle == CPU_NOT_IDLE)
7665 * ASYM_PACKING needs to move all the work to the highest
7666 * prority CPUs in the group, therefore mark all groups
7667 * of lower priority than ourself as busy.
7669 if (sgs->sum_nr_running &&
7670 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7674 /* Prefer to move from lowest priority cpu's work */
7675 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7676 sg->asym_prefer_cpu))
7683 #ifdef CONFIG_NUMA_BALANCING
7684 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7686 if (sgs->sum_nr_running > sgs->nr_numa_running)
7688 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7693 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7695 if (rq->nr_running > rq->nr_numa_running)
7697 if (rq->nr_running > rq->nr_preferred_running)
7702 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7707 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7711 #endif /* CONFIG_NUMA_BALANCING */
7714 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7715 * @env: The load balancing environment.
7716 * @sds: variable to hold the statistics for this sched_domain.
7718 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7720 struct sched_domain_shared *shared = env->sd->shared;
7721 struct sched_domain *child = env->sd->child;
7722 struct sched_group *sg = env->sd->groups;
7723 struct sg_lb_stats *local = &sds->local_stat;
7724 struct sg_lb_stats tmp_sgs;
7725 int load_idx, prefer_sibling = 0;
7726 bool overload = false;
7728 if (child && child->flags & SD_PREFER_SIBLING)
7731 load_idx = get_sd_load_idx(env->sd, env->idle);
7734 struct sg_lb_stats *sgs = &tmp_sgs;
7737 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7742 if (env->idle != CPU_NEWLY_IDLE ||
7743 time_after_eq(jiffies, sg->sgc->next_update))
7744 update_group_capacity(env->sd, env->dst_cpu);
7747 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7754 * In case the child domain prefers tasks go to siblings
7755 * first, lower the sg capacity so that we'll try
7756 * and move all the excess tasks away. We lower the capacity
7757 * of a group only if the local group has the capacity to fit
7758 * these excess tasks. The extra check prevents the case where
7759 * you always pull from the heaviest group when it is already
7760 * under-utilized (possible with a large weight task outweighs
7761 * the tasks on the system).
7763 if (prefer_sibling && sds->local &&
7764 group_has_capacity(env, local) &&
7765 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7766 sgs->group_no_capacity = 1;
7767 sgs->group_type = group_classify(sg, sgs);
7770 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7772 sds->busiest_stat = *sgs;
7776 /* Now, start updating sd_lb_stats */
7777 sds->total_running += sgs->sum_nr_running;
7778 sds->total_load += sgs->group_load;
7779 sds->total_capacity += sgs->group_capacity;
7782 } while (sg != env->sd->groups);
7784 if (env->sd->flags & SD_NUMA)
7785 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7787 if (!env->sd->parent) {
7788 /* update overload indicator if we are at root domain */
7789 if (env->dst_rq->rd->overload != overload)
7790 env->dst_rq->rd->overload = overload;
7797 * Since these are sums over groups they can contain some CPUs
7798 * multiple times for the NUMA domains.
7800 * Currently only wake_affine_llc() and find_busiest_group()
7801 * uses these numbers, only the last is affected by this problem.
7805 WRITE_ONCE(shared->nr_running, sds->total_running);
7806 WRITE_ONCE(shared->load, sds->total_load);
7807 WRITE_ONCE(shared->capacity, sds->total_capacity);
7811 * check_asym_packing - Check to see if the group is packed into the
7814 * This is primarily intended to used at the sibling level. Some
7815 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7816 * case of POWER7, it can move to lower SMT modes only when higher
7817 * threads are idle. When in lower SMT modes, the threads will
7818 * perform better since they share less core resources. Hence when we
7819 * have idle threads, we want them to be the higher ones.
7821 * This packing function is run on idle threads. It checks to see if
7822 * the busiest CPU in this domain (core in the P7 case) has a higher
7823 * CPU number than the packing function is being run on. Here we are
7824 * assuming lower CPU number will be equivalent to lower a SMT thread
7827 * Return: 1 when packing is required and a task should be moved to
7828 * this CPU. The amount of the imbalance is returned in env->imbalance.
7830 * @env: The load balancing environment.
7831 * @sds: Statistics of the sched_domain which is to be packed
7833 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7837 if (!(env->sd->flags & SD_ASYM_PACKING))
7840 if (env->idle == CPU_NOT_IDLE)
7846 busiest_cpu = sds->busiest->asym_prefer_cpu;
7847 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7850 env->imbalance = DIV_ROUND_CLOSEST(
7851 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7852 SCHED_CAPACITY_SCALE);
7858 * fix_small_imbalance - Calculate the minor imbalance that exists
7859 * amongst the groups of a sched_domain, during
7861 * @env: The load balancing environment.
7862 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7865 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7867 unsigned long tmp, capa_now = 0, capa_move = 0;
7868 unsigned int imbn = 2;
7869 unsigned long scaled_busy_load_per_task;
7870 struct sg_lb_stats *local, *busiest;
7872 local = &sds->local_stat;
7873 busiest = &sds->busiest_stat;
7875 if (!local->sum_nr_running)
7876 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7877 else if (busiest->load_per_task > local->load_per_task)
7880 scaled_busy_load_per_task =
7881 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7882 busiest->group_capacity;
7884 if (busiest->avg_load + scaled_busy_load_per_task >=
7885 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7886 env->imbalance = busiest->load_per_task;
7891 * OK, we don't have enough imbalance to justify moving tasks,
7892 * however we may be able to increase total CPU capacity used by
7896 capa_now += busiest->group_capacity *
7897 min(busiest->load_per_task, busiest->avg_load);
7898 capa_now += local->group_capacity *
7899 min(local->load_per_task, local->avg_load);
7900 capa_now /= SCHED_CAPACITY_SCALE;
7902 /* Amount of load we'd subtract */
7903 if (busiest->avg_load > scaled_busy_load_per_task) {
7904 capa_move += busiest->group_capacity *
7905 min(busiest->load_per_task,
7906 busiest->avg_load - scaled_busy_load_per_task);
7909 /* Amount of load we'd add */
7910 if (busiest->avg_load * busiest->group_capacity <
7911 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7912 tmp = (busiest->avg_load * busiest->group_capacity) /
7913 local->group_capacity;
7915 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7916 local->group_capacity;
7918 capa_move += local->group_capacity *
7919 min(local->load_per_task, local->avg_load + tmp);
7920 capa_move /= SCHED_CAPACITY_SCALE;
7922 /* Move if we gain throughput */
7923 if (capa_move > capa_now)
7924 env->imbalance = busiest->load_per_task;
7928 * calculate_imbalance - Calculate the amount of imbalance present within the
7929 * groups of a given sched_domain during load balance.
7930 * @env: load balance environment
7931 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7933 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7935 unsigned long max_pull, load_above_capacity = ~0UL;
7936 struct sg_lb_stats *local, *busiest;
7938 local = &sds->local_stat;
7939 busiest = &sds->busiest_stat;
7941 if (busiest->group_type == group_imbalanced) {
7943 * In the group_imb case we cannot rely on group-wide averages
7944 * to ensure cpu-load equilibrium, look at wider averages. XXX
7946 busiest->load_per_task =
7947 min(busiest->load_per_task, sds->avg_load);
7951 * Avg load of busiest sg can be less and avg load of local sg can
7952 * be greater than avg load across all sgs of sd because avg load
7953 * factors in sg capacity and sgs with smaller group_type are
7954 * skipped when updating the busiest sg:
7956 if (busiest->avg_load <= sds->avg_load ||
7957 local->avg_load >= sds->avg_load) {
7959 return fix_small_imbalance(env, sds);
7963 * If there aren't any idle cpus, avoid creating some.
7965 if (busiest->group_type == group_overloaded &&
7966 local->group_type == group_overloaded) {
7967 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7968 if (load_above_capacity > busiest->group_capacity) {
7969 load_above_capacity -= busiest->group_capacity;
7970 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7971 load_above_capacity /= busiest->group_capacity;
7973 load_above_capacity = ~0UL;
7977 * We're trying to get all the cpus to the average_load, so we don't
7978 * want to push ourselves above the average load, nor do we wish to
7979 * reduce the max loaded cpu below the average load. At the same time,
7980 * we also don't want to reduce the group load below the group
7981 * capacity. Thus we look for the minimum possible imbalance.
7983 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7985 /* How much load to actually move to equalise the imbalance */
7986 env->imbalance = min(
7987 max_pull * busiest->group_capacity,
7988 (sds->avg_load - local->avg_load) * local->group_capacity
7989 ) / SCHED_CAPACITY_SCALE;
7992 * if *imbalance is less than the average load per runnable task
7993 * there is no guarantee that any tasks will be moved so we'll have
7994 * a think about bumping its value to force at least one task to be
7997 if (env->imbalance < busiest->load_per_task)
7998 return fix_small_imbalance(env, sds);
8001 /******* find_busiest_group() helpers end here *********************/
8004 * find_busiest_group - Returns the busiest group within the sched_domain
8005 * if there is an imbalance.
8007 * Also calculates the amount of weighted load which should be moved
8008 * to restore balance.
8010 * @env: The load balancing environment.
8012 * Return: - The busiest group if imbalance exists.
8014 static struct sched_group *find_busiest_group(struct lb_env *env)
8016 struct sg_lb_stats *local, *busiest;
8017 struct sd_lb_stats sds;
8019 init_sd_lb_stats(&sds);
8022 * Compute the various statistics relavent for load balancing at
8025 update_sd_lb_stats(env, &sds);
8026 local = &sds.local_stat;
8027 busiest = &sds.busiest_stat;
8029 /* ASYM feature bypasses nice load balance check */
8030 if (check_asym_packing(env, &sds))
8033 /* There is no busy sibling group to pull tasks from */
8034 if (!sds.busiest || busiest->sum_nr_running == 0)
8037 /* XXX broken for overlapping NUMA groups */
8038 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8039 / sds.total_capacity;
8042 * If the busiest group is imbalanced the below checks don't
8043 * work because they assume all things are equal, which typically
8044 * isn't true due to cpus_allowed constraints and the like.
8046 if (busiest->group_type == group_imbalanced)
8049 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8050 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8051 busiest->group_no_capacity)
8055 * If the local group is busier than the selected busiest group
8056 * don't try and pull any tasks.
8058 if (local->avg_load >= busiest->avg_load)
8062 * Don't pull any tasks if this group is already above the domain
8065 if (local->avg_load >= sds.avg_load)
8068 if (env->idle == CPU_IDLE) {
8070 * This cpu is idle. If the busiest group is not overloaded
8071 * and there is no imbalance between this and busiest group
8072 * wrt idle cpus, it is balanced. The imbalance becomes
8073 * significant if the diff is greater than 1 otherwise we
8074 * might end up to just move the imbalance on another group
8076 if ((busiest->group_type != group_overloaded) &&
8077 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8081 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8082 * imbalance_pct to be conservative.
8084 if (100 * busiest->avg_load <=
8085 env->sd->imbalance_pct * local->avg_load)
8090 /* Looks like there is an imbalance. Compute it */
8091 calculate_imbalance(env, &sds);
8100 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8102 static struct rq *find_busiest_queue(struct lb_env *env,
8103 struct sched_group *group)
8105 struct rq *busiest = NULL, *rq;
8106 unsigned long busiest_load = 0, busiest_capacity = 1;
8109 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8110 unsigned long capacity, wl;
8114 rt = fbq_classify_rq(rq);
8117 * We classify groups/runqueues into three groups:
8118 * - regular: there are !numa tasks
8119 * - remote: there are numa tasks that run on the 'wrong' node
8120 * - all: there is no distinction
8122 * In order to avoid migrating ideally placed numa tasks,
8123 * ignore those when there's better options.
8125 * If we ignore the actual busiest queue to migrate another
8126 * task, the next balance pass can still reduce the busiest
8127 * queue by moving tasks around inside the node.
8129 * If we cannot move enough load due to this classification
8130 * the next pass will adjust the group classification and
8131 * allow migration of more tasks.
8133 * Both cases only affect the total convergence complexity.
8135 if (rt > env->fbq_type)
8138 capacity = capacity_of(i);
8140 wl = weighted_cpuload(rq);
8143 * When comparing with imbalance, use weighted_cpuload()
8144 * which is not scaled with the cpu capacity.
8147 if (rq->nr_running == 1 && wl > env->imbalance &&
8148 !check_cpu_capacity(rq, env->sd))
8152 * For the load comparisons with the other cpu's, consider
8153 * the weighted_cpuload() scaled with the cpu capacity, so
8154 * that the load can be moved away from the cpu that is
8155 * potentially running at a lower capacity.
8157 * Thus we're looking for max(wl_i / capacity_i), crosswise
8158 * multiplication to rid ourselves of the division works out
8159 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8160 * our previous maximum.
8162 if (wl * busiest_capacity > busiest_load * capacity) {
8164 busiest_capacity = capacity;
8173 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8174 * so long as it is large enough.
8176 #define MAX_PINNED_INTERVAL 512
8178 static int need_active_balance(struct lb_env *env)
8180 struct sched_domain *sd = env->sd;
8182 if (env->idle == CPU_NEWLY_IDLE) {
8185 * ASYM_PACKING needs to force migrate tasks from busy but
8186 * lower priority CPUs in order to pack all tasks in the
8187 * highest priority CPUs.
8189 if ((sd->flags & SD_ASYM_PACKING) &&
8190 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8195 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8196 * It's worth migrating the task if the src_cpu's capacity is reduced
8197 * because of other sched_class or IRQs if more capacity stays
8198 * available on dst_cpu.
8200 if ((env->idle != CPU_NOT_IDLE) &&
8201 (env->src_rq->cfs.h_nr_running == 1)) {
8202 if ((check_cpu_capacity(env->src_rq, sd)) &&
8203 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8207 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8210 static int active_load_balance_cpu_stop(void *data);
8212 static int should_we_balance(struct lb_env *env)
8214 struct sched_group *sg = env->sd->groups;
8215 int cpu, balance_cpu = -1;
8218 * In the newly idle case, we will allow all the cpu's
8219 * to do the newly idle load balance.
8221 if (env->idle == CPU_NEWLY_IDLE)
8224 /* Try to find first idle cpu */
8225 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8233 if (balance_cpu == -1)
8234 balance_cpu = group_balance_cpu(sg);
8237 * First idle cpu or the first cpu(busiest) in this sched group
8238 * is eligible for doing load balancing at this and above domains.
8240 return balance_cpu == env->dst_cpu;
8244 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8245 * tasks if there is an imbalance.
8247 static int load_balance(int this_cpu, struct rq *this_rq,
8248 struct sched_domain *sd, enum cpu_idle_type idle,
8249 int *continue_balancing)
8251 int ld_moved, cur_ld_moved, active_balance = 0;
8252 struct sched_domain *sd_parent = sd->parent;
8253 struct sched_group *group;
8256 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8258 struct lb_env env = {
8260 .dst_cpu = this_cpu,
8262 .dst_grpmask = sched_group_span(sd->groups),
8264 .loop_break = sched_nr_migrate_break,
8267 .tasks = LIST_HEAD_INIT(env.tasks),
8270 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8272 schedstat_inc(sd->lb_count[idle]);
8275 if (!should_we_balance(&env)) {
8276 *continue_balancing = 0;
8280 group = find_busiest_group(&env);
8282 schedstat_inc(sd->lb_nobusyg[idle]);
8286 busiest = find_busiest_queue(&env, group);
8288 schedstat_inc(sd->lb_nobusyq[idle]);
8292 BUG_ON(busiest == env.dst_rq);
8294 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8296 env.src_cpu = busiest->cpu;
8297 env.src_rq = busiest;
8300 if (busiest->nr_running > 1) {
8302 * Attempt to move tasks. If find_busiest_group has found
8303 * an imbalance but busiest->nr_running <= 1, the group is
8304 * still unbalanced. ld_moved simply stays zero, so it is
8305 * correctly treated as an imbalance.
8307 env.flags |= LBF_ALL_PINNED;
8308 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8311 rq_lock_irqsave(busiest, &rf);
8312 update_rq_clock(busiest);
8315 * cur_ld_moved - load moved in current iteration
8316 * ld_moved - cumulative load moved across iterations
8318 cur_ld_moved = detach_tasks(&env);
8321 * We've detached some tasks from busiest_rq. Every
8322 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8323 * unlock busiest->lock, and we are able to be sure
8324 * that nobody can manipulate the tasks in parallel.
8325 * See task_rq_lock() family for the details.
8328 rq_unlock(busiest, &rf);
8332 ld_moved += cur_ld_moved;
8335 local_irq_restore(rf.flags);
8337 if (env.flags & LBF_NEED_BREAK) {
8338 env.flags &= ~LBF_NEED_BREAK;
8343 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8344 * us and move them to an alternate dst_cpu in our sched_group
8345 * where they can run. The upper limit on how many times we
8346 * iterate on same src_cpu is dependent on number of cpus in our
8349 * This changes load balance semantics a bit on who can move
8350 * load to a given_cpu. In addition to the given_cpu itself
8351 * (or a ilb_cpu acting on its behalf where given_cpu is
8352 * nohz-idle), we now have balance_cpu in a position to move
8353 * load to given_cpu. In rare situations, this may cause
8354 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8355 * _independently_ and at _same_ time to move some load to
8356 * given_cpu) causing exceess load to be moved to given_cpu.
8357 * This however should not happen so much in practice and
8358 * moreover subsequent load balance cycles should correct the
8359 * excess load moved.
8361 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8363 /* Prevent to re-select dst_cpu via env's cpus */
8364 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8366 env.dst_rq = cpu_rq(env.new_dst_cpu);
8367 env.dst_cpu = env.new_dst_cpu;
8368 env.flags &= ~LBF_DST_PINNED;
8370 env.loop_break = sched_nr_migrate_break;
8373 * Go back to "more_balance" rather than "redo" since we
8374 * need to continue with same src_cpu.
8380 * We failed to reach balance because of affinity.
8383 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8385 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8386 *group_imbalance = 1;
8389 /* All tasks on this runqueue were pinned by CPU affinity */
8390 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8391 cpumask_clear_cpu(cpu_of(busiest), cpus);
8393 * Attempting to continue load balancing at the current
8394 * sched_domain level only makes sense if there are
8395 * active CPUs remaining as possible busiest CPUs to
8396 * pull load from which are not contained within the
8397 * destination group that is receiving any migrated
8400 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8402 env.loop_break = sched_nr_migrate_break;
8405 goto out_all_pinned;
8410 schedstat_inc(sd->lb_failed[idle]);
8412 * Increment the failure counter only on periodic balance.
8413 * We do not want newidle balance, which can be very
8414 * frequent, pollute the failure counter causing
8415 * excessive cache_hot migrations and active balances.
8417 if (idle != CPU_NEWLY_IDLE)
8418 sd->nr_balance_failed++;
8420 if (need_active_balance(&env)) {
8421 unsigned long flags;
8423 raw_spin_lock_irqsave(&busiest->lock, flags);
8425 /* don't kick the active_load_balance_cpu_stop,
8426 * if the curr task on busiest cpu can't be
8429 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8430 raw_spin_unlock_irqrestore(&busiest->lock,
8432 env.flags |= LBF_ALL_PINNED;
8433 goto out_one_pinned;
8437 * ->active_balance synchronizes accesses to
8438 * ->active_balance_work. Once set, it's cleared
8439 * only after active load balance is finished.
8441 if (!busiest->active_balance) {
8442 busiest->active_balance = 1;
8443 busiest->push_cpu = this_cpu;
8446 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8448 if (active_balance) {
8449 stop_one_cpu_nowait(cpu_of(busiest),
8450 active_load_balance_cpu_stop, busiest,
8451 &busiest->active_balance_work);
8454 /* We've kicked active balancing, force task migration. */
8455 sd->nr_balance_failed = sd->cache_nice_tries+1;
8458 sd->nr_balance_failed = 0;
8460 if (likely(!active_balance)) {
8461 /* We were unbalanced, so reset the balancing interval */
8462 sd->balance_interval = sd->min_interval;
8465 * If we've begun active balancing, start to back off. This
8466 * case may not be covered by the all_pinned logic if there
8467 * is only 1 task on the busy runqueue (because we don't call
8470 if (sd->balance_interval < sd->max_interval)
8471 sd->balance_interval *= 2;
8478 * We reach balance although we may have faced some affinity
8479 * constraints. Clear the imbalance flag if it was set.
8482 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8484 if (*group_imbalance)
8485 *group_imbalance = 0;
8490 * We reach balance because all tasks are pinned at this level so
8491 * we can't migrate them. Let the imbalance flag set so parent level
8492 * can try to migrate them.
8494 schedstat_inc(sd->lb_balanced[idle]);
8496 sd->nr_balance_failed = 0;
8499 /* tune up the balancing interval */
8500 if (((env.flags & LBF_ALL_PINNED) &&
8501 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8502 (sd->balance_interval < sd->max_interval))
8503 sd->balance_interval *= 2;
8510 static inline unsigned long
8511 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8513 unsigned long interval = sd->balance_interval;
8516 interval *= sd->busy_factor;
8518 /* scale ms to jiffies */
8519 interval = msecs_to_jiffies(interval);
8520 interval = clamp(interval, 1UL, max_load_balance_interval);
8526 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8528 unsigned long interval, next;
8530 /* used by idle balance, so cpu_busy = 0 */
8531 interval = get_sd_balance_interval(sd, 0);
8532 next = sd->last_balance + interval;
8534 if (time_after(*next_balance, next))
8535 *next_balance = next;
8539 * idle_balance is called by schedule() if this_cpu is about to become
8540 * idle. Attempts to pull tasks from other CPUs.
8542 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8544 unsigned long next_balance = jiffies + HZ;
8545 int this_cpu = this_rq->cpu;
8546 struct sched_domain *sd;
8547 int pulled_task = 0;
8551 * We must set idle_stamp _before_ calling idle_balance(), such that we
8552 * measure the duration of idle_balance() as idle time.
8554 this_rq->idle_stamp = rq_clock(this_rq);
8557 * Do not pull tasks towards !active CPUs...
8559 if (!cpu_active(this_cpu))
8563 * This is OK, because current is on_cpu, which avoids it being picked
8564 * for load-balance and preemption/IRQs are still disabled avoiding
8565 * further scheduler activity on it and we're being very careful to
8566 * re-start the picking loop.
8568 rq_unpin_lock(this_rq, rf);
8570 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8571 !this_rq->rd->overload) {
8573 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8575 update_next_balance(sd, &next_balance);
8581 raw_spin_unlock(&this_rq->lock);
8583 update_blocked_averages(this_cpu);
8585 for_each_domain(this_cpu, sd) {
8586 int continue_balancing = 1;
8587 u64 t0, domain_cost;
8589 if (!(sd->flags & SD_LOAD_BALANCE))
8592 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8593 update_next_balance(sd, &next_balance);
8597 if (sd->flags & SD_BALANCE_NEWIDLE) {
8598 t0 = sched_clock_cpu(this_cpu);
8600 pulled_task = load_balance(this_cpu, this_rq,
8602 &continue_balancing);
8604 domain_cost = sched_clock_cpu(this_cpu) - t0;
8605 if (domain_cost > sd->max_newidle_lb_cost)
8606 sd->max_newidle_lb_cost = domain_cost;
8608 curr_cost += domain_cost;
8611 update_next_balance(sd, &next_balance);
8614 * Stop searching for tasks to pull if there are
8615 * now runnable tasks on this rq.
8617 if (pulled_task || this_rq->nr_running > 0)
8622 raw_spin_lock(&this_rq->lock);
8624 if (curr_cost > this_rq->max_idle_balance_cost)
8625 this_rq->max_idle_balance_cost = curr_cost;
8628 * While browsing the domains, we released the rq lock, a task could
8629 * have been enqueued in the meantime. Since we're not going idle,
8630 * pretend we pulled a task.
8632 if (this_rq->cfs.h_nr_running && !pulled_task)
8636 /* Move the next balance forward */
8637 if (time_after(this_rq->next_balance, next_balance))
8638 this_rq->next_balance = next_balance;
8640 /* Is there a task of a high priority class? */
8641 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8645 this_rq->idle_stamp = 0;
8647 rq_repin_lock(this_rq, rf);
8653 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8654 * running tasks off the busiest CPU onto idle CPUs. It requires at
8655 * least 1 task to be running on each physical CPU where possible, and
8656 * avoids physical / logical imbalances.
8658 static int active_load_balance_cpu_stop(void *data)
8660 struct rq *busiest_rq = data;
8661 int busiest_cpu = cpu_of(busiest_rq);
8662 int target_cpu = busiest_rq->push_cpu;
8663 struct rq *target_rq = cpu_rq(target_cpu);
8664 struct sched_domain *sd;
8665 struct task_struct *p = NULL;
8668 rq_lock_irq(busiest_rq, &rf);
8670 * Between queueing the stop-work and running it is a hole in which
8671 * CPUs can become inactive. We should not move tasks from or to
8674 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8677 /* make sure the requested cpu hasn't gone down in the meantime */
8678 if (unlikely(busiest_cpu != smp_processor_id() ||
8679 !busiest_rq->active_balance))
8682 /* Is there any task to move? */
8683 if (busiest_rq->nr_running <= 1)
8687 * This condition is "impossible", if it occurs
8688 * we need to fix it. Originally reported by
8689 * Bjorn Helgaas on a 128-cpu setup.
8691 BUG_ON(busiest_rq == target_rq);
8693 /* Search for an sd spanning us and the target CPU. */
8695 for_each_domain(target_cpu, sd) {
8696 if ((sd->flags & SD_LOAD_BALANCE) &&
8697 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8702 struct lb_env env = {
8704 .dst_cpu = target_cpu,
8705 .dst_rq = target_rq,
8706 .src_cpu = busiest_rq->cpu,
8707 .src_rq = busiest_rq,
8710 * can_migrate_task() doesn't need to compute new_dst_cpu
8711 * for active balancing. Since we have CPU_IDLE, but no
8712 * @dst_grpmask we need to make that test go away with lying
8715 .flags = LBF_DST_PINNED,
8718 schedstat_inc(sd->alb_count);
8719 update_rq_clock(busiest_rq);
8721 p = detach_one_task(&env);
8723 schedstat_inc(sd->alb_pushed);
8724 /* Active balancing done, reset the failure counter. */
8725 sd->nr_balance_failed = 0;
8727 schedstat_inc(sd->alb_failed);
8732 busiest_rq->active_balance = 0;
8733 rq_unlock(busiest_rq, &rf);
8736 attach_one_task(target_rq, p);
8743 static inline int on_null_domain(struct rq *rq)
8745 return unlikely(!rcu_dereference_sched(rq->sd));
8748 #ifdef CONFIG_NO_HZ_COMMON
8750 * idle load balancing details
8751 * - When one of the busy CPUs notice that there may be an idle rebalancing
8752 * needed, they will kick the idle load balancer, which then does idle
8753 * load balancing for all the idle CPUs.
8756 cpumask_var_t idle_cpus_mask;
8758 unsigned long next_balance; /* in jiffy units */
8759 } nohz ____cacheline_aligned;
8761 static inline int find_new_ilb(void)
8763 int ilb = cpumask_first(nohz.idle_cpus_mask);
8765 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8772 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8773 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8774 * CPU (if there is one).
8776 static void nohz_balancer_kick(void)
8780 nohz.next_balance++;
8782 ilb_cpu = find_new_ilb();
8784 if (ilb_cpu >= nr_cpu_ids)
8787 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8790 * Use smp_send_reschedule() instead of resched_cpu().
8791 * This way we generate a sched IPI on the target cpu which
8792 * is idle. And the softirq performing nohz idle load balance
8793 * will be run before returning from the IPI.
8795 smp_send_reschedule(ilb_cpu);
8799 void nohz_balance_exit_idle(unsigned int cpu)
8801 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8803 * Completely isolated CPUs don't ever set, so we must test.
8805 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8806 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8807 atomic_dec(&nohz.nr_cpus);
8809 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8813 static inline void set_cpu_sd_state_busy(void)
8815 struct sched_domain *sd;
8816 int cpu = smp_processor_id();
8819 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8821 if (!sd || !sd->nohz_idle)
8825 atomic_inc(&sd->shared->nr_busy_cpus);
8830 void set_cpu_sd_state_idle(void)
8832 struct sched_domain *sd;
8833 int cpu = smp_processor_id();
8836 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8838 if (!sd || sd->nohz_idle)
8842 atomic_dec(&sd->shared->nr_busy_cpus);
8848 * This routine will record that the cpu is going idle with tick stopped.
8849 * This info will be used in performing idle load balancing in the future.
8851 void nohz_balance_enter_idle(int cpu)
8854 * If this cpu is going down, then nothing needs to be done.
8856 if (!cpu_active(cpu))
8859 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8860 if (!is_housekeeping_cpu(cpu))
8863 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8867 * If we're a completely isolated CPU, we don't play.
8869 if (on_null_domain(cpu_rq(cpu)))
8872 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8873 atomic_inc(&nohz.nr_cpus);
8874 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8878 static DEFINE_SPINLOCK(balancing);
8881 * Scale the max load_balance interval with the number of CPUs in the system.
8882 * This trades load-balance latency on larger machines for less cross talk.
8884 void update_max_interval(void)
8886 max_load_balance_interval = HZ*num_online_cpus()/10;
8890 * It checks each scheduling domain to see if it is due to be balanced,
8891 * and initiates a balancing operation if so.
8893 * Balancing parameters are set up in init_sched_domains.
8895 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8897 int continue_balancing = 1;
8899 unsigned long interval;
8900 struct sched_domain *sd;
8901 /* Earliest time when we have to do rebalance again */
8902 unsigned long next_balance = jiffies + 60*HZ;
8903 int update_next_balance = 0;
8904 int need_serialize, need_decay = 0;
8907 update_blocked_averages(cpu);
8910 for_each_domain(cpu, sd) {
8912 * Decay the newidle max times here because this is a regular
8913 * visit to all the domains. Decay ~1% per second.
8915 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8916 sd->max_newidle_lb_cost =
8917 (sd->max_newidle_lb_cost * 253) / 256;
8918 sd->next_decay_max_lb_cost = jiffies + HZ;
8921 max_cost += sd->max_newidle_lb_cost;
8923 if (!(sd->flags & SD_LOAD_BALANCE))
8927 * Stop the load balance at this level. There is another
8928 * CPU in our sched group which is doing load balancing more
8931 if (!continue_balancing) {
8937 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8939 need_serialize = sd->flags & SD_SERIALIZE;
8940 if (need_serialize) {
8941 if (!spin_trylock(&balancing))
8945 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8946 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8948 * The LBF_DST_PINNED logic could have changed
8949 * env->dst_cpu, so we can't know our idle
8950 * state even if we migrated tasks. Update it.
8952 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8954 sd->last_balance = jiffies;
8955 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8958 spin_unlock(&balancing);
8960 if (time_after(next_balance, sd->last_balance + interval)) {
8961 next_balance = sd->last_balance + interval;
8962 update_next_balance = 1;
8967 * Ensure the rq-wide value also decays but keep it at a
8968 * reasonable floor to avoid funnies with rq->avg_idle.
8970 rq->max_idle_balance_cost =
8971 max((u64)sysctl_sched_migration_cost, max_cost);
8976 * next_balance will be updated only when there is a need.
8977 * When the cpu is attached to null domain for ex, it will not be
8980 if (likely(update_next_balance)) {
8981 rq->next_balance = next_balance;
8983 #ifdef CONFIG_NO_HZ_COMMON
8985 * If this CPU has been elected to perform the nohz idle
8986 * balance. Other idle CPUs have already rebalanced with
8987 * nohz_idle_balance() and nohz.next_balance has been
8988 * updated accordingly. This CPU is now running the idle load
8989 * balance for itself and we need to update the
8990 * nohz.next_balance accordingly.
8992 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8993 nohz.next_balance = rq->next_balance;
8998 #ifdef CONFIG_NO_HZ_COMMON
9000 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9001 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9003 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9005 int this_cpu = this_rq->cpu;
9008 /* Earliest time when we have to do rebalance again */
9009 unsigned long next_balance = jiffies + 60*HZ;
9010 int update_next_balance = 0;
9012 if (idle != CPU_IDLE ||
9013 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9016 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9017 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9021 * If this cpu gets work to do, stop the load balancing
9022 * work being done for other cpus. Next load
9023 * balancing owner will pick it up.
9028 rq = cpu_rq(balance_cpu);
9031 * If time for next balance is due,
9034 if (time_after_eq(jiffies, rq->next_balance)) {
9037 rq_lock_irq(rq, &rf);
9038 update_rq_clock(rq);
9039 cpu_load_update_idle(rq);
9040 rq_unlock_irq(rq, &rf);
9042 rebalance_domains(rq, CPU_IDLE);
9045 if (time_after(next_balance, rq->next_balance)) {
9046 next_balance = rq->next_balance;
9047 update_next_balance = 1;
9052 * next_balance will be updated only when there is a need.
9053 * When the CPU is attached to null domain for ex, it will not be
9056 if (likely(update_next_balance))
9057 nohz.next_balance = next_balance;
9059 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9063 * Current heuristic for kicking the idle load balancer in the presence
9064 * of an idle cpu in the system.
9065 * - This rq has more than one task.
9066 * - This rq has at least one CFS task and the capacity of the CPU is
9067 * significantly reduced because of RT tasks or IRQs.
9068 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9069 * multiple busy cpu.
9070 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9071 * domain span are idle.
9073 static inline bool nohz_kick_needed(struct rq *rq)
9075 unsigned long now = jiffies;
9076 struct sched_domain_shared *sds;
9077 struct sched_domain *sd;
9078 int nr_busy, i, cpu = rq->cpu;
9081 if (unlikely(rq->idle_balance))
9085 * We may be recently in ticked or tickless idle mode. At the first
9086 * busy tick after returning from idle, we will update the busy stats.
9088 set_cpu_sd_state_busy();
9089 nohz_balance_exit_idle(cpu);
9092 * None are in tickless mode and hence no need for NOHZ idle load
9095 if (likely(!atomic_read(&nohz.nr_cpus)))
9098 if (time_before(now, nohz.next_balance))
9101 if (rq->nr_running >= 2)
9105 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9108 * XXX: write a coherent comment on why we do this.
9109 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9111 nr_busy = atomic_read(&sds->nr_busy_cpus);
9119 sd = rcu_dereference(rq->sd);
9121 if ((rq->cfs.h_nr_running >= 1) &&
9122 check_cpu_capacity(rq, sd)) {
9128 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9130 for_each_cpu(i, sched_domain_span(sd)) {
9132 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9135 if (sched_asym_prefer(i, cpu)) {
9146 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9150 * run_rebalance_domains is triggered when needed from the scheduler tick.
9151 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9153 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9155 struct rq *this_rq = this_rq();
9156 enum cpu_idle_type idle = this_rq->idle_balance ?
9157 CPU_IDLE : CPU_NOT_IDLE;
9160 * If this cpu has a pending nohz_balance_kick, then do the
9161 * balancing on behalf of the other idle cpus whose ticks are
9162 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9163 * give the idle cpus a chance to load balance. Else we may
9164 * load balance only within the local sched_domain hierarchy
9165 * and abort nohz_idle_balance altogether if we pull some load.
9167 nohz_idle_balance(this_rq, idle);
9168 rebalance_domains(this_rq, idle);
9172 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9174 void trigger_load_balance(struct rq *rq)
9176 /* Don't need to rebalance while attached to NULL domain */
9177 if (unlikely(on_null_domain(rq)))
9180 if (time_after_eq(jiffies, rq->next_balance))
9181 raise_softirq(SCHED_SOFTIRQ);
9182 #ifdef CONFIG_NO_HZ_COMMON
9183 if (nohz_kick_needed(rq))
9184 nohz_balancer_kick();
9188 static void rq_online_fair(struct rq *rq)
9192 update_runtime_enabled(rq);
9195 static void rq_offline_fair(struct rq *rq)
9199 /* Ensure any throttled groups are reachable by pick_next_task */
9200 unthrottle_offline_cfs_rqs(rq);
9203 #endif /* CONFIG_SMP */
9206 * scheduler tick hitting a task of our scheduling class:
9208 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9210 struct cfs_rq *cfs_rq;
9211 struct sched_entity *se = &curr->se;
9213 for_each_sched_entity(se) {
9214 cfs_rq = cfs_rq_of(se);
9215 entity_tick(cfs_rq, se, queued);
9218 if (static_branch_unlikely(&sched_numa_balancing))
9219 task_tick_numa(rq, curr);
9223 * called on fork with the child task as argument from the parent's context
9224 * - child not yet on the tasklist
9225 * - preemption disabled
9227 static void task_fork_fair(struct task_struct *p)
9229 struct cfs_rq *cfs_rq;
9230 struct sched_entity *se = &p->se, *curr;
9231 struct rq *rq = this_rq();
9235 update_rq_clock(rq);
9237 cfs_rq = task_cfs_rq(current);
9238 curr = cfs_rq->curr;
9240 update_curr(cfs_rq);
9241 se->vruntime = curr->vruntime;
9243 place_entity(cfs_rq, se, 1);
9245 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9247 * Upon rescheduling, sched_class::put_prev_task() will place
9248 * 'current' within the tree based on its new key value.
9250 swap(curr->vruntime, se->vruntime);
9254 se->vruntime -= cfs_rq->min_vruntime;
9259 * Priority of the task has changed. Check to see if we preempt
9263 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9265 if (!task_on_rq_queued(p))
9269 * Reschedule if we are currently running on this runqueue and
9270 * our priority decreased, or if we are not currently running on
9271 * this runqueue and our priority is higher than the current's
9273 if (rq->curr == p) {
9274 if (p->prio > oldprio)
9277 check_preempt_curr(rq, p, 0);
9280 static inline bool vruntime_normalized(struct task_struct *p)
9282 struct sched_entity *se = &p->se;
9285 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9286 * the dequeue_entity(.flags=0) will already have normalized the
9293 * When !on_rq, vruntime of the task has usually NOT been normalized.
9294 * But there are some cases where it has already been normalized:
9296 * - A forked child which is waiting for being woken up by
9297 * wake_up_new_task().
9298 * - A task which has been woken up by try_to_wake_up() and
9299 * waiting for actually being woken up by sched_ttwu_pending().
9301 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9307 #ifdef CONFIG_FAIR_GROUP_SCHED
9309 * Propagate the changes of the sched_entity across the tg tree to make it
9310 * visible to the root
9312 static void propagate_entity_cfs_rq(struct sched_entity *se)
9314 struct cfs_rq *cfs_rq;
9316 /* Start to propagate at parent */
9319 for_each_sched_entity(se) {
9320 cfs_rq = cfs_rq_of(se);
9322 if (cfs_rq_throttled(cfs_rq))
9325 update_load_avg(cfs_rq, se, UPDATE_TG);
9329 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9332 static void detach_entity_cfs_rq(struct sched_entity *se)
9334 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9336 /* Catch up with the cfs_rq and remove our load when we leave */
9337 update_load_avg(cfs_rq, se, 0);
9338 detach_entity_load_avg(cfs_rq, se);
9339 update_tg_load_avg(cfs_rq, false);
9340 propagate_entity_cfs_rq(se);
9343 static void attach_entity_cfs_rq(struct sched_entity *se)
9345 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9347 #ifdef CONFIG_FAIR_GROUP_SCHED
9349 * Since the real-depth could have been changed (only FAIR
9350 * class maintain depth value), reset depth properly.
9352 se->depth = se->parent ? se->parent->depth + 1 : 0;
9355 /* Synchronize entity with its cfs_rq */
9356 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9357 attach_entity_load_avg(cfs_rq, se);
9358 update_tg_load_avg(cfs_rq, false);
9359 propagate_entity_cfs_rq(se);
9362 static void detach_task_cfs_rq(struct task_struct *p)
9364 struct sched_entity *se = &p->se;
9365 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9367 if (!vruntime_normalized(p)) {
9369 * Fix up our vruntime so that the current sleep doesn't
9370 * cause 'unlimited' sleep bonus.
9372 place_entity(cfs_rq, se, 0);
9373 se->vruntime -= cfs_rq->min_vruntime;
9376 detach_entity_cfs_rq(se);
9379 static void attach_task_cfs_rq(struct task_struct *p)
9381 struct sched_entity *se = &p->se;
9382 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9384 attach_entity_cfs_rq(se);
9386 if (!vruntime_normalized(p))
9387 se->vruntime += cfs_rq->min_vruntime;
9390 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9392 detach_task_cfs_rq(p);
9395 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9397 attach_task_cfs_rq(p);
9399 if (task_on_rq_queued(p)) {
9401 * We were most likely switched from sched_rt, so
9402 * kick off the schedule if running, otherwise just see
9403 * if we can still preempt the current task.
9408 check_preempt_curr(rq, p, 0);
9412 /* Account for a task changing its policy or group.
9414 * This routine is mostly called to set cfs_rq->curr field when a task
9415 * migrates between groups/classes.
9417 static void set_curr_task_fair(struct rq *rq)
9419 struct sched_entity *se = &rq->curr->se;
9421 for_each_sched_entity(se) {
9422 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9424 set_next_entity(cfs_rq, se);
9425 /* ensure bandwidth has been allocated on our new cfs_rq */
9426 account_cfs_rq_runtime(cfs_rq, 0);
9430 void init_cfs_rq(struct cfs_rq *cfs_rq)
9432 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9433 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9434 #ifndef CONFIG_64BIT
9435 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9438 #ifdef CONFIG_FAIR_GROUP_SCHED
9439 cfs_rq->propagate_avg = 0;
9441 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9442 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9446 #ifdef CONFIG_FAIR_GROUP_SCHED
9447 static void task_set_group_fair(struct task_struct *p)
9449 struct sched_entity *se = &p->se;
9451 set_task_rq(p, task_cpu(p));
9452 se->depth = se->parent ? se->parent->depth + 1 : 0;
9455 static void task_move_group_fair(struct task_struct *p)
9457 detach_task_cfs_rq(p);
9458 set_task_rq(p, task_cpu(p));
9461 /* Tell se's cfs_rq has been changed -- migrated */
9462 p->se.avg.last_update_time = 0;
9464 attach_task_cfs_rq(p);
9467 static void task_change_group_fair(struct task_struct *p, int type)
9470 case TASK_SET_GROUP:
9471 task_set_group_fair(p);
9474 case TASK_MOVE_GROUP:
9475 task_move_group_fair(p);
9480 void free_fair_sched_group(struct task_group *tg)
9484 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9486 for_each_possible_cpu(i) {
9488 kfree(tg->cfs_rq[i]);
9497 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9499 struct sched_entity *se;
9500 struct cfs_rq *cfs_rq;
9503 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9506 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9510 tg->shares = NICE_0_LOAD;
9512 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9514 for_each_possible_cpu(i) {
9515 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9516 GFP_KERNEL, cpu_to_node(i));
9520 se = kzalloc_node(sizeof(struct sched_entity),
9521 GFP_KERNEL, cpu_to_node(i));
9525 init_cfs_rq(cfs_rq);
9526 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9527 init_entity_runnable_average(se);
9538 void online_fair_sched_group(struct task_group *tg)
9540 struct sched_entity *se;
9544 for_each_possible_cpu(i) {
9548 raw_spin_lock_irq(&rq->lock);
9549 update_rq_clock(rq);
9550 attach_entity_cfs_rq(se);
9551 sync_throttle(tg, i);
9552 raw_spin_unlock_irq(&rq->lock);
9556 void unregister_fair_sched_group(struct task_group *tg)
9558 unsigned long flags;
9562 for_each_possible_cpu(cpu) {
9564 remove_entity_load_avg(tg->se[cpu]);
9567 * Only empty task groups can be destroyed; so we can speculatively
9568 * check on_list without danger of it being re-added.
9570 if (!tg->cfs_rq[cpu]->on_list)
9575 raw_spin_lock_irqsave(&rq->lock, flags);
9576 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9577 raw_spin_unlock_irqrestore(&rq->lock, flags);
9581 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9582 struct sched_entity *se, int cpu,
9583 struct sched_entity *parent)
9585 struct rq *rq = cpu_rq(cpu);
9589 init_cfs_rq_runtime(cfs_rq);
9591 tg->cfs_rq[cpu] = cfs_rq;
9594 /* se could be NULL for root_task_group */
9599 se->cfs_rq = &rq->cfs;
9602 se->cfs_rq = parent->my_q;
9603 se->depth = parent->depth + 1;
9607 /* guarantee group entities always have weight */
9608 update_load_set(&se->load, NICE_0_LOAD);
9609 se->parent = parent;
9612 static DEFINE_MUTEX(shares_mutex);
9614 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9619 * We can't change the weight of the root cgroup.
9624 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9626 mutex_lock(&shares_mutex);
9627 if (tg->shares == shares)
9630 tg->shares = shares;
9631 for_each_possible_cpu(i) {
9632 struct rq *rq = cpu_rq(i);
9633 struct sched_entity *se = tg->se[i];
9636 /* Propagate contribution to hierarchy */
9637 rq_lock_irqsave(rq, &rf);
9638 update_rq_clock(rq);
9639 for_each_sched_entity(se) {
9640 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9641 update_cfs_shares(se);
9643 rq_unlock_irqrestore(rq, &rf);
9647 mutex_unlock(&shares_mutex);
9650 #else /* CONFIG_FAIR_GROUP_SCHED */
9652 void free_fair_sched_group(struct task_group *tg) { }
9654 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9659 void online_fair_sched_group(struct task_group *tg) { }
9661 void unregister_fair_sched_group(struct task_group *tg) { }
9663 #endif /* CONFIG_FAIR_GROUP_SCHED */
9666 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9668 struct sched_entity *se = &task->se;
9669 unsigned int rr_interval = 0;
9672 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9675 if (rq->cfs.load.weight)
9676 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9682 * All the scheduling class methods:
9684 const struct sched_class fair_sched_class = {
9685 .next = &idle_sched_class,
9686 .enqueue_task = enqueue_task_fair,
9687 .dequeue_task = dequeue_task_fair,
9688 .yield_task = yield_task_fair,
9689 .yield_to_task = yield_to_task_fair,
9691 .check_preempt_curr = check_preempt_wakeup,
9693 .pick_next_task = pick_next_task_fair,
9694 .put_prev_task = put_prev_task_fair,
9697 .select_task_rq = select_task_rq_fair,
9698 .migrate_task_rq = migrate_task_rq_fair,
9700 .rq_online = rq_online_fair,
9701 .rq_offline = rq_offline_fair,
9703 .task_dead = task_dead_fair,
9704 .set_cpus_allowed = set_cpus_allowed_common,
9707 .set_curr_task = set_curr_task_fair,
9708 .task_tick = task_tick_fair,
9709 .task_fork = task_fork_fair,
9711 .prio_changed = prio_changed_fair,
9712 .switched_from = switched_from_fair,
9713 .switched_to = switched_to_fair,
9715 .get_rr_interval = get_rr_interval_fair,
9717 .update_curr = update_curr_fair,
9719 #ifdef CONFIG_FAIR_GROUP_SCHED
9720 .task_change_group = task_change_group_fair,
9724 #ifdef CONFIG_SCHED_DEBUG
9725 void print_cfs_stats(struct seq_file *m, int cpu)
9727 struct cfs_rq *cfs_rq, *pos;
9730 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9731 print_cfs_rq(m, cpu, cfs_rq);
9735 #ifdef CONFIG_NUMA_BALANCING
9736 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9739 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9741 for_each_online_node(node) {
9742 if (p->numa_faults) {
9743 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9744 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9746 if (p->numa_group) {
9747 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9748 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9750 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9753 #endif /* CONFIG_NUMA_BALANCING */
9754 #endif /* CONFIG_SCHED_DEBUG */
9756 __init void init_sched_fair_class(void)
9759 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9761 #ifdef CONFIG_NO_HZ_COMMON
9762 nohz.next_balance = jiffies;
9763 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);