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;
517 u64 vruntime = cfs_rq->min_vruntime;
521 vruntime = curr->vruntime;
526 if (cfs_rq->rb_leftmost) {
527 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
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_node;
551 struct rb_node *parent = NULL;
552 struct sched_entity *entry;
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;
574 * Maintain a cache of leftmost tree entries (it is frequently
578 cfs_rq->rb_leftmost = &se->run_node;
580 rb_link_node(&se->run_node, parent, link);
581 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
584 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
586 if (cfs_rq->rb_leftmost == &se->run_node) {
587 struct rb_node *next_node;
589 next_node = rb_next(&se->run_node);
590 cfs_rq->rb_leftmost = next_node;
593 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
596 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
598 struct rb_node *left = cfs_rq->rb_leftmost;
603 return rb_entry(left, struct sched_entity, run_node);
606 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
608 struct rb_node *next = rb_next(&se->run_node);
613 return rb_entry(next, struct sched_entity, run_node);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
619 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
624 return rb_entry(last, struct sched_entity, run_node);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_proc_update_handler(struct ctl_table *table, int write,
632 void __user *buffer, size_t *lenp,
635 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
636 unsigned int factor = get_update_sysctl_factor();
641 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
642 sysctl_sched_min_granularity);
644 #define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity);
647 WRT_SYSCTL(sched_latency);
648 WRT_SYSCTL(sched_wakeup_granularity);
658 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
660 if (unlikely(se->load.weight != NICE_0_LOAD))
661 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
667 * The idea is to set a period in which each task runs once.
669 * When there are too many tasks (sched_nr_latency) we have to stretch
670 * this period because otherwise the slices get too small.
672 * p = (nr <= nl) ? l : l*nr/nl
674 static u64 __sched_period(unsigned long nr_running)
676 if (unlikely(nr_running > sched_nr_latency))
677 return nr_running * sysctl_sched_min_granularity;
679 return sysctl_sched_latency;
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
688 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
692 for_each_sched_entity(se) {
693 struct load_weight *load;
694 struct load_weight lw;
696 cfs_rq = cfs_rq_of(se);
697 load = &cfs_rq->load;
699 if (unlikely(!se->on_rq)) {
702 update_load_add(&lw, se->load.weight);
705 slice = __calc_delta(slice, se->load.weight, load);
711 * We calculate the vruntime slice of a to-be-inserted task.
715 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
717 return calc_delta_fair(sched_slice(cfs_rq, se), se);
722 #include "sched-pelt.h"
724 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
725 static unsigned long task_h_load(struct task_struct *p);
727 /* Give new sched_entity start runnable values to heavy its load in infant time */
728 void init_entity_runnable_average(struct sched_entity *se)
730 struct sched_avg *sa = &se->avg;
732 sa->last_update_time = 0;
734 * sched_avg's period_contrib should be strictly less then 1024, so
735 * we give it 1023 to make sure it is almost a period (1024us), and
736 * will definitely be update (after enqueue).
738 sa->period_contrib = 1023;
740 * Tasks are intialized with full load to be seen as heavy tasks until
741 * they get a chance to stabilize to their real load level.
742 * Group entities are intialized with zero load to reflect the fact that
743 * nothing has been attached to the task group yet.
745 if (entity_is_task(se))
746 sa->load_avg = scale_load_down(se->load.weight);
747 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
749 * At this point, util_avg won't be used in select_task_rq_fair anyway
753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
756 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
757 static void attach_entity_cfs_rq(struct sched_entity *se);
760 * With new tasks being created, their initial util_avgs are extrapolated
761 * based on the cfs_rq's current util_avg:
763 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
765 * However, in many cases, the above util_avg does not give a desired
766 * value. Moreover, the sum of the util_avgs may be divergent, such
767 * as when the series is a harmonic series.
769 * To solve this problem, we also cap the util_avg of successive tasks to
770 * only 1/2 of the left utilization budget:
772 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
774 * where n denotes the nth task.
776 * For example, a simplest series from the beginning would be like:
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
784 void post_init_entity_util_avg(struct sched_entity *se)
786 struct cfs_rq *cfs_rq = cfs_rq_of(se);
787 struct sched_avg *sa = &se->avg;
788 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
791 if (cfs_rq->avg.util_avg != 0) {
792 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
793 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
795 if (sa->util_avg > cap)
800 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
803 if (entity_is_task(se)) {
804 struct task_struct *p = task_of(se);
805 if (p->sched_class != &fair_sched_class) {
807 * For !fair tasks do:
809 update_cfs_rq_load_avg(now, cfs_rq);
810 attach_entity_load_avg(cfs_rq, se);
811 switched_from_fair(rq, p);
813 * such that the next switched_to_fair() has the
816 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
821 attach_entity_cfs_rq(se);
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity *se)
828 void post_init_entity_util_avg(struct sched_entity *se)
831 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
834 #endif /* CONFIG_SMP */
837 * Update the current task's runtime statistics.
839 static void update_curr(struct cfs_rq *cfs_rq)
841 struct sched_entity *curr = cfs_rq->curr;
842 u64 now = rq_clock_task(rq_of(cfs_rq));
848 delta_exec = now - curr->exec_start;
849 if (unlikely((s64)delta_exec <= 0))
852 curr->exec_start = now;
854 schedstat_set(curr->statistics.exec_max,
855 max(delta_exec, curr->statistics.exec_max));
857 curr->sum_exec_runtime += delta_exec;
858 schedstat_add(cfs_rq->exec_clock, delta_exec);
860 curr->vruntime += calc_delta_fair(delta_exec, curr);
861 update_min_vruntime(cfs_rq);
863 if (entity_is_task(curr)) {
864 struct task_struct *curtask = task_of(curr);
866 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
867 cpuacct_charge(curtask, delta_exec);
868 account_group_exec_runtime(curtask, delta_exec);
871 account_cfs_rq_runtime(cfs_rq, delta_exec);
874 static void update_curr_fair(struct rq *rq)
876 update_curr(cfs_rq_of(&rq->curr->se));
880 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
882 u64 wait_start, prev_wait_start;
884 if (!schedstat_enabled())
887 wait_start = rq_clock(rq_of(cfs_rq));
888 prev_wait_start = schedstat_val(se->statistics.wait_start);
890 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
891 likely(wait_start > prev_wait_start))
892 wait_start -= prev_wait_start;
894 schedstat_set(se->statistics.wait_start, wait_start);
898 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
900 struct task_struct *p;
903 if (!schedstat_enabled())
906 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
908 if (entity_is_task(se)) {
910 if (task_on_rq_migrating(p)) {
912 * Preserve migrating task's wait time so wait_start
913 * time stamp can be adjusted to accumulate wait time
914 * prior to migration.
916 schedstat_set(se->statistics.wait_start, delta);
919 trace_sched_stat_wait(p, delta);
922 schedstat_set(se->statistics.wait_max,
923 max(schedstat_val(se->statistics.wait_max), delta));
924 schedstat_inc(se->statistics.wait_count);
925 schedstat_add(se->statistics.wait_sum, delta);
926 schedstat_set(se->statistics.wait_start, 0);
930 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
932 struct task_struct *tsk = NULL;
933 u64 sleep_start, block_start;
935 if (!schedstat_enabled())
938 sleep_start = schedstat_val(se->statistics.sleep_start);
939 block_start = schedstat_val(se->statistics.block_start);
941 if (entity_is_task(se))
945 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
950 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
951 schedstat_set(se->statistics.sleep_max, delta);
953 schedstat_set(se->statistics.sleep_start, 0);
954 schedstat_add(se->statistics.sum_sleep_runtime, delta);
957 account_scheduler_latency(tsk, delta >> 10, 1);
958 trace_sched_stat_sleep(tsk, delta);
962 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
967 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
968 schedstat_set(se->statistics.block_max, delta);
970 schedstat_set(se->statistics.block_start, 0);
971 schedstat_add(se->statistics.sum_sleep_runtime, delta);
974 if (tsk->in_iowait) {
975 schedstat_add(se->statistics.iowait_sum, delta);
976 schedstat_inc(se->statistics.iowait_count);
977 trace_sched_stat_iowait(tsk, delta);
980 trace_sched_stat_blocked(tsk, delta);
983 * Blocking time is in units of nanosecs, so shift by
984 * 20 to get a milliseconds-range estimation of the
985 * amount of time that the task spent sleeping:
987 if (unlikely(prof_on == SLEEP_PROFILING)) {
988 profile_hits(SLEEP_PROFILING,
989 (void *)get_wchan(tsk),
992 account_scheduler_latency(tsk, delta >> 10, 0);
998 * Task is being enqueued - update stats:
1001 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1003 if (!schedstat_enabled())
1007 * Are we enqueueing a waiting task? (for current tasks
1008 * a dequeue/enqueue event is a NOP)
1010 if (se != cfs_rq->curr)
1011 update_stats_wait_start(cfs_rq, se);
1013 if (flags & ENQUEUE_WAKEUP)
1014 update_stats_enqueue_sleeper(cfs_rq, se);
1018 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1021 if (!schedstat_enabled())
1025 * Mark the end of the wait period if dequeueing a
1028 if (se != cfs_rq->curr)
1029 update_stats_wait_end(cfs_rq, se);
1031 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1032 struct task_struct *tsk = task_of(se);
1034 if (tsk->state & TASK_INTERRUPTIBLE)
1035 schedstat_set(se->statistics.sleep_start,
1036 rq_clock(rq_of(cfs_rq)));
1037 if (tsk->state & TASK_UNINTERRUPTIBLE)
1038 schedstat_set(se->statistics.block_start,
1039 rq_clock(rq_of(cfs_rq)));
1044 * We are picking a new current task - update its stats:
1047 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1050 * We are starting a new run period:
1052 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1055 /**************************************************
1056 * Scheduling class queueing methods:
1059 #ifdef CONFIG_NUMA_BALANCING
1061 * Approximate time to scan a full NUMA task in ms. The task scan period is
1062 * calculated based on the tasks virtual memory size and
1063 * numa_balancing_scan_size.
1065 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1066 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1068 /* Portion of address space to scan in MB */
1069 unsigned int sysctl_numa_balancing_scan_size = 256;
1071 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1077 spinlock_t lock; /* nr_tasks, tasks */
1082 struct rcu_head rcu;
1083 unsigned long total_faults;
1084 unsigned long max_faults_cpu;
1086 * Faults_cpu is used to decide whether memory should move
1087 * towards the CPU. As a consequence, these stats are weighted
1088 * more by CPU use than by memory faults.
1090 unsigned long *faults_cpu;
1091 unsigned long faults[0];
1094 static inline unsigned long group_faults_priv(struct numa_group *ng);
1095 static inline unsigned long group_faults_shared(struct numa_group *ng);
1097 static unsigned int task_nr_scan_windows(struct task_struct *p)
1099 unsigned long rss = 0;
1100 unsigned long nr_scan_pages;
1103 * Calculations based on RSS as non-present and empty pages are skipped
1104 * by the PTE scanner and NUMA hinting faults should be trapped based
1107 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1108 rss = get_mm_rss(p->mm);
1110 rss = nr_scan_pages;
1112 rss = round_up(rss, nr_scan_pages);
1113 return rss / nr_scan_pages;
1116 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1117 #define MAX_SCAN_WINDOW 2560
1119 static unsigned int task_scan_min(struct task_struct *p)
1121 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1122 unsigned int scan, floor;
1123 unsigned int windows = 1;
1125 if (scan_size < MAX_SCAN_WINDOW)
1126 windows = MAX_SCAN_WINDOW / scan_size;
1127 floor = 1000 / windows;
1129 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1130 return max_t(unsigned int, floor, scan);
1133 static unsigned int task_scan_start(struct task_struct *p)
1135 unsigned long smin = task_scan_min(p);
1136 unsigned long period = smin;
1138 /* Scale the maximum scan period with the amount of shared memory. */
1139 if (p->numa_group) {
1140 struct numa_group *ng = p->numa_group;
1141 unsigned long shared = group_faults_shared(ng);
1142 unsigned long private = group_faults_priv(ng);
1144 period *= atomic_read(&ng->refcount);
1145 period *= shared + 1;
1146 period /= private + shared + 1;
1149 return max(smin, period);
1152 static unsigned int task_scan_max(struct task_struct *p)
1154 unsigned long smin = task_scan_min(p);
1157 /* Watch for min being lower than max due to floor calculations */
1158 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1160 /* Scale the maximum scan period with the amount of shared memory. */
1161 if (p->numa_group) {
1162 struct numa_group *ng = p->numa_group;
1163 unsigned long shared = group_faults_shared(ng);
1164 unsigned long private = group_faults_priv(ng);
1165 unsigned long period = smax;
1167 period *= atomic_read(&ng->refcount);
1168 period *= shared + 1;
1169 period /= private + shared + 1;
1171 smax = max(smax, period);
1174 return max(smin, smax);
1177 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1179 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1180 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1183 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1185 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1186 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1189 /* Shared or private faults. */
1190 #define NR_NUMA_HINT_FAULT_TYPES 2
1192 /* Memory and CPU locality */
1193 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1195 /* Averaged statistics, and temporary buffers. */
1196 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1198 pid_t task_numa_group_id(struct task_struct *p)
1200 return p->numa_group ? p->numa_group->gid : 0;
1204 * The averaged statistics, shared & private, memory & cpu,
1205 * occupy the first half of the array. The second half of the
1206 * array is for current counters, which are averaged into the
1207 * first set by task_numa_placement.
1209 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1211 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1214 static inline unsigned long task_faults(struct task_struct *p, int nid)
1216 if (!p->numa_faults)
1219 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1220 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1223 static inline unsigned long group_faults(struct task_struct *p, int nid)
1228 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1229 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1232 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1234 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1235 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1238 static inline unsigned long group_faults_priv(struct numa_group *ng)
1240 unsigned long faults = 0;
1243 for_each_online_node(node) {
1244 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1250 static inline unsigned long group_faults_shared(struct numa_group *ng)
1252 unsigned long faults = 0;
1255 for_each_online_node(node) {
1256 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1263 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1264 * considered part of a numa group's pseudo-interleaving set. Migrations
1265 * between these nodes are slowed down, to allow things to settle down.
1267 #define ACTIVE_NODE_FRACTION 3
1269 static bool numa_is_active_node(int nid, struct numa_group *ng)
1271 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1274 /* Handle placement on systems where not all nodes are directly connected. */
1275 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1276 int maxdist, bool task)
1278 unsigned long score = 0;
1282 * All nodes are directly connected, and the same distance
1283 * from each other. No need for fancy placement algorithms.
1285 if (sched_numa_topology_type == NUMA_DIRECT)
1289 * This code is called for each node, introducing N^2 complexity,
1290 * which should be ok given the number of nodes rarely exceeds 8.
1292 for_each_online_node(node) {
1293 unsigned long faults;
1294 int dist = node_distance(nid, node);
1297 * The furthest away nodes in the system are not interesting
1298 * for placement; nid was already counted.
1300 if (dist == sched_max_numa_distance || node == nid)
1304 * On systems with a backplane NUMA topology, compare groups
1305 * of nodes, and move tasks towards the group with the most
1306 * memory accesses. When comparing two nodes at distance
1307 * "hoplimit", only nodes closer by than "hoplimit" are part
1308 * of each group. Skip other nodes.
1310 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1314 /* Add up the faults from nearby nodes. */
1316 faults = task_faults(p, node);
1318 faults = group_faults(p, node);
1321 * On systems with a glueless mesh NUMA topology, there are
1322 * no fixed "groups of nodes". Instead, nodes that are not
1323 * directly connected bounce traffic through intermediate
1324 * nodes; a numa_group can occupy any set of nodes.
1325 * The further away a node is, the less the faults count.
1326 * This seems to result in good task placement.
1328 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1329 faults *= (sched_max_numa_distance - dist);
1330 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1340 * These return the fraction of accesses done by a particular task, or
1341 * task group, on a particular numa node. The group weight is given a
1342 * larger multiplier, in order to group tasks together that are almost
1343 * evenly spread out between numa nodes.
1345 static inline unsigned long task_weight(struct task_struct *p, int nid,
1348 unsigned long faults, total_faults;
1350 if (!p->numa_faults)
1353 total_faults = p->total_numa_faults;
1358 faults = task_faults(p, nid);
1359 faults += score_nearby_nodes(p, nid, dist, true);
1361 return 1000 * faults / total_faults;
1364 static inline unsigned long group_weight(struct task_struct *p, int nid,
1367 unsigned long faults, total_faults;
1372 total_faults = p->numa_group->total_faults;
1377 faults = group_faults(p, nid);
1378 faults += score_nearby_nodes(p, nid, dist, false);
1380 return 1000 * faults / total_faults;
1383 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1384 int src_nid, int dst_cpu)
1386 struct numa_group *ng = p->numa_group;
1387 int dst_nid = cpu_to_node(dst_cpu);
1388 int last_cpupid, this_cpupid;
1390 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1393 * Multi-stage node selection is used in conjunction with a periodic
1394 * migration fault to build a temporal task<->page relation. By using
1395 * a two-stage filter we remove short/unlikely relations.
1397 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1398 * a task's usage of a particular page (n_p) per total usage of this
1399 * page (n_t) (in a given time-span) to a probability.
1401 * Our periodic faults will sample this probability and getting the
1402 * same result twice in a row, given these samples are fully
1403 * independent, is then given by P(n)^2, provided our sample period
1404 * is sufficiently short compared to the usage pattern.
1406 * This quadric squishes small probabilities, making it less likely we
1407 * act on an unlikely task<->page relation.
1409 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1410 if (!cpupid_pid_unset(last_cpupid) &&
1411 cpupid_to_nid(last_cpupid) != dst_nid)
1414 /* Always allow migrate on private faults */
1415 if (cpupid_match_pid(p, last_cpupid))
1418 /* A shared fault, but p->numa_group has not been set up yet. */
1423 * Destination node is much more heavily used than the source
1424 * node? Allow migration.
1426 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1427 ACTIVE_NODE_FRACTION)
1431 * Distribute memory according to CPU & memory use on each node,
1432 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1434 * faults_cpu(dst) 3 faults_cpu(src)
1435 * --------------- * - > ---------------
1436 * faults_mem(dst) 4 faults_mem(src)
1438 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1439 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1442 static unsigned long weighted_cpuload(struct rq *rq);
1443 static unsigned long source_load(int cpu, int type);
1444 static unsigned long target_load(int cpu, int type);
1445 static unsigned long capacity_of(int cpu);
1447 /* Cached statistics for all CPUs within a node */
1449 unsigned long nr_running;
1452 /* Total compute capacity of CPUs on a node */
1453 unsigned long compute_capacity;
1455 /* Approximate capacity in terms of runnable tasks on a node */
1456 unsigned long task_capacity;
1457 int has_free_capacity;
1461 * XXX borrowed from update_sg_lb_stats
1463 static void update_numa_stats(struct numa_stats *ns, int nid)
1465 int smt, cpu, cpus = 0;
1466 unsigned long capacity;
1468 memset(ns, 0, sizeof(*ns));
1469 for_each_cpu(cpu, cpumask_of_node(nid)) {
1470 struct rq *rq = cpu_rq(cpu);
1472 ns->nr_running += rq->nr_running;
1473 ns->load += weighted_cpuload(rq);
1474 ns->compute_capacity += capacity_of(cpu);
1480 * If we raced with hotplug and there are no CPUs left in our mask
1481 * the @ns structure is NULL'ed and task_numa_compare() will
1482 * not find this node attractive.
1484 * We'll either bail at !has_free_capacity, or we'll detect a huge
1485 * imbalance and bail there.
1490 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1491 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1492 capacity = cpus / smt; /* cores */
1494 ns->task_capacity = min_t(unsigned, capacity,
1495 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1496 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1499 struct task_numa_env {
1500 struct task_struct *p;
1502 int src_cpu, src_nid;
1503 int dst_cpu, dst_nid;
1505 struct numa_stats src_stats, dst_stats;
1510 struct task_struct *best_task;
1515 static void task_numa_assign(struct task_numa_env *env,
1516 struct task_struct *p, long imp)
1519 put_task_struct(env->best_task);
1524 env->best_imp = imp;
1525 env->best_cpu = env->dst_cpu;
1528 static bool load_too_imbalanced(long src_load, long dst_load,
1529 struct task_numa_env *env)
1532 long orig_src_load, orig_dst_load;
1533 long src_capacity, dst_capacity;
1536 * The load is corrected for the CPU capacity available on each node.
1539 * ------------ vs ---------
1540 * src_capacity dst_capacity
1542 src_capacity = env->src_stats.compute_capacity;
1543 dst_capacity = env->dst_stats.compute_capacity;
1545 /* We care about the slope of the imbalance, not the direction. */
1546 if (dst_load < src_load)
1547 swap(dst_load, src_load);
1549 /* Is the difference below the threshold? */
1550 imb = dst_load * src_capacity * 100 -
1551 src_load * dst_capacity * env->imbalance_pct;
1556 * The imbalance is above the allowed threshold.
1557 * Compare it with the old imbalance.
1559 orig_src_load = env->src_stats.load;
1560 orig_dst_load = env->dst_stats.load;
1562 if (orig_dst_load < orig_src_load)
1563 swap(orig_dst_load, orig_src_load);
1565 old_imb = orig_dst_load * src_capacity * 100 -
1566 orig_src_load * dst_capacity * env->imbalance_pct;
1568 /* Would this change make things worse? */
1569 return (imb > old_imb);
1573 * This checks if the overall compute and NUMA accesses of the system would
1574 * be improved if the source tasks was migrated to the target dst_cpu taking
1575 * into account that it might be best if task running on the dst_cpu should
1576 * be exchanged with the source task
1578 static void task_numa_compare(struct task_numa_env *env,
1579 long taskimp, long groupimp)
1581 struct rq *src_rq = cpu_rq(env->src_cpu);
1582 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1583 struct task_struct *cur;
1584 long src_load, dst_load;
1586 long imp = env->p->numa_group ? groupimp : taskimp;
1588 int dist = env->dist;
1591 cur = task_rcu_dereference(&dst_rq->curr);
1592 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1596 * Because we have preemption enabled we can get migrated around and
1597 * end try selecting ourselves (current == env->p) as a swap candidate.
1603 * "imp" is the fault differential for the source task between the
1604 * source and destination node. Calculate the total differential for
1605 * the source task and potential destination task. The more negative
1606 * the value is, the more rmeote accesses that would be expected to
1607 * be incurred if the tasks were swapped.
1610 /* Skip this swap candidate if cannot move to the source cpu */
1611 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1615 * If dst and source tasks are in the same NUMA group, or not
1616 * in any group then look only at task weights.
1618 if (cur->numa_group == env->p->numa_group) {
1619 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1620 task_weight(cur, env->dst_nid, dist);
1622 * Add some hysteresis to prevent swapping the
1623 * tasks within a group over tiny differences.
1625 if (cur->numa_group)
1629 * Compare the group weights. If a task is all by
1630 * itself (not part of a group), use the task weight
1633 if (cur->numa_group)
1634 imp += group_weight(cur, env->src_nid, dist) -
1635 group_weight(cur, env->dst_nid, dist);
1637 imp += task_weight(cur, env->src_nid, dist) -
1638 task_weight(cur, env->dst_nid, dist);
1642 if (imp <= env->best_imp && moveimp <= env->best_imp)
1646 /* Is there capacity at our destination? */
1647 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1648 !env->dst_stats.has_free_capacity)
1654 /* Balance doesn't matter much if we're running a task per cpu */
1655 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1656 dst_rq->nr_running == 1)
1660 * In the overloaded case, try and keep the load balanced.
1663 load = task_h_load(env->p);
1664 dst_load = env->dst_stats.load + load;
1665 src_load = env->src_stats.load - load;
1667 if (moveimp > imp && moveimp > env->best_imp) {
1669 * If the improvement from just moving env->p direction is
1670 * better than swapping tasks around, check if a move is
1671 * possible. Store a slightly smaller score than moveimp,
1672 * so an actually idle CPU will win.
1674 if (!load_too_imbalanced(src_load, dst_load, env)) {
1681 if (imp <= env->best_imp)
1685 load = task_h_load(cur);
1690 if (load_too_imbalanced(src_load, dst_load, env))
1694 * One idle CPU per node is evaluated for a task numa move.
1695 * Call select_idle_sibling to maybe find a better one.
1699 * select_idle_siblings() uses an per-cpu cpumask that
1700 * can be used from IRQ context.
1702 local_irq_disable();
1703 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1709 task_numa_assign(env, cur, imp);
1714 static void task_numa_find_cpu(struct task_numa_env *env,
1715 long taskimp, long groupimp)
1719 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1720 /* Skip this CPU if the source task cannot migrate */
1721 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1725 task_numa_compare(env, taskimp, groupimp);
1729 /* Only move tasks to a NUMA node less busy than the current node. */
1730 static bool numa_has_capacity(struct task_numa_env *env)
1732 struct numa_stats *src = &env->src_stats;
1733 struct numa_stats *dst = &env->dst_stats;
1735 if (src->has_free_capacity && !dst->has_free_capacity)
1739 * Only consider a task move if the source has a higher load
1740 * than the destination, corrected for CPU capacity on each node.
1742 * src->load dst->load
1743 * --------------------- vs ---------------------
1744 * src->compute_capacity dst->compute_capacity
1746 if (src->load * dst->compute_capacity * env->imbalance_pct >
1748 dst->load * src->compute_capacity * 100)
1754 static int task_numa_migrate(struct task_struct *p)
1756 struct task_numa_env env = {
1759 .src_cpu = task_cpu(p),
1760 .src_nid = task_node(p),
1762 .imbalance_pct = 112,
1768 struct sched_domain *sd;
1769 unsigned long taskweight, groupweight;
1771 long taskimp, groupimp;
1774 * Pick the lowest SD_NUMA domain, as that would have the smallest
1775 * imbalance and would be the first to start moving tasks about.
1777 * And we want to avoid any moving of tasks about, as that would create
1778 * random movement of tasks -- counter the numa conditions we're trying
1782 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1784 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1788 * Cpusets can break the scheduler domain tree into smaller
1789 * balance domains, some of which do not cross NUMA boundaries.
1790 * Tasks that are "trapped" in such domains cannot be migrated
1791 * elsewhere, so there is no point in (re)trying.
1793 if (unlikely(!sd)) {
1794 p->numa_preferred_nid = task_node(p);
1798 env.dst_nid = p->numa_preferred_nid;
1799 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1800 taskweight = task_weight(p, env.src_nid, dist);
1801 groupweight = group_weight(p, env.src_nid, dist);
1802 update_numa_stats(&env.src_stats, env.src_nid);
1803 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1804 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1805 update_numa_stats(&env.dst_stats, env.dst_nid);
1807 /* Try to find a spot on the preferred nid. */
1808 if (numa_has_capacity(&env))
1809 task_numa_find_cpu(&env, taskimp, groupimp);
1812 * Look at other nodes in these cases:
1813 * - there is no space available on the preferred_nid
1814 * - the task is part of a numa_group that is interleaved across
1815 * multiple NUMA nodes; in order to better consolidate the group,
1816 * we need to check other locations.
1818 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1819 for_each_online_node(nid) {
1820 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1823 dist = node_distance(env.src_nid, env.dst_nid);
1824 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1826 taskweight = task_weight(p, env.src_nid, dist);
1827 groupweight = group_weight(p, env.src_nid, dist);
1830 /* Only consider nodes where both task and groups benefit */
1831 taskimp = task_weight(p, nid, dist) - taskweight;
1832 groupimp = group_weight(p, nid, dist) - groupweight;
1833 if (taskimp < 0 && groupimp < 0)
1838 update_numa_stats(&env.dst_stats, env.dst_nid);
1839 if (numa_has_capacity(&env))
1840 task_numa_find_cpu(&env, taskimp, groupimp);
1845 * If the task is part of a workload that spans multiple NUMA nodes,
1846 * and is migrating into one of the workload's active nodes, remember
1847 * this node as the task's preferred numa node, so the workload can
1849 * A task that migrated to a second choice node will be better off
1850 * trying for a better one later. Do not set the preferred node here.
1852 if (p->numa_group) {
1853 struct numa_group *ng = p->numa_group;
1855 if (env.best_cpu == -1)
1860 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1861 sched_setnuma(p, env.dst_nid);
1864 /* No better CPU than the current one was found. */
1865 if (env.best_cpu == -1)
1869 * Reset the scan period if the task is being rescheduled on an
1870 * alternative node to recheck if the tasks is now properly placed.
1872 p->numa_scan_period = task_scan_start(p);
1874 if (env.best_task == NULL) {
1875 ret = migrate_task_to(p, env.best_cpu);
1877 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1881 ret = migrate_swap(p, env.best_task);
1883 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1884 put_task_struct(env.best_task);
1888 /* Attempt to migrate a task to a CPU on the preferred node. */
1889 static void numa_migrate_preferred(struct task_struct *p)
1891 unsigned long interval = HZ;
1893 /* This task has no NUMA fault statistics yet */
1894 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1897 /* Periodically retry migrating the task to the preferred node */
1898 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1899 p->numa_migrate_retry = jiffies + interval;
1901 /* Success if task is already running on preferred CPU */
1902 if (task_node(p) == p->numa_preferred_nid)
1905 /* Otherwise, try migrate to a CPU on the preferred node */
1906 task_numa_migrate(p);
1910 * Find out how many nodes on the workload is actively running on. Do this by
1911 * tracking the nodes from which NUMA hinting faults are triggered. This can
1912 * be different from the set of nodes where the workload's memory is currently
1915 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1917 unsigned long faults, max_faults = 0;
1918 int nid, active_nodes = 0;
1920 for_each_online_node(nid) {
1921 faults = group_faults_cpu(numa_group, nid);
1922 if (faults > max_faults)
1923 max_faults = faults;
1926 for_each_online_node(nid) {
1927 faults = group_faults_cpu(numa_group, nid);
1928 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1932 numa_group->max_faults_cpu = max_faults;
1933 numa_group->active_nodes = active_nodes;
1937 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1938 * increments. The more local the fault statistics are, the higher the scan
1939 * period will be for the next scan window. If local/(local+remote) ratio is
1940 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1941 * the scan period will decrease. Aim for 70% local accesses.
1943 #define NUMA_PERIOD_SLOTS 10
1944 #define NUMA_PERIOD_THRESHOLD 7
1947 * Increase the scan period (slow down scanning) if the majority of
1948 * our memory is already on our local node, or if the majority of
1949 * the page accesses are shared with other processes.
1950 * Otherwise, decrease the scan period.
1952 static void update_task_scan_period(struct task_struct *p,
1953 unsigned long shared, unsigned long private)
1955 unsigned int period_slot;
1956 int lr_ratio, ps_ratio;
1959 unsigned long remote = p->numa_faults_locality[0];
1960 unsigned long local = p->numa_faults_locality[1];
1963 * If there were no record hinting faults then either the task is
1964 * completely idle or all activity is areas that are not of interest
1965 * to automatic numa balancing. Related to that, if there were failed
1966 * migration then it implies we are migrating too quickly or the local
1967 * node is overloaded. In either case, scan slower
1969 if (local + shared == 0 || p->numa_faults_locality[2]) {
1970 p->numa_scan_period = min(p->numa_scan_period_max,
1971 p->numa_scan_period << 1);
1973 p->mm->numa_next_scan = jiffies +
1974 msecs_to_jiffies(p->numa_scan_period);
1980 * Prepare to scale scan period relative to the current period.
1981 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1982 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1983 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1985 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1986 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1987 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1989 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1991 * Most memory accesses are local. There is no need to
1992 * do fast NUMA scanning, since memory is already local.
1994 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1997 diff = slot * period_slot;
1998 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2000 * Most memory accesses are shared with other tasks.
2001 * There is no point in continuing fast NUMA scanning,
2002 * since other tasks may just move the memory elsewhere.
2004 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2007 diff = slot * period_slot;
2010 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2011 * yet they are not on the local NUMA node. Speed up
2012 * NUMA scanning to get the memory moved over.
2014 int ratio = max(lr_ratio, ps_ratio);
2015 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2018 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2019 task_scan_min(p), task_scan_max(p));
2020 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2024 * Get the fraction of time the task has been running since the last
2025 * NUMA placement cycle. The scheduler keeps similar statistics, but
2026 * decays those on a 32ms period, which is orders of magnitude off
2027 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2028 * stats only if the task is so new there are no NUMA statistics yet.
2030 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2032 u64 runtime, delta, now;
2033 /* Use the start of this time slice to avoid calculations. */
2034 now = p->se.exec_start;
2035 runtime = p->se.sum_exec_runtime;
2037 if (p->last_task_numa_placement) {
2038 delta = runtime - p->last_sum_exec_runtime;
2039 *period = now - p->last_task_numa_placement;
2041 delta = p->se.avg.load_sum / p->se.load.weight;
2042 *period = LOAD_AVG_MAX;
2045 p->last_sum_exec_runtime = runtime;
2046 p->last_task_numa_placement = now;
2052 * Determine the preferred nid for a task in a numa_group. This needs to
2053 * be done in a way that produces consistent results with group_weight,
2054 * otherwise workloads might not converge.
2056 static int preferred_group_nid(struct task_struct *p, int nid)
2061 /* Direct connections between all NUMA nodes. */
2062 if (sched_numa_topology_type == NUMA_DIRECT)
2066 * On a system with glueless mesh NUMA topology, group_weight
2067 * scores nodes according to the number of NUMA hinting faults on
2068 * both the node itself, and on nearby nodes.
2070 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2071 unsigned long score, max_score = 0;
2072 int node, max_node = nid;
2074 dist = sched_max_numa_distance;
2076 for_each_online_node(node) {
2077 score = group_weight(p, node, dist);
2078 if (score > max_score) {
2087 * Finding the preferred nid in a system with NUMA backplane
2088 * interconnect topology is more involved. The goal is to locate
2089 * tasks from numa_groups near each other in the system, and
2090 * untangle workloads from different sides of the system. This requires
2091 * searching down the hierarchy of node groups, recursively searching
2092 * inside the highest scoring group of nodes. The nodemask tricks
2093 * keep the complexity of the search down.
2095 nodes = node_online_map;
2096 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2097 unsigned long max_faults = 0;
2098 nodemask_t max_group = NODE_MASK_NONE;
2101 /* Are there nodes at this distance from each other? */
2102 if (!find_numa_distance(dist))
2105 for_each_node_mask(a, nodes) {
2106 unsigned long faults = 0;
2107 nodemask_t this_group;
2108 nodes_clear(this_group);
2110 /* Sum group's NUMA faults; includes a==b case. */
2111 for_each_node_mask(b, nodes) {
2112 if (node_distance(a, b) < dist) {
2113 faults += group_faults(p, b);
2114 node_set(b, this_group);
2115 node_clear(b, nodes);
2119 /* Remember the top group. */
2120 if (faults > max_faults) {
2121 max_faults = faults;
2122 max_group = this_group;
2124 * subtle: at the smallest distance there is
2125 * just one node left in each "group", the
2126 * winner is the preferred nid.
2131 /* Next round, evaluate the nodes within max_group. */
2139 static void task_numa_placement(struct task_struct *p)
2141 int seq, nid, max_nid = -1, max_group_nid = -1;
2142 unsigned long max_faults = 0, max_group_faults = 0;
2143 unsigned long fault_types[2] = { 0, 0 };
2144 unsigned long total_faults;
2145 u64 runtime, period;
2146 spinlock_t *group_lock = NULL;
2149 * The p->mm->numa_scan_seq field gets updated without
2150 * exclusive access. Use READ_ONCE() here to ensure
2151 * that the field is read in a single access:
2153 seq = READ_ONCE(p->mm->numa_scan_seq);
2154 if (p->numa_scan_seq == seq)
2156 p->numa_scan_seq = seq;
2157 p->numa_scan_period_max = task_scan_max(p);
2159 total_faults = p->numa_faults_locality[0] +
2160 p->numa_faults_locality[1];
2161 runtime = numa_get_avg_runtime(p, &period);
2163 /* If the task is part of a group prevent parallel updates to group stats */
2164 if (p->numa_group) {
2165 group_lock = &p->numa_group->lock;
2166 spin_lock_irq(group_lock);
2169 /* Find the node with the highest number of faults */
2170 for_each_online_node(nid) {
2171 /* Keep track of the offsets in numa_faults array */
2172 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2173 unsigned long faults = 0, group_faults = 0;
2176 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2177 long diff, f_diff, f_weight;
2179 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2180 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2181 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2182 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2184 /* Decay existing window, copy faults since last scan */
2185 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2186 fault_types[priv] += p->numa_faults[membuf_idx];
2187 p->numa_faults[membuf_idx] = 0;
2190 * Normalize the faults_from, so all tasks in a group
2191 * count according to CPU use, instead of by the raw
2192 * number of faults. Tasks with little runtime have
2193 * little over-all impact on throughput, and thus their
2194 * faults are less important.
2196 f_weight = div64_u64(runtime << 16, period + 1);
2197 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2199 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2200 p->numa_faults[cpubuf_idx] = 0;
2202 p->numa_faults[mem_idx] += diff;
2203 p->numa_faults[cpu_idx] += f_diff;
2204 faults += p->numa_faults[mem_idx];
2205 p->total_numa_faults += diff;
2206 if (p->numa_group) {
2208 * safe because we can only change our own group
2210 * mem_idx represents the offset for a given
2211 * nid and priv in a specific region because it
2212 * is at the beginning of the numa_faults array.
2214 p->numa_group->faults[mem_idx] += diff;
2215 p->numa_group->faults_cpu[mem_idx] += f_diff;
2216 p->numa_group->total_faults += diff;
2217 group_faults += p->numa_group->faults[mem_idx];
2221 if (faults > max_faults) {
2222 max_faults = faults;
2226 if (group_faults > max_group_faults) {
2227 max_group_faults = group_faults;
2228 max_group_nid = nid;
2232 update_task_scan_period(p, fault_types[0], fault_types[1]);
2234 if (p->numa_group) {
2235 numa_group_count_active_nodes(p->numa_group);
2236 spin_unlock_irq(group_lock);
2237 max_nid = preferred_group_nid(p, max_group_nid);
2241 /* Set the new preferred node */
2242 if (max_nid != p->numa_preferred_nid)
2243 sched_setnuma(p, max_nid);
2245 if (task_node(p) != p->numa_preferred_nid)
2246 numa_migrate_preferred(p);
2250 static inline int get_numa_group(struct numa_group *grp)
2252 return atomic_inc_not_zero(&grp->refcount);
2255 static inline void put_numa_group(struct numa_group *grp)
2257 if (atomic_dec_and_test(&grp->refcount))
2258 kfree_rcu(grp, rcu);
2261 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2264 struct numa_group *grp, *my_grp;
2265 struct task_struct *tsk;
2267 int cpu = cpupid_to_cpu(cpupid);
2270 if (unlikely(!p->numa_group)) {
2271 unsigned int size = sizeof(struct numa_group) +
2272 4*nr_node_ids*sizeof(unsigned long);
2274 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2278 atomic_set(&grp->refcount, 1);
2279 grp->active_nodes = 1;
2280 grp->max_faults_cpu = 0;
2281 spin_lock_init(&grp->lock);
2283 /* Second half of the array tracks nids where faults happen */
2284 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2287 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2288 grp->faults[i] = p->numa_faults[i];
2290 grp->total_faults = p->total_numa_faults;
2293 rcu_assign_pointer(p->numa_group, grp);
2297 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2299 if (!cpupid_match_pid(tsk, cpupid))
2302 grp = rcu_dereference(tsk->numa_group);
2306 my_grp = p->numa_group;
2311 * Only join the other group if its bigger; if we're the bigger group,
2312 * the other task will join us.
2314 if (my_grp->nr_tasks > grp->nr_tasks)
2318 * Tie-break on the grp address.
2320 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2323 /* Always join threads in the same process. */
2324 if (tsk->mm == current->mm)
2327 /* Simple filter to avoid false positives due to PID collisions */
2328 if (flags & TNF_SHARED)
2331 /* Update priv based on whether false sharing was detected */
2334 if (join && !get_numa_group(grp))
2342 BUG_ON(irqs_disabled());
2343 double_lock_irq(&my_grp->lock, &grp->lock);
2345 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2346 my_grp->faults[i] -= p->numa_faults[i];
2347 grp->faults[i] += p->numa_faults[i];
2349 my_grp->total_faults -= p->total_numa_faults;
2350 grp->total_faults += p->total_numa_faults;
2355 spin_unlock(&my_grp->lock);
2356 spin_unlock_irq(&grp->lock);
2358 rcu_assign_pointer(p->numa_group, grp);
2360 put_numa_group(my_grp);
2368 void task_numa_free(struct task_struct *p)
2370 struct numa_group *grp = p->numa_group;
2371 void *numa_faults = p->numa_faults;
2372 unsigned long flags;
2376 spin_lock_irqsave(&grp->lock, flags);
2377 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2378 grp->faults[i] -= p->numa_faults[i];
2379 grp->total_faults -= p->total_numa_faults;
2382 spin_unlock_irqrestore(&grp->lock, flags);
2383 RCU_INIT_POINTER(p->numa_group, NULL);
2384 put_numa_group(grp);
2387 p->numa_faults = NULL;
2392 * Got a PROT_NONE fault for a page on @node.
2394 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2396 struct task_struct *p = current;
2397 bool migrated = flags & TNF_MIGRATED;
2398 int cpu_node = task_node(current);
2399 int local = !!(flags & TNF_FAULT_LOCAL);
2400 struct numa_group *ng;
2403 if (!static_branch_likely(&sched_numa_balancing))
2406 /* for example, ksmd faulting in a user's mm */
2410 /* Allocate buffer to track faults on a per-node basis */
2411 if (unlikely(!p->numa_faults)) {
2412 int size = sizeof(*p->numa_faults) *
2413 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2415 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2416 if (!p->numa_faults)
2419 p->total_numa_faults = 0;
2420 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2424 * First accesses are treated as private, otherwise consider accesses
2425 * to be private if the accessing pid has not changed
2427 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2430 priv = cpupid_match_pid(p, last_cpupid);
2431 if (!priv && !(flags & TNF_NO_GROUP))
2432 task_numa_group(p, last_cpupid, flags, &priv);
2436 * If a workload spans multiple NUMA nodes, a shared fault that
2437 * occurs wholly within the set of nodes that the workload is
2438 * actively using should be counted as local. This allows the
2439 * scan rate to slow down when a workload has settled down.
2442 if (!priv && !local && ng && ng->active_nodes > 1 &&
2443 numa_is_active_node(cpu_node, ng) &&
2444 numa_is_active_node(mem_node, ng))
2447 task_numa_placement(p);
2450 * Retry task to preferred node migration periodically, in case it
2451 * case it previously failed, or the scheduler moved us.
2453 if (time_after(jiffies, p->numa_migrate_retry))
2454 numa_migrate_preferred(p);
2457 p->numa_pages_migrated += pages;
2458 if (flags & TNF_MIGRATE_FAIL)
2459 p->numa_faults_locality[2] += pages;
2461 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2462 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2463 p->numa_faults_locality[local] += pages;
2466 static void reset_ptenuma_scan(struct task_struct *p)
2469 * We only did a read acquisition of the mmap sem, so
2470 * p->mm->numa_scan_seq is written to without exclusive access
2471 * and the update is not guaranteed to be atomic. That's not
2472 * much of an issue though, since this is just used for
2473 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2474 * expensive, to avoid any form of compiler optimizations:
2476 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2477 p->mm->numa_scan_offset = 0;
2481 * The expensive part of numa migration is done from task_work context.
2482 * Triggered from task_tick_numa().
2484 void task_numa_work(struct callback_head *work)
2486 unsigned long migrate, next_scan, now = jiffies;
2487 struct task_struct *p = current;
2488 struct mm_struct *mm = p->mm;
2489 u64 runtime = p->se.sum_exec_runtime;
2490 struct vm_area_struct *vma;
2491 unsigned long start, end;
2492 unsigned long nr_pte_updates = 0;
2493 long pages, virtpages;
2495 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2497 work->next = work; /* protect against double add */
2499 * Who cares about NUMA placement when they're dying.
2501 * NOTE: make sure not to dereference p->mm before this check,
2502 * exit_task_work() happens _after_ exit_mm() so we could be called
2503 * without p->mm even though we still had it when we enqueued this
2506 if (p->flags & PF_EXITING)
2509 if (!mm->numa_next_scan) {
2510 mm->numa_next_scan = now +
2511 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2515 * Enforce maximal scan/migration frequency..
2517 migrate = mm->numa_next_scan;
2518 if (time_before(now, migrate))
2521 if (p->numa_scan_period == 0) {
2522 p->numa_scan_period_max = task_scan_max(p);
2523 p->numa_scan_period = task_scan_start(p);
2526 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2527 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2531 * Delay this task enough that another task of this mm will likely win
2532 * the next time around.
2534 p->node_stamp += 2 * TICK_NSEC;
2536 start = mm->numa_scan_offset;
2537 pages = sysctl_numa_balancing_scan_size;
2538 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2539 virtpages = pages * 8; /* Scan up to this much virtual space */
2544 if (!down_read_trylock(&mm->mmap_sem))
2546 vma = find_vma(mm, start);
2548 reset_ptenuma_scan(p);
2552 for (; vma; vma = vma->vm_next) {
2553 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2554 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2559 * Shared library pages mapped by multiple processes are not
2560 * migrated as it is expected they are cache replicated. Avoid
2561 * hinting faults in read-only file-backed mappings or the vdso
2562 * as migrating the pages will be of marginal benefit.
2565 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2569 * Skip inaccessible VMAs to avoid any confusion between
2570 * PROT_NONE and NUMA hinting ptes
2572 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2576 start = max(start, vma->vm_start);
2577 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2578 end = min(end, vma->vm_end);
2579 nr_pte_updates = change_prot_numa(vma, start, end);
2582 * Try to scan sysctl_numa_balancing_size worth of
2583 * hpages that have at least one present PTE that
2584 * is not already pte-numa. If the VMA contains
2585 * areas that are unused or already full of prot_numa
2586 * PTEs, scan up to virtpages, to skip through those
2590 pages -= (end - start) >> PAGE_SHIFT;
2591 virtpages -= (end - start) >> PAGE_SHIFT;
2594 if (pages <= 0 || virtpages <= 0)
2598 } while (end != vma->vm_end);
2603 * It is possible to reach the end of the VMA list but the last few
2604 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2605 * would find the !migratable VMA on the next scan but not reset the
2606 * scanner to the start so check it now.
2609 mm->numa_scan_offset = start;
2611 reset_ptenuma_scan(p);
2612 up_read(&mm->mmap_sem);
2615 * Make sure tasks use at least 32x as much time to run other code
2616 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2617 * Usually update_task_scan_period slows down scanning enough; on an
2618 * overloaded system we need to limit overhead on a per task basis.
2620 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2621 u64 diff = p->se.sum_exec_runtime - runtime;
2622 p->node_stamp += 32 * diff;
2627 * Drive the periodic memory faults..
2629 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2631 struct callback_head *work = &curr->numa_work;
2635 * We don't care about NUMA placement if we don't have memory.
2637 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2641 * Using runtime rather than walltime has the dual advantage that
2642 * we (mostly) drive the selection from busy threads and that the
2643 * task needs to have done some actual work before we bother with
2646 now = curr->se.sum_exec_runtime;
2647 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2649 if (now > curr->node_stamp + period) {
2650 if (!curr->node_stamp)
2651 curr->numa_scan_period = task_scan_start(curr);
2652 curr->node_stamp += period;
2654 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2655 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2656 task_work_add(curr, work, true);
2662 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2666 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2670 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2674 #endif /* CONFIG_NUMA_BALANCING */
2677 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2679 update_load_add(&cfs_rq->load, se->load.weight);
2680 if (!parent_entity(se))
2681 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2683 if (entity_is_task(se)) {
2684 struct rq *rq = rq_of(cfs_rq);
2686 account_numa_enqueue(rq, task_of(se));
2687 list_add(&se->group_node, &rq->cfs_tasks);
2690 cfs_rq->nr_running++;
2694 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2696 update_load_sub(&cfs_rq->load, se->load.weight);
2697 if (!parent_entity(se))
2698 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2700 if (entity_is_task(se)) {
2701 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2702 list_del_init(&se->group_node);
2705 cfs_rq->nr_running--;
2708 #ifdef CONFIG_FAIR_GROUP_SCHED
2710 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2712 long tg_weight, load, shares;
2715 * This really should be: cfs_rq->avg.load_avg, but instead we use
2716 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2717 * the shares for small weight interactive tasks.
2719 load = scale_load_down(cfs_rq->load.weight);
2721 tg_weight = atomic_long_read(&tg->load_avg);
2723 /* Ensure tg_weight >= load */
2724 tg_weight -= cfs_rq->tg_load_avg_contrib;
2727 shares = (tg->shares * load);
2729 shares /= tg_weight;
2732 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2733 * of a group with small tg->shares value. It is a floor value which is
2734 * assigned as a minimum load.weight to the sched_entity representing
2735 * the group on a CPU.
2737 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2738 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2739 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2740 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2743 if (shares < MIN_SHARES)
2744 shares = MIN_SHARES;
2745 if (shares > tg->shares)
2746 shares = tg->shares;
2750 # else /* CONFIG_SMP */
2751 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2755 # endif /* CONFIG_SMP */
2757 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2758 unsigned long weight)
2761 /* commit outstanding execution time */
2762 if (cfs_rq->curr == se)
2763 update_curr(cfs_rq);
2764 account_entity_dequeue(cfs_rq, se);
2767 update_load_set(&se->load, weight);
2770 account_entity_enqueue(cfs_rq, se);
2773 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2775 static void update_cfs_shares(struct sched_entity *se)
2777 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2778 struct task_group *tg;
2784 if (throttled_hierarchy(cfs_rq))
2790 if (likely(se->load.weight == tg->shares))
2793 shares = calc_cfs_shares(cfs_rq, tg);
2795 reweight_entity(cfs_rq_of(se), se, shares);
2798 #else /* CONFIG_FAIR_GROUP_SCHED */
2799 static inline void update_cfs_shares(struct sched_entity *se)
2802 #endif /* CONFIG_FAIR_GROUP_SCHED */
2804 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2806 struct rq *rq = rq_of(cfs_rq);
2808 if (&rq->cfs == cfs_rq) {
2810 * There are a few boundary cases this might miss but it should
2811 * get called often enough that that should (hopefully) not be
2812 * a real problem -- added to that it only calls on the local
2813 * CPU, so if we enqueue remotely we'll miss an update, but
2814 * the next tick/schedule should update.
2816 * It will not get called when we go idle, because the idle
2817 * thread is a different class (!fair), nor will the utilization
2818 * number include things like RT tasks.
2820 * As is, the util number is not freq-invariant (we'd have to
2821 * implement arch_scale_freq_capacity() for that).
2825 cpufreq_update_util(rq, 0);
2832 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2834 static u64 decay_load(u64 val, u64 n)
2836 unsigned int local_n;
2838 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2841 /* after bounds checking we can collapse to 32-bit */
2845 * As y^PERIOD = 1/2, we can combine
2846 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2847 * With a look-up table which covers y^n (n<PERIOD)
2849 * To achieve constant time decay_load.
2851 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2852 val >>= local_n / LOAD_AVG_PERIOD;
2853 local_n %= LOAD_AVG_PERIOD;
2856 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2860 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2862 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2867 c1 = decay_load((u64)d1, periods);
2871 * c2 = 1024 \Sum y^n
2875 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2878 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2880 return c1 + c2 + c3;
2883 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2886 * Accumulate the three separate parts of the sum; d1 the remainder
2887 * of the last (incomplete) period, d2 the span of full periods and d3
2888 * the remainder of the (incomplete) current period.
2893 * |<->|<----------------->|<--->|
2894 * ... |---x---|------| ... |------|-----x (now)
2897 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2900 * = u y^p + (Step 1)
2903 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2906 static __always_inline u32
2907 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2908 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2910 unsigned long scale_freq, scale_cpu;
2911 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2914 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2915 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2917 delta += sa->period_contrib;
2918 periods = delta / 1024; /* A period is 1024us (~1ms) */
2921 * Step 1: decay old *_sum if we crossed period boundaries.
2924 sa->load_sum = decay_load(sa->load_sum, periods);
2926 cfs_rq->runnable_load_sum =
2927 decay_load(cfs_rq->runnable_load_sum, periods);
2929 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2935 contrib = __accumulate_pelt_segments(periods,
2936 1024 - sa->period_contrib, delta);
2938 sa->period_contrib = delta;
2940 contrib = cap_scale(contrib, scale_freq);
2942 sa->load_sum += weight * contrib;
2944 cfs_rq->runnable_load_sum += weight * contrib;
2947 sa->util_sum += contrib * scale_cpu;
2953 * We can represent the historical contribution to runnable average as the
2954 * coefficients of a geometric series. To do this we sub-divide our runnable
2955 * history into segments of approximately 1ms (1024us); label the segment that
2956 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2958 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2960 * (now) (~1ms ago) (~2ms ago)
2962 * Let u_i denote the fraction of p_i that the entity was runnable.
2964 * We then designate the fractions u_i as our co-efficients, yielding the
2965 * following representation of historical load:
2966 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2968 * We choose y based on the with of a reasonably scheduling period, fixing:
2971 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2972 * approximately half as much as the contribution to load within the last ms
2975 * When a period "rolls over" and we have new u_0`, multiplying the previous
2976 * sum again by y is sufficient to update:
2977 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2978 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2980 static __always_inline int
2981 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2982 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2986 delta = now - sa->last_update_time;
2988 * This should only happen when time goes backwards, which it
2989 * unfortunately does during sched clock init when we swap over to TSC.
2991 if ((s64)delta < 0) {
2992 sa->last_update_time = now;
2997 * Use 1024ns as the unit of measurement since it's a reasonable
2998 * approximation of 1us and fast to compute.
3004 sa->last_update_time += delta << 10;
3007 * running is a subset of runnable (weight) so running can't be set if
3008 * runnable is clear. But there are some corner cases where the current
3009 * se has been already dequeued but cfs_rq->curr still points to it.
3010 * This means that weight will be 0 but not running for a sched_entity
3011 * but also for a cfs_rq if the latter becomes idle. As an example,
3012 * this happens during idle_balance() which calls
3013 * update_blocked_averages()
3019 * Now we know we crossed measurement unit boundaries. The *_avg
3020 * accrues by two steps:
3022 * Step 1: accumulate *_sum since last_update_time. If we haven't
3023 * crossed period boundaries, finish.
3025 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
3029 * Step 2: update *_avg.
3031 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3033 cfs_rq->runnable_load_avg =
3034 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
3036 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
3042 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3044 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
3048 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3050 return ___update_load_avg(now, cpu, &se->avg,
3051 se->on_rq * scale_load_down(se->load.weight),
3052 cfs_rq->curr == se, NULL);
3056 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3058 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3059 scale_load_down(cfs_rq->load.weight),
3060 cfs_rq->curr != NULL, cfs_rq);
3064 * Signed add and clamp on underflow.
3066 * Explicitly do a load-store to ensure the intermediate value never hits
3067 * memory. This allows lockless observations without ever seeing the negative
3070 #define add_positive(_ptr, _val) do { \
3071 typeof(_ptr) ptr = (_ptr); \
3072 typeof(_val) val = (_val); \
3073 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3077 if (val < 0 && res > var) \
3080 WRITE_ONCE(*ptr, res); \
3083 #ifdef CONFIG_FAIR_GROUP_SCHED
3085 * update_tg_load_avg - update the tg's load avg
3086 * @cfs_rq: the cfs_rq whose avg changed
3087 * @force: update regardless of how small the difference
3089 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3090 * However, because tg->load_avg is a global value there are performance
3093 * In order to avoid having to look at the other cfs_rq's, we use a
3094 * differential update where we store the last value we propagated. This in
3095 * turn allows skipping updates if the differential is 'small'.
3097 * Updating tg's load_avg is necessary before update_cfs_share().
3099 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3101 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3104 * No need to update load_avg for root_task_group as it is not used.
3106 if (cfs_rq->tg == &root_task_group)
3109 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3110 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3111 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3116 * Called within set_task_rq() right before setting a task's cpu. The
3117 * caller only guarantees p->pi_lock is held; no other assumptions,
3118 * including the state of rq->lock, should be made.
3120 void set_task_rq_fair(struct sched_entity *se,
3121 struct cfs_rq *prev, struct cfs_rq *next)
3123 u64 p_last_update_time;
3124 u64 n_last_update_time;
3126 if (!sched_feat(ATTACH_AGE_LOAD))
3130 * We are supposed to update the task to "current" time, then its up to
3131 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3132 * getting what current time is, so simply throw away the out-of-date
3133 * time. This will result in the wakee task is less decayed, but giving
3134 * the wakee more load sounds not bad.
3136 if (!(se->avg.last_update_time && prev))
3139 #ifndef CONFIG_64BIT
3141 u64 p_last_update_time_copy;
3142 u64 n_last_update_time_copy;
3145 p_last_update_time_copy = prev->load_last_update_time_copy;
3146 n_last_update_time_copy = next->load_last_update_time_copy;
3150 p_last_update_time = prev->avg.last_update_time;
3151 n_last_update_time = next->avg.last_update_time;
3153 } while (p_last_update_time != p_last_update_time_copy ||
3154 n_last_update_time != n_last_update_time_copy);
3157 p_last_update_time = prev->avg.last_update_time;
3158 n_last_update_time = next->avg.last_update_time;
3160 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3161 se->avg.last_update_time = n_last_update_time;
3164 /* Take into account change of utilization of a child task group */
3166 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3168 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3169 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3171 /* Nothing to update */
3175 /* Set new sched_entity's utilization */
3176 se->avg.util_avg = gcfs_rq->avg.util_avg;
3177 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3179 /* Update parent cfs_rq utilization */
3180 add_positive(&cfs_rq->avg.util_avg, delta);
3181 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3184 /* Take into account change of load of a child task group */
3186 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3188 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3189 long delta, load = gcfs_rq->avg.load_avg;
3192 * If the load of group cfs_rq is null, the load of the
3193 * sched_entity will also be null so we can skip the formula
3198 /* Get tg's load and ensure tg_load > 0 */
3199 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3201 /* Ensure tg_load >= load and updated with current load*/
3202 tg_load -= gcfs_rq->tg_load_avg_contrib;
3206 * We need to compute a correction term in the case that the
3207 * task group is consuming more CPU than a task of equal
3208 * weight. A task with a weight equals to tg->shares will have
3209 * a load less or equal to scale_load_down(tg->shares).
3210 * Similarly, the sched_entities that represent the task group
3211 * at parent level, can't have a load higher than
3212 * scale_load_down(tg->shares). And the Sum of sched_entities'
3213 * load must be <= scale_load_down(tg->shares).
3215 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3216 /* scale gcfs_rq's load into tg's shares*/
3217 load *= scale_load_down(gcfs_rq->tg->shares);
3222 delta = load - se->avg.load_avg;
3224 /* Nothing to update */
3228 /* Set new sched_entity's load */
3229 se->avg.load_avg = load;
3230 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3232 /* Update parent cfs_rq load */
3233 add_positive(&cfs_rq->avg.load_avg, delta);
3234 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3237 * If the sched_entity is already enqueued, we also have to update the
3238 * runnable load avg.
3241 /* Update parent cfs_rq runnable_load_avg */
3242 add_positive(&cfs_rq->runnable_load_avg, delta);
3243 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3247 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3249 cfs_rq->propagate_avg = 1;
3252 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3254 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3256 if (!cfs_rq->propagate_avg)
3259 cfs_rq->propagate_avg = 0;
3263 /* Update task and its cfs_rq load average */
3264 static inline int propagate_entity_load_avg(struct sched_entity *se)
3266 struct cfs_rq *cfs_rq;
3268 if (entity_is_task(se))
3271 if (!test_and_clear_tg_cfs_propagate(se))
3274 cfs_rq = cfs_rq_of(se);
3276 set_tg_cfs_propagate(cfs_rq);
3278 update_tg_cfs_util(cfs_rq, se);
3279 update_tg_cfs_load(cfs_rq, se);
3285 * Check if we need to update the load and the utilization of a blocked
3288 static inline bool skip_blocked_update(struct sched_entity *se)
3290 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3293 * If sched_entity still have not zero load or utilization, we have to
3296 if (se->avg.load_avg || se->avg.util_avg)
3300 * If there is a pending propagation, we have to update the load and
3301 * the utilization of the sched_entity:
3303 if (gcfs_rq->propagate_avg)
3307 * Otherwise, the load and the utilization of the sched_entity is
3308 * already zero and there is no pending propagation, so it will be a
3309 * waste of time to try to decay it:
3314 #else /* CONFIG_FAIR_GROUP_SCHED */
3316 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3318 static inline int propagate_entity_load_avg(struct sched_entity *se)
3323 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3325 #endif /* CONFIG_FAIR_GROUP_SCHED */
3328 * Unsigned subtract and clamp on underflow.
3330 * Explicitly do a load-store to ensure the intermediate value never hits
3331 * memory. This allows lockless observations without ever seeing the negative
3334 #define sub_positive(_ptr, _val) do { \
3335 typeof(_ptr) ptr = (_ptr); \
3336 typeof(*ptr) val = (_val); \
3337 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3341 WRITE_ONCE(*ptr, res); \
3345 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3346 * @now: current time, as per cfs_rq_clock_task()
3347 * @cfs_rq: cfs_rq to update
3349 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3350 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3351 * post_init_entity_util_avg().
3353 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3355 * Returns true if the load decayed or we removed load.
3357 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3358 * call update_tg_load_avg() when this function returns true.
3361 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3363 struct sched_avg *sa = &cfs_rq->avg;
3364 int decayed, removed_load = 0, removed_util = 0;
3366 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3367 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3368 sub_positive(&sa->load_avg, r);
3369 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3371 set_tg_cfs_propagate(cfs_rq);
3374 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3375 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3376 sub_positive(&sa->util_avg, r);
3377 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3379 set_tg_cfs_propagate(cfs_rq);
3382 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3384 #ifndef CONFIG_64BIT
3386 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3389 if (decayed || removed_util)
3390 cfs_rq_util_change(cfs_rq);
3392 return decayed || removed_load;
3396 * Optional action to be done while updating the load average
3398 #define UPDATE_TG 0x1
3399 #define SKIP_AGE_LOAD 0x2
3401 /* Update task and its cfs_rq load average */
3402 static inline void update_load_avg(struct sched_entity *se, int flags)
3404 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3405 u64 now = cfs_rq_clock_task(cfs_rq);
3406 struct rq *rq = rq_of(cfs_rq);
3407 int cpu = cpu_of(rq);
3411 * Track task load average for carrying it to new CPU after migrated, and
3412 * track group sched_entity load average for task_h_load calc in migration
3414 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3415 __update_load_avg_se(now, cpu, cfs_rq, se);
3417 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3418 decayed |= propagate_entity_load_avg(se);
3420 if (decayed && (flags & UPDATE_TG))
3421 update_tg_load_avg(cfs_rq, 0);
3425 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3426 * @cfs_rq: cfs_rq to attach to
3427 * @se: sched_entity to attach
3429 * Must call update_cfs_rq_load_avg() before this, since we rely on
3430 * cfs_rq->avg.last_update_time being current.
3432 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3434 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3435 cfs_rq->avg.load_avg += se->avg.load_avg;
3436 cfs_rq->avg.load_sum += se->avg.load_sum;
3437 cfs_rq->avg.util_avg += se->avg.util_avg;
3438 cfs_rq->avg.util_sum += se->avg.util_sum;
3439 set_tg_cfs_propagate(cfs_rq);
3441 cfs_rq_util_change(cfs_rq);
3445 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3446 * @cfs_rq: cfs_rq to detach from
3447 * @se: sched_entity to detach
3449 * Must call update_cfs_rq_load_avg() before this, since we rely on
3450 * cfs_rq->avg.last_update_time being current.
3452 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3455 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3456 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3457 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3458 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3459 set_tg_cfs_propagate(cfs_rq);
3461 cfs_rq_util_change(cfs_rq);
3464 /* Add the load generated by se into cfs_rq's load average */
3466 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3468 struct sched_avg *sa = &se->avg;
3470 cfs_rq->runnable_load_avg += sa->load_avg;
3471 cfs_rq->runnable_load_sum += sa->load_sum;
3473 if (!sa->last_update_time) {
3474 attach_entity_load_avg(cfs_rq, se);
3475 update_tg_load_avg(cfs_rq, 0);
3479 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3481 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3483 cfs_rq->runnable_load_avg =
3484 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3485 cfs_rq->runnable_load_sum =
3486 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3489 #ifndef CONFIG_64BIT
3490 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3492 u64 last_update_time_copy;
3493 u64 last_update_time;
3496 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3498 last_update_time = cfs_rq->avg.last_update_time;
3499 } while (last_update_time != last_update_time_copy);
3501 return last_update_time;
3504 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3506 return cfs_rq->avg.last_update_time;
3511 * Synchronize entity load avg of dequeued entity without locking
3514 void sync_entity_load_avg(struct sched_entity *se)
3516 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3517 u64 last_update_time;
3519 last_update_time = cfs_rq_last_update_time(cfs_rq);
3520 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3524 * Task first catches up with cfs_rq, and then subtract
3525 * itself from the cfs_rq (task must be off the queue now).
3527 void remove_entity_load_avg(struct sched_entity *se)
3529 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3532 * tasks cannot exit without having gone through wake_up_new_task() ->
3533 * post_init_entity_util_avg() which will have added things to the
3534 * cfs_rq, so we can remove unconditionally.
3536 * Similarly for groups, they will have passed through
3537 * post_init_entity_util_avg() before unregister_sched_fair_group()
3541 sync_entity_load_avg(se);
3542 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3543 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3546 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3548 return cfs_rq->runnable_load_avg;
3551 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3553 return cfs_rq->avg.load_avg;
3556 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3558 #else /* CONFIG_SMP */
3561 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3566 #define UPDATE_TG 0x0
3567 #define SKIP_AGE_LOAD 0x0
3569 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3571 cfs_rq_util_change(cfs_rq_of(se));
3575 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3577 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3578 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3581 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3583 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3585 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3590 #endif /* CONFIG_SMP */
3592 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3594 #ifdef CONFIG_SCHED_DEBUG
3595 s64 d = se->vruntime - cfs_rq->min_vruntime;
3600 if (d > 3*sysctl_sched_latency)
3601 schedstat_inc(cfs_rq->nr_spread_over);
3606 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3608 u64 vruntime = cfs_rq->min_vruntime;
3611 * The 'current' period is already promised to the current tasks,
3612 * however the extra weight of the new task will slow them down a
3613 * little, place the new task so that it fits in the slot that
3614 * stays open at the end.
3616 if (initial && sched_feat(START_DEBIT))
3617 vruntime += sched_vslice(cfs_rq, se);
3619 /* sleeps up to a single latency don't count. */
3621 unsigned long thresh = sysctl_sched_latency;
3624 * Halve their sleep time's effect, to allow
3625 * for a gentler effect of sleepers:
3627 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3633 /* ensure we never gain time by being placed backwards. */
3634 se->vruntime = max_vruntime(se->vruntime, vruntime);
3637 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3639 static inline void check_schedstat_required(void)
3641 #ifdef CONFIG_SCHEDSTATS
3642 if (schedstat_enabled())
3645 /* Force schedstat enabled if a dependent tracepoint is active */
3646 if (trace_sched_stat_wait_enabled() ||
3647 trace_sched_stat_sleep_enabled() ||
3648 trace_sched_stat_iowait_enabled() ||
3649 trace_sched_stat_blocked_enabled() ||
3650 trace_sched_stat_runtime_enabled()) {
3651 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3652 "stat_blocked and stat_runtime require the "
3653 "kernel parameter schedstats=enable or "
3654 "kernel.sched_schedstats=1\n");
3665 * update_min_vruntime()
3666 * vruntime -= min_vruntime
3670 * update_min_vruntime()
3671 * vruntime += min_vruntime
3673 * this way the vruntime transition between RQs is done when both
3674 * min_vruntime are up-to-date.
3678 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3679 * vruntime -= min_vruntime
3683 * update_min_vruntime()
3684 * vruntime += min_vruntime
3686 * this way we don't have the most up-to-date min_vruntime on the originating
3687 * CPU and an up-to-date min_vruntime on the destination CPU.
3691 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3693 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3694 bool curr = cfs_rq->curr == se;
3697 * If we're the current task, we must renormalise before calling
3701 se->vruntime += cfs_rq->min_vruntime;
3703 update_curr(cfs_rq);
3706 * Otherwise, renormalise after, such that we're placed at the current
3707 * moment in time, instead of some random moment in the past. Being
3708 * placed in the past could significantly boost this task to the
3709 * fairness detriment of existing tasks.
3711 if (renorm && !curr)
3712 se->vruntime += cfs_rq->min_vruntime;
3715 * When enqueuing a sched_entity, we must:
3716 * - Update loads to have both entity and cfs_rq synced with now.
3717 * - Add its load to cfs_rq->runnable_avg
3718 * - For group_entity, update its weight to reflect the new share of
3720 * - Add its new weight to cfs_rq->load.weight
3722 update_load_avg(se, UPDATE_TG);
3723 enqueue_entity_load_avg(cfs_rq, se);
3724 update_cfs_shares(se);
3725 account_entity_enqueue(cfs_rq, se);
3727 if (flags & ENQUEUE_WAKEUP)
3728 place_entity(cfs_rq, se, 0);
3730 check_schedstat_required();
3731 update_stats_enqueue(cfs_rq, se, flags);
3732 check_spread(cfs_rq, se);
3734 __enqueue_entity(cfs_rq, se);
3737 if (cfs_rq->nr_running == 1) {
3738 list_add_leaf_cfs_rq(cfs_rq);
3739 check_enqueue_throttle(cfs_rq);
3743 static void __clear_buddies_last(struct sched_entity *se)
3745 for_each_sched_entity(se) {
3746 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3747 if (cfs_rq->last != se)
3750 cfs_rq->last = NULL;
3754 static void __clear_buddies_next(struct sched_entity *se)
3756 for_each_sched_entity(se) {
3757 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3758 if (cfs_rq->next != se)
3761 cfs_rq->next = NULL;
3765 static void __clear_buddies_skip(struct sched_entity *se)
3767 for_each_sched_entity(se) {
3768 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3769 if (cfs_rq->skip != se)
3772 cfs_rq->skip = NULL;
3776 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3778 if (cfs_rq->last == se)
3779 __clear_buddies_last(se);
3781 if (cfs_rq->next == se)
3782 __clear_buddies_next(se);
3784 if (cfs_rq->skip == se)
3785 __clear_buddies_skip(se);
3788 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3791 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3794 * Update run-time statistics of the 'current'.
3796 update_curr(cfs_rq);
3799 * When dequeuing a sched_entity, we must:
3800 * - Update loads to have both entity and cfs_rq synced with now.
3801 * - Substract its load from the cfs_rq->runnable_avg.
3802 * - Substract its previous weight from cfs_rq->load.weight.
3803 * - For group entity, update its weight to reflect the new share
3804 * of its group cfs_rq.
3806 update_load_avg(se, UPDATE_TG);
3807 dequeue_entity_load_avg(cfs_rq, se);
3809 update_stats_dequeue(cfs_rq, se, flags);
3811 clear_buddies(cfs_rq, se);
3813 if (se != cfs_rq->curr)
3814 __dequeue_entity(cfs_rq, se);
3816 account_entity_dequeue(cfs_rq, se);
3819 * Normalize after update_curr(); which will also have moved
3820 * min_vruntime if @se is the one holding it back. But before doing
3821 * update_min_vruntime() again, which will discount @se's position and
3822 * can move min_vruntime forward still more.
3824 if (!(flags & DEQUEUE_SLEEP))
3825 se->vruntime -= cfs_rq->min_vruntime;
3827 /* return excess runtime on last dequeue */
3828 return_cfs_rq_runtime(cfs_rq);
3830 update_cfs_shares(se);
3833 * Now advance min_vruntime if @se was the entity holding it back,
3834 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3835 * put back on, and if we advance min_vruntime, we'll be placed back
3836 * further than we started -- ie. we'll be penalized.
3838 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3839 update_min_vruntime(cfs_rq);
3843 * Preempt the current task with a newly woken task if needed:
3846 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3848 unsigned long ideal_runtime, delta_exec;
3849 struct sched_entity *se;
3852 ideal_runtime = sched_slice(cfs_rq, curr);
3853 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3854 if (delta_exec > ideal_runtime) {
3855 resched_curr(rq_of(cfs_rq));
3857 * The current task ran long enough, ensure it doesn't get
3858 * re-elected due to buddy favours.
3860 clear_buddies(cfs_rq, curr);
3865 * Ensure that a task that missed wakeup preemption by a
3866 * narrow margin doesn't have to wait for a full slice.
3867 * This also mitigates buddy induced latencies under load.
3869 if (delta_exec < sysctl_sched_min_granularity)
3872 se = __pick_first_entity(cfs_rq);
3873 delta = curr->vruntime - se->vruntime;
3878 if (delta > ideal_runtime)
3879 resched_curr(rq_of(cfs_rq));
3883 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3885 /* 'current' is not kept within the tree. */
3888 * Any task has to be enqueued before it get to execute on
3889 * a CPU. So account for the time it spent waiting on the
3892 update_stats_wait_end(cfs_rq, se);
3893 __dequeue_entity(cfs_rq, se);
3894 update_load_avg(se, UPDATE_TG);
3897 update_stats_curr_start(cfs_rq, se);
3901 * Track our maximum slice length, if the CPU's load is at
3902 * least twice that of our own weight (i.e. dont track it
3903 * when there are only lesser-weight tasks around):
3905 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3906 schedstat_set(se->statistics.slice_max,
3907 max((u64)schedstat_val(se->statistics.slice_max),
3908 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3911 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3915 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3918 * Pick the next process, keeping these things in mind, in this order:
3919 * 1) keep things fair between processes/task groups
3920 * 2) pick the "next" process, since someone really wants that to run
3921 * 3) pick the "last" process, for cache locality
3922 * 4) do not run the "skip" process, if something else is available
3924 static struct sched_entity *
3925 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3927 struct sched_entity *left = __pick_first_entity(cfs_rq);
3928 struct sched_entity *se;
3931 * If curr is set we have to see if its left of the leftmost entity
3932 * still in the tree, provided there was anything in the tree at all.
3934 if (!left || (curr && entity_before(curr, left)))
3937 se = left; /* ideally we run the leftmost entity */
3940 * Avoid running the skip buddy, if running something else can
3941 * be done without getting too unfair.
3943 if (cfs_rq->skip == se) {
3944 struct sched_entity *second;
3947 second = __pick_first_entity(cfs_rq);
3949 second = __pick_next_entity(se);
3950 if (!second || (curr && entity_before(curr, second)))
3954 if (second && wakeup_preempt_entity(second, left) < 1)
3959 * Prefer last buddy, try to return the CPU to a preempted task.
3961 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3965 * Someone really wants this to run. If it's not unfair, run it.
3967 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3970 clear_buddies(cfs_rq, se);
3975 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3977 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3980 * If still on the runqueue then deactivate_task()
3981 * was not called and update_curr() has to be done:
3984 update_curr(cfs_rq);
3986 /* throttle cfs_rqs exceeding runtime */
3987 check_cfs_rq_runtime(cfs_rq);
3989 check_spread(cfs_rq, prev);
3992 update_stats_wait_start(cfs_rq, prev);
3993 /* Put 'current' back into the tree. */
3994 __enqueue_entity(cfs_rq, prev);
3995 /* in !on_rq case, update occurred at dequeue */
3996 update_load_avg(prev, 0);
3998 cfs_rq->curr = NULL;
4002 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4005 * Update run-time statistics of the 'current'.
4007 update_curr(cfs_rq);
4010 * Ensure that runnable average is periodically updated.
4012 update_load_avg(curr, UPDATE_TG);
4013 update_cfs_shares(curr);
4015 #ifdef CONFIG_SCHED_HRTICK
4017 * queued ticks are scheduled to match the slice, so don't bother
4018 * validating it and just reschedule.
4021 resched_curr(rq_of(cfs_rq));
4025 * don't let the period tick interfere with the hrtick preemption
4027 if (!sched_feat(DOUBLE_TICK) &&
4028 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4032 if (cfs_rq->nr_running > 1)
4033 check_preempt_tick(cfs_rq, curr);
4037 /**************************************************
4038 * CFS bandwidth control machinery
4041 #ifdef CONFIG_CFS_BANDWIDTH
4043 #ifdef HAVE_JUMP_LABEL
4044 static struct static_key __cfs_bandwidth_used;
4046 static inline bool cfs_bandwidth_used(void)
4048 return static_key_false(&__cfs_bandwidth_used);
4051 void cfs_bandwidth_usage_inc(void)
4053 static_key_slow_inc(&__cfs_bandwidth_used);
4056 void cfs_bandwidth_usage_dec(void)
4058 static_key_slow_dec(&__cfs_bandwidth_used);
4060 #else /* HAVE_JUMP_LABEL */
4061 static bool cfs_bandwidth_used(void)
4066 void cfs_bandwidth_usage_inc(void) {}
4067 void cfs_bandwidth_usage_dec(void) {}
4068 #endif /* HAVE_JUMP_LABEL */
4071 * default period for cfs group bandwidth.
4072 * default: 0.1s, units: nanoseconds
4074 static inline u64 default_cfs_period(void)
4076 return 100000000ULL;
4079 static inline u64 sched_cfs_bandwidth_slice(void)
4081 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4085 * Replenish runtime according to assigned quota and update expiration time.
4086 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4087 * additional synchronization around rq->lock.
4089 * requires cfs_b->lock
4091 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4095 if (cfs_b->quota == RUNTIME_INF)
4098 now = sched_clock_cpu(smp_processor_id());
4099 cfs_b->runtime = cfs_b->quota;
4100 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4103 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4105 return &tg->cfs_bandwidth;
4108 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4109 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4111 if (unlikely(cfs_rq->throttle_count))
4112 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4114 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4117 /* returns 0 on failure to allocate runtime */
4118 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4120 struct task_group *tg = cfs_rq->tg;
4121 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4122 u64 amount = 0, min_amount, expires;
4124 /* note: this is a positive sum as runtime_remaining <= 0 */
4125 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4127 raw_spin_lock(&cfs_b->lock);
4128 if (cfs_b->quota == RUNTIME_INF)
4129 amount = min_amount;
4131 start_cfs_bandwidth(cfs_b);
4133 if (cfs_b->runtime > 0) {
4134 amount = min(cfs_b->runtime, min_amount);
4135 cfs_b->runtime -= amount;
4139 expires = cfs_b->runtime_expires;
4140 raw_spin_unlock(&cfs_b->lock);
4142 cfs_rq->runtime_remaining += amount;
4144 * we may have advanced our local expiration to account for allowed
4145 * spread between our sched_clock and the one on which runtime was
4148 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4149 cfs_rq->runtime_expires = expires;
4151 return cfs_rq->runtime_remaining > 0;
4155 * Note: This depends on the synchronization provided by sched_clock and the
4156 * fact that rq->clock snapshots this value.
4158 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4160 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4162 /* if the deadline is ahead of our clock, nothing to do */
4163 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4166 if (cfs_rq->runtime_remaining < 0)
4170 * If the local deadline has passed we have to consider the
4171 * possibility that our sched_clock is 'fast' and the global deadline
4172 * has not truly expired.
4174 * Fortunately we can check determine whether this the case by checking
4175 * whether the global deadline has advanced. It is valid to compare
4176 * cfs_b->runtime_expires without any locks since we only care about
4177 * exact equality, so a partial write will still work.
4180 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4181 /* extend local deadline, drift is bounded above by 2 ticks */
4182 cfs_rq->runtime_expires += TICK_NSEC;
4184 /* global deadline is ahead, expiration has passed */
4185 cfs_rq->runtime_remaining = 0;
4189 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4191 /* dock delta_exec before expiring quota (as it could span periods) */
4192 cfs_rq->runtime_remaining -= delta_exec;
4193 expire_cfs_rq_runtime(cfs_rq);
4195 if (likely(cfs_rq->runtime_remaining > 0))
4199 * if we're unable to extend our runtime we resched so that the active
4200 * hierarchy can be throttled
4202 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4203 resched_curr(rq_of(cfs_rq));
4206 static __always_inline
4207 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4209 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4212 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4215 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4217 return cfs_bandwidth_used() && cfs_rq->throttled;
4220 /* check whether cfs_rq, or any parent, is throttled */
4221 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4223 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4227 * Ensure that neither of the group entities corresponding to src_cpu or
4228 * dest_cpu are members of a throttled hierarchy when performing group
4229 * load-balance operations.
4231 static inline int throttled_lb_pair(struct task_group *tg,
4232 int src_cpu, int dest_cpu)
4234 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4236 src_cfs_rq = tg->cfs_rq[src_cpu];
4237 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4239 return throttled_hierarchy(src_cfs_rq) ||
4240 throttled_hierarchy(dest_cfs_rq);
4243 /* updated child weight may affect parent so we have to do this bottom up */
4244 static int tg_unthrottle_up(struct task_group *tg, void *data)
4246 struct rq *rq = data;
4247 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4249 cfs_rq->throttle_count--;
4250 if (!cfs_rq->throttle_count) {
4251 /* adjust cfs_rq_clock_task() */
4252 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4253 cfs_rq->throttled_clock_task;
4259 static int tg_throttle_down(struct task_group *tg, void *data)
4261 struct rq *rq = data;
4262 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4264 /* group is entering throttled state, stop time */
4265 if (!cfs_rq->throttle_count)
4266 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4267 cfs_rq->throttle_count++;
4272 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4274 struct rq *rq = rq_of(cfs_rq);
4275 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4276 struct sched_entity *se;
4277 long task_delta, dequeue = 1;
4280 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4282 /* freeze hierarchy runnable averages while throttled */
4284 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4287 task_delta = cfs_rq->h_nr_running;
4288 for_each_sched_entity(se) {
4289 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4290 /* throttled entity or throttle-on-deactivate */
4295 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4296 qcfs_rq->h_nr_running -= task_delta;
4298 if (qcfs_rq->load.weight)
4303 sub_nr_running(rq, task_delta);
4305 cfs_rq->throttled = 1;
4306 cfs_rq->throttled_clock = rq_clock(rq);
4307 raw_spin_lock(&cfs_b->lock);
4308 empty = list_empty(&cfs_b->throttled_cfs_rq);
4311 * Add to the _head_ of the list, so that an already-started
4312 * distribute_cfs_runtime will not see us
4314 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4317 * If we're the first throttled task, make sure the bandwidth
4321 start_cfs_bandwidth(cfs_b);
4323 raw_spin_unlock(&cfs_b->lock);
4326 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4328 struct rq *rq = rq_of(cfs_rq);
4329 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4330 struct sched_entity *se;
4334 se = cfs_rq->tg->se[cpu_of(rq)];
4336 cfs_rq->throttled = 0;
4338 update_rq_clock(rq);
4340 raw_spin_lock(&cfs_b->lock);
4341 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4342 list_del_rcu(&cfs_rq->throttled_list);
4343 raw_spin_unlock(&cfs_b->lock);
4345 /* update hierarchical throttle state */
4346 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4348 if (!cfs_rq->load.weight)
4351 task_delta = cfs_rq->h_nr_running;
4352 for_each_sched_entity(se) {
4356 cfs_rq = cfs_rq_of(se);
4358 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4359 cfs_rq->h_nr_running += task_delta;
4361 if (cfs_rq_throttled(cfs_rq))
4366 add_nr_running(rq, task_delta);
4368 /* determine whether we need to wake up potentially idle cpu */
4369 if (rq->curr == rq->idle && rq->cfs.nr_running)
4373 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4374 u64 remaining, u64 expires)
4376 struct cfs_rq *cfs_rq;
4378 u64 starting_runtime = remaining;
4381 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4383 struct rq *rq = rq_of(cfs_rq);
4387 if (!cfs_rq_throttled(cfs_rq))
4390 runtime = -cfs_rq->runtime_remaining + 1;
4391 if (runtime > remaining)
4392 runtime = remaining;
4393 remaining -= runtime;
4395 cfs_rq->runtime_remaining += runtime;
4396 cfs_rq->runtime_expires = expires;
4398 /* we check whether we're throttled above */
4399 if (cfs_rq->runtime_remaining > 0)
4400 unthrottle_cfs_rq(cfs_rq);
4410 return starting_runtime - remaining;
4414 * Responsible for refilling a task_group's bandwidth and unthrottling its
4415 * cfs_rqs as appropriate. If there has been no activity within the last
4416 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4417 * used to track this state.
4419 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4421 u64 runtime, runtime_expires;
4424 /* no need to continue the timer with no bandwidth constraint */
4425 if (cfs_b->quota == RUNTIME_INF)
4426 goto out_deactivate;
4428 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4429 cfs_b->nr_periods += overrun;
4432 * idle depends on !throttled (for the case of a large deficit), and if
4433 * we're going inactive then everything else can be deferred
4435 if (cfs_b->idle && !throttled)
4436 goto out_deactivate;
4438 __refill_cfs_bandwidth_runtime(cfs_b);
4441 /* mark as potentially idle for the upcoming period */
4446 /* account preceding periods in which throttling occurred */
4447 cfs_b->nr_throttled += overrun;
4449 runtime_expires = cfs_b->runtime_expires;
4452 * This check is repeated as we are holding onto the new bandwidth while
4453 * we unthrottle. This can potentially race with an unthrottled group
4454 * trying to acquire new bandwidth from the global pool. This can result
4455 * in us over-using our runtime if it is all used during this loop, but
4456 * only by limited amounts in that extreme case.
4458 while (throttled && cfs_b->runtime > 0) {
4459 runtime = cfs_b->runtime;
4460 raw_spin_unlock(&cfs_b->lock);
4461 /* we can't nest cfs_b->lock while distributing bandwidth */
4462 runtime = distribute_cfs_runtime(cfs_b, runtime,
4464 raw_spin_lock(&cfs_b->lock);
4466 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4468 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4472 * While we are ensured activity in the period following an
4473 * unthrottle, this also covers the case in which the new bandwidth is
4474 * insufficient to cover the existing bandwidth deficit. (Forcing the
4475 * timer to remain active while there are any throttled entities.)
4485 /* a cfs_rq won't donate quota below this amount */
4486 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4487 /* minimum remaining period time to redistribute slack quota */
4488 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4489 /* how long we wait to gather additional slack before distributing */
4490 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4493 * Are we near the end of the current quota period?
4495 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4496 * hrtimer base being cleared by hrtimer_start. In the case of
4497 * migrate_hrtimers, base is never cleared, so we are fine.
4499 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4501 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4504 /* if the call-back is running a quota refresh is already occurring */
4505 if (hrtimer_callback_running(refresh_timer))
4508 /* is a quota refresh about to occur? */
4509 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4510 if (remaining < min_expire)
4516 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4518 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4520 /* if there's a quota refresh soon don't bother with slack */
4521 if (runtime_refresh_within(cfs_b, min_left))
4524 hrtimer_start(&cfs_b->slack_timer,
4525 ns_to_ktime(cfs_bandwidth_slack_period),
4529 /* we know any runtime found here is valid as update_curr() precedes return */
4530 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4532 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4533 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4535 if (slack_runtime <= 0)
4538 raw_spin_lock(&cfs_b->lock);
4539 if (cfs_b->quota != RUNTIME_INF &&
4540 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4541 cfs_b->runtime += slack_runtime;
4543 /* we are under rq->lock, defer unthrottling using a timer */
4544 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4545 !list_empty(&cfs_b->throttled_cfs_rq))
4546 start_cfs_slack_bandwidth(cfs_b);
4548 raw_spin_unlock(&cfs_b->lock);
4550 /* even if it's not valid for return we don't want to try again */
4551 cfs_rq->runtime_remaining -= slack_runtime;
4554 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4556 if (!cfs_bandwidth_used())
4559 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4562 __return_cfs_rq_runtime(cfs_rq);
4566 * This is done with a timer (instead of inline with bandwidth return) since
4567 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4569 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4571 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4574 /* confirm we're still not at a refresh boundary */
4575 raw_spin_lock(&cfs_b->lock);
4576 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4577 raw_spin_unlock(&cfs_b->lock);
4581 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4582 runtime = cfs_b->runtime;
4584 expires = cfs_b->runtime_expires;
4585 raw_spin_unlock(&cfs_b->lock);
4590 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4592 raw_spin_lock(&cfs_b->lock);
4593 if (expires == cfs_b->runtime_expires)
4594 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4595 raw_spin_unlock(&cfs_b->lock);
4599 * When a group wakes up we want to make sure that its quota is not already
4600 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4601 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4603 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4605 if (!cfs_bandwidth_used())
4608 /* an active group must be handled by the update_curr()->put() path */
4609 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4612 /* ensure the group is not already throttled */
4613 if (cfs_rq_throttled(cfs_rq))
4616 /* update runtime allocation */
4617 account_cfs_rq_runtime(cfs_rq, 0);
4618 if (cfs_rq->runtime_remaining <= 0)
4619 throttle_cfs_rq(cfs_rq);
4622 static void sync_throttle(struct task_group *tg, int cpu)
4624 struct cfs_rq *pcfs_rq, *cfs_rq;
4626 if (!cfs_bandwidth_used())
4632 cfs_rq = tg->cfs_rq[cpu];
4633 pcfs_rq = tg->parent->cfs_rq[cpu];
4635 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4636 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4639 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4640 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4642 if (!cfs_bandwidth_used())
4645 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4649 * it's possible for a throttled entity to be forced into a running
4650 * state (e.g. set_curr_task), in this case we're finished.
4652 if (cfs_rq_throttled(cfs_rq))
4655 throttle_cfs_rq(cfs_rq);
4659 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4661 struct cfs_bandwidth *cfs_b =
4662 container_of(timer, struct cfs_bandwidth, slack_timer);
4664 do_sched_cfs_slack_timer(cfs_b);
4666 return HRTIMER_NORESTART;
4669 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4671 struct cfs_bandwidth *cfs_b =
4672 container_of(timer, struct cfs_bandwidth, period_timer);
4676 raw_spin_lock(&cfs_b->lock);
4678 overrun = hrtimer_forward_now(timer, cfs_b->period);
4682 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4685 cfs_b->period_active = 0;
4686 raw_spin_unlock(&cfs_b->lock);
4688 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4691 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4693 raw_spin_lock_init(&cfs_b->lock);
4695 cfs_b->quota = RUNTIME_INF;
4696 cfs_b->period = ns_to_ktime(default_cfs_period());
4698 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4699 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4700 cfs_b->period_timer.function = sched_cfs_period_timer;
4701 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4702 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4705 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4707 cfs_rq->runtime_enabled = 0;
4708 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4711 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4713 lockdep_assert_held(&cfs_b->lock);
4715 if (!cfs_b->period_active) {
4716 cfs_b->period_active = 1;
4717 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4718 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4722 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4724 /* init_cfs_bandwidth() was not called */
4725 if (!cfs_b->throttled_cfs_rq.next)
4728 hrtimer_cancel(&cfs_b->period_timer);
4729 hrtimer_cancel(&cfs_b->slack_timer);
4733 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4735 * The race is harmless, since modifying bandwidth settings of unhooked group
4736 * bits doesn't do much.
4739 /* cpu online calback */
4740 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4742 struct task_group *tg;
4744 lockdep_assert_held(&rq->lock);
4747 list_for_each_entry_rcu(tg, &task_groups, list) {
4748 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4749 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4751 raw_spin_lock(&cfs_b->lock);
4752 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4753 raw_spin_unlock(&cfs_b->lock);
4758 /* cpu offline callback */
4759 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4761 struct task_group *tg;
4763 lockdep_assert_held(&rq->lock);
4766 list_for_each_entry_rcu(tg, &task_groups, list) {
4767 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4769 if (!cfs_rq->runtime_enabled)
4773 * clock_task is not advancing so we just need to make sure
4774 * there's some valid quota amount
4776 cfs_rq->runtime_remaining = 1;
4778 * Offline rq is schedulable till cpu is completely disabled
4779 * in take_cpu_down(), so we prevent new cfs throttling here.
4781 cfs_rq->runtime_enabled = 0;
4783 if (cfs_rq_throttled(cfs_rq))
4784 unthrottle_cfs_rq(cfs_rq);
4789 #else /* CONFIG_CFS_BANDWIDTH */
4790 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4792 return rq_clock_task(rq_of(cfs_rq));
4795 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4796 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4797 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4798 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4799 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4801 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4806 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4811 static inline int throttled_lb_pair(struct task_group *tg,
4812 int src_cpu, int dest_cpu)
4817 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4819 #ifdef CONFIG_FAIR_GROUP_SCHED
4820 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4823 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4827 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4828 static inline void update_runtime_enabled(struct rq *rq) {}
4829 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4831 #endif /* CONFIG_CFS_BANDWIDTH */
4833 /**************************************************
4834 * CFS operations on tasks:
4837 #ifdef CONFIG_SCHED_HRTICK
4838 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4840 struct sched_entity *se = &p->se;
4841 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4843 SCHED_WARN_ON(task_rq(p) != rq);
4845 if (rq->cfs.h_nr_running > 1) {
4846 u64 slice = sched_slice(cfs_rq, se);
4847 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4848 s64 delta = slice - ran;
4855 hrtick_start(rq, delta);
4860 * called from enqueue/dequeue and updates the hrtick when the
4861 * current task is from our class and nr_running is low enough
4864 static void hrtick_update(struct rq *rq)
4866 struct task_struct *curr = rq->curr;
4868 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4871 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4872 hrtick_start_fair(rq, curr);
4874 #else /* !CONFIG_SCHED_HRTICK */
4876 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4880 static inline void hrtick_update(struct rq *rq)
4886 * The enqueue_task method is called before nr_running is
4887 * increased. Here we update the fair scheduling stats and
4888 * then put the task into the rbtree:
4891 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4893 struct cfs_rq *cfs_rq;
4894 struct sched_entity *se = &p->se;
4897 * If in_iowait is set, the code below may not trigger any cpufreq
4898 * utilization updates, so do it here explicitly with the IOWAIT flag
4902 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
4904 for_each_sched_entity(se) {
4907 cfs_rq = cfs_rq_of(se);
4908 enqueue_entity(cfs_rq, se, flags);
4911 * end evaluation on encountering a throttled cfs_rq
4913 * note: in the case of encountering a throttled cfs_rq we will
4914 * post the final h_nr_running increment below.
4916 if (cfs_rq_throttled(cfs_rq))
4918 cfs_rq->h_nr_running++;
4920 flags = ENQUEUE_WAKEUP;
4923 for_each_sched_entity(se) {
4924 cfs_rq = cfs_rq_of(se);
4925 cfs_rq->h_nr_running++;
4927 if (cfs_rq_throttled(cfs_rq))
4930 update_load_avg(se, UPDATE_TG);
4931 update_cfs_shares(se);
4935 add_nr_running(rq, 1);
4940 static void set_next_buddy(struct sched_entity *se);
4943 * The dequeue_task method is called before nr_running is
4944 * decreased. We remove the task from the rbtree and
4945 * update the fair scheduling stats:
4947 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4949 struct cfs_rq *cfs_rq;
4950 struct sched_entity *se = &p->se;
4951 int task_sleep = flags & DEQUEUE_SLEEP;
4953 for_each_sched_entity(se) {
4954 cfs_rq = cfs_rq_of(se);
4955 dequeue_entity(cfs_rq, se, flags);
4958 * end evaluation on encountering a throttled cfs_rq
4960 * note: in the case of encountering a throttled cfs_rq we will
4961 * post the final h_nr_running decrement below.
4963 if (cfs_rq_throttled(cfs_rq))
4965 cfs_rq->h_nr_running--;
4967 /* Don't dequeue parent if it has other entities besides us */
4968 if (cfs_rq->load.weight) {
4969 /* Avoid re-evaluating load for this entity: */
4970 se = parent_entity(se);
4972 * Bias pick_next to pick a task from this cfs_rq, as
4973 * p is sleeping when it is within its sched_slice.
4975 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4979 flags |= DEQUEUE_SLEEP;
4982 for_each_sched_entity(se) {
4983 cfs_rq = cfs_rq_of(se);
4984 cfs_rq->h_nr_running--;
4986 if (cfs_rq_throttled(cfs_rq))
4989 update_load_avg(se, UPDATE_TG);
4990 update_cfs_shares(se);
4994 sub_nr_running(rq, 1);
5001 /* Working cpumask for: load_balance, load_balance_newidle. */
5002 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5003 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5005 #ifdef CONFIG_NO_HZ_COMMON
5007 * per rq 'load' arrray crap; XXX kill this.
5011 * The exact cpuload calculated at every tick would be:
5013 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5015 * If a cpu misses updates for n ticks (as it was idle) and update gets
5016 * called on the n+1-th tick when cpu may be busy, then we have:
5018 * load_n = (1 - 1/2^i)^n * load_0
5019 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5021 * decay_load_missed() below does efficient calculation of
5023 * load' = (1 - 1/2^i)^n * load
5025 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5026 * This allows us to precompute the above in said factors, thereby allowing the
5027 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5028 * fixed_power_int())
5030 * The calculation is approximated on a 128 point scale.
5032 #define DEGRADE_SHIFT 7
5034 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5035 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5036 { 0, 0, 0, 0, 0, 0, 0, 0 },
5037 { 64, 32, 8, 0, 0, 0, 0, 0 },
5038 { 96, 72, 40, 12, 1, 0, 0, 0 },
5039 { 112, 98, 75, 43, 15, 1, 0, 0 },
5040 { 120, 112, 98, 76, 45, 16, 2, 0 }
5044 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5045 * would be when CPU is idle and so we just decay the old load without
5046 * adding any new load.
5048 static unsigned long
5049 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5053 if (!missed_updates)
5056 if (missed_updates >= degrade_zero_ticks[idx])
5060 return load >> missed_updates;
5062 while (missed_updates) {
5063 if (missed_updates % 2)
5064 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5066 missed_updates >>= 1;
5071 #endif /* CONFIG_NO_HZ_COMMON */
5074 * __cpu_load_update - update the rq->cpu_load[] statistics
5075 * @this_rq: The rq to update statistics for
5076 * @this_load: The current load
5077 * @pending_updates: The number of missed updates
5079 * Update rq->cpu_load[] statistics. This function is usually called every
5080 * scheduler tick (TICK_NSEC).
5082 * This function computes a decaying average:
5084 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5086 * Because of NOHZ it might not get called on every tick which gives need for
5087 * the @pending_updates argument.
5089 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5090 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5091 * = A * (A * load[i]_n-2 + B) + B
5092 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5093 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5094 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5095 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5096 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5098 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5099 * any change in load would have resulted in the tick being turned back on.
5101 * For regular NOHZ, this reduces to:
5103 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5105 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5108 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5109 unsigned long pending_updates)
5111 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5114 this_rq->nr_load_updates++;
5116 /* Update our load: */
5117 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5118 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5119 unsigned long old_load, new_load;
5121 /* scale is effectively 1 << i now, and >> i divides by scale */
5123 old_load = this_rq->cpu_load[i];
5124 #ifdef CONFIG_NO_HZ_COMMON
5125 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5126 if (tickless_load) {
5127 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5129 * old_load can never be a negative value because a
5130 * decayed tickless_load cannot be greater than the
5131 * original tickless_load.
5133 old_load += tickless_load;
5136 new_load = this_load;
5138 * Round up the averaging division if load is increasing. This
5139 * prevents us from getting stuck on 9 if the load is 10, for
5142 if (new_load > old_load)
5143 new_load += scale - 1;
5145 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5148 sched_avg_update(this_rq);
5151 /* Used instead of source_load when we know the type == 0 */
5152 static unsigned long weighted_cpuload(struct rq *rq)
5154 return cfs_rq_runnable_load_avg(&rq->cfs);
5157 #ifdef CONFIG_NO_HZ_COMMON
5159 * There is no sane way to deal with nohz on smp when using jiffies because the
5160 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5161 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5163 * Therefore we need to avoid the delta approach from the regular tick when
5164 * possible since that would seriously skew the load calculation. This is why we
5165 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5166 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5167 * loop exit, nohz_idle_balance, nohz full exit...)
5169 * This means we might still be one tick off for nohz periods.
5172 static void cpu_load_update_nohz(struct rq *this_rq,
5173 unsigned long curr_jiffies,
5176 unsigned long pending_updates;
5178 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5179 if (pending_updates) {
5180 this_rq->last_load_update_tick = curr_jiffies;
5182 * In the regular NOHZ case, we were idle, this means load 0.
5183 * In the NOHZ_FULL case, we were non-idle, we should consider
5184 * its weighted load.
5186 cpu_load_update(this_rq, load, pending_updates);
5191 * Called from nohz_idle_balance() to update the load ratings before doing the
5194 static void cpu_load_update_idle(struct rq *this_rq)
5197 * bail if there's load or we're actually up-to-date.
5199 if (weighted_cpuload(this_rq))
5202 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5206 * Record CPU load on nohz entry so we know the tickless load to account
5207 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5208 * than other cpu_load[idx] but it should be fine as cpu_load readers
5209 * shouldn't rely into synchronized cpu_load[*] updates.
5211 void cpu_load_update_nohz_start(void)
5213 struct rq *this_rq = this_rq();
5216 * This is all lockless but should be fine. If weighted_cpuload changes
5217 * concurrently we'll exit nohz. And cpu_load write can race with
5218 * cpu_load_update_idle() but both updater would be writing the same.
5220 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5224 * Account the tickless load in the end of a nohz frame.
5226 void cpu_load_update_nohz_stop(void)
5228 unsigned long curr_jiffies = READ_ONCE(jiffies);
5229 struct rq *this_rq = this_rq();
5233 if (curr_jiffies == this_rq->last_load_update_tick)
5236 load = weighted_cpuload(this_rq);
5237 rq_lock(this_rq, &rf);
5238 update_rq_clock(this_rq);
5239 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5240 rq_unlock(this_rq, &rf);
5242 #else /* !CONFIG_NO_HZ_COMMON */
5243 static inline void cpu_load_update_nohz(struct rq *this_rq,
5244 unsigned long curr_jiffies,
5245 unsigned long load) { }
5246 #endif /* CONFIG_NO_HZ_COMMON */
5248 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5250 #ifdef CONFIG_NO_HZ_COMMON
5251 /* See the mess around cpu_load_update_nohz(). */
5252 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5254 cpu_load_update(this_rq, load, 1);
5258 * Called from scheduler_tick()
5260 void cpu_load_update_active(struct rq *this_rq)
5262 unsigned long load = weighted_cpuload(this_rq);
5264 if (tick_nohz_tick_stopped())
5265 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5267 cpu_load_update_periodic(this_rq, load);
5271 * Return a low guess at the load of a migration-source cpu weighted
5272 * according to the scheduling class and "nice" value.
5274 * We want to under-estimate the load of migration sources, to
5275 * balance conservatively.
5277 static unsigned long source_load(int cpu, int type)
5279 struct rq *rq = cpu_rq(cpu);
5280 unsigned long total = weighted_cpuload(rq);
5282 if (type == 0 || !sched_feat(LB_BIAS))
5285 return min(rq->cpu_load[type-1], total);
5289 * Return a high guess at the load of a migration-target cpu weighted
5290 * according to the scheduling class and "nice" value.
5292 static unsigned long target_load(int cpu, int type)
5294 struct rq *rq = cpu_rq(cpu);
5295 unsigned long total = weighted_cpuload(rq);
5297 if (type == 0 || !sched_feat(LB_BIAS))
5300 return max(rq->cpu_load[type-1], total);
5303 static unsigned long capacity_of(int cpu)
5305 return cpu_rq(cpu)->cpu_capacity;
5308 static unsigned long capacity_orig_of(int cpu)
5310 return cpu_rq(cpu)->cpu_capacity_orig;
5313 static unsigned long cpu_avg_load_per_task(int cpu)
5315 struct rq *rq = cpu_rq(cpu);
5316 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5317 unsigned long load_avg = weighted_cpuload(rq);
5320 return load_avg / nr_running;
5325 static void record_wakee(struct task_struct *p)
5328 * Only decay a single time; tasks that have less then 1 wakeup per
5329 * jiffy will not have built up many flips.
5331 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5332 current->wakee_flips >>= 1;
5333 current->wakee_flip_decay_ts = jiffies;
5336 if (current->last_wakee != p) {
5337 current->last_wakee = p;
5338 current->wakee_flips++;
5343 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5345 * A waker of many should wake a different task than the one last awakened
5346 * at a frequency roughly N times higher than one of its wakees.
5348 * In order to determine whether we should let the load spread vs consolidating
5349 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5350 * partner, and a factor of lls_size higher frequency in the other.
5352 * With both conditions met, we can be relatively sure that the relationship is
5353 * non-monogamous, with partner count exceeding socket size.
5355 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5356 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5359 static int wake_wide(struct task_struct *p)
5361 unsigned int master = current->wakee_flips;
5362 unsigned int slave = p->wakee_flips;
5363 int factor = this_cpu_read(sd_llc_size);
5366 swap(master, slave);
5367 if (slave < factor || master < slave * factor)
5373 unsigned long nr_running;
5375 unsigned long capacity;
5379 static bool get_llc_stats(struct llc_stats *stats, int cpu)
5381 struct sched_domain_shared *sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5386 stats->nr_running = READ_ONCE(sds->nr_running);
5387 stats->load = READ_ONCE(sds->load);
5388 stats->capacity = READ_ONCE(sds->capacity);
5389 stats->has_capacity = stats->nr_running < per_cpu(sd_llc_size, cpu);
5395 * Can a task be moved from prev_cpu to this_cpu without causing a load
5396 * imbalance that would trigger the load balancer?
5398 * Since we're running on 'stale' values, we might in fact create an imbalance
5399 * but recomputing these values is expensive, as that'd mean iteration 2 cache
5400 * domains worth of CPUs.
5403 wake_affine_llc(struct sched_domain *sd, struct task_struct *p,
5404 int this_cpu, int prev_cpu, int sync)
5406 struct llc_stats prev_stats, this_stats;
5407 s64 this_eff_load, prev_eff_load;
5408 unsigned long task_load;
5410 if (!get_llc_stats(&prev_stats, prev_cpu) ||
5411 !get_llc_stats(&this_stats, this_cpu))
5415 * If sync wakeup then subtract the (maximum possible)
5416 * effect of the currently running task from the load
5417 * of the current LLC.
5420 unsigned long current_load = task_h_load(current);
5422 /* in this case load hits 0 and this LLC is considered 'idle' */
5423 if (current_load > this_stats.load)
5426 this_stats.load -= current_load;
5430 * The has_capacity stuff is not SMT aware, but by trying to balance
5431 * the nr_running on both ends we try and fill the domain at equal
5432 * rates, thereby first consuming cores before siblings.
5435 /* if the old cache has capacity, stay there */
5436 if (prev_stats.has_capacity && prev_stats.nr_running < this_stats.nr_running+1)
5439 /* if this cache has capacity, come here */
5440 if (this_stats.has_capacity && this_stats.nr_running < prev_stats.nr_running+1)
5444 * Check to see if we can move the load without causing too much
5447 task_load = task_h_load(p);
5449 this_eff_load = 100;
5450 this_eff_load *= prev_stats.capacity;
5452 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5453 prev_eff_load *= this_stats.capacity;
5455 this_eff_load *= this_stats.load + task_load;
5456 prev_eff_load *= prev_stats.load - task_load;
5458 return this_eff_load <= prev_eff_load;
5461 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5462 int prev_cpu, int sync)
5464 int this_cpu = smp_processor_id();
5468 * Default to no affine wakeups; wake_affine() should not effect a task
5469 * placement the load-balancer feels inclined to undo. The conservative
5470 * option is therefore to not move tasks when they wake up.
5475 * If the wakeup is across cache domains, try to evaluate if movement
5476 * makes sense, otherwise rely on select_idle_siblings() to do
5477 * placement inside the cache domain.
5479 if (!cpus_share_cache(prev_cpu, this_cpu))
5480 affine = wake_affine_llc(sd, p, this_cpu, prev_cpu, sync);
5482 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5484 schedstat_inc(sd->ttwu_move_affine);
5485 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5491 static inline int task_util(struct task_struct *p);
5492 static int cpu_util_wake(int cpu, struct task_struct *p);
5494 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5496 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5500 * find_idlest_group finds and returns the least busy CPU group within the
5503 static struct sched_group *
5504 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5505 int this_cpu, int sd_flag)
5507 struct sched_group *idlest = NULL, *group = sd->groups;
5508 struct sched_group *most_spare_sg = NULL;
5509 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5510 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5511 unsigned long most_spare = 0, this_spare = 0;
5512 int load_idx = sd->forkexec_idx;
5513 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5514 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5515 (sd->imbalance_pct-100) / 100;
5517 if (sd_flag & SD_BALANCE_WAKE)
5518 load_idx = sd->wake_idx;
5521 unsigned long load, avg_load, runnable_load;
5522 unsigned long spare_cap, max_spare_cap;
5526 /* Skip over this group if it has no CPUs allowed */
5527 if (!cpumask_intersects(sched_group_span(group),
5531 local_group = cpumask_test_cpu(this_cpu,
5532 sched_group_span(group));
5535 * Tally up the load of all CPUs in the group and find
5536 * the group containing the CPU with most spare capacity.
5542 for_each_cpu(i, sched_group_span(group)) {
5543 /* Bias balancing toward cpus of our domain */
5545 load = source_load(i, load_idx);
5547 load = target_load(i, load_idx);
5549 runnable_load += load;
5551 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5553 spare_cap = capacity_spare_wake(i, p);
5555 if (spare_cap > max_spare_cap)
5556 max_spare_cap = spare_cap;
5559 /* Adjust by relative CPU capacity of the group */
5560 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5561 group->sgc->capacity;
5562 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5563 group->sgc->capacity;
5566 this_runnable_load = runnable_load;
5567 this_avg_load = avg_load;
5568 this_spare = max_spare_cap;
5570 if (min_runnable_load > (runnable_load + imbalance)) {
5572 * The runnable load is significantly smaller
5573 * so we can pick this new cpu
5575 min_runnable_load = runnable_load;
5576 min_avg_load = avg_load;
5578 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5579 (100*min_avg_load > imbalance_scale*avg_load)) {
5581 * The runnable loads are close so take the
5582 * blocked load into account through avg_load.
5584 min_avg_load = avg_load;
5588 if (most_spare < max_spare_cap) {
5589 most_spare = max_spare_cap;
5590 most_spare_sg = group;
5593 } while (group = group->next, group != sd->groups);
5596 * The cross-over point between using spare capacity or least load
5597 * is too conservative for high utilization tasks on partially
5598 * utilized systems if we require spare_capacity > task_util(p),
5599 * so we allow for some task stuffing by using
5600 * spare_capacity > task_util(p)/2.
5602 * Spare capacity can't be used for fork because the utilization has
5603 * not been set yet, we must first select a rq to compute the initial
5606 if (sd_flag & SD_BALANCE_FORK)
5609 if (this_spare > task_util(p) / 2 &&
5610 imbalance_scale*this_spare > 100*most_spare)
5613 if (most_spare > task_util(p) / 2)
5614 return most_spare_sg;
5620 if (min_runnable_load > (this_runnable_load + imbalance))
5623 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5624 (100*this_avg_load < imbalance_scale*min_avg_load))
5631 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5634 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5636 unsigned long load, min_load = ULONG_MAX;
5637 unsigned int min_exit_latency = UINT_MAX;
5638 u64 latest_idle_timestamp = 0;
5639 int least_loaded_cpu = this_cpu;
5640 int shallowest_idle_cpu = -1;
5643 /* Check if we have any choice: */
5644 if (group->group_weight == 1)
5645 return cpumask_first(sched_group_span(group));
5647 /* Traverse only the allowed CPUs */
5648 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5650 struct rq *rq = cpu_rq(i);
5651 struct cpuidle_state *idle = idle_get_state(rq);
5652 if (idle && idle->exit_latency < min_exit_latency) {
5654 * We give priority to a CPU whose idle state
5655 * has the smallest exit latency irrespective
5656 * of any idle timestamp.
5658 min_exit_latency = idle->exit_latency;
5659 latest_idle_timestamp = rq->idle_stamp;
5660 shallowest_idle_cpu = i;
5661 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5662 rq->idle_stamp > latest_idle_timestamp) {
5664 * If equal or no active idle state, then
5665 * the most recently idled CPU might have
5668 latest_idle_timestamp = rq->idle_stamp;
5669 shallowest_idle_cpu = i;
5671 } else if (shallowest_idle_cpu == -1) {
5672 load = weighted_cpuload(cpu_rq(i));
5673 if (load < min_load || (load == min_load && i == this_cpu)) {
5675 least_loaded_cpu = i;
5680 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5683 #ifdef CONFIG_SCHED_SMT
5685 static inline void set_idle_cores(int cpu, int val)
5687 struct sched_domain_shared *sds;
5689 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5691 WRITE_ONCE(sds->has_idle_cores, val);
5694 static inline bool test_idle_cores(int cpu, bool def)
5696 struct sched_domain_shared *sds;
5698 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5700 return READ_ONCE(sds->has_idle_cores);
5706 * Scans the local SMT mask to see if the entire core is idle, and records this
5707 * information in sd_llc_shared->has_idle_cores.
5709 * Since SMT siblings share all cache levels, inspecting this limited remote
5710 * state should be fairly cheap.
5712 void __update_idle_core(struct rq *rq)
5714 int core = cpu_of(rq);
5718 if (test_idle_cores(core, true))
5721 for_each_cpu(cpu, cpu_smt_mask(core)) {
5729 set_idle_cores(core, 1);
5735 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5736 * there are no idle cores left in the system; tracked through
5737 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5739 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5741 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5744 if (!static_branch_likely(&sched_smt_present))
5747 if (!test_idle_cores(target, false))
5750 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5752 for_each_cpu_wrap(core, cpus, target) {
5755 for_each_cpu(cpu, cpu_smt_mask(core)) {
5756 cpumask_clear_cpu(cpu, cpus);
5766 * Failed to find an idle core; stop looking for one.
5768 set_idle_cores(target, 0);
5774 * Scan the local SMT mask for idle CPUs.
5776 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5780 if (!static_branch_likely(&sched_smt_present))
5783 for_each_cpu(cpu, cpu_smt_mask(target)) {
5784 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5793 #else /* CONFIG_SCHED_SMT */
5795 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5800 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5805 #endif /* CONFIG_SCHED_SMT */
5808 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5809 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5810 * average idle time for this rq (as found in rq->avg_idle).
5812 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5814 struct sched_domain *this_sd;
5815 u64 avg_cost, avg_idle;
5818 int cpu, nr = INT_MAX;
5820 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5825 * Due to large variance we need a large fuzz factor; hackbench in
5826 * particularly is sensitive here.
5828 avg_idle = this_rq()->avg_idle / 512;
5829 avg_cost = this_sd->avg_scan_cost + 1;
5831 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5834 if (sched_feat(SIS_PROP)) {
5835 u64 span_avg = sd->span_weight * avg_idle;
5836 if (span_avg > 4*avg_cost)
5837 nr = div_u64(span_avg, avg_cost);
5842 time = local_clock();
5844 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5847 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5853 time = local_clock() - time;
5854 cost = this_sd->avg_scan_cost;
5855 delta = (s64)(time - cost) / 8;
5856 this_sd->avg_scan_cost += delta;
5862 * Try and locate an idle core/thread in the LLC cache domain.
5864 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5866 struct sched_domain *sd;
5869 if (idle_cpu(target))
5873 * If the previous cpu is cache affine and idle, don't be stupid.
5875 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5878 sd = rcu_dereference(per_cpu(sd_llc, target));
5882 i = select_idle_core(p, sd, target);
5883 if ((unsigned)i < nr_cpumask_bits)
5886 i = select_idle_cpu(p, sd, target);
5887 if ((unsigned)i < nr_cpumask_bits)
5890 i = select_idle_smt(p, sd, target);
5891 if ((unsigned)i < nr_cpumask_bits)
5898 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5899 * tasks. The unit of the return value must be the one of capacity so we can
5900 * compare the utilization with the capacity of the CPU that is available for
5901 * CFS task (ie cpu_capacity).
5903 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5904 * recent utilization of currently non-runnable tasks on a CPU. It represents
5905 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5906 * capacity_orig is the cpu_capacity available at the highest frequency
5907 * (arch_scale_freq_capacity()).
5908 * The utilization of a CPU converges towards a sum equal to or less than the
5909 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5910 * the running time on this CPU scaled by capacity_curr.
5912 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5913 * higher than capacity_orig because of unfortunate rounding in
5914 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5915 * the average stabilizes with the new running time. We need to check that the
5916 * utilization stays within the range of [0..capacity_orig] and cap it if
5917 * necessary. Without utilization capping, a group could be seen as overloaded
5918 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5919 * available capacity. We allow utilization to overshoot capacity_curr (but not
5920 * capacity_orig) as it useful for predicting the capacity required after task
5921 * migrations (scheduler-driven DVFS).
5923 static int cpu_util(int cpu)
5925 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5926 unsigned long capacity = capacity_orig_of(cpu);
5928 return (util >= capacity) ? capacity : util;
5931 static inline int task_util(struct task_struct *p)
5933 return p->se.avg.util_avg;
5937 * cpu_util_wake: Compute cpu utilization with any contributions from
5938 * the waking task p removed.
5940 static int cpu_util_wake(int cpu, struct task_struct *p)
5942 unsigned long util, capacity;
5944 /* Task has no contribution or is new */
5945 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5946 return cpu_util(cpu);
5948 capacity = capacity_orig_of(cpu);
5949 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5951 return (util >= capacity) ? capacity : util;
5955 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5956 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5958 * In that case WAKE_AFFINE doesn't make sense and we'll let
5959 * BALANCE_WAKE sort things out.
5961 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5963 long min_cap, max_cap;
5965 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5966 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5968 /* Minimum capacity is close to max, no need to abort wake_affine */
5969 if (max_cap - min_cap < max_cap >> 3)
5972 /* Bring task utilization in sync with prev_cpu */
5973 sync_entity_load_avg(&p->se);
5975 return min_cap * 1024 < task_util(p) * capacity_margin;
5979 * select_task_rq_fair: Select target runqueue for the waking task in domains
5980 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5981 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5983 * Balances load by selecting the idlest cpu in the idlest group, or under
5984 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5986 * Returns the target cpu number.
5988 * preempt must be disabled.
5991 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5993 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5994 int cpu = smp_processor_id();
5995 int new_cpu = prev_cpu;
5996 int want_affine = 0;
5997 int sync = wake_flags & WF_SYNC;
5999 if (sd_flag & SD_BALANCE_WAKE) {
6001 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6002 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6006 for_each_domain(cpu, tmp) {
6007 if (!(tmp->flags & SD_LOAD_BALANCE))
6011 * If both cpu and prev_cpu are part of this domain,
6012 * cpu is a valid SD_WAKE_AFFINE target.
6014 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6015 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6020 if (tmp->flags & sd_flag)
6022 else if (!want_affine)
6027 sd = NULL; /* Prefer wake_affine over balance flags */
6028 if (cpu == prev_cpu)
6031 if (wake_affine(affine_sd, p, prev_cpu, sync))
6037 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6038 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6041 struct sched_group *group;
6044 if (!(sd->flags & sd_flag)) {
6049 group = find_idlest_group(sd, p, cpu, sd_flag);
6055 new_cpu = find_idlest_cpu(group, p, cpu);
6056 if (new_cpu == -1 || new_cpu == cpu) {
6057 /* Now try balancing at a lower domain level of cpu */
6062 /* Now try balancing at a lower domain level of new_cpu */
6064 weight = sd->span_weight;
6066 for_each_domain(cpu, tmp) {
6067 if (weight <= tmp->span_weight)
6069 if (tmp->flags & sd_flag)
6072 /* while loop will break here if sd == NULL */
6080 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6081 * cfs_rq_of(p) references at time of call are still valid and identify the
6082 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6084 static void migrate_task_rq_fair(struct task_struct *p)
6087 * As blocked tasks retain absolute vruntime the migration needs to
6088 * deal with this by subtracting the old and adding the new
6089 * min_vruntime -- the latter is done by enqueue_entity() when placing
6090 * the task on the new runqueue.
6092 if (p->state == TASK_WAKING) {
6093 struct sched_entity *se = &p->se;
6094 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6097 #ifndef CONFIG_64BIT
6098 u64 min_vruntime_copy;
6101 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6103 min_vruntime = cfs_rq->min_vruntime;
6104 } while (min_vruntime != min_vruntime_copy);
6106 min_vruntime = cfs_rq->min_vruntime;
6109 se->vruntime -= min_vruntime;
6113 * We are supposed to update the task to "current" time, then its up to date
6114 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6115 * what current time is, so simply throw away the out-of-date time. This
6116 * will result in the wakee task is less decayed, but giving the wakee more
6117 * load sounds not bad.
6119 remove_entity_load_avg(&p->se);
6121 /* Tell new CPU we are migrated */
6122 p->se.avg.last_update_time = 0;
6124 /* We have migrated, no longer consider this task hot */
6125 p->se.exec_start = 0;
6128 static void task_dead_fair(struct task_struct *p)
6130 remove_entity_load_avg(&p->se);
6132 #endif /* CONFIG_SMP */
6134 static unsigned long
6135 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6137 unsigned long gran = sysctl_sched_wakeup_granularity;
6140 * Since its curr running now, convert the gran from real-time
6141 * to virtual-time in his units.
6143 * By using 'se' instead of 'curr' we penalize light tasks, so
6144 * they get preempted easier. That is, if 'se' < 'curr' then
6145 * the resulting gran will be larger, therefore penalizing the
6146 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6147 * be smaller, again penalizing the lighter task.
6149 * This is especially important for buddies when the leftmost
6150 * task is higher priority than the buddy.
6152 return calc_delta_fair(gran, se);
6156 * Should 'se' preempt 'curr'.
6170 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6172 s64 gran, vdiff = curr->vruntime - se->vruntime;
6177 gran = wakeup_gran(curr, se);
6184 static void set_last_buddy(struct sched_entity *se)
6186 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6189 for_each_sched_entity(se) {
6190 if (SCHED_WARN_ON(!se->on_rq))
6192 cfs_rq_of(se)->last = se;
6196 static void set_next_buddy(struct sched_entity *se)
6198 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6201 for_each_sched_entity(se) {
6202 if (SCHED_WARN_ON(!se->on_rq))
6204 cfs_rq_of(se)->next = se;
6208 static void set_skip_buddy(struct sched_entity *se)
6210 for_each_sched_entity(se)
6211 cfs_rq_of(se)->skip = se;
6215 * Preempt the current task with a newly woken task if needed:
6217 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6219 struct task_struct *curr = rq->curr;
6220 struct sched_entity *se = &curr->se, *pse = &p->se;
6221 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6222 int scale = cfs_rq->nr_running >= sched_nr_latency;
6223 int next_buddy_marked = 0;
6225 if (unlikely(se == pse))
6229 * This is possible from callers such as attach_tasks(), in which we
6230 * unconditionally check_prempt_curr() after an enqueue (which may have
6231 * lead to a throttle). This both saves work and prevents false
6232 * next-buddy nomination below.
6234 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6237 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6238 set_next_buddy(pse);
6239 next_buddy_marked = 1;
6243 * We can come here with TIF_NEED_RESCHED already set from new task
6246 * Note: this also catches the edge-case of curr being in a throttled
6247 * group (e.g. via set_curr_task), since update_curr() (in the
6248 * enqueue of curr) will have resulted in resched being set. This
6249 * prevents us from potentially nominating it as a false LAST_BUDDY
6252 if (test_tsk_need_resched(curr))
6255 /* Idle tasks are by definition preempted by non-idle tasks. */
6256 if (unlikely(curr->policy == SCHED_IDLE) &&
6257 likely(p->policy != SCHED_IDLE))
6261 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6262 * is driven by the tick):
6264 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6267 find_matching_se(&se, &pse);
6268 update_curr(cfs_rq_of(se));
6270 if (wakeup_preempt_entity(se, pse) == 1) {
6272 * Bias pick_next to pick the sched entity that is
6273 * triggering this preemption.
6275 if (!next_buddy_marked)
6276 set_next_buddy(pse);
6285 * Only set the backward buddy when the current task is still
6286 * on the rq. This can happen when a wakeup gets interleaved
6287 * with schedule on the ->pre_schedule() or idle_balance()
6288 * point, either of which can * drop the rq lock.
6290 * Also, during early boot the idle thread is in the fair class,
6291 * for obvious reasons its a bad idea to schedule back to it.
6293 if (unlikely(!se->on_rq || curr == rq->idle))
6296 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6300 static struct task_struct *
6301 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6303 struct cfs_rq *cfs_rq = &rq->cfs;
6304 struct sched_entity *se;
6305 struct task_struct *p;
6309 if (!cfs_rq->nr_running)
6312 #ifdef CONFIG_FAIR_GROUP_SCHED
6313 if (prev->sched_class != &fair_sched_class)
6317 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6318 * likely that a next task is from the same cgroup as the current.
6320 * Therefore attempt to avoid putting and setting the entire cgroup
6321 * hierarchy, only change the part that actually changes.
6325 struct sched_entity *curr = cfs_rq->curr;
6328 * Since we got here without doing put_prev_entity() we also
6329 * have to consider cfs_rq->curr. If it is still a runnable
6330 * entity, update_curr() will update its vruntime, otherwise
6331 * forget we've ever seen it.
6335 update_curr(cfs_rq);
6340 * This call to check_cfs_rq_runtime() will do the
6341 * throttle and dequeue its entity in the parent(s).
6342 * Therefore the nr_running test will indeed
6345 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6348 if (!cfs_rq->nr_running)
6355 se = pick_next_entity(cfs_rq, curr);
6356 cfs_rq = group_cfs_rq(se);
6362 * Since we haven't yet done put_prev_entity and if the selected task
6363 * is a different task than we started out with, try and touch the
6364 * least amount of cfs_rqs.
6367 struct sched_entity *pse = &prev->se;
6369 while (!(cfs_rq = is_same_group(se, pse))) {
6370 int se_depth = se->depth;
6371 int pse_depth = pse->depth;
6373 if (se_depth <= pse_depth) {
6374 put_prev_entity(cfs_rq_of(pse), pse);
6375 pse = parent_entity(pse);
6377 if (se_depth >= pse_depth) {
6378 set_next_entity(cfs_rq_of(se), se);
6379 se = parent_entity(se);
6383 put_prev_entity(cfs_rq, pse);
6384 set_next_entity(cfs_rq, se);
6387 if (hrtick_enabled(rq))
6388 hrtick_start_fair(rq, p);
6394 put_prev_task(rq, prev);
6397 se = pick_next_entity(cfs_rq, NULL);
6398 set_next_entity(cfs_rq, se);
6399 cfs_rq = group_cfs_rq(se);
6404 if (hrtick_enabled(rq))
6405 hrtick_start_fair(rq, p);
6410 new_tasks = idle_balance(rq, rf);
6413 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6414 * possible for any higher priority task to appear. In that case we
6415 * must re-start the pick_next_entity() loop.
6427 * Account for a descheduled task:
6429 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6431 struct sched_entity *se = &prev->se;
6432 struct cfs_rq *cfs_rq;
6434 for_each_sched_entity(se) {
6435 cfs_rq = cfs_rq_of(se);
6436 put_prev_entity(cfs_rq, se);
6441 * sched_yield() is very simple
6443 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6445 static void yield_task_fair(struct rq *rq)
6447 struct task_struct *curr = rq->curr;
6448 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6449 struct sched_entity *se = &curr->se;
6452 * Are we the only task in the tree?
6454 if (unlikely(rq->nr_running == 1))
6457 clear_buddies(cfs_rq, se);
6459 if (curr->policy != SCHED_BATCH) {
6460 update_rq_clock(rq);
6462 * Update run-time statistics of the 'current'.
6464 update_curr(cfs_rq);
6466 * Tell update_rq_clock() that we've just updated,
6467 * so we don't do microscopic update in schedule()
6468 * and double the fastpath cost.
6470 rq_clock_skip_update(rq, true);
6476 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6478 struct sched_entity *se = &p->se;
6480 /* throttled hierarchies are not runnable */
6481 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6484 /* Tell the scheduler that we'd really like pse to run next. */
6487 yield_task_fair(rq);
6493 /**************************************************
6494 * Fair scheduling class load-balancing methods.
6498 * The purpose of load-balancing is to achieve the same basic fairness the
6499 * per-cpu scheduler provides, namely provide a proportional amount of compute
6500 * time to each task. This is expressed in the following equation:
6502 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6504 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6505 * W_i,0 is defined as:
6507 * W_i,0 = \Sum_j w_i,j (2)
6509 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6510 * is derived from the nice value as per sched_prio_to_weight[].
6512 * The weight average is an exponential decay average of the instantaneous
6515 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6517 * C_i is the compute capacity of cpu i, typically it is the
6518 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6519 * can also include other factors [XXX].
6521 * To achieve this balance we define a measure of imbalance which follows
6522 * directly from (1):
6524 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6526 * We them move tasks around to minimize the imbalance. In the continuous
6527 * function space it is obvious this converges, in the discrete case we get
6528 * a few fun cases generally called infeasible weight scenarios.
6531 * - infeasible weights;
6532 * - local vs global optima in the discrete case. ]
6537 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6538 * for all i,j solution, we create a tree of cpus that follows the hardware
6539 * topology where each level pairs two lower groups (or better). This results
6540 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6541 * tree to only the first of the previous level and we decrease the frequency
6542 * of load-balance at each level inv. proportional to the number of cpus in
6548 * \Sum { --- * --- * 2^i } = O(n) (5)
6550 * `- size of each group
6551 * | | `- number of cpus doing load-balance
6553 * `- sum over all levels
6555 * Coupled with a limit on how many tasks we can migrate every balance pass,
6556 * this makes (5) the runtime complexity of the balancer.
6558 * An important property here is that each CPU is still (indirectly) connected
6559 * to every other cpu in at most O(log n) steps:
6561 * The adjacency matrix of the resulting graph is given by:
6564 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6567 * And you'll find that:
6569 * A^(log_2 n)_i,j != 0 for all i,j (7)
6571 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6572 * The task movement gives a factor of O(m), giving a convergence complexity
6575 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6580 * In order to avoid CPUs going idle while there's still work to do, new idle
6581 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6582 * tree itself instead of relying on other CPUs to bring it work.
6584 * This adds some complexity to both (5) and (8) but it reduces the total idle
6592 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6595 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6600 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6602 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6604 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6607 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6608 * rewrite all of this once again.]
6611 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6613 enum fbq_type { regular, remote, all };
6615 #define LBF_ALL_PINNED 0x01
6616 #define LBF_NEED_BREAK 0x02
6617 #define LBF_DST_PINNED 0x04
6618 #define LBF_SOME_PINNED 0x08
6621 struct sched_domain *sd;
6629 struct cpumask *dst_grpmask;
6631 enum cpu_idle_type idle;
6633 /* The set of CPUs under consideration for load-balancing */
6634 struct cpumask *cpus;
6639 unsigned int loop_break;
6640 unsigned int loop_max;
6642 enum fbq_type fbq_type;
6643 struct list_head tasks;
6647 * Is this task likely cache-hot:
6649 static int task_hot(struct task_struct *p, struct lb_env *env)
6653 lockdep_assert_held(&env->src_rq->lock);
6655 if (p->sched_class != &fair_sched_class)
6658 if (unlikely(p->policy == SCHED_IDLE))
6662 * Buddy candidates are cache hot:
6664 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6665 (&p->se == cfs_rq_of(&p->se)->next ||
6666 &p->se == cfs_rq_of(&p->se)->last))
6669 if (sysctl_sched_migration_cost == -1)
6671 if (sysctl_sched_migration_cost == 0)
6674 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6676 return delta < (s64)sysctl_sched_migration_cost;
6679 #ifdef CONFIG_NUMA_BALANCING
6681 * Returns 1, if task migration degrades locality
6682 * Returns 0, if task migration improves locality i.e migration preferred.
6683 * Returns -1, if task migration is not affected by locality.
6685 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6687 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6688 unsigned long src_faults, dst_faults;
6689 int src_nid, dst_nid;
6691 if (!static_branch_likely(&sched_numa_balancing))
6694 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6697 src_nid = cpu_to_node(env->src_cpu);
6698 dst_nid = cpu_to_node(env->dst_cpu);
6700 if (src_nid == dst_nid)
6703 /* Migrating away from the preferred node is always bad. */
6704 if (src_nid == p->numa_preferred_nid) {
6705 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6711 /* Encourage migration to the preferred node. */
6712 if (dst_nid == p->numa_preferred_nid)
6715 /* Leaving a core idle is often worse than degrading locality. */
6716 if (env->idle != CPU_NOT_IDLE)
6720 src_faults = group_faults(p, src_nid);
6721 dst_faults = group_faults(p, dst_nid);
6723 src_faults = task_faults(p, src_nid);
6724 dst_faults = task_faults(p, dst_nid);
6727 return dst_faults < src_faults;
6731 static inline int migrate_degrades_locality(struct task_struct *p,
6739 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6742 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6746 lockdep_assert_held(&env->src_rq->lock);
6749 * We do not migrate tasks that are:
6750 * 1) throttled_lb_pair, or
6751 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6752 * 3) running (obviously), or
6753 * 4) are cache-hot on their current CPU.
6755 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6758 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6761 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6763 env->flags |= LBF_SOME_PINNED;
6766 * Remember if this task can be migrated to any other cpu in
6767 * our sched_group. We may want to revisit it if we couldn't
6768 * meet load balance goals by pulling other tasks on src_cpu.
6770 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6771 * already computed one in current iteration.
6773 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6776 /* Prevent to re-select dst_cpu via env's cpus */
6777 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6778 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6779 env->flags |= LBF_DST_PINNED;
6780 env->new_dst_cpu = cpu;
6788 /* Record that we found atleast one task that could run on dst_cpu */
6789 env->flags &= ~LBF_ALL_PINNED;
6791 if (task_running(env->src_rq, p)) {
6792 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6797 * Aggressive migration if:
6798 * 1) destination numa is preferred
6799 * 2) task is cache cold, or
6800 * 3) too many balance attempts have failed.
6802 tsk_cache_hot = migrate_degrades_locality(p, env);
6803 if (tsk_cache_hot == -1)
6804 tsk_cache_hot = task_hot(p, env);
6806 if (tsk_cache_hot <= 0 ||
6807 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6808 if (tsk_cache_hot == 1) {
6809 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6810 schedstat_inc(p->se.statistics.nr_forced_migrations);
6815 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6820 * detach_task() -- detach the task for the migration specified in env
6822 static void detach_task(struct task_struct *p, struct lb_env *env)
6824 lockdep_assert_held(&env->src_rq->lock);
6826 p->on_rq = TASK_ON_RQ_MIGRATING;
6827 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6828 set_task_cpu(p, env->dst_cpu);
6832 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6833 * part of active balancing operations within "domain".
6835 * Returns a task if successful and NULL otherwise.
6837 static struct task_struct *detach_one_task(struct lb_env *env)
6839 struct task_struct *p, *n;
6841 lockdep_assert_held(&env->src_rq->lock);
6843 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6844 if (!can_migrate_task(p, env))
6847 detach_task(p, env);
6850 * Right now, this is only the second place where
6851 * lb_gained[env->idle] is updated (other is detach_tasks)
6852 * so we can safely collect stats here rather than
6853 * inside detach_tasks().
6855 schedstat_inc(env->sd->lb_gained[env->idle]);
6861 static const unsigned int sched_nr_migrate_break = 32;
6864 * detach_tasks() -- tries to detach up to imbalance weighted load from
6865 * busiest_rq, as part of a balancing operation within domain "sd".
6867 * Returns number of detached tasks if successful and 0 otherwise.
6869 static int detach_tasks(struct lb_env *env)
6871 struct list_head *tasks = &env->src_rq->cfs_tasks;
6872 struct task_struct *p;
6876 lockdep_assert_held(&env->src_rq->lock);
6878 if (env->imbalance <= 0)
6881 while (!list_empty(tasks)) {
6883 * We don't want to steal all, otherwise we may be treated likewise,
6884 * which could at worst lead to a livelock crash.
6886 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6889 p = list_first_entry(tasks, struct task_struct, se.group_node);
6892 /* We've more or less seen every task there is, call it quits */
6893 if (env->loop > env->loop_max)
6896 /* take a breather every nr_migrate tasks */
6897 if (env->loop > env->loop_break) {
6898 env->loop_break += sched_nr_migrate_break;
6899 env->flags |= LBF_NEED_BREAK;
6903 if (!can_migrate_task(p, env))
6906 load = task_h_load(p);
6908 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6911 if ((load / 2) > env->imbalance)
6914 detach_task(p, env);
6915 list_add(&p->se.group_node, &env->tasks);
6918 env->imbalance -= load;
6920 #ifdef CONFIG_PREEMPT
6922 * NEWIDLE balancing is a source of latency, so preemptible
6923 * kernels will stop after the first task is detached to minimize
6924 * the critical section.
6926 if (env->idle == CPU_NEWLY_IDLE)
6931 * We only want to steal up to the prescribed amount of
6934 if (env->imbalance <= 0)
6939 list_move_tail(&p->se.group_node, tasks);
6943 * Right now, this is one of only two places we collect this stat
6944 * so we can safely collect detach_one_task() stats here rather
6945 * than inside detach_one_task().
6947 schedstat_add(env->sd->lb_gained[env->idle], detached);
6953 * attach_task() -- attach the task detached by detach_task() to its new rq.
6955 static void attach_task(struct rq *rq, struct task_struct *p)
6957 lockdep_assert_held(&rq->lock);
6959 BUG_ON(task_rq(p) != rq);
6960 activate_task(rq, p, ENQUEUE_NOCLOCK);
6961 p->on_rq = TASK_ON_RQ_QUEUED;
6962 check_preempt_curr(rq, p, 0);
6966 * attach_one_task() -- attaches the task returned from detach_one_task() to
6969 static void attach_one_task(struct rq *rq, struct task_struct *p)
6974 update_rq_clock(rq);
6980 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6983 static void attach_tasks(struct lb_env *env)
6985 struct list_head *tasks = &env->tasks;
6986 struct task_struct *p;
6989 rq_lock(env->dst_rq, &rf);
6990 update_rq_clock(env->dst_rq);
6992 while (!list_empty(tasks)) {
6993 p = list_first_entry(tasks, struct task_struct, se.group_node);
6994 list_del_init(&p->se.group_node);
6996 attach_task(env->dst_rq, p);
6999 rq_unlock(env->dst_rq, &rf);
7002 #ifdef CONFIG_FAIR_GROUP_SCHED
7004 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7006 if (cfs_rq->load.weight)
7009 if (cfs_rq->avg.load_sum)
7012 if (cfs_rq->avg.util_sum)
7015 if (cfs_rq->runnable_load_sum)
7021 static void update_blocked_averages(int cpu)
7023 struct rq *rq = cpu_rq(cpu);
7024 struct cfs_rq *cfs_rq, *pos;
7027 rq_lock_irqsave(rq, &rf);
7028 update_rq_clock(rq);
7031 * Iterates the task_group tree in a bottom up fashion, see
7032 * list_add_leaf_cfs_rq() for details.
7034 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7035 struct sched_entity *se;
7037 /* throttled entities do not contribute to load */
7038 if (throttled_hierarchy(cfs_rq))
7041 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7042 update_tg_load_avg(cfs_rq, 0);
7044 /* Propagate pending load changes to the parent, if any: */
7045 se = cfs_rq->tg->se[cpu];
7046 if (se && !skip_blocked_update(se))
7047 update_load_avg(se, 0);
7050 * There can be a lot of idle CPU cgroups. Don't let fully
7051 * decayed cfs_rqs linger on the list.
7053 if (cfs_rq_is_decayed(cfs_rq))
7054 list_del_leaf_cfs_rq(cfs_rq);
7056 rq_unlock_irqrestore(rq, &rf);
7060 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7061 * This needs to be done in a top-down fashion because the load of a child
7062 * group is a fraction of its parents load.
7064 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7066 struct rq *rq = rq_of(cfs_rq);
7067 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7068 unsigned long now = jiffies;
7071 if (cfs_rq->last_h_load_update == now)
7074 cfs_rq->h_load_next = NULL;
7075 for_each_sched_entity(se) {
7076 cfs_rq = cfs_rq_of(se);
7077 cfs_rq->h_load_next = se;
7078 if (cfs_rq->last_h_load_update == now)
7083 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7084 cfs_rq->last_h_load_update = now;
7087 while ((se = cfs_rq->h_load_next) != NULL) {
7088 load = cfs_rq->h_load;
7089 load = div64_ul(load * se->avg.load_avg,
7090 cfs_rq_load_avg(cfs_rq) + 1);
7091 cfs_rq = group_cfs_rq(se);
7092 cfs_rq->h_load = load;
7093 cfs_rq->last_h_load_update = now;
7097 static unsigned long task_h_load(struct task_struct *p)
7099 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7101 update_cfs_rq_h_load(cfs_rq);
7102 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7103 cfs_rq_load_avg(cfs_rq) + 1);
7106 static inline void update_blocked_averages(int cpu)
7108 struct rq *rq = cpu_rq(cpu);
7109 struct cfs_rq *cfs_rq = &rq->cfs;
7112 rq_lock_irqsave(rq, &rf);
7113 update_rq_clock(rq);
7114 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7115 rq_unlock_irqrestore(rq, &rf);
7118 static unsigned long task_h_load(struct task_struct *p)
7120 return p->se.avg.load_avg;
7124 /********** Helpers for find_busiest_group ************************/
7133 * sg_lb_stats - stats of a sched_group required for load_balancing
7135 struct sg_lb_stats {
7136 unsigned long avg_load; /*Avg load across the CPUs of the group */
7137 unsigned long group_load; /* Total load over the CPUs of the group */
7138 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7139 unsigned long load_per_task;
7140 unsigned long group_capacity;
7141 unsigned long group_util; /* Total utilization of the group */
7142 unsigned int sum_nr_running; /* Nr tasks running in the group */
7143 unsigned int idle_cpus;
7144 unsigned int group_weight;
7145 enum group_type group_type;
7146 int group_no_capacity;
7147 #ifdef CONFIG_NUMA_BALANCING
7148 unsigned int nr_numa_running;
7149 unsigned int nr_preferred_running;
7154 * sd_lb_stats - Structure to store the statistics of a sched_domain
7155 * during load balancing.
7157 struct sd_lb_stats {
7158 struct sched_group *busiest; /* Busiest group in this sd */
7159 struct sched_group *local; /* Local group in this sd */
7160 unsigned long total_running;
7161 unsigned long total_load; /* Total load of all groups in sd */
7162 unsigned long total_capacity; /* Total capacity of all groups in sd */
7163 unsigned long avg_load; /* Average load across all groups in sd */
7165 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7166 struct sg_lb_stats local_stat; /* Statistics of the local group */
7169 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7172 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7173 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7174 * We must however clear busiest_stat::avg_load because
7175 * update_sd_pick_busiest() reads this before assignment.
7177 *sds = (struct sd_lb_stats){
7180 .total_running = 0UL,
7182 .total_capacity = 0UL,
7185 .sum_nr_running = 0,
7186 .group_type = group_other,
7192 * get_sd_load_idx - Obtain the load index for a given sched domain.
7193 * @sd: The sched_domain whose load_idx is to be obtained.
7194 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7196 * Return: The load index.
7198 static inline int get_sd_load_idx(struct sched_domain *sd,
7199 enum cpu_idle_type idle)
7205 load_idx = sd->busy_idx;
7208 case CPU_NEWLY_IDLE:
7209 load_idx = sd->newidle_idx;
7212 load_idx = sd->idle_idx;
7219 static unsigned long scale_rt_capacity(int cpu)
7221 struct rq *rq = cpu_rq(cpu);
7222 u64 total, used, age_stamp, avg;
7226 * Since we're reading these variables without serialization make sure
7227 * we read them once before doing sanity checks on them.
7229 age_stamp = READ_ONCE(rq->age_stamp);
7230 avg = READ_ONCE(rq->rt_avg);
7231 delta = __rq_clock_broken(rq) - age_stamp;
7233 if (unlikely(delta < 0))
7236 total = sched_avg_period() + delta;
7238 used = div_u64(avg, total);
7240 if (likely(used < SCHED_CAPACITY_SCALE))
7241 return SCHED_CAPACITY_SCALE - used;
7246 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7248 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7249 struct sched_group *sdg = sd->groups;
7251 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7253 capacity *= scale_rt_capacity(cpu);
7254 capacity >>= SCHED_CAPACITY_SHIFT;
7259 cpu_rq(cpu)->cpu_capacity = capacity;
7260 sdg->sgc->capacity = capacity;
7261 sdg->sgc->min_capacity = capacity;
7264 void update_group_capacity(struct sched_domain *sd, int cpu)
7266 struct sched_domain *child = sd->child;
7267 struct sched_group *group, *sdg = sd->groups;
7268 unsigned long capacity, min_capacity;
7269 unsigned long interval;
7271 interval = msecs_to_jiffies(sd->balance_interval);
7272 interval = clamp(interval, 1UL, max_load_balance_interval);
7273 sdg->sgc->next_update = jiffies + interval;
7276 update_cpu_capacity(sd, cpu);
7281 min_capacity = ULONG_MAX;
7283 if (child->flags & SD_OVERLAP) {
7285 * SD_OVERLAP domains cannot assume that child groups
7286 * span the current group.
7289 for_each_cpu(cpu, sched_group_span(sdg)) {
7290 struct sched_group_capacity *sgc;
7291 struct rq *rq = cpu_rq(cpu);
7294 * build_sched_domains() -> init_sched_groups_capacity()
7295 * gets here before we've attached the domains to the
7298 * Use capacity_of(), which is set irrespective of domains
7299 * in update_cpu_capacity().
7301 * This avoids capacity from being 0 and
7302 * causing divide-by-zero issues on boot.
7304 if (unlikely(!rq->sd)) {
7305 capacity += capacity_of(cpu);
7307 sgc = rq->sd->groups->sgc;
7308 capacity += sgc->capacity;
7311 min_capacity = min(capacity, min_capacity);
7315 * !SD_OVERLAP domains can assume that child groups
7316 * span the current group.
7319 group = child->groups;
7321 struct sched_group_capacity *sgc = group->sgc;
7323 capacity += sgc->capacity;
7324 min_capacity = min(sgc->min_capacity, min_capacity);
7325 group = group->next;
7326 } while (group != child->groups);
7329 sdg->sgc->capacity = capacity;
7330 sdg->sgc->min_capacity = min_capacity;
7334 * Check whether the capacity of the rq has been noticeably reduced by side
7335 * activity. The imbalance_pct is used for the threshold.
7336 * Return true is the capacity is reduced
7339 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7341 return ((rq->cpu_capacity * sd->imbalance_pct) <
7342 (rq->cpu_capacity_orig * 100));
7346 * Group imbalance indicates (and tries to solve) the problem where balancing
7347 * groups is inadequate due to ->cpus_allowed constraints.
7349 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7350 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7353 * { 0 1 2 3 } { 4 5 6 7 }
7356 * If we were to balance group-wise we'd place two tasks in the first group and
7357 * two tasks in the second group. Clearly this is undesired as it will overload
7358 * cpu 3 and leave one of the cpus in the second group unused.
7360 * The current solution to this issue is detecting the skew in the first group
7361 * by noticing the lower domain failed to reach balance and had difficulty
7362 * moving tasks due to affinity constraints.
7364 * When this is so detected; this group becomes a candidate for busiest; see
7365 * update_sd_pick_busiest(). And calculate_imbalance() and
7366 * find_busiest_group() avoid some of the usual balance conditions to allow it
7367 * to create an effective group imbalance.
7369 * This is a somewhat tricky proposition since the next run might not find the
7370 * group imbalance and decide the groups need to be balanced again. A most
7371 * subtle and fragile situation.
7374 static inline int sg_imbalanced(struct sched_group *group)
7376 return group->sgc->imbalance;
7380 * group_has_capacity returns true if the group has spare capacity that could
7381 * be used by some tasks.
7382 * We consider that a group has spare capacity if the * number of task is
7383 * smaller than the number of CPUs or if the utilization is lower than the
7384 * available capacity for CFS tasks.
7385 * For the latter, we use a threshold to stabilize the state, to take into
7386 * account the variance of the tasks' load and to return true if the available
7387 * capacity in meaningful for the load balancer.
7388 * As an example, an available capacity of 1% can appear but it doesn't make
7389 * any benefit for the load balance.
7392 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7394 if (sgs->sum_nr_running < sgs->group_weight)
7397 if ((sgs->group_capacity * 100) >
7398 (sgs->group_util * env->sd->imbalance_pct))
7405 * group_is_overloaded returns true if the group has more tasks than it can
7407 * group_is_overloaded is not equals to !group_has_capacity because a group
7408 * with the exact right number of tasks, has no more spare capacity but is not
7409 * overloaded so both group_has_capacity and group_is_overloaded return
7413 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7415 if (sgs->sum_nr_running <= sgs->group_weight)
7418 if ((sgs->group_capacity * 100) <
7419 (sgs->group_util * env->sd->imbalance_pct))
7426 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7427 * per-CPU capacity than sched_group ref.
7430 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7432 return sg->sgc->min_capacity * capacity_margin <
7433 ref->sgc->min_capacity * 1024;
7437 group_type group_classify(struct sched_group *group,
7438 struct sg_lb_stats *sgs)
7440 if (sgs->group_no_capacity)
7441 return group_overloaded;
7443 if (sg_imbalanced(group))
7444 return group_imbalanced;
7450 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7451 * @env: The load balancing environment.
7452 * @group: sched_group whose statistics are to be updated.
7453 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7454 * @local_group: Does group contain this_cpu.
7455 * @sgs: variable to hold the statistics for this group.
7456 * @overload: Indicate more than one runnable task for any CPU.
7458 static inline void update_sg_lb_stats(struct lb_env *env,
7459 struct sched_group *group, int load_idx,
7460 int local_group, struct sg_lb_stats *sgs,
7466 memset(sgs, 0, sizeof(*sgs));
7468 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7469 struct rq *rq = cpu_rq(i);
7471 /* Bias balancing toward cpus of our domain */
7473 load = target_load(i, load_idx);
7475 load = source_load(i, load_idx);
7477 sgs->group_load += load;
7478 sgs->group_util += cpu_util(i);
7479 sgs->sum_nr_running += rq->cfs.h_nr_running;
7481 nr_running = rq->nr_running;
7485 #ifdef CONFIG_NUMA_BALANCING
7486 sgs->nr_numa_running += rq->nr_numa_running;
7487 sgs->nr_preferred_running += rq->nr_preferred_running;
7489 sgs->sum_weighted_load += weighted_cpuload(rq);
7491 * No need to call idle_cpu() if nr_running is not 0
7493 if (!nr_running && idle_cpu(i))
7497 /* Adjust by relative CPU capacity of the group */
7498 sgs->group_capacity = group->sgc->capacity;
7499 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7501 if (sgs->sum_nr_running)
7502 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7504 sgs->group_weight = group->group_weight;
7506 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7507 sgs->group_type = group_classify(group, sgs);
7511 * update_sd_pick_busiest - return 1 on busiest group
7512 * @env: The load balancing environment.
7513 * @sds: sched_domain statistics
7514 * @sg: sched_group candidate to be checked for being the busiest
7515 * @sgs: sched_group statistics
7517 * Determine if @sg is a busier group than the previously selected
7520 * Return: %true if @sg is a busier group than the previously selected
7521 * busiest group. %false otherwise.
7523 static bool update_sd_pick_busiest(struct lb_env *env,
7524 struct sd_lb_stats *sds,
7525 struct sched_group *sg,
7526 struct sg_lb_stats *sgs)
7528 struct sg_lb_stats *busiest = &sds->busiest_stat;
7530 if (sgs->group_type > busiest->group_type)
7533 if (sgs->group_type < busiest->group_type)
7536 if (sgs->avg_load <= busiest->avg_load)
7539 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7543 * Candidate sg has no more than one task per CPU and
7544 * has higher per-CPU capacity. Migrating tasks to less
7545 * capable CPUs may harm throughput. Maximize throughput,
7546 * power/energy consequences are not considered.
7548 if (sgs->sum_nr_running <= sgs->group_weight &&
7549 group_smaller_cpu_capacity(sds->local, sg))
7553 /* This is the busiest node in its class. */
7554 if (!(env->sd->flags & SD_ASYM_PACKING))
7557 /* No ASYM_PACKING if target cpu is already busy */
7558 if (env->idle == CPU_NOT_IDLE)
7561 * ASYM_PACKING needs to move all the work to the highest
7562 * prority CPUs in the group, therefore mark all groups
7563 * of lower priority than ourself as busy.
7565 if (sgs->sum_nr_running &&
7566 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7570 /* Prefer to move from lowest priority cpu's work */
7571 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7572 sg->asym_prefer_cpu))
7579 #ifdef CONFIG_NUMA_BALANCING
7580 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7582 if (sgs->sum_nr_running > sgs->nr_numa_running)
7584 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7589 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7591 if (rq->nr_running > rq->nr_numa_running)
7593 if (rq->nr_running > rq->nr_preferred_running)
7598 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7603 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7607 #endif /* CONFIG_NUMA_BALANCING */
7610 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7611 * @env: The load balancing environment.
7612 * @sds: variable to hold the statistics for this sched_domain.
7614 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7616 struct sched_domain_shared *shared = env->sd->shared;
7617 struct sched_domain *child = env->sd->child;
7618 struct sched_group *sg = env->sd->groups;
7619 struct sg_lb_stats *local = &sds->local_stat;
7620 struct sg_lb_stats tmp_sgs;
7621 int load_idx, prefer_sibling = 0;
7622 bool overload = false;
7624 if (child && child->flags & SD_PREFER_SIBLING)
7627 load_idx = get_sd_load_idx(env->sd, env->idle);
7630 struct sg_lb_stats *sgs = &tmp_sgs;
7633 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7638 if (env->idle != CPU_NEWLY_IDLE ||
7639 time_after_eq(jiffies, sg->sgc->next_update))
7640 update_group_capacity(env->sd, env->dst_cpu);
7643 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7650 * In case the child domain prefers tasks go to siblings
7651 * first, lower the sg capacity so that we'll try
7652 * and move all the excess tasks away. We lower the capacity
7653 * of a group only if the local group has the capacity to fit
7654 * these excess tasks. The extra check prevents the case where
7655 * you always pull from the heaviest group when it is already
7656 * under-utilized (possible with a large weight task outweighs
7657 * the tasks on the system).
7659 if (prefer_sibling && sds->local &&
7660 group_has_capacity(env, local) &&
7661 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7662 sgs->group_no_capacity = 1;
7663 sgs->group_type = group_classify(sg, sgs);
7666 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7668 sds->busiest_stat = *sgs;
7672 /* Now, start updating sd_lb_stats */
7673 sds->total_running += sgs->sum_nr_running;
7674 sds->total_load += sgs->group_load;
7675 sds->total_capacity += sgs->group_capacity;
7678 } while (sg != env->sd->groups);
7680 if (env->sd->flags & SD_NUMA)
7681 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7683 if (!env->sd->parent) {
7684 /* update overload indicator if we are at root domain */
7685 if (env->dst_rq->rd->overload != overload)
7686 env->dst_rq->rd->overload = overload;
7693 * Since these are sums over groups they can contain some CPUs
7694 * multiple times for the NUMA domains.
7696 * Currently only wake_affine_llc() and find_busiest_group()
7697 * uses these numbers, only the last is affected by this problem.
7701 WRITE_ONCE(shared->nr_running, sds->total_running);
7702 WRITE_ONCE(shared->load, sds->total_load);
7703 WRITE_ONCE(shared->capacity, sds->total_capacity);
7707 * check_asym_packing - Check to see if the group is packed into the
7710 * This is primarily intended to used at the sibling level. Some
7711 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7712 * case of POWER7, it can move to lower SMT modes only when higher
7713 * threads are idle. When in lower SMT modes, the threads will
7714 * perform better since they share less core resources. Hence when we
7715 * have idle threads, we want them to be the higher ones.
7717 * This packing function is run on idle threads. It checks to see if
7718 * the busiest CPU in this domain (core in the P7 case) has a higher
7719 * CPU number than the packing function is being run on. Here we are
7720 * assuming lower CPU number will be equivalent to lower a SMT thread
7723 * Return: 1 when packing is required and a task should be moved to
7724 * this CPU. The amount of the imbalance is returned in *imbalance.
7726 * @env: The load balancing environment.
7727 * @sds: Statistics of the sched_domain which is to be packed
7729 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7733 if (!(env->sd->flags & SD_ASYM_PACKING))
7736 if (env->idle == CPU_NOT_IDLE)
7742 busiest_cpu = sds->busiest->asym_prefer_cpu;
7743 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7746 env->imbalance = DIV_ROUND_CLOSEST(
7747 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7748 SCHED_CAPACITY_SCALE);
7754 * fix_small_imbalance - Calculate the minor imbalance that exists
7755 * amongst the groups of a sched_domain, during
7757 * @env: The load balancing environment.
7758 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7761 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7763 unsigned long tmp, capa_now = 0, capa_move = 0;
7764 unsigned int imbn = 2;
7765 unsigned long scaled_busy_load_per_task;
7766 struct sg_lb_stats *local, *busiest;
7768 local = &sds->local_stat;
7769 busiest = &sds->busiest_stat;
7771 if (!local->sum_nr_running)
7772 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7773 else if (busiest->load_per_task > local->load_per_task)
7776 scaled_busy_load_per_task =
7777 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7778 busiest->group_capacity;
7780 if (busiest->avg_load + scaled_busy_load_per_task >=
7781 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7782 env->imbalance = busiest->load_per_task;
7787 * OK, we don't have enough imbalance to justify moving tasks,
7788 * however we may be able to increase total CPU capacity used by
7792 capa_now += busiest->group_capacity *
7793 min(busiest->load_per_task, busiest->avg_load);
7794 capa_now += local->group_capacity *
7795 min(local->load_per_task, local->avg_load);
7796 capa_now /= SCHED_CAPACITY_SCALE;
7798 /* Amount of load we'd subtract */
7799 if (busiest->avg_load > scaled_busy_load_per_task) {
7800 capa_move += busiest->group_capacity *
7801 min(busiest->load_per_task,
7802 busiest->avg_load - scaled_busy_load_per_task);
7805 /* Amount of load we'd add */
7806 if (busiest->avg_load * busiest->group_capacity <
7807 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7808 tmp = (busiest->avg_load * busiest->group_capacity) /
7809 local->group_capacity;
7811 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7812 local->group_capacity;
7814 capa_move += local->group_capacity *
7815 min(local->load_per_task, local->avg_load + tmp);
7816 capa_move /= SCHED_CAPACITY_SCALE;
7818 /* Move if we gain throughput */
7819 if (capa_move > capa_now)
7820 env->imbalance = busiest->load_per_task;
7824 * calculate_imbalance - Calculate the amount of imbalance present within the
7825 * groups of a given sched_domain during load balance.
7826 * @env: load balance environment
7827 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7829 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7831 unsigned long max_pull, load_above_capacity = ~0UL;
7832 struct sg_lb_stats *local, *busiest;
7834 local = &sds->local_stat;
7835 busiest = &sds->busiest_stat;
7837 if (busiest->group_type == group_imbalanced) {
7839 * In the group_imb case we cannot rely on group-wide averages
7840 * to ensure cpu-load equilibrium, look at wider averages. XXX
7842 busiest->load_per_task =
7843 min(busiest->load_per_task, sds->avg_load);
7847 * Avg load of busiest sg can be less and avg load of local sg can
7848 * be greater than avg load across all sgs of sd because avg load
7849 * factors in sg capacity and sgs with smaller group_type are
7850 * skipped when updating the busiest sg:
7852 if (busiest->avg_load <= sds->avg_load ||
7853 local->avg_load >= sds->avg_load) {
7855 return fix_small_imbalance(env, sds);
7859 * If there aren't any idle cpus, avoid creating some.
7861 if (busiest->group_type == group_overloaded &&
7862 local->group_type == group_overloaded) {
7863 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7864 if (load_above_capacity > busiest->group_capacity) {
7865 load_above_capacity -= busiest->group_capacity;
7866 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7867 load_above_capacity /= busiest->group_capacity;
7869 load_above_capacity = ~0UL;
7873 * We're trying to get all the cpus to the average_load, so we don't
7874 * want to push ourselves above the average load, nor do we wish to
7875 * reduce the max loaded cpu below the average load. At the same time,
7876 * we also don't want to reduce the group load below the group
7877 * capacity. Thus we look for the minimum possible imbalance.
7879 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7881 /* How much load to actually move to equalise the imbalance */
7882 env->imbalance = min(
7883 max_pull * busiest->group_capacity,
7884 (sds->avg_load - local->avg_load) * local->group_capacity
7885 ) / SCHED_CAPACITY_SCALE;
7888 * if *imbalance is less than the average load per runnable task
7889 * there is no guarantee that any tasks will be moved so we'll have
7890 * a think about bumping its value to force at least one task to be
7893 if (env->imbalance < busiest->load_per_task)
7894 return fix_small_imbalance(env, sds);
7897 /******* find_busiest_group() helpers end here *********************/
7900 * find_busiest_group - Returns the busiest group within the sched_domain
7901 * if there is an imbalance.
7903 * Also calculates the amount of weighted load which should be moved
7904 * to restore balance.
7906 * @env: The load balancing environment.
7908 * Return: - The busiest group if imbalance exists.
7910 static struct sched_group *find_busiest_group(struct lb_env *env)
7912 struct sg_lb_stats *local, *busiest;
7913 struct sd_lb_stats sds;
7915 init_sd_lb_stats(&sds);
7918 * Compute the various statistics relavent for load balancing at
7921 update_sd_lb_stats(env, &sds);
7922 local = &sds.local_stat;
7923 busiest = &sds.busiest_stat;
7925 /* ASYM feature bypasses nice load balance check */
7926 if (check_asym_packing(env, &sds))
7929 /* There is no busy sibling group to pull tasks from */
7930 if (!sds.busiest || busiest->sum_nr_running == 0)
7933 /* XXX broken for overlapping NUMA groups */
7934 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7935 / sds.total_capacity;
7938 * If the busiest group is imbalanced the below checks don't
7939 * work because they assume all things are equal, which typically
7940 * isn't true due to cpus_allowed constraints and the like.
7942 if (busiest->group_type == group_imbalanced)
7945 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7946 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7947 busiest->group_no_capacity)
7951 * If the local group is busier than the selected busiest group
7952 * don't try and pull any tasks.
7954 if (local->avg_load >= busiest->avg_load)
7958 * Don't pull any tasks if this group is already above the domain
7961 if (local->avg_load >= sds.avg_load)
7964 if (env->idle == CPU_IDLE) {
7966 * This cpu is idle. If the busiest group is not overloaded
7967 * and there is no imbalance between this and busiest group
7968 * wrt idle cpus, it is balanced. The imbalance becomes
7969 * significant if the diff is greater than 1 otherwise we
7970 * might end up to just move the imbalance on another group
7972 if ((busiest->group_type != group_overloaded) &&
7973 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7977 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7978 * imbalance_pct to be conservative.
7980 if (100 * busiest->avg_load <=
7981 env->sd->imbalance_pct * local->avg_load)
7986 /* Looks like there is an imbalance. Compute it */
7987 calculate_imbalance(env, &sds);
7996 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7998 static struct rq *find_busiest_queue(struct lb_env *env,
7999 struct sched_group *group)
8001 struct rq *busiest = NULL, *rq;
8002 unsigned long busiest_load = 0, busiest_capacity = 1;
8005 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8006 unsigned long capacity, wl;
8010 rt = fbq_classify_rq(rq);
8013 * We classify groups/runqueues into three groups:
8014 * - regular: there are !numa tasks
8015 * - remote: there are numa tasks that run on the 'wrong' node
8016 * - all: there is no distinction
8018 * In order to avoid migrating ideally placed numa tasks,
8019 * ignore those when there's better options.
8021 * If we ignore the actual busiest queue to migrate another
8022 * task, the next balance pass can still reduce the busiest
8023 * queue by moving tasks around inside the node.
8025 * If we cannot move enough load due to this classification
8026 * the next pass will adjust the group classification and
8027 * allow migration of more tasks.
8029 * Both cases only affect the total convergence complexity.
8031 if (rt > env->fbq_type)
8034 capacity = capacity_of(i);
8036 wl = weighted_cpuload(rq);
8039 * When comparing with imbalance, use weighted_cpuload()
8040 * which is not scaled with the cpu capacity.
8043 if (rq->nr_running == 1 && wl > env->imbalance &&
8044 !check_cpu_capacity(rq, env->sd))
8048 * For the load comparisons with the other cpu's, consider
8049 * the weighted_cpuload() scaled with the cpu capacity, so
8050 * that the load can be moved away from the cpu that is
8051 * potentially running at a lower capacity.
8053 * Thus we're looking for max(wl_i / capacity_i), crosswise
8054 * multiplication to rid ourselves of the division works out
8055 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8056 * our previous maximum.
8058 if (wl * busiest_capacity > busiest_load * capacity) {
8060 busiest_capacity = capacity;
8069 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8070 * so long as it is large enough.
8072 #define MAX_PINNED_INTERVAL 512
8074 static int need_active_balance(struct lb_env *env)
8076 struct sched_domain *sd = env->sd;
8078 if (env->idle == CPU_NEWLY_IDLE) {
8081 * ASYM_PACKING needs to force migrate tasks from busy but
8082 * lower priority CPUs in order to pack all tasks in the
8083 * highest priority CPUs.
8085 if ((sd->flags & SD_ASYM_PACKING) &&
8086 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8091 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8092 * It's worth migrating the task if the src_cpu's capacity is reduced
8093 * because of other sched_class or IRQs if more capacity stays
8094 * available on dst_cpu.
8096 if ((env->idle != CPU_NOT_IDLE) &&
8097 (env->src_rq->cfs.h_nr_running == 1)) {
8098 if ((check_cpu_capacity(env->src_rq, sd)) &&
8099 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8103 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8106 static int active_load_balance_cpu_stop(void *data);
8108 static int should_we_balance(struct lb_env *env)
8110 struct sched_group *sg = env->sd->groups;
8111 int cpu, balance_cpu = -1;
8114 * In the newly idle case, we will allow all the cpu's
8115 * to do the newly idle load balance.
8117 if (env->idle == CPU_NEWLY_IDLE)
8120 /* Try to find first idle cpu */
8121 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8129 if (balance_cpu == -1)
8130 balance_cpu = group_balance_cpu(sg);
8133 * First idle cpu or the first cpu(busiest) in this sched group
8134 * is eligible for doing load balancing at this and above domains.
8136 return balance_cpu == env->dst_cpu;
8140 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8141 * tasks if there is an imbalance.
8143 static int load_balance(int this_cpu, struct rq *this_rq,
8144 struct sched_domain *sd, enum cpu_idle_type idle,
8145 int *continue_balancing)
8147 int ld_moved, cur_ld_moved, active_balance = 0;
8148 struct sched_domain *sd_parent = sd->parent;
8149 struct sched_group *group;
8152 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8154 struct lb_env env = {
8156 .dst_cpu = this_cpu,
8158 .dst_grpmask = sched_group_span(sd->groups),
8160 .loop_break = sched_nr_migrate_break,
8163 .tasks = LIST_HEAD_INIT(env.tasks),
8166 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8168 schedstat_inc(sd->lb_count[idle]);
8171 if (!should_we_balance(&env)) {
8172 *continue_balancing = 0;
8176 group = find_busiest_group(&env);
8178 schedstat_inc(sd->lb_nobusyg[idle]);
8182 busiest = find_busiest_queue(&env, group);
8184 schedstat_inc(sd->lb_nobusyq[idle]);
8188 BUG_ON(busiest == env.dst_rq);
8190 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8192 env.src_cpu = busiest->cpu;
8193 env.src_rq = busiest;
8196 if (busiest->nr_running > 1) {
8198 * Attempt to move tasks. If find_busiest_group has found
8199 * an imbalance but busiest->nr_running <= 1, the group is
8200 * still unbalanced. ld_moved simply stays zero, so it is
8201 * correctly treated as an imbalance.
8203 env.flags |= LBF_ALL_PINNED;
8204 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8207 rq_lock_irqsave(busiest, &rf);
8208 update_rq_clock(busiest);
8211 * cur_ld_moved - load moved in current iteration
8212 * ld_moved - cumulative load moved across iterations
8214 cur_ld_moved = detach_tasks(&env);
8217 * We've detached some tasks from busiest_rq. Every
8218 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8219 * unlock busiest->lock, and we are able to be sure
8220 * that nobody can manipulate the tasks in parallel.
8221 * See task_rq_lock() family for the details.
8224 rq_unlock(busiest, &rf);
8228 ld_moved += cur_ld_moved;
8231 local_irq_restore(rf.flags);
8233 if (env.flags & LBF_NEED_BREAK) {
8234 env.flags &= ~LBF_NEED_BREAK;
8239 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8240 * us and move them to an alternate dst_cpu in our sched_group
8241 * where they can run. The upper limit on how many times we
8242 * iterate on same src_cpu is dependent on number of cpus in our
8245 * This changes load balance semantics a bit on who can move
8246 * load to a given_cpu. In addition to the given_cpu itself
8247 * (or a ilb_cpu acting on its behalf where given_cpu is
8248 * nohz-idle), we now have balance_cpu in a position to move
8249 * load to given_cpu. In rare situations, this may cause
8250 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8251 * _independently_ and at _same_ time to move some load to
8252 * given_cpu) causing exceess load to be moved to given_cpu.
8253 * This however should not happen so much in practice and
8254 * moreover subsequent load balance cycles should correct the
8255 * excess load moved.
8257 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8259 /* Prevent to re-select dst_cpu via env's cpus */
8260 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8262 env.dst_rq = cpu_rq(env.new_dst_cpu);
8263 env.dst_cpu = env.new_dst_cpu;
8264 env.flags &= ~LBF_DST_PINNED;
8266 env.loop_break = sched_nr_migrate_break;
8269 * Go back to "more_balance" rather than "redo" since we
8270 * need to continue with same src_cpu.
8276 * We failed to reach balance because of affinity.
8279 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8281 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8282 *group_imbalance = 1;
8285 /* All tasks on this runqueue were pinned by CPU affinity */
8286 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8287 cpumask_clear_cpu(cpu_of(busiest), cpus);
8289 * Attempting to continue load balancing at the current
8290 * sched_domain level only makes sense if there are
8291 * active CPUs remaining as possible busiest CPUs to
8292 * pull load from which are not contained within the
8293 * destination group that is receiving any migrated
8296 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8298 env.loop_break = sched_nr_migrate_break;
8301 goto out_all_pinned;
8306 schedstat_inc(sd->lb_failed[idle]);
8308 * Increment the failure counter only on periodic balance.
8309 * We do not want newidle balance, which can be very
8310 * frequent, pollute the failure counter causing
8311 * excessive cache_hot migrations and active balances.
8313 if (idle != CPU_NEWLY_IDLE)
8314 sd->nr_balance_failed++;
8316 if (need_active_balance(&env)) {
8317 unsigned long flags;
8319 raw_spin_lock_irqsave(&busiest->lock, flags);
8321 /* don't kick the active_load_balance_cpu_stop,
8322 * if the curr task on busiest cpu can't be
8325 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8326 raw_spin_unlock_irqrestore(&busiest->lock,
8328 env.flags |= LBF_ALL_PINNED;
8329 goto out_one_pinned;
8333 * ->active_balance synchronizes accesses to
8334 * ->active_balance_work. Once set, it's cleared
8335 * only after active load balance is finished.
8337 if (!busiest->active_balance) {
8338 busiest->active_balance = 1;
8339 busiest->push_cpu = this_cpu;
8342 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8344 if (active_balance) {
8345 stop_one_cpu_nowait(cpu_of(busiest),
8346 active_load_balance_cpu_stop, busiest,
8347 &busiest->active_balance_work);
8350 /* We've kicked active balancing, force task migration. */
8351 sd->nr_balance_failed = sd->cache_nice_tries+1;
8354 sd->nr_balance_failed = 0;
8356 if (likely(!active_balance)) {
8357 /* We were unbalanced, so reset the balancing interval */
8358 sd->balance_interval = sd->min_interval;
8361 * If we've begun active balancing, start to back off. This
8362 * case may not be covered by the all_pinned logic if there
8363 * is only 1 task on the busy runqueue (because we don't call
8366 if (sd->balance_interval < sd->max_interval)
8367 sd->balance_interval *= 2;
8374 * We reach balance although we may have faced some affinity
8375 * constraints. Clear the imbalance flag if it was set.
8378 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8380 if (*group_imbalance)
8381 *group_imbalance = 0;
8386 * We reach balance because all tasks are pinned at this level so
8387 * we can't migrate them. Let the imbalance flag set so parent level
8388 * can try to migrate them.
8390 schedstat_inc(sd->lb_balanced[idle]);
8392 sd->nr_balance_failed = 0;
8395 /* tune up the balancing interval */
8396 if (((env.flags & LBF_ALL_PINNED) &&
8397 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8398 (sd->balance_interval < sd->max_interval))
8399 sd->balance_interval *= 2;
8406 static inline unsigned long
8407 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8409 unsigned long interval = sd->balance_interval;
8412 interval *= sd->busy_factor;
8414 /* scale ms to jiffies */
8415 interval = msecs_to_jiffies(interval);
8416 interval = clamp(interval, 1UL, max_load_balance_interval);
8422 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8424 unsigned long interval, next;
8426 /* used by idle balance, so cpu_busy = 0 */
8427 interval = get_sd_balance_interval(sd, 0);
8428 next = sd->last_balance + interval;
8430 if (time_after(*next_balance, next))
8431 *next_balance = next;
8435 * idle_balance is called by schedule() if this_cpu is about to become
8436 * idle. Attempts to pull tasks from other CPUs.
8438 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8440 unsigned long next_balance = jiffies + HZ;
8441 int this_cpu = this_rq->cpu;
8442 struct sched_domain *sd;
8443 int pulled_task = 0;
8447 * We must set idle_stamp _before_ calling idle_balance(), such that we
8448 * measure the duration of idle_balance() as idle time.
8450 this_rq->idle_stamp = rq_clock(this_rq);
8453 * This is OK, because current is on_cpu, which avoids it being picked
8454 * for load-balance and preemption/IRQs are still disabled avoiding
8455 * further scheduler activity on it and we're being very careful to
8456 * re-start the picking loop.
8458 rq_unpin_lock(this_rq, rf);
8460 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8461 !this_rq->rd->overload) {
8463 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8465 update_next_balance(sd, &next_balance);
8471 raw_spin_unlock(&this_rq->lock);
8473 update_blocked_averages(this_cpu);
8475 for_each_domain(this_cpu, sd) {
8476 int continue_balancing = 1;
8477 u64 t0, domain_cost;
8479 if (!(sd->flags & SD_LOAD_BALANCE))
8482 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8483 update_next_balance(sd, &next_balance);
8487 if (sd->flags & SD_BALANCE_NEWIDLE) {
8488 t0 = sched_clock_cpu(this_cpu);
8490 pulled_task = load_balance(this_cpu, this_rq,
8492 &continue_balancing);
8494 domain_cost = sched_clock_cpu(this_cpu) - t0;
8495 if (domain_cost > sd->max_newidle_lb_cost)
8496 sd->max_newidle_lb_cost = domain_cost;
8498 curr_cost += domain_cost;
8501 update_next_balance(sd, &next_balance);
8504 * Stop searching for tasks to pull if there are
8505 * now runnable tasks on this rq.
8507 if (pulled_task || this_rq->nr_running > 0)
8512 raw_spin_lock(&this_rq->lock);
8514 if (curr_cost > this_rq->max_idle_balance_cost)
8515 this_rq->max_idle_balance_cost = curr_cost;
8518 * While browsing the domains, we released the rq lock, a task could
8519 * have been enqueued in the meantime. Since we're not going idle,
8520 * pretend we pulled a task.
8522 if (this_rq->cfs.h_nr_running && !pulled_task)
8526 /* Move the next balance forward */
8527 if (time_after(this_rq->next_balance, next_balance))
8528 this_rq->next_balance = next_balance;
8530 /* Is there a task of a high priority class? */
8531 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8535 this_rq->idle_stamp = 0;
8537 rq_repin_lock(this_rq, rf);
8543 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8544 * running tasks off the busiest CPU onto idle CPUs. It requires at
8545 * least 1 task to be running on each physical CPU where possible, and
8546 * avoids physical / logical imbalances.
8548 static int active_load_balance_cpu_stop(void *data)
8550 struct rq *busiest_rq = data;
8551 int busiest_cpu = cpu_of(busiest_rq);
8552 int target_cpu = busiest_rq->push_cpu;
8553 struct rq *target_rq = cpu_rq(target_cpu);
8554 struct sched_domain *sd;
8555 struct task_struct *p = NULL;
8558 rq_lock_irq(busiest_rq, &rf);
8560 /* make sure the requested cpu hasn't gone down in the meantime */
8561 if (unlikely(busiest_cpu != smp_processor_id() ||
8562 !busiest_rq->active_balance))
8565 /* Is there any task to move? */
8566 if (busiest_rq->nr_running <= 1)
8570 * This condition is "impossible", if it occurs
8571 * we need to fix it. Originally reported by
8572 * Bjorn Helgaas on a 128-cpu setup.
8574 BUG_ON(busiest_rq == target_rq);
8576 /* Search for an sd spanning us and the target CPU. */
8578 for_each_domain(target_cpu, sd) {
8579 if ((sd->flags & SD_LOAD_BALANCE) &&
8580 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8585 struct lb_env env = {
8587 .dst_cpu = target_cpu,
8588 .dst_rq = target_rq,
8589 .src_cpu = busiest_rq->cpu,
8590 .src_rq = busiest_rq,
8593 * can_migrate_task() doesn't need to compute new_dst_cpu
8594 * for active balancing. Since we have CPU_IDLE, but no
8595 * @dst_grpmask we need to make that test go away with lying
8598 .flags = LBF_DST_PINNED,
8601 schedstat_inc(sd->alb_count);
8602 update_rq_clock(busiest_rq);
8604 p = detach_one_task(&env);
8606 schedstat_inc(sd->alb_pushed);
8607 /* Active balancing done, reset the failure counter. */
8608 sd->nr_balance_failed = 0;
8610 schedstat_inc(sd->alb_failed);
8615 busiest_rq->active_balance = 0;
8616 rq_unlock(busiest_rq, &rf);
8619 attach_one_task(target_rq, p);
8626 static inline int on_null_domain(struct rq *rq)
8628 return unlikely(!rcu_dereference_sched(rq->sd));
8631 #ifdef CONFIG_NO_HZ_COMMON
8633 * idle load balancing details
8634 * - When one of the busy CPUs notice that there may be an idle rebalancing
8635 * needed, they will kick the idle load balancer, which then does idle
8636 * load balancing for all the idle CPUs.
8639 cpumask_var_t idle_cpus_mask;
8641 unsigned long next_balance; /* in jiffy units */
8642 } nohz ____cacheline_aligned;
8644 static inline int find_new_ilb(void)
8646 int ilb = cpumask_first(nohz.idle_cpus_mask);
8648 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8655 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8656 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8657 * CPU (if there is one).
8659 static void nohz_balancer_kick(void)
8663 nohz.next_balance++;
8665 ilb_cpu = find_new_ilb();
8667 if (ilb_cpu >= nr_cpu_ids)
8670 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8673 * Use smp_send_reschedule() instead of resched_cpu().
8674 * This way we generate a sched IPI on the target cpu which
8675 * is idle. And the softirq performing nohz idle load balance
8676 * will be run before returning from the IPI.
8678 smp_send_reschedule(ilb_cpu);
8682 void nohz_balance_exit_idle(unsigned int cpu)
8684 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8686 * Completely isolated CPUs don't ever set, so we must test.
8688 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8689 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8690 atomic_dec(&nohz.nr_cpus);
8692 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8696 static inline void set_cpu_sd_state_busy(void)
8698 struct sched_domain *sd;
8699 int cpu = smp_processor_id();
8702 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8704 if (!sd || !sd->nohz_idle)
8708 atomic_inc(&sd->shared->nr_busy_cpus);
8713 void set_cpu_sd_state_idle(void)
8715 struct sched_domain *sd;
8716 int cpu = smp_processor_id();
8719 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8721 if (!sd || sd->nohz_idle)
8725 atomic_dec(&sd->shared->nr_busy_cpus);
8731 * This routine will record that the cpu is going idle with tick stopped.
8732 * This info will be used in performing idle load balancing in the future.
8734 void nohz_balance_enter_idle(int cpu)
8737 * If this cpu is going down, then nothing needs to be done.
8739 if (!cpu_active(cpu))
8742 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8743 if (!is_housekeeping_cpu(cpu))
8746 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8750 * If we're a completely isolated CPU, we don't play.
8752 if (on_null_domain(cpu_rq(cpu)))
8755 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8756 atomic_inc(&nohz.nr_cpus);
8757 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8761 static DEFINE_SPINLOCK(balancing);
8764 * Scale the max load_balance interval with the number of CPUs in the system.
8765 * This trades load-balance latency on larger machines for less cross talk.
8767 void update_max_interval(void)
8769 max_load_balance_interval = HZ*num_online_cpus()/10;
8773 * It checks each scheduling domain to see if it is due to be balanced,
8774 * and initiates a balancing operation if so.
8776 * Balancing parameters are set up in init_sched_domains.
8778 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8780 int continue_balancing = 1;
8782 unsigned long interval;
8783 struct sched_domain *sd;
8784 /* Earliest time when we have to do rebalance again */
8785 unsigned long next_balance = jiffies + 60*HZ;
8786 int update_next_balance = 0;
8787 int need_serialize, need_decay = 0;
8790 update_blocked_averages(cpu);
8793 for_each_domain(cpu, sd) {
8795 * Decay the newidle max times here because this is a regular
8796 * visit to all the domains. Decay ~1% per second.
8798 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8799 sd->max_newidle_lb_cost =
8800 (sd->max_newidle_lb_cost * 253) / 256;
8801 sd->next_decay_max_lb_cost = jiffies + HZ;
8804 max_cost += sd->max_newidle_lb_cost;
8806 if (!(sd->flags & SD_LOAD_BALANCE))
8810 * Stop the load balance at this level. There is another
8811 * CPU in our sched group which is doing load balancing more
8814 if (!continue_balancing) {
8820 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8822 need_serialize = sd->flags & SD_SERIALIZE;
8823 if (need_serialize) {
8824 if (!spin_trylock(&balancing))
8828 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8829 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8831 * The LBF_DST_PINNED logic could have changed
8832 * env->dst_cpu, so we can't know our idle
8833 * state even if we migrated tasks. Update it.
8835 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8837 sd->last_balance = jiffies;
8838 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8841 spin_unlock(&balancing);
8843 if (time_after(next_balance, sd->last_balance + interval)) {
8844 next_balance = sd->last_balance + interval;
8845 update_next_balance = 1;
8850 * Ensure the rq-wide value also decays but keep it at a
8851 * reasonable floor to avoid funnies with rq->avg_idle.
8853 rq->max_idle_balance_cost =
8854 max((u64)sysctl_sched_migration_cost, max_cost);
8859 * next_balance will be updated only when there is a need.
8860 * When the cpu is attached to null domain for ex, it will not be
8863 if (likely(update_next_balance)) {
8864 rq->next_balance = next_balance;
8866 #ifdef CONFIG_NO_HZ_COMMON
8868 * If this CPU has been elected to perform the nohz idle
8869 * balance. Other idle CPUs have already rebalanced with
8870 * nohz_idle_balance() and nohz.next_balance has been
8871 * updated accordingly. This CPU is now running the idle load
8872 * balance for itself and we need to update the
8873 * nohz.next_balance accordingly.
8875 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8876 nohz.next_balance = rq->next_balance;
8881 #ifdef CONFIG_NO_HZ_COMMON
8883 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8884 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8886 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8888 int this_cpu = this_rq->cpu;
8891 /* Earliest time when we have to do rebalance again */
8892 unsigned long next_balance = jiffies + 60*HZ;
8893 int update_next_balance = 0;
8895 if (idle != CPU_IDLE ||
8896 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8899 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8900 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8904 * If this cpu gets work to do, stop the load balancing
8905 * work being done for other cpus. Next load
8906 * balancing owner will pick it up.
8911 rq = cpu_rq(balance_cpu);
8914 * If time for next balance is due,
8917 if (time_after_eq(jiffies, rq->next_balance)) {
8920 rq_lock_irq(rq, &rf);
8921 update_rq_clock(rq);
8922 cpu_load_update_idle(rq);
8923 rq_unlock_irq(rq, &rf);
8925 rebalance_domains(rq, CPU_IDLE);
8928 if (time_after(next_balance, rq->next_balance)) {
8929 next_balance = rq->next_balance;
8930 update_next_balance = 1;
8935 * next_balance will be updated only when there is a need.
8936 * When the CPU is attached to null domain for ex, it will not be
8939 if (likely(update_next_balance))
8940 nohz.next_balance = next_balance;
8942 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8946 * Current heuristic for kicking the idle load balancer in the presence
8947 * of an idle cpu in the system.
8948 * - This rq has more than one task.
8949 * - This rq has at least one CFS task and the capacity of the CPU is
8950 * significantly reduced because of RT tasks or IRQs.
8951 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8952 * multiple busy cpu.
8953 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8954 * domain span are idle.
8956 static inline bool nohz_kick_needed(struct rq *rq)
8958 unsigned long now = jiffies;
8959 struct sched_domain_shared *sds;
8960 struct sched_domain *sd;
8961 int nr_busy, i, cpu = rq->cpu;
8964 if (unlikely(rq->idle_balance))
8968 * We may be recently in ticked or tickless idle mode. At the first
8969 * busy tick after returning from idle, we will update the busy stats.
8971 set_cpu_sd_state_busy();
8972 nohz_balance_exit_idle(cpu);
8975 * None are in tickless mode and hence no need for NOHZ idle load
8978 if (likely(!atomic_read(&nohz.nr_cpus)))
8981 if (time_before(now, nohz.next_balance))
8984 if (rq->nr_running >= 2)
8988 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8991 * XXX: write a coherent comment on why we do this.
8992 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8994 nr_busy = atomic_read(&sds->nr_busy_cpus);
9002 sd = rcu_dereference(rq->sd);
9004 if ((rq->cfs.h_nr_running >= 1) &&
9005 check_cpu_capacity(rq, sd)) {
9011 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9013 for_each_cpu(i, sched_domain_span(sd)) {
9015 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9018 if (sched_asym_prefer(i, cpu)) {
9029 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9033 * run_rebalance_domains is triggered when needed from the scheduler tick.
9034 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9036 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9038 struct rq *this_rq = this_rq();
9039 enum cpu_idle_type idle = this_rq->idle_balance ?
9040 CPU_IDLE : CPU_NOT_IDLE;
9043 * If this cpu has a pending nohz_balance_kick, then do the
9044 * balancing on behalf of the other idle cpus whose ticks are
9045 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9046 * give the idle cpus a chance to load balance. Else we may
9047 * load balance only within the local sched_domain hierarchy
9048 * and abort nohz_idle_balance altogether if we pull some load.
9050 nohz_idle_balance(this_rq, idle);
9051 rebalance_domains(this_rq, idle);
9055 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9057 void trigger_load_balance(struct rq *rq)
9059 /* Don't need to rebalance while attached to NULL domain */
9060 if (unlikely(on_null_domain(rq)))
9063 if (time_after_eq(jiffies, rq->next_balance))
9064 raise_softirq(SCHED_SOFTIRQ);
9065 #ifdef CONFIG_NO_HZ_COMMON
9066 if (nohz_kick_needed(rq))
9067 nohz_balancer_kick();
9071 static void rq_online_fair(struct rq *rq)
9075 update_runtime_enabled(rq);
9078 static void rq_offline_fair(struct rq *rq)
9082 /* Ensure any throttled groups are reachable by pick_next_task */
9083 unthrottle_offline_cfs_rqs(rq);
9086 #endif /* CONFIG_SMP */
9089 * scheduler tick hitting a task of our scheduling class:
9091 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9093 struct cfs_rq *cfs_rq;
9094 struct sched_entity *se = &curr->se;
9096 for_each_sched_entity(se) {
9097 cfs_rq = cfs_rq_of(se);
9098 entity_tick(cfs_rq, se, queued);
9101 if (static_branch_unlikely(&sched_numa_balancing))
9102 task_tick_numa(rq, curr);
9106 * called on fork with the child task as argument from the parent's context
9107 * - child not yet on the tasklist
9108 * - preemption disabled
9110 static void task_fork_fair(struct task_struct *p)
9112 struct cfs_rq *cfs_rq;
9113 struct sched_entity *se = &p->se, *curr;
9114 struct rq *rq = this_rq();
9118 update_rq_clock(rq);
9120 cfs_rq = task_cfs_rq(current);
9121 curr = cfs_rq->curr;
9123 update_curr(cfs_rq);
9124 se->vruntime = curr->vruntime;
9126 place_entity(cfs_rq, se, 1);
9128 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9130 * Upon rescheduling, sched_class::put_prev_task() will place
9131 * 'current' within the tree based on its new key value.
9133 swap(curr->vruntime, se->vruntime);
9137 se->vruntime -= cfs_rq->min_vruntime;
9142 * Priority of the task has changed. Check to see if we preempt
9146 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9148 if (!task_on_rq_queued(p))
9152 * Reschedule if we are currently running on this runqueue and
9153 * our priority decreased, or if we are not currently running on
9154 * this runqueue and our priority is higher than the current's
9156 if (rq->curr == p) {
9157 if (p->prio > oldprio)
9160 check_preempt_curr(rq, p, 0);
9163 static inline bool vruntime_normalized(struct task_struct *p)
9165 struct sched_entity *se = &p->se;
9168 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9169 * the dequeue_entity(.flags=0) will already have normalized the
9176 * When !on_rq, vruntime of the task has usually NOT been normalized.
9177 * But there are some cases where it has already been normalized:
9179 * - A forked child which is waiting for being woken up by
9180 * wake_up_new_task().
9181 * - A task which has been woken up by try_to_wake_up() and
9182 * waiting for actually being woken up by sched_ttwu_pending().
9184 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9192 * Propagate the changes of the sched_entity across the tg tree to make it
9193 * visible to the root
9195 static void propagate_entity_cfs_rq(struct sched_entity *se)
9197 struct cfs_rq *cfs_rq;
9199 /* Start to propagate at parent */
9202 for_each_sched_entity(se) {
9203 cfs_rq = cfs_rq_of(se);
9205 if (cfs_rq_throttled(cfs_rq))
9208 update_load_avg(se, UPDATE_TG);
9212 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9215 static void detach_entity_cfs_rq(struct sched_entity *se)
9217 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9219 /* Catch up with the cfs_rq and remove our load when we leave */
9220 update_load_avg(se, 0);
9221 detach_entity_load_avg(cfs_rq, se);
9222 update_tg_load_avg(cfs_rq, false);
9223 propagate_entity_cfs_rq(se);
9226 static void attach_entity_cfs_rq(struct sched_entity *se)
9228 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9230 #ifdef CONFIG_FAIR_GROUP_SCHED
9232 * Since the real-depth could have been changed (only FAIR
9233 * class maintain depth value), reset depth properly.
9235 se->depth = se->parent ? se->parent->depth + 1 : 0;
9238 /* Synchronize entity with its cfs_rq */
9239 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9240 attach_entity_load_avg(cfs_rq, se);
9241 update_tg_load_avg(cfs_rq, false);
9242 propagate_entity_cfs_rq(se);
9245 static void detach_task_cfs_rq(struct task_struct *p)
9247 struct sched_entity *se = &p->se;
9248 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9250 if (!vruntime_normalized(p)) {
9252 * Fix up our vruntime so that the current sleep doesn't
9253 * cause 'unlimited' sleep bonus.
9255 place_entity(cfs_rq, se, 0);
9256 se->vruntime -= cfs_rq->min_vruntime;
9259 detach_entity_cfs_rq(se);
9262 static void attach_task_cfs_rq(struct task_struct *p)
9264 struct sched_entity *se = &p->se;
9265 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9267 attach_entity_cfs_rq(se);
9269 if (!vruntime_normalized(p))
9270 se->vruntime += cfs_rq->min_vruntime;
9273 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9275 detach_task_cfs_rq(p);
9278 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9280 attach_task_cfs_rq(p);
9282 if (task_on_rq_queued(p)) {
9284 * We were most likely switched from sched_rt, so
9285 * kick off the schedule if running, otherwise just see
9286 * if we can still preempt the current task.
9291 check_preempt_curr(rq, p, 0);
9295 /* Account for a task changing its policy or group.
9297 * This routine is mostly called to set cfs_rq->curr field when a task
9298 * migrates between groups/classes.
9300 static void set_curr_task_fair(struct rq *rq)
9302 struct sched_entity *se = &rq->curr->se;
9304 for_each_sched_entity(se) {
9305 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9307 set_next_entity(cfs_rq, se);
9308 /* ensure bandwidth has been allocated on our new cfs_rq */
9309 account_cfs_rq_runtime(cfs_rq, 0);
9313 void init_cfs_rq(struct cfs_rq *cfs_rq)
9315 cfs_rq->tasks_timeline = RB_ROOT;
9316 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9317 #ifndef CONFIG_64BIT
9318 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9322 cfs_rq->propagate_avg = 0;
9324 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9325 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9329 #ifdef CONFIG_FAIR_GROUP_SCHED
9330 static void task_set_group_fair(struct task_struct *p)
9332 struct sched_entity *se = &p->se;
9334 set_task_rq(p, task_cpu(p));
9335 se->depth = se->parent ? se->parent->depth + 1 : 0;
9338 static void task_move_group_fair(struct task_struct *p)
9340 detach_task_cfs_rq(p);
9341 set_task_rq(p, task_cpu(p));
9344 /* Tell se's cfs_rq has been changed -- migrated */
9345 p->se.avg.last_update_time = 0;
9347 attach_task_cfs_rq(p);
9350 static void task_change_group_fair(struct task_struct *p, int type)
9353 case TASK_SET_GROUP:
9354 task_set_group_fair(p);
9357 case TASK_MOVE_GROUP:
9358 task_move_group_fair(p);
9363 void free_fair_sched_group(struct task_group *tg)
9367 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9369 for_each_possible_cpu(i) {
9371 kfree(tg->cfs_rq[i]);
9380 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9382 struct sched_entity *se;
9383 struct cfs_rq *cfs_rq;
9386 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9389 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9393 tg->shares = NICE_0_LOAD;
9395 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9397 for_each_possible_cpu(i) {
9398 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9399 GFP_KERNEL, cpu_to_node(i));
9403 se = kzalloc_node(sizeof(struct sched_entity),
9404 GFP_KERNEL, cpu_to_node(i));
9408 init_cfs_rq(cfs_rq);
9409 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9410 init_entity_runnable_average(se);
9421 void online_fair_sched_group(struct task_group *tg)
9423 struct sched_entity *se;
9427 for_each_possible_cpu(i) {
9431 raw_spin_lock_irq(&rq->lock);
9432 update_rq_clock(rq);
9433 attach_entity_cfs_rq(se);
9434 sync_throttle(tg, i);
9435 raw_spin_unlock_irq(&rq->lock);
9439 void unregister_fair_sched_group(struct task_group *tg)
9441 unsigned long flags;
9445 for_each_possible_cpu(cpu) {
9447 remove_entity_load_avg(tg->se[cpu]);
9450 * Only empty task groups can be destroyed; so we can speculatively
9451 * check on_list without danger of it being re-added.
9453 if (!tg->cfs_rq[cpu]->on_list)
9458 raw_spin_lock_irqsave(&rq->lock, flags);
9459 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9460 raw_spin_unlock_irqrestore(&rq->lock, flags);
9464 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9465 struct sched_entity *se, int cpu,
9466 struct sched_entity *parent)
9468 struct rq *rq = cpu_rq(cpu);
9472 init_cfs_rq_runtime(cfs_rq);
9474 tg->cfs_rq[cpu] = cfs_rq;
9477 /* se could be NULL for root_task_group */
9482 se->cfs_rq = &rq->cfs;
9485 se->cfs_rq = parent->my_q;
9486 se->depth = parent->depth + 1;
9490 /* guarantee group entities always have weight */
9491 update_load_set(&se->load, NICE_0_LOAD);
9492 se->parent = parent;
9495 static DEFINE_MUTEX(shares_mutex);
9497 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9502 * We can't change the weight of the root cgroup.
9507 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9509 mutex_lock(&shares_mutex);
9510 if (tg->shares == shares)
9513 tg->shares = shares;
9514 for_each_possible_cpu(i) {
9515 struct rq *rq = cpu_rq(i);
9516 struct sched_entity *se = tg->se[i];
9519 /* Propagate contribution to hierarchy */
9520 rq_lock_irqsave(rq, &rf);
9521 update_rq_clock(rq);
9522 for_each_sched_entity(se) {
9523 update_load_avg(se, UPDATE_TG);
9524 update_cfs_shares(se);
9526 rq_unlock_irqrestore(rq, &rf);
9530 mutex_unlock(&shares_mutex);
9533 #else /* CONFIG_FAIR_GROUP_SCHED */
9535 void free_fair_sched_group(struct task_group *tg) { }
9537 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9542 void online_fair_sched_group(struct task_group *tg) { }
9544 void unregister_fair_sched_group(struct task_group *tg) { }
9546 #endif /* CONFIG_FAIR_GROUP_SCHED */
9549 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9551 struct sched_entity *se = &task->se;
9552 unsigned int rr_interval = 0;
9555 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9558 if (rq->cfs.load.weight)
9559 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9565 * All the scheduling class methods:
9567 const struct sched_class fair_sched_class = {
9568 .next = &idle_sched_class,
9569 .enqueue_task = enqueue_task_fair,
9570 .dequeue_task = dequeue_task_fair,
9571 .yield_task = yield_task_fair,
9572 .yield_to_task = yield_to_task_fair,
9574 .check_preempt_curr = check_preempt_wakeup,
9576 .pick_next_task = pick_next_task_fair,
9577 .put_prev_task = put_prev_task_fair,
9580 .select_task_rq = select_task_rq_fair,
9581 .migrate_task_rq = migrate_task_rq_fair,
9583 .rq_online = rq_online_fair,
9584 .rq_offline = rq_offline_fair,
9586 .task_dead = task_dead_fair,
9587 .set_cpus_allowed = set_cpus_allowed_common,
9590 .set_curr_task = set_curr_task_fair,
9591 .task_tick = task_tick_fair,
9592 .task_fork = task_fork_fair,
9594 .prio_changed = prio_changed_fair,
9595 .switched_from = switched_from_fair,
9596 .switched_to = switched_to_fair,
9598 .get_rr_interval = get_rr_interval_fair,
9600 .update_curr = update_curr_fair,
9602 #ifdef CONFIG_FAIR_GROUP_SCHED
9603 .task_change_group = task_change_group_fair,
9607 #ifdef CONFIG_SCHED_DEBUG
9608 void print_cfs_stats(struct seq_file *m, int cpu)
9610 struct cfs_rq *cfs_rq, *pos;
9613 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9614 print_cfs_rq(m, cpu, cfs_rq);
9618 #ifdef CONFIG_NUMA_BALANCING
9619 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9622 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9624 for_each_online_node(node) {
9625 if (p->numa_faults) {
9626 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9627 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9629 if (p->numa_group) {
9630 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9631 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9633 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9636 #endif /* CONFIG_NUMA_BALANCING */
9637 #endif /* CONFIG_SCHED_DEBUG */
9639 __init void init_sched_fair_class(void)
9642 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9644 #ifdef CONFIG_NO_HZ_COMMON
9645 nohz.next_balance = jiffies;
9646 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);