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, false);
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;
1074 static unsigned int task_nr_scan_windows(struct task_struct *p)
1076 unsigned long rss = 0;
1077 unsigned long nr_scan_pages;
1080 * Calculations based on RSS as non-present and empty pages are skipped
1081 * by the PTE scanner and NUMA hinting faults should be trapped based
1084 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1085 rss = get_mm_rss(p->mm);
1087 rss = nr_scan_pages;
1089 rss = round_up(rss, nr_scan_pages);
1090 return rss / nr_scan_pages;
1093 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1094 #define MAX_SCAN_WINDOW 2560
1096 static unsigned int task_scan_min(struct task_struct *p)
1098 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1099 unsigned int scan, floor;
1100 unsigned int windows = 1;
1102 if (scan_size < MAX_SCAN_WINDOW)
1103 windows = MAX_SCAN_WINDOW / scan_size;
1104 floor = 1000 / windows;
1106 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1107 return max_t(unsigned int, floor, scan);
1110 static unsigned int task_scan_max(struct task_struct *p)
1112 unsigned int smin = task_scan_min(p);
1115 /* Watch for min being lower than max due to floor calculations */
1116 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1117 return max(smin, smax);
1120 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1122 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1123 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1126 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1128 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1129 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1135 spinlock_t lock; /* nr_tasks, tasks */
1140 struct rcu_head rcu;
1141 unsigned long total_faults;
1142 unsigned long max_faults_cpu;
1144 * Faults_cpu is used to decide whether memory should move
1145 * towards the CPU. As a consequence, these stats are weighted
1146 * more by CPU use than by memory faults.
1148 unsigned long *faults_cpu;
1149 unsigned long faults[0];
1152 /* Shared or private faults. */
1153 #define NR_NUMA_HINT_FAULT_TYPES 2
1155 /* Memory and CPU locality */
1156 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1158 /* Averaged statistics, and temporary buffers. */
1159 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1161 pid_t task_numa_group_id(struct task_struct *p)
1163 return p->numa_group ? p->numa_group->gid : 0;
1167 * The averaged statistics, shared & private, memory & cpu,
1168 * occupy the first half of the array. The second half of the
1169 * array is for current counters, which are averaged into the
1170 * first set by task_numa_placement.
1172 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1174 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1177 static inline unsigned long task_faults(struct task_struct *p, int nid)
1179 if (!p->numa_faults)
1182 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1183 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1186 static inline unsigned long group_faults(struct task_struct *p, int nid)
1191 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1192 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1195 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1197 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1198 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1202 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1203 * considered part of a numa group's pseudo-interleaving set. Migrations
1204 * between these nodes are slowed down, to allow things to settle down.
1206 #define ACTIVE_NODE_FRACTION 3
1208 static bool numa_is_active_node(int nid, struct numa_group *ng)
1210 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1213 /* Handle placement on systems where not all nodes are directly connected. */
1214 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1215 int maxdist, bool task)
1217 unsigned long score = 0;
1221 * All nodes are directly connected, and the same distance
1222 * from each other. No need for fancy placement algorithms.
1224 if (sched_numa_topology_type == NUMA_DIRECT)
1228 * This code is called for each node, introducing N^2 complexity,
1229 * which should be ok given the number of nodes rarely exceeds 8.
1231 for_each_online_node(node) {
1232 unsigned long faults;
1233 int dist = node_distance(nid, node);
1236 * The furthest away nodes in the system are not interesting
1237 * for placement; nid was already counted.
1239 if (dist == sched_max_numa_distance || node == nid)
1243 * On systems with a backplane NUMA topology, compare groups
1244 * of nodes, and move tasks towards the group with the most
1245 * memory accesses. When comparing two nodes at distance
1246 * "hoplimit", only nodes closer by than "hoplimit" are part
1247 * of each group. Skip other nodes.
1249 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1253 /* Add up the faults from nearby nodes. */
1255 faults = task_faults(p, node);
1257 faults = group_faults(p, node);
1260 * On systems with a glueless mesh NUMA topology, there are
1261 * no fixed "groups of nodes". Instead, nodes that are not
1262 * directly connected bounce traffic through intermediate
1263 * nodes; a numa_group can occupy any set of nodes.
1264 * The further away a node is, the less the faults count.
1265 * This seems to result in good task placement.
1267 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1268 faults *= (sched_max_numa_distance - dist);
1269 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1279 * These return the fraction of accesses done by a particular task, or
1280 * task group, on a particular numa node. The group weight is given a
1281 * larger multiplier, in order to group tasks together that are almost
1282 * evenly spread out between numa nodes.
1284 static inline unsigned long task_weight(struct task_struct *p, int nid,
1287 unsigned long faults, total_faults;
1289 if (!p->numa_faults)
1292 total_faults = p->total_numa_faults;
1297 faults = task_faults(p, nid);
1298 faults += score_nearby_nodes(p, nid, dist, true);
1300 return 1000 * faults / total_faults;
1303 static inline unsigned long group_weight(struct task_struct *p, int nid,
1306 unsigned long faults, total_faults;
1311 total_faults = p->numa_group->total_faults;
1316 faults = group_faults(p, nid);
1317 faults += score_nearby_nodes(p, nid, dist, false);
1319 return 1000 * faults / total_faults;
1322 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1323 int src_nid, int dst_cpu)
1325 struct numa_group *ng = p->numa_group;
1326 int dst_nid = cpu_to_node(dst_cpu);
1327 int last_cpupid, this_cpupid;
1329 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1332 * Multi-stage node selection is used in conjunction with a periodic
1333 * migration fault to build a temporal task<->page relation. By using
1334 * a two-stage filter we remove short/unlikely relations.
1336 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1337 * a task's usage of a particular page (n_p) per total usage of this
1338 * page (n_t) (in a given time-span) to a probability.
1340 * Our periodic faults will sample this probability and getting the
1341 * same result twice in a row, given these samples are fully
1342 * independent, is then given by P(n)^2, provided our sample period
1343 * is sufficiently short compared to the usage pattern.
1345 * This quadric squishes small probabilities, making it less likely we
1346 * act on an unlikely task<->page relation.
1348 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1349 if (!cpupid_pid_unset(last_cpupid) &&
1350 cpupid_to_nid(last_cpupid) != dst_nid)
1353 /* Always allow migrate on private faults */
1354 if (cpupid_match_pid(p, last_cpupid))
1357 /* A shared fault, but p->numa_group has not been set up yet. */
1362 * Destination node is much more heavily used than the source
1363 * node? Allow migration.
1365 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1366 ACTIVE_NODE_FRACTION)
1370 * Distribute memory according to CPU & memory use on each node,
1371 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1373 * faults_cpu(dst) 3 faults_cpu(src)
1374 * --------------- * - > ---------------
1375 * faults_mem(dst) 4 faults_mem(src)
1377 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1378 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1381 static unsigned long weighted_cpuload(const int cpu);
1382 static unsigned long source_load(int cpu, int type);
1383 static unsigned long target_load(int cpu, int type);
1384 static unsigned long capacity_of(int cpu);
1386 /* Cached statistics for all CPUs within a node */
1388 unsigned long nr_running;
1391 /* Total compute capacity of CPUs on a node */
1392 unsigned long compute_capacity;
1394 /* Approximate capacity in terms of runnable tasks on a node */
1395 unsigned long task_capacity;
1396 int has_free_capacity;
1400 * XXX borrowed from update_sg_lb_stats
1402 static void update_numa_stats(struct numa_stats *ns, int nid)
1404 int smt, cpu, cpus = 0;
1405 unsigned long capacity;
1407 memset(ns, 0, sizeof(*ns));
1408 for_each_cpu(cpu, cpumask_of_node(nid)) {
1409 struct rq *rq = cpu_rq(cpu);
1411 ns->nr_running += rq->nr_running;
1412 ns->load += weighted_cpuload(cpu);
1413 ns->compute_capacity += capacity_of(cpu);
1419 * If we raced with hotplug and there are no CPUs left in our mask
1420 * the @ns structure is NULL'ed and task_numa_compare() will
1421 * not find this node attractive.
1423 * We'll either bail at !has_free_capacity, or we'll detect a huge
1424 * imbalance and bail there.
1429 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1430 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1431 capacity = cpus / smt; /* cores */
1433 ns->task_capacity = min_t(unsigned, capacity,
1434 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1435 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1438 struct task_numa_env {
1439 struct task_struct *p;
1441 int src_cpu, src_nid;
1442 int dst_cpu, dst_nid;
1444 struct numa_stats src_stats, dst_stats;
1449 struct task_struct *best_task;
1454 static void task_numa_assign(struct task_numa_env *env,
1455 struct task_struct *p, long imp)
1458 put_task_struct(env->best_task);
1463 env->best_imp = imp;
1464 env->best_cpu = env->dst_cpu;
1467 static bool load_too_imbalanced(long src_load, long dst_load,
1468 struct task_numa_env *env)
1471 long orig_src_load, orig_dst_load;
1472 long src_capacity, dst_capacity;
1475 * The load is corrected for the CPU capacity available on each node.
1478 * ------------ vs ---------
1479 * src_capacity dst_capacity
1481 src_capacity = env->src_stats.compute_capacity;
1482 dst_capacity = env->dst_stats.compute_capacity;
1484 /* We care about the slope of the imbalance, not the direction. */
1485 if (dst_load < src_load)
1486 swap(dst_load, src_load);
1488 /* Is the difference below the threshold? */
1489 imb = dst_load * src_capacity * 100 -
1490 src_load * dst_capacity * env->imbalance_pct;
1495 * The imbalance is above the allowed threshold.
1496 * Compare it with the old imbalance.
1498 orig_src_load = env->src_stats.load;
1499 orig_dst_load = env->dst_stats.load;
1501 if (orig_dst_load < orig_src_load)
1502 swap(orig_dst_load, orig_src_load);
1504 old_imb = orig_dst_load * src_capacity * 100 -
1505 orig_src_load * dst_capacity * env->imbalance_pct;
1507 /* Would this change make things worse? */
1508 return (imb > old_imb);
1512 * This checks if the overall compute and NUMA accesses of the system would
1513 * be improved if the source tasks was migrated to the target dst_cpu taking
1514 * into account that it might be best if task running on the dst_cpu should
1515 * be exchanged with the source task
1517 static void task_numa_compare(struct task_numa_env *env,
1518 long taskimp, long groupimp)
1520 struct rq *src_rq = cpu_rq(env->src_cpu);
1521 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1522 struct task_struct *cur;
1523 long src_load, dst_load;
1525 long imp = env->p->numa_group ? groupimp : taskimp;
1527 int dist = env->dist;
1530 cur = task_rcu_dereference(&dst_rq->curr);
1531 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1535 * Because we have preemption enabled we can get migrated around and
1536 * end try selecting ourselves (current == env->p) as a swap candidate.
1542 * "imp" is the fault differential for the source task between the
1543 * source and destination node. Calculate the total differential for
1544 * the source task and potential destination task. The more negative
1545 * the value is, the more rmeote accesses that would be expected to
1546 * be incurred if the tasks were swapped.
1549 /* Skip this swap candidate if cannot move to the source cpu */
1550 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1554 * If dst and source tasks are in the same NUMA group, or not
1555 * in any group then look only at task weights.
1557 if (cur->numa_group == env->p->numa_group) {
1558 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1559 task_weight(cur, env->dst_nid, dist);
1561 * Add some hysteresis to prevent swapping the
1562 * tasks within a group over tiny differences.
1564 if (cur->numa_group)
1568 * Compare the group weights. If a task is all by
1569 * itself (not part of a group), use the task weight
1572 if (cur->numa_group)
1573 imp += group_weight(cur, env->src_nid, dist) -
1574 group_weight(cur, env->dst_nid, dist);
1576 imp += task_weight(cur, env->src_nid, dist) -
1577 task_weight(cur, env->dst_nid, dist);
1581 if (imp <= env->best_imp && moveimp <= env->best_imp)
1585 /* Is there capacity at our destination? */
1586 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1587 !env->dst_stats.has_free_capacity)
1593 /* Balance doesn't matter much if we're running a task per cpu */
1594 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1595 dst_rq->nr_running == 1)
1599 * In the overloaded case, try and keep the load balanced.
1602 load = task_h_load(env->p);
1603 dst_load = env->dst_stats.load + load;
1604 src_load = env->src_stats.load - load;
1606 if (moveimp > imp && moveimp > env->best_imp) {
1608 * If the improvement from just moving env->p direction is
1609 * better than swapping tasks around, check if a move is
1610 * possible. Store a slightly smaller score than moveimp,
1611 * so an actually idle CPU will win.
1613 if (!load_too_imbalanced(src_load, dst_load, env)) {
1620 if (imp <= env->best_imp)
1624 load = task_h_load(cur);
1629 if (load_too_imbalanced(src_load, dst_load, env))
1633 * One idle CPU per node is evaluated for a task numa move.
1634 * Call select_idle_sibling to maybe find a better one.
1638 * select_idle_siblings() uses an per-cpu cpumask that
1639 * can be used from IRQ context.
1641 local_irq_disable();
1642 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1648 task_numa_assign(env, cur, imp);
1653 static void task_numa_find_cpu(struct task_numa_env *env,
1654 long taskimp, long groupimp)
1658 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1659 /* Skip this CPU if the source task cannot migrate */
1660 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1664 task_numa_compare(env, taskimp, groupimp);
1668 /* Only move tasks to a NUMA node less busy than the current node. */
1669 static bool numa_has_capacity(struct task_numa_env *env)
1671 struct numa_stats *src = &env->src_stats;
1672 struct numa_stats *dst = &env->dst_stats;
1674 if (src->has_free_capacity && !dst->has_free_capacity)
1678 * Only consider a task move if the source has a higher load
1679 * than the destination, corrected for CPU capacity on each node.
1681 * src->load dst->load
1682 * --------------------- vs ---------------------
1683 * src->compute_capacity dst->compute_capacity
1685 if (src->load * dst->compute_capacity * env->imbalance_pct >
1687 dst->load * src->compute_capacity * 100)
1693 static int task_numa_migrate(struct task_struct *p)
1695 struct task_numa_env env = {
1698 .src_cpu = task_cpu(p),
1699 .src_nid = task_node(p),
1701 .imbalance_pct = 112,
1707 struct sched_domain *sd;
1708 unsigned long taskweight, groupweight;
1710 long taskimp, groupimp;
1713 * Pick the lowest SD_NUMA domain, as that would have the smallest
1714 * imbalance and would be the first to start moving tasks about.
1716 * And we want to avoid any moving of tasks about, as that would create
1717 * random movement of tasks -- counter the numa conditions we're trying
1721 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1723 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1727 * Cpusets can break the scheduler domain tree into smaller
1728 * balance domains, some of which do not cross NUMA boundaries.
1729 * Tasks that are "trapped" in such domains cannot be migrated
1730 * elsewhere, so there is no point in (re)trying.
1732 if (unlikely(!sd)) {
1733 p->numa_preferred_nid = task_node(p);
1737 env.dst_nid = p->numa_preferred_nid;
1738 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1739 taskweight = task_weight(p, env.src_nid, dist);
1740 groupweight = group_weight(p, env.src_nid, dist);
1741 update_numa_stats(&env.src_stats, env.src_nid);
1742 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1743 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1744 update_numa_stats(&env.dst_stats, env.dst_nid);
1746 /* Try to find a spot on the preferred nid. */
1747 if (numa_has_capacity(&env))
1748 task_numa_find_cpu(&env, taskimp, groupimp);
1751 * Look at other nodes in these cases:
1752 * - there is no space available on the preferred_nid
1753 * - the task is part of a numa_group that is interleaved across
1754 * multiple NUMA nodes; in order to better consolidate the group,
1755 * we need to check other locations.
1757 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1758 for_each_online_node(nid) {
1759 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1762 dist = node_distance(env.src_nid, env.dst_nid);
1763 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1765 taskweight = task_weight(p, env.src_nid, dist);
1766 groupweight = group_weight(p, env.src_nid, dist);
1769 /* Only consider nodes where both task and groups benefit */
1770 taskimp = task_weight(p, nid, dist) - taskweight;
1771 groupimp = group_weight(p, nid, dist) - groupweight;
1772 if (taskimp < 0 && groupimp < 0)
1777 update_numa_stats(&env.dst_stats, env.dst_nid);
1778 if (numa_has_capacity(&env))
1779 task_numa_find_cpu(&env, taskimp, groupimp);
1784 * If the task is part of a workload that spans multiple NUMA nodes,
1785 * and is migrating into one of the workload's active nodes, remember
1786 * this node as the task's preferred numa node, so the workload can
1788 * A task that migrated to a second choice node will be better off
1789 * trying for a better one later. Do not set the preferred node here.
1791 if (p->numa_group) {
1792 struct numa_group *ng = p->numa_group;
1794 if (env.best_cpu == -1)
1799 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1800 sched_setnuma(p, env.dst_nid);
1803 /* No better CPU than the current one was found. */
1804 if (env.best_cpu == -1)
1808 * Reset the scan period if the task is being rescheduled on an
1809 * alternative node to recheck if the tasks is now properly placed.
1811 p->numa_scan_period = task_scan_min(p);
1813 if (env.best_task == NULL) {
1814 ret = migrate_task_to(p, env.best_cpu);
1816 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1820 ret = migrate_swap(p, env.best_task);
1822 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1823 put_task_struct(env.best_task);
1827 /* Attempt to migrate a task to a CPU on the preferred node. */
1828 static void numa_migrate_preferred(struct task_struct *p)
1830 unsigned long interval = HZ;
1832 /* This task has no NUMA fault statistics yet */
1833 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1836 /* Periodically retry migrating the task to the preferred node */
1837 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1838 p->numa_migrate_retry = jiffies + interval;
1840 /* Success if task is already running on preferred CPU */
1841 if (task_node(p) == p->numa_preferred_nid)
1844 /* Otherwise, try migrate to a CPU on the preferred node */
1845 task_numa_migrate(p);
1849 * Find out how many nodes on the workload is actively running on. Do this by
1850 * tracking the nodes from which NUMA hinting faults are triggered. This can
1851 * be different from the set of nodes where the workload's memory is currently
1854 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1856 unsigned long faults, max_faults = 0;
1857 int nid, active_nodes = 0;
1859 for_each_online_node(nid) {
1860 faults = group_faults_cpu(numa_group, nid);
1861 if (faults > max_faults)
1862 max_faults = faults;
1865 for_each_online_node(nid) {
1866 faults = group_faults_cpu(numa_group, nid);
1867 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1871 numa_group->max_faults_cpu = max_faults;
1872 numa_group->active_nodes = active_nodes;
1876 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1877 * increments. The more local the fault statistics are, the higher the scan
1878 * period will be for the next scan window. If local/(local+remote) ratio is
1879 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1880 * the scan period will decrease. Aim for 70% local accesses.
1882 #define NUMA_PERIOD_SLOTS 10
1883 #define NUMA_PERIOD_THRESHOLD 7
1886 * Increase the scan period (slow down scanning) if the majority of
1887 * our memory is already on our local node, or if the majority of
1888 * the page accesses are shared with other processes.
1889 * Otherwise, decrease the scan period.
1891 static void update_task_scan_period(struct task_struct *p,
1892 unsigned long shared, unsigned long private)
1894 unsigned int period_slot;
1898 unsigned long remote = p->numa_faults_locality[0];
1899 unsigned long local = p->numa_faults_locality[1];
1902 * If there were no record hinting faults then either the task is
1903 * completely idle or all activity is areas that are not of interest
1904 * to automatic numa balancing. Related to that, if there were failed
1905 * migration then it implies we are migrating too quickly or the local
1906 * node is overloaded. In either case, scan slower
1908 if (local + shared == 0 || p->numa_faults_locality[2]) {
1909 p->numa_scan_period = min(p->numa_scan_period_max,
1910 p->numa_scan_period << 1);
1912 p->mm->numa_next_scan = jiffies +
1913 msecs_to_jiffies(p->numa_scan_period);
1919 * Prepare to scale scan period relative to the current period.
1920 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1921 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1922 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1924 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1925 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1926 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1927 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1930 diff = slot * period_slot;
1932 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1935 * Scale scan rate increases based on sharing. There is an
1936 * inverse relationship between the degree of sharing and
1937 * the adjustment made to the scanning period. Broadly
1938 * speaking the intent is that there is little point
1939 * scanning faster if shared accesses dominate as it may
1940 * simply bounce migrations uselessly
1942 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1943 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1946 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1947 task_scan_min(p), task_scan_max(p));
1948 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1952 * Get the fraction of time the task has been running since the last
1953 * NUMA placement cycle. The scheduler keeps similar statistics, but
1954 * decays those on a 32ms period, which is orders of magnitude off
1955 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1956 * stats only if the task is so new there are no NUMA statistics yet.
1958 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1960 u64 runtime, delta, now;
1961 /* Use the start of this time slice to avoid calculations. */
1962 now = p->se.exec_start;
1963 runtime = p->se.sum_exec_runtime;
1965 if (p->last_task_numa_placement) {
1966 delta = runtime - p->last_sum_exec_runtime;
1967 *period = now - p->last_task_numa_placement;
1969 delta = p->se.avg.load_sum / p->se.load.weight;
1970 *period = LOAD_AVG_MAX;
1973 p->last_sum_exec_runtime = runtime;
1974 p->last_task_numa_placement = now;
1980 * Determine the preferred nid for a task in a numa_group. This needs to
1981 * be done in a way that produces consistent results with group_weight,
1982 * otherwise workloads might not converge.
1984 static int preferred_group_nid(struct task_struct *p, int nid)
1989 /* Direct connections between all NUMA nodes. */
1990 if (sched_numa_topology_type == NUMA_DIRECT)
1994 * On a system with glueless mesh NUMA topology, group_weight
1995 * scores nodes according to the number of NUMA hinting faults on
1996 * both the node itself, and on nearby nodes.
1998 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1999 unsigned long score, max_score = 0;
2000 int node, max_node = nid;
2002 dist = sched_max_numa_distance;
2004 for_each_online_node(node) {
2005 score = group_weight(p, node, dist);
2006 if (score > max_score) {
2015 * Finding the preferred nid in a system with NUMA backplane
2016 * interconnect topology is more involved. The goal is to locate
2017 * tasks from numa_groups near each other in the system, and
2018 * untangle workloads from different sides of the system. This requires
2019 * searching down the hierarchy of node groups, recursively searching
2020 * inside the highest scoring group of nodes. The nodemask tricks
2021 * keep the complexity of the search down.
2023 nodes = node_online_map;
2024 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2025 unsigned long max_faults = 0;
2026 nodemask_t max_group = NODE_MASK_NONE;
2029 /* Are there nodes at this distance from each other? */
2030 if (!find_numa_distance(dist))
2033 for_each_node_mask(a, nodes) {
2034 unsigned long faults = 0;
2035 nodemask_t this_group;
2036 nodes_clear(this_group);
2038 /* Sum group's NUMA faults; includes a==b case. */
2039 for_each_node_mask(b, nodes) {
2040 if (node_distance(a, b) < dist) {
2041 faults += group_faults(p, b);
2042 node_set(b, this_group);
2043 node_clear(b, nodes);
2047 /* Remember the top group. */
2048 if (faults > max_faults) {
2049 max_faults = faults;
2050 max_group = this_group;
2052 * subtle: at the smallest distance there is
2053 * just one node left in each "group", the
2054 * winner is the preferred nid.
2059 /* Next round, evaluate the nodes within max_group. */
2067 static void task_numa_placement(struct task_struct *p)
2069 int seq, nid, max_nid = -1, max_group_nid = -1;
2070 unsigned long max_faults = 0, max_group_faults = 0;
2071 unsigned long fault_types[2] = { 0, 0 };
2072 unsigned long total_faults;
2073 u64 runtime, period;
2074 spinlock_t *group_lock = NULL;
2077 * The p->mm->numa_scan_seq field gets updated without
2078 * exclusive access. Use READ_ONCE() here to ensure
2079 * that the field is read in a single access:
2081 seq = READ_ONCE(p->mm->numa_scan_seq);
2082 if (p->numa_scan_seq == seq)
2084 p->numa_scan_seq = seq;
2085 p->numa_scan_period_max = task_scan_max(p);
2087 total_faults = p->numa_faults_locality[0] +
2088 p->numa_faults_locality[1];
2089 runtime = numa_get_avg_runtime(p, &period);
2091 /* If the task is part of a group prevent parallel updates to group stats */
2092 if (p->numa_group) {
2093 group_lock = &p->numa_group->lock;
2094 spin_lock_irq(group_lock);
2097 /* Find the node with the highest number of faults */
2098 for_each_online_node(nid) {
2099 /* Keep track of the offsets in numa_faults array */
2100 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2101 unsigned long faults = 0, group_faults = 0;
2104 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2105 long diff, f_diff, f_weight;
2107 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2108 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2109 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2110 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2112 /* Decay existing window, copy faults since last scan */
2113 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2114 fault_types[priv] += p->numa_faults[membuf_idx];
2115 p->numa_faults[membuf_idx] = 0;
2118 * Normalize the faults_from, so all tasks in a group
2119 * count according to CPU use, instead of by the raw
2120 * number of faults. Tasks with little runtime have
2121 * little over-all impact on throughput, and thus their
2122 * faults are less important.
2124 f_weight = div64_u64(runtime << 16, period + 1);
2125 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2127 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2128 p->numa_faults[cpubuf_idx] = 0;
2130 p->numa_faults[mem_idx] += diff;
2131 p->numa_faults[cpu_idx] += f_diff;
2132 faults += p->numa_faults[mem_idx];
2133 p->total_numa_faults += diff;
2134 if (p->numa_group) {
2136 * safe because we can only change our own group
2138 * mem_idx represents the offset for a given
2139 * nid and priv in a specific region because it
2140 * is at the beginning of the numa_faults array.
2142 p->numa_group->faults[mem_idx] += diff;
2143 p->numa_group->faults_cpu[mem_idx] += f_diff;
2144 p->numa_group->total_faults += diff;
2145 group_faults += p->numa_group->faults[mem_idx];
2149 if (faults > max_faults) {
2150 max_faults = faults;
2154 if (group_faults > max_group_faults) {
2155 max_group_faults = group_faults;
2156 max_group_nid = nid;
2160 update_task_scan_period(p, fault_types[0], fault_types[1]);
2162 if (p->numa_group) {
2163 numa_group_count_active_nodes(p->numa_group);
2164 spin_unlock_irq(group_lock);
2165 max_nid = preferred_group_nid(p, max_group_nid);
2169 /* Set the new preferred node */
2170 if (max_nid != p->numa_preferred_nid)
2171 sched_setnuma(p, max_nid);
2173 if (task_node(p) != p->numa_preferred_nid)
2174 numa_migrate_preferred(p);
2178 static inline int get_numa_group(struct numa_group *grp)
2180 return atomic_inc_not_zero(&grp->refcount);
2183 static inline void put_numa_group(struct numa_group *grp)
2185 if (atomic_dec_and_test(&grp->refcount))
2186 kfree_rcu(grp, rcu);
2189 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2192 struct numa_group *grp, *my_grp;
2193 struct task_struct *tsk;
2195 int cpu = cpupid_to_cpu(cpupid);
2198 if (unlikely(!p->numa_group)) {
2199 unsigned int size = sizeof(struct numa_group) +
2200 4*nr_node_ids*sizeof(unsigned long);
2202 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2206 atomic_set(&grp->refcount, 1);
2207 grp->active_nodes = 1;
2208 grp->max_faults_cpu = 0;
2209 spin_lock_init(&grp->lock);
2211 /* Second half of the array tracks nids where faults happen */
2212 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2215 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2216 grp->faults[i] = p->numa_faults[i];
2218 grp->total_faults = p->total_numa_faults;
2221 rcu_assign_pointer(p->numa_group, grp);
2225 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2227 if (!cpupid_match_pid(tsk, cpupid))
2230 grp = rcu_dereference(tsk->numa_group);
2234 my_grp = p->numa_group;
2239 * Only join the other group if its bigger; if we're the bigger group,
2240 * the other task will join us.
2242 if (my_grp->nr_tasks > grp->nr_tasks)
2246 * Tie-break on the grp address.
2248 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2251 /* Always join threads in the same process. */
2252 if (tsk->mm == current->mm)
2255 /* Simple filter to avoid false positives due to PID collisions */
2256 if (flags & TNF_SHARED)
2259 /* Update priv based on whether false sharing was detected */
2262 if (join && !get_numa_group(grp))
2270 BUG_ON(irqs_disabled());
2271 double_lock_irq(&my_grp->lock, &grp->lock);
2273 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2274 my_grp->faults[i] -= p->numa_faults[i];
2275 grp->faults[i] += p->numa_faults[i];
2277 my_grp->total_faults -= p->total_numa_faults;
2278 grp->total_faults += p->total_numa_faults;
2283 spin_unlock(&my_grp->lock);
2284 spin_unlock_irq(&grp->lock);
2286 rcu_assign_pointer(p->numa_group, grp);
2288 put_numa_group(my_grp);
2296 void task_numa_free(struct task_struct *p)
2298 struct numa_group *grp = p->numa_group;
2299 void *numa_faults = p->numa_faults;
2300 unsigned long flags;
2304 spin_lock_irqsave(&grp->lock, flags);
2305 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2306 grp->faults[i] -= p->numa_faults[i];
2307 grp->total_faults -= p->total_numa_faults;
2310 spin_unlock_irqrestore(&grp->lock, flags);
2311 RCU_INIT_POINTER(p->numa_group, NULL);
2312 put_numa_group(grp);
2315 p->numa_faults = NULL;
2320 * Got a PROT_NONE fault for a page on @node.
2322 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2324 struct task_struct *p = current;
2325 bool migrated = flags & TNF_MIGRATED;
2326 int cpu_node = task_node(current);
2327 int local = !!(flags & TNF_FAULT_LOCAL);
2328 struct numa_group *ng;
2331 if (!static_branch_likely(&sched_numa_balancing))
2334 /* for example, ksmd faulting in a user's mm */
2338 /* Allocate buffer to track faults on a per-node basis */
2339 if (unlikely(!p->numa_faults)) {
2340 int size = sizeof(*p->numa_faults) *
2341 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2343 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2344 if (!p->numa_faults)
2347 p->total_numa_faults = 0;
2348 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2352 * First accesses are treated as private, otherwise consider accesses
2353 * to be private if the accessing pid has not changed
2355 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2358 priv = cpupid_match_pid(p, last_cpupid);
2359 if (!priv && !(flags & TNF_NO_GROUP))
2360 task_numa_group(p, last_cpupid, flags, &priv);
2364 * If a workload spans multiple NUMA nodes, a shared fault that
2365 * occurs wholly within the set of nodes that the workload is
2366 * actively using should be counted as local. This allows the
2367 * scan rate to slow down when a workload has settled down.
2370 if (!priv && !local && ng && ng->active_nodes > 1 &&
2371 numa_is_active_node(cpu_node, ng) &&
2372 numa_is_active_node(mem_node, ng))
2375 task_numa_placement(p);
2378 * Retry task to preferred node migration periodically, in case it
2379 * case it previously failed, or the scheduler moved us.
2381 if (time_after(jiffies, p->numa_migrate_retry))
2382 numa_migrate_preferred(p);
2385 p->numa_pages_migrated += pages;
2386 if (flags & TNF_MIGRATE_FAIL)
2387 p->numa_faults_locality[2] += pages;
2389 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2390 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2391 p->numa_faults_locality[local] += pages;
2394 static void reset_ptenuma_scan(struct task_struct *p)
2397 * We only did a read acquisition of the mmap sem, so
2398 * p->mm->numa_scan_seq is written to without exclusive access
2399 * and the update is not guaranteed to be atomic. That's not
2400 * much of an issue though, since this is just used for
2401 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2402 * expensive, to avoid any form of compiler optimizations:
2404 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2405 p->mm->numa_scan_offset = 0;
2409 * The expensive part of numa migration is done from task_work context.
2410 * Triggered from task_tick_numa().
2412 void task_numa_work(struct callback_head *work)
2414 unsigned long migrate, next_scan, now = jiffies;
2415 struct task_struct *p = current;
2416 struct mm_struct *mm = p->mm;
2417 u64 runtime = p->se.sum_exec_runtime;
2418 struct vm_area_struct *vma;
2419 unsigned long start, end;
2420 unsigned long nr_pte_updates = 0;
2421 long pages, virtpages;
2423 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2425 work->next = work; /* protect against double add */
2427 * Who cares about NUMA placement when they're dying.
2429 * NOTE: make sure not to dereference p->mm before this check,
2430 * exit_task_work() happens _after_ exit_mm() so we could be called
2431 * without p->mm even though we still had it when we enqueued this
2434 if (p->flags & PF_EXITING)
2437 if (!mm->numa_next_scan) {
2438 mm->numa_next_scan = now +
2439 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2443 * Enforce maximal scan/migration frequency..
2445 migrate = mm->numa_next_scan;
2446 if (time_before(now, migrate))
2449 if (p->numa_scan_period == 0) {
2450 p->numa_scan_period_max = task_scan_max(p);
2451 p->numa_scan_period = task_scan_min(p);
2454 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2455 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2459 * Delay this task enough that another task of this mm will likely win
2460 * the next time around.
2462 p->node_stamp += 2 * TICK_NSEC;
2464 start = mm->numa_scan_offset;
2465 pages = sysctl_numa_balancing_scan_size;
2466 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2467 virtpages = pages * 8; /* Scan up to this much virtual space */
2472 if (!down_read_trylock(&mm->mmap_sem))
2474 vma = find_vma(mm, start);
2476 reset_ptenuma_scan(p);
2480 for (; vma; vma = vma->vm_next) {
2481 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2482 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2487 * Shared library pages mapped by multiple processes are not
2488 * migrated as it is expected they are cache replicated. Avoid
2489 * hinting faults in read-only file-backed mappings or the vdso
2490 * as migrating the pages will be of marginal benefit.
2493 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2497 * Skip inaccessible VMAs to avoid any confusion between
2498 * PROT_NONE and NUMA hinting ptes
2500 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2504 start = max(start, vma->vm_start);
2505 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2506 end = min(end, vma->vm_end);
2507 nr_pte_updates = change_prot_numa(vma, start, end);
2510 * Try to scan sysctl_numa_balancing_size worth of
2511 * hpages that have at least one present PTE that
2512 * is not already pte-numa. If the VMA contains
2513 * areas that are unused or already full of prot_numa
2514 * PTEs, scan up to virtpages, to skip through those
2518 pages -= (end - start) >> PAGE_SHIFT;
2519 virtpages -= (end - start) >> PAGE_SHIFT;
2522 if (pages <= 0 || virtpages <= 0)
2526 } while (end != vma->vm_end);
2531 * It is possible to reach the end of the VMA list but the last few
2532 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2533 * would find the !migratable VMA on the next scan but not reset the
2534 * scanner to the start so check it now.
2537 mm->numa_scan_offset = start;
2539 reset_ptenuma_scan(p);
2540 up_read(&mm->mmap_sem);
2543 * Make sure tasks use at least 32x as much time to run other code
2544 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2545 * Usually update_task_scan_period slows down scanning enough; on an
2546 * overloaded system we need to limit overhead on a per task basis.
2548 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2549 u64 diff = p->se.sum_exec_runtime - runtime;
2550 p->node_stamp += 32 * diff;
2555 * Drive the periodic memory faults..
2557 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2559 struct callback_head *work = &curr->numa_work;
2563 * We don't care about NUMA placement if we don't have memory.
2565 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2569 * Using runtime rather than walltime has the dual advantage that
2570 * we (mostly) drive the selection from busy threads and that the
2571 * task needs to have done some actual work before we bother with
2574 now = curr->se.sum_exec_runtime;
2575 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2577 if (now > curr->node_stamp + period) {
2578 if (!curr->node_stamp)
2579 curr->numa_scan_period = task_scan_min(curr);
2580 curr->node_stamp += period;
2582 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2583 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2584 task_work_add(curr, work, true);
2590 * Can a task be moved from prev_cpu to this_cpu without causing a load
2591 * imbalance that would trigger the load balancer?
2593 static inline bool numa_wake_affine(struct sched_domain *sd,
2594 struct task_struct *p, int this_cpu,
2595 int prev_cpu, int sync)
2597 struct numa_stats prev_load, this_load;
2598 s64 this_eff_load, prev_eff_load;
2600 update_numa_stats(&prev_load, cpu_to_node(prev_cpu));
2601 update_numa_stats(&this_load, cpu_to_node(this_cpu));
2604 * If sync wakeup then subtract the (maximum possible)
2605 * effect of the currently running task from the load
2606 * of the current CPU:
2609 unsigned long current_load = task_h_load(current);
2611 if (this_load.load > current_load)
2612 this_load.load -= current_load;
2618 * In low-load situations, where this_cpu's node is idle due to the
2619 * sync cause above having dropped this_load.load to 0, move the task.
2620 * Moving to an idle socket will not create a bad imbalance.
2622 * Otherwise check if the nodes are near enough in load to allow this
2623 * task to be woken on this_cpu's node.
2625 if (this_load.load > 0) {
2626 unsigned long task_load = task_h_load(p);
2628 this_eff_load = 100;
2629 this_eff_load *= prev_load.compute_capacity;
2631 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2632 prev_eff_load *= this_load.compute_capacity;
2634 this_eff_load *= this_load.load + task_load;
2635 prev_eff_load *= prev_load.load - task_load;
2637 return this_eff_load <= prev_eff_load;
2643 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2647 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2651 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2656 static inline bool numa_wake_affine(struct sched_domain *sd,
2657 struct task_struct *p, int this_cpu,
2658 int prev_cpu, int sync)
2663 #endif /* CONFIG_NUMA_BALANCING */
2666 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2668 update_load_add(&cfs_rq->load, se->load.weight);
2669 if (!parent_entity(se))
2670 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2672 if (entity_is_task(se)) {
2673 struct rq *rq = rq_of(cfs_rq);
2675 account_numa_enqueue(rq, task_of(se));
2676 list_add(&se->group_node, &rq->cfs_tasks);
2679 cfs_rq->nr_running++;
2683 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2685 update_load_sub(&cfs_rq->load, se->load.weight);
2686 if (!parent_entity(se))
2687 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2689 if (entity_is_task(se)) {
2690 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2691 list_del_init(&se->group_node);
2694 cfs_rq->nr_running--;
2697 #ifdef CONFIG_FAIR_GROUP_SCHED
2699 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2701 long tg_weight, load, shares;
2704 * This really should be: cfs_rq->avg.load_avg, but instead we use
2705 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2706 * the shares for small weight interactive tasks.
2708 load = scale_load_down(cfs_rq->load.weight);
2710 tg_weight = atomic_long_read(&tg->load_avg);
2712 /* Ensure tg_weight >= load */
2713 tg_weight -= cfs_rq->tg_load_avg_contrib;
2716 shares = (tg->shares * load);
2718 shares /= tg_weight;
2721 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2722 * of a group with small tg->shares value. It is a floor value which is
2723 * assigned as a minimum load.weight to the sched_entity representing
2724 * the group on a CPU.
2726 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2727 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2728 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2729 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2732 if (shares < MIN_SHARES)
2733 shares = MIN_SHARES;
2734 if (shares > tg->shares)
2735 shares = tg->shares;
2739 # else /* CONFIG_SMP */
2740 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2744 # endif /* CONFIG_SMP */
2746 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2747 unsigned long weight)
2750 /* commit outstanding execution time */
2751 if (cfs_rq->curr == se)
2752 update_curr(cfs_rq);
2753 account_entity_dequeue(cfs_rq, se);
2756 update_load_set(&se->load, weight);
2759 account_entity_enqueue(cfs_rq, se);
2762 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2764 static void update_cfs_shares(struct sched_entity *se)
2766 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2767 struct task_group *tg;
2773 if (throttled_hierarchy(cfs_rq))
2779 if (likely(se->load.weight == tg->shares))
2782 shares = calc_cfs_shares(cfs_rq, tg);
2784 reweight_entity(cfs_rq_of(se), se, shares);
2787 #else /* CONFIG_FAIR_GROUP_SCHED */
2788 static inline void update_cfs_shares(struct sched_entity *se)
2791 #endif /* CONFIG_FAIR_GROUP_SCHED */
2796 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2798 static u64 decay_load(u64 val, u64 n)
2800 unsigned int local_n;
2802 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2805 /* after bounds checking we can collapse to 32-bit */
2809 * As y^PERIOD = 1/2, we can combine
2810 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2811 * With a look-up table which covers y^n (n<PERIOD)
2813 * To achieve constant time decay_load.
2815 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2816 val >>= local_n / LOAD_AVG_PERIOD;
2817 local_n %= LOAD_AVG_PERIOD;
2820 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2824 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2826 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2831 c1 = decay_load((u64)d1, periods);
2835 * c2 = 1024 \Sum y^n
2839 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2842 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2844 return c1 + c2 + c3;
2847 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2850 * Accumulate the three separate parts of the sum; d1 the remainder
2851 * of the last (incomplete) period, d2 the span of full periods and d3
2852 * the remainder of the (incomplete) current period.
2857 * |<->|<----------------->|<--->|
2858 * ... |---x---|------| ... |------|-----x (now)
2861 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2864 * = u y^p + (Step 1)
2867 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2870 static __always_inline u32
2871 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2872 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2874 unsigned long scale_freq, scale_cpu;
2875 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2878 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2879 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2881 delta += sa->period_contrib;
2882 periods = delta / 1024; /* A period is 1024us (~1ms) */
2885 * Step 1: decay old *_sum if we crossed period boundaries.
2888 sa->load_sum = decay_load(sa->load_sum, periods);
2890 cfs_rq->runnable_load_sum =
2891 decay_load(cfs_rq->runnable_load_sum, periods);
2893 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2899 contrib = __accumulate_pelt_segments(periods,
2900 1024 - sa->period_contrib, delta);
2902 sa->period_contrib = delta;
2904 contrib = cap_scale(contrib, scale_freq);
2906 sa->load_sum += weight * contrib;
2908 cfs_rq->runnable_load_sum += weight * contrib;
2911 sa->util_sum += contrib * scale_cpu;
2917 * We can represent the historical contribution to runnable average as the
2918 * coefficients of a geometric series. To do this we sub-divide our runnable
2919 * history into segments of approximately 1ms (1024us); label the segment that
2920 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2922 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2924 * (now) (~1ms ago) (~2ms ago)
2926 * Let u_i denote the fraction of p_i that the entity was runnable.
2928 * We then designate the fractions u_i as our co-efficients, yielding the
2929 * following representation of historical load:
2930 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2932 * We choose y based on the with of a reasonably scheduling period, fixing:
2935 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2936 * approximately half as much as the contribution to load within the last ms
2939 * When a period "rolls over" and we have new u_0`, multiplying the previous
2940 * sum again by y is sufficient to update:
2941 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2942 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2944 static __always_inline int
2945 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2946 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2950 delta = now - sa->last_update_time;
2952 * This should only happen when time goes backwards, which it
2953 * unfortunately does during sched clock init when we swap over to TSC.
2955 if ((s64)delta < 0) {
2956 sa->last_update_time = now;
2961 * Use 1024ns as the unit of measurement since it's a reasonable
2962 * approximation of 1us and fast to compute.
2968 sa->last_update_time += delta << 10;
2971 * Now we know we crossed measurement unit boundaries. The *_avg
2972 * accrues by two steps:
2974 * Step 1: accumulate *_sum since last_update_time. If we haven't
2975 * crossed period boundaries, finish.
2977 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
2981 * Step 2: update *_avg.
2983 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2985 cfs_rq->runnable_load_avg =
2986 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2988 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
2994 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
2996 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
3000 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3002 return ___update_load_avg(now, cpu, &se->avg,
3003 se->on_rq * scale_load_down(se->load.weight),
3004 cfs_rq->curr == se, NULL);
3008 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3010 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3011 scale_load_down(cfs_rq->load.weight),
3012 cfs_rq->curr != NULL, cfs_rq);
3016 * Signed add and clamp on underflow.
3018 * Explicitly do a load-store to ensure the intermediate value never hits
3019 * memory. This allows lockless observations without ever seeing the negative
3022 #define add_positive(_ptr, _val) do { \
3023 typeof(_ptr) ptr = (_ptr); \
3024 typeof(_val) val = (_val); \
3025 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3029 if (val < 0 && res > var) \
3032 WRITE_ONCE(*ptr, res); \
3035 #ifdef CONFIG_FAIR_GROUP_SCHED
3037 * update_tg_load_avg - update the tg's load avg
3038 * @cfs_rq: the cfs_rq whose avg changed
3039 * @force: update regardless of how small the difference
3041 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3042 * However, because tg->load_avg is a global value there are performance
3045 * In order to avoid having to look at the other cfs_rq's, we use a
3046 * differential update where we store the last value we propagated. This in
3047 * turn allows skipping updates if the differential is 'small'.
3049 * Updating tg's load_avg is necessary before update_cfs_share().
3051 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3053 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3056 * No need to update load_avg for root_task_group as it is not used.
3058 if (cfs_rq->tg == &root_task_group)
3061 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3062 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3063 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3068 * Called within set_task_rq() right before setting a task's cpu. The
3069 * caller only guarantees p->pi_lock is held; no other assumptions,
3070 * including the state of rq->lock, should be made.
3072 void set_task_rq_fair(struct sched_entity *se,
3073 struct cfs_rq *prev, struct cfs_rq *next)
3075 u64 p_last_update_time;
3076 u64 n_last_update_time;
3078 if (!sched_feat(ATTACH_AGE_LOAD))
3082 * We are supposed to update the task to "current" time, then its up to
3083 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3084 * getting what current time is, so simply throw away the out-of-date
3085 * time. This will result in the wakee task is less decayed, but giving
3086 * the wakee more load sounds not bad.
3088 if (!(se->avg.last_update_time && prev))
3091 #ifndef CONFIG_64BIT
3093 u64 p_last_update_time_copy;
3094 u64 n_last_update_time_copy;
3097 p_last_update_time_copy = prev->load_last_update_time_copy;
3098 n_last_update_time_copy = next->load_last_update_time_copy;
3102 p_last_update_time = prev->avg.last_update_time;
3103 n_last_update_time = next->avg.last_update_time;
3105 } while (p_last_update_time != p_last_update_time_copy ||
3106 n_last_update_time != n_last_update_time_copy);
3109 p_last_update_time = prev->avg.last_update_time;
3110 n_last_update_time = next->avg.last_update_time;
3112 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3113 se->avg.last_update_time = n_last_update_time;
3116 /* Take into account change of utilization of a child task group */
3118 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3120 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3121 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3123 /* Nothing to update */
3127 /* Set new sched_entity's utilization */
3128 se->avg.util_avg = gcfs_rq->avg.util_avg;
3129 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3131 /* Update parent cfs_rq utilization */
3132 add_positive(&cfs_rq->avg.util_avg, delta);
3133 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3136 /* Take into account change of load of a child task group */
3138 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3140 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3141 long delta, load = gcfs_rq->avg.load_avg;
3144 * If the load of group cfs_rq is null, the load of the
3145 * sched_entity will also be null so we can skip the formula
3150 /* Get tg's load and ensure tg_load > 0 */
3151 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3153 /* Ensure tg_load >= load and updated with current load*/
3154 tg_load -= gcfs_rq->tg_load_avg_contrib;
3158 * We need to compute a correction term in the case that the
3159 * task group is consuming more CPU than a task of equal
3160 * weight. A task with a weight equals to tg->shares will have
3161 * a load less or equal to scale_load_down(tg->shares).
3162 * Similarly, the sched_entities that represent the task group
3163 * at parent level, can't have a load higher than
3164 * scale_load_down(tg->shares). And the Sum of sched_entities'
3165 * load must be <= scale_load_down(tg->shares).
3167 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3168 /* scale gcfs_rq's load into tg's shares*/
3169 load *= scale_load_down(gcfs_rq->tg->shares);
3174 delta = load - se->avg.load_avg;
3176 /* Nothing to update */
3180 /* Set new sched_entity's load */
3181 se->avg.load_avg = load;
3182 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3184 /* Update parent cfs_rq load */
3185 add_positive(&cfs_rq->avg.load_avg, delta);
3186 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3189 * If the sched_entity is already enqueued, we also have to update the
3190 * runnable load avg.
3193 /* Update parent cfs_rq runnable_load_avg */
3194 add_positive(&cfs_rq->runnable_load_avg, delta);
3195 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3199 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3201 cfs_rq->propagate_avg = 1;
3204 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3206 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3208 if (!cfs_rq->propagate_avg)
3211 cfs_rq->propagate_avg = 0;
3215 /* Update task and its cfs_rq load average */
3216 static inline int propagate_entity_load_avg(struct sched_entity *se)
3218 struct cfs_rq *cfs_rq;
3220 if (entity_is_task(se))
3223 if (!test_and_clear_tg_cfs_propagate(se))
3226 cfs_rq = cfs_rq_of(se);
3228 set_tg_cfs_propagate(cfs_rq);
3230 update_tg_cfs_util(cfs_rq, se);
3231 update_tg_cfs_load(cfs_rq, se);
3237 * Check if we need to update the load and the utilization of a blocked
3240 static inline bool skip_blocked_update(struct sched_entity *se)
3242 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3245 * If sched_entity still have not zero load or utilization, we have to
3248 if (se->avg.load_avg || se->avg.util_avg)
3252 * If there is a pending propagation, we have to update the load and
3253 * the utilization of the sched_entity:
3255 if (gcfs_rq->propagate_avg)
3259 * Otherwise, the load and the utilization of the sched_entity is
3260 * already zero and there is no pending propagation, so it will be a
3261 * waste of time to try to decay it:
3266 #else /* CONFIG_FAIR_GROUP_SCHED */
3268 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3270 static inline int propagate_entity_load_avg(struct sched_entity *se)
3275 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3277 #endif /* CONFIG_FAIR_GROUP_SCHED */
3279 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3281 struct rq *rq = rq_of(cfs_rq);
3283 if (&rq->cfs == cfs_rq) {
3285 * There are a few boundary cases this might miss but it should
3286 * get called often enough that that should (hopefully) not be
3287 * a real problem -- added to that it only calls on the local
3288 * CPU, so if we enqueue remotely we'll miss an update, but
3289 * the next tick/schedule should update.
3291 * It will not get called when we go idle, because the idle
3292 * thread is a different class (!fair), nor will the utilization
3293 * number include things like RT tasks.
3295 * As is, the util number is not freq-invariant (we'd have to
3296 * implement arch_scale_freq_capacity() for that).
3300 cpufreq_update_util(rq, 0);
3305 * Unsigned subtract and clamp on underflow.
3307 * Explicitly do a load-store to ensure the intermediate value never hits
3308 * memory. This allows lockless observations without ever seeing the negative
3311 #define sub_positive(_ptr, _val) do { \
3312 typeof(_ptr) ptr = (_ptr); \
3313 typeof(*ptr) val = (_val); \
3314 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3318 WRITE_ONCE(*ptr, res); \
3322 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3323 * @now: current time, as per cfs_rq_clock_task()
3324 * @cfs_rq: cfs_rq to update
3325 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3327 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3328 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3329 * post_init_entity_util_avg().
3331 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3333 * Returns true if the load decayed or we removed load.
3335 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3336 * call update_tg_load_avg() when this function returns true.
3339 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3341 struct sched_avg *sa = &cfs_rq->avg;
3342 int decayed, removed_load = 0, removed_util = 0;
3344 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3345 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3346 sub_positive(&sa->load_avg, r);
3347 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3349 set_tg_cfs_propagate(cfs_rq);
3352 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3353 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3354 sub_positive(&sa->util_avg, r);
3355 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3357 set_tg_cfs_propagate(cfs_rq);
3360 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3362 #ifndef CONFIG_64BIT
3364 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3367 if (update_freq && (decayed || removed_util))
3368 cfs_rq_util_change(cfs_rq);
3370 return decayed || removed_load;
3374 * Optional action to be done while updating the load average
3376 #define UPDATE_TG 0x1
3377 #define SKIP_AGE_LOAD 0x2
3379 /* Update task and its cfs_rq load average */
3380 static inline void update_load_avg(struct sched_entity *se, int flags)
3382 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3383 u64 now = cfs_rq_clock_task(cfs_rq);
3384 struct rq *rq = rq_of(cfs_rq);
3385 int cpu = cpu_of(rq);
3389 * Track task load average for carrying it to new CPU after migrated, and
3390 * track group sched_entity load average for task_h_load calc in migration
3392 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3393 __update_load_avg_se(now, cpu, cfs_rq, se);
3395 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3396 decayed |= propagate_entity_load_avg(se);
3398 if (decayed && (flags & UPDATE_TG))
3399 update_tg_load_avg(cfs_rq, 0);
3403 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3404 * @cfs_rq: cfs_rq to attach to
3405 * @se: sched_entity to attach
3407 * Must call update_cfs_rq_load_avg() before this, since we rely on
3408 * cfs_rq->avg.last_update_time being current.
3410 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3412 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3413 cfs_rq->avg.load_avg += se->avg.load_avg;
3414 cfs_rq->avg.load_sum += se->avg.load_sum;
3415 cfs_rq->avg.util_avg += se->avg.util_avg;
3416 cfs_rq->avg.util_sum += se->avg.util_sum;
3417 set_tg_cfs_propagate(cfs_rq);
3419 cfs_rq_util_change(cfs_rq);
3423 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3424 * @cfs_rq: cfs_rq to detach from
3425 * @se: sched_entity to detach
3427 * Must call update_cfs_rq_load_avg() before this, since we rely on
3428 * cfs_rq->avg.last_update_time being current.
3430 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3433 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3434 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3435 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3436 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3437 set_tg_cfs_propagate(cfs_rq);
3439 cfs_rq_util_change(cfs_rq);
3442 /* Add the load generated by se into cfs_rq's load average */
3444 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3446 struct sched_avg *sa = &se->avg;
3448 cfs_rq->runnable_load_avg += sa->load_avg;
3449 cfs_rq->runnable_load_sum += sa->load_sum;
3451 if (!sa->last_update_time) {
3452 attach_entity_load_avg(cfs_rq, se);
3453 update_tg_load_avg(cfs_rq, 0);
3457 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3459 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3461 cfs_rq->runnable_load_avg =
3462 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3463 cfs_rq->runnable_load_sum =
3464 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3467 #ifndef CONFIG_64BIT
3468 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3470 u64 last_update_time_copy;
3471 u64 last_update_time;
3474 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3476 last_update_time = cfs_rq->avg.last_update_time;
3477 } while (last_update_time != last_update_time_copy);
3479 return last_update_time;
3482 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3484 return cfs_rq->avg.last_update_time;
3489 * Synchronize entity load avg of dequeued entity without locking
3492 void sync_entity_load_avg(struct sched_entity *se)
3494 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3495 u64 last_update_time;
3497 last_update_time = cfs_rq_last_update_time(cfs_rq);
3498 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3502 * Task first catches up with cfs_rq, and then subtract
3503 * itself from the cfs_rq (task must be off the queue now).
3505 void remove_entity_load_avg(struct sched_entity *se)
3507 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3510 * tasks cannot exit without having gone through wake_up_new_task() ->
3511 * post_init_entity_util_avg() which will have added things to the
3512 * cfs_rq, so we can remove unconditionally.
3514 * Similarly for groups, they will have passed through
3515 * post_init_entity_util_avg() before unregister_sched_fair_group()
3519 sync_entity_load_avg(se);
3520 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3521 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3524 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3526 return cfs_rq->runnable_load_avg;
3529 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3531 return cfs_rq->avg.load_avg;
3534 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3536 #else /* CONFIG_SMP */
3539 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3544 #define UPDATE_TG 0x0
3545 #define SKIP_AGE_LOAD 0x0
3547 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3549 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3553 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3555 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3556 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3559 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3561 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3563 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3568 #endif /* CONFIG_SMP */
3570 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3572 #ifdef CONFIG_SCHED_DEBUG
3573 s64 d = se->vruntime - cfs_rq->min_vruntime;
3578 if (d > 3*sysctl_sched_latency)
3579 schedstat_inc(cfs_rq->nr_spread_over);
3584 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3586 u64 vruntime = cfs_rq->min_vruntime;
3589 * The 'current' period is already promised to the current tasks,
3590 * however the extra weight of the new task will slow them down a
3591 * little, place the new task so that it fits in the slot that
3592 * stays open at the end.
3594 if (initial && sched_feat(START_DEBIT))
3595 vruntime += sched_vslice(cfs_rq, se);
3597 /* sleeps up to a single latency don't count. */
3599 unsigned long thresh = sysctl_sched_latency;
3602 * Halve their sleep time's effect, to allow
3603 * for a gentler effect of sleepers:
3605 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3611 /* ensure we never gain time by being placed backwards. */
3612 se->vruntime = max_vruntime(se->vruntime, vruntime);
3615 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3617 static inline void check_schedstat_required(void)
3619 #ifdef CONFIG_SCHEDSTATS
3620 if (schedstat_enabled())
3623 /* Force schedstat enabled if a dependent tracepoint is active */
3624 if (trace_sched_stat_wait_enabled() ||
3625 trace_sched_stat_sleep_enabled() ||
3626 trace_sched_stat_iowait_enabled() ||
3627 trace_sched_stat_blocked_enabled() ||
3628 trace_sched_stat_runtime_enabled()) {
3629 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3630 "stat_blocked and stat_runtime require the "
3631 "kernel parameter schedstats=enable or "
3632 "kernel.sched_schedstats=1\n");
3643 * update_min_vruntime()
3644 * vruntime -= min_vruntime
3648 * update_min_vruntime()
3649 * vruntime += min_vruntime
3651 * this way the vruntime transition between RQs is done when both
3652 * min_vruntime are up-to-date.
3656 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3657 * vruntime -= min_vruntime
3661 * update_min_vruntime()
3662 * vruntime += min_vruntime
3664 * this way we don't have the most up-to-date min_vruntime on the originating
3665 * CPU and an up-to-date min_vruntime on the destination CPU.
3669 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3671 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3672 bool curr = cfs_rq->curr == se;
3675 * If we're the current task, we must renormalise before calling
3679 se->vruntime += cfs_rq->min_vruntime;
3681 update_curr(cfs_rq);
3684 * Otherwise, renormalise after, such that we're placed at the current
3685 * moment in time, instead of some random moment in the past. Being
3686 * placed in the past could significantly boost this task to the
3687 * fairness detriment of existing tasks.
3689 if (renorm && !curr)
3690 se->vruntime += cfs_rq->min_vruntime;
3693 * When enqueuing a sched_entity, we must:
3694 * - Update loads to have both entity and cfs_rq synced with now.
3695 * - Add its load to cfs_rq->runnable_avg
3696 * - For group_entity, update its weight to reflect the new share of
3698 * - Add its new weight to cfs_rq->load.weight
3700 update_load_avg(se, UPDATE_TG);
3701 enqueue_entity_load_avg(cfs_rq, se);
3702 update_cfs_shares(se);
3703 account_entity_enqueue(cfs_rq, se);
3705 if (flags & ENQUEUE_WAKEUP)
3706 place_entity(cfs_rq, se, 0);
3708 check_schedstat_required();
3709 update_stats_enqueue(cfs_rq, se, flags);
3710 check_spread(cfs_rq, se);
3712 __enqueue_entity(cfs_rq, se);
3715 if (cfs_rq->nr_running == 1) {
3716 list_add_leaf_cfs_rq(cfs_rq);
3717 check_enqueue_throttle(cfs_rq);
3721 static void __clear_buddies_last(struct sched_entity *se)
3723 for_each_sched_entity(se) {
3724 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3725 if (cfs_rq->last != se)
3728 cfs_rq->last = NULL;
3732 static void __clear_buddies_next(struct sched_entity *se)
3734 for_each_sched_entity(se) {
3735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3736 if (cfs_rq->next != se)
3739 cfs_rq->next = NULL;
3743 static void __clear_buddies_skip(struct sched_entity *se)
3745 for_each_sched_entity(se) {
3746 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3747 if (cfs_rq->skip != se)
3750 cfs_rq->skip = NULL;
3754 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3756 if (cfs_rq->last == se)
3757 __clear_buddies_last(se);
3759 if (cfs_rq->next == se)
3760 __clear_buddies_next(se);
3762 if (cfs_rq->skip == se)
3763 __clear_buddies_skip(se);
3766 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3769 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3772 * Update run-time statistics of the 'current'.
3774 update_curr(cfs_rq);
3777 * When dequeuing a sched_entity, we must:
3778 * - Update loads to have both entity and cfs_rq synced with now.
3779 * - Substract its load from the cfs_rq->runnable_avg.
3780 * - Substract its previous weight from cfs_rq->load.weight.
3781 * - For group entity, update its weight to reflect the new share
3782 * of its group cfs_rq.
3784 update_load_avg(se, UPDATE_TG);
3785 dequeue_entity_load_avg(cfs_rq, se);
3787 update_stats_dequeue(cfs_rq, se, flags);
3789 clear_buddies(cfs_rq, se);
3791 if (se != cfs_rq->curr)
3792 __dequeue_entity(cfs_rq, se);
3794 account_entity_dequeue(cfs_rq, se);
3797 * Normalize after update_curr(); which will also have moved
3798 * min_vruntime if @se is the one holding it back. But before doing
3799 * update_min_vruntime() again, which will discount @se's position and
3800 * can move min_vruntime forward still more.
3802 if (!(flags & DEQUEUE_SLEEP))
3803 se->vruntime -= cfs_rq->min_vruntime;
3805 /* return excess runtime on last dequeue */
3806 return_cfs_rq_runtime(cfs_rq);
3808 update_cfs_shares(se);
3811 * Now advance min_vruntime if @se was the entity holding it back,
3812 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3813 * put back on, and if we advance min_vruntime, we'll be placed back
3814 * further than we started -- ie. we'll be penalized.
3816 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3817 update_min_vruntime(cfs_rq);
3821 * Preempt the current task with a newly woken task if needed:
3824 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3826 unsigned long ideal_runtime, delta_exec;
3827 struct sched_entity *se;
3830 ideal_runtime = sched_slice(cfs_rq, curr);
3831 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3832 if (delta_exec > ideal_runtime) {
3833 resched_curr(rq_of(cfs_rq));
3835 * The current task ran long enough, ensure it doesn't get
3836 * re-elected due to buddy favours.
3838 clear_buddies(cfs_rq, curr);
3843 * Ensure that a task that missed wakeup preemption by a
3844 * narrow margin doesn't have to wait for a full slice.
3845 * This also mitigates buddy induced latencies under load.
3847 if (delta_exec < sysctl_sched_min_granularity)
3850 se = __pick_first_entity(cfs_rq);
3851 delta = curr->vruntime - se->vruntime;
3856 if (delta > ideal_runtime)
3857 resched_curr(rq_of(cfs_rq));
3861 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3863 /* 'current' is not kept within the tree. */
3866 * Any task has to be enqueued before it get to execute on
3867 * a CPU. So account for the time it spent waiting on the
3870 update_stats_wait_end(cfs_rq, se);
3871 __dequeue_entity(cfs_rq, se);
3872 update_load_avg(se, UPDATE_TG);
3875 update_stats_curr_start(cfs_rq, se);
3879 * Track our maximum slice length, if the CPU's load is at
3880 * least twice that of our own weight (i.e. dont track it
3881 * when there are only lesser-weight tasks around):
3883 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3884 schedstat_set(se->statistics.slice_max,
3885 max((u64)schedstat_val(se->statistics.slice_max),
3886 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3889 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3893 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3896 * Pick the next process, keeping these things in mind, in this order:
3897 * 1) keep things fair between processes/task groups
3898 * 2) pick the "next" process, since someone really wants that to run
3899 * 3) pick the "last" process, for cache locality
3900 * 4) do not run the "skip" process, if something else is available
3902 static struct sched_entity *
3903 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3905 struct sched_entity *left = __pick_first_entity(cfs_rq);
3906 struct sched_entity *se;
3909 * If curr is set we have to see if its left of the leftmost entity
3910 * still in the tree, provided there was anything in the tree at all.
3912 if (!left || (curr && entity_before(curr, left)))
3915 se = left; /* ideally we run the leftmost entity */
3918 * Avoid running the skip buddy, if running something else can
3919 * be done without getting too unfair.
3921 if (cfs_rq->skip == se) {
3922 struct sched_entity *second;
3925 second = __pick_first_entity(cfs_rq);
3927 second = __pick_next_entity(se);
3928 if (!second || (curr && entity_before(curr, second)))
3932 if (second && wakeup_preempt_entity(second, left) < 1)
3937 * Prefer last buddy, try to return the CPU to a preempted task.
3939 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3943 * Someone really wants this to run. If it's not unfair, run it.
3945 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3948 clear_buddies(cfs_rq, se);
3953 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3955 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3958 * If still on the runqueue then deactivate_task()
3959 * was not called and update_curr() has to be done:
3962 update_curr(cfs_rq);
3964 /* throttle cfs_rqs exceeding runtime */
3965 check_cfs_rq_runtime(cfs_rq);
3967 check_spread(cfs_rq, prev);
3970 update_stats_wait_start(cfs_rq, prev);
3971 /* Put 'current' back into the tree. */
3972 __enqueue_entity(cfs_rq, prev);
3973 /* in !on_rq case, update occurred at dequeue */
3974 update_load_avg(prev, 0);
3976 cfs_rq->curr = NULL;
3980 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3983 * Update run-time statistics of the 'current'.
3985 update_curr(cfs_rq);
3988 * Ensure that runnable average is periodically updated.
3990 update_load_avg(curr, UPDATE_TG);
3991 update_cfs_shares(curr);
3993 #ifdef CONFIG_SCHED_HRTICK
3995 * queued ticks are scheduled to match the slice, so don't bother
3996 * validating it and just reschedule.
3999 resched_curr(rq_of(cfs_rq));
4003 * don't let the period tick interfere with the hrtick preemption
4005 if (!sched_feat(DOUBLE_TICK) &&
4006 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4010 if (cfs_rq->nr_running > 1)
4011 check_preempt_tick(cfs_rq, curr);
4015 /**************************************************
4016 * CFS bandwidth control machinery
4019 #ifdef CONFIG_CFS_BANDWIDTH
4021 #ifdef HAVE_JUMP_LABEL
4022 static struct static_key __cfs_bandwidth_used;
4024 static inline bool cfs_bandwidth_used(void)
4026 return static_key_false(&__cfs_bandwidth_used);
4029 void cfs_bandwidth_usage_inc(void)
4031 static_key_slow_inc(&__cfs_bandwidth_used);
4034 void cfs_bandwidth_usage_dec(void)
4036 static_key_slow_dec(&__cfs_bandwidth_used);
4038 #else /* HAVE_JUMP_LABEL */
4039 static bool cfs_bandwidth_used(void)
4044 void cfs_bandwidth_usage_inc(void) {}
4045 void cfs_bandwidth_usage_dec(void) {}
4046 #endif /* HAVE_JUMP_LABEL */
4049 * default period for cfs group bandwidth.
4050 * default: 0.1s, units: nanoseconds
4052 static inline u64 default_cfs_period(void)
4054 return 100000000ULL;
4057 static inline u64 sched_cfs_bandwidth_slice(void)
4059 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4063 * Replenish runtime according to assigned quota and update expiration time.
4064 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4065 * additional synchronization around rq->lock.
4067 * requires cfs_b->lock
4069 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4073 if (cfs_b->quota == RUNTIME_INF)
4076 now = sched_clock_cpu(smp_processor_id());
4077 cfs_b->runtime = cfs_b->quota;
4078 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4081 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4083 return &tg->cfs_bandwidth;
4086 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4087 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4089 if (unlikely(cfs_rq->throttle_count))
4090 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4092 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4095 /* returns 0 on failure to allocate runtime */
4096 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4098 struct task_group *tg = cfs_rq->tg;
4099 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4100 u64 amount = 0, min_amount, expires;
4102 /* note: this is a positive sum as runtime_remaining <= 0 */
4103 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4105 raw_spin_lock(&cfs_b->lock);
4106 if (cfs_b->quota == RUNTIME_INF)
4107 amount = min_amount;
4109 start_cfs_bandwidth(cfs_b);
4111 if (cfs_b->runtime > 0) {
4112 amount = min(cfs_b->runtime, min_amount);
4113 cfs_b->runtime -= amount;
4117 expires = cfs_b->runtime_expires;
4118 raw_spin_unlock(&cfs_b->lock);
4120 cfs_rq->runtime_remaining += amount;
4122 * we may have advanced our local expiration to account for allowed
4123 * spread between our sched_clock and the one on which runtime was
4126 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4127 cfs_rq->runtime_expires = expires;
4129 return cfs_rq->runtime_remaining > 0;
4133 * Note: This depends on the synchronization provided by sched_clock and the
4134 * fact that rq->clock snapshots this value.
4136 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4138 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4140 /* if the deadline is ahead of our clock, nothing to do */
4141 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4144 if (cfs_rq->runtime_remaining < 0)
4148 * If the local deadline has passed we have to consider the
4149 * possibility that our sched_clock is 'fast' and the global deadline
4150 * has not truly expired.
4152 * Fortunately we can check determine whether this the case by checking
4153 * whether the global deadline has advanced. It is valid to compare
4154 * cfs_b->runtime_expires without any locks since we only care about
4155 * exact equality, so a partial write will still work.
4158 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4159 /* extend local deadline, drift is bounded above by 2 ticks */
4160 cfs_rq->runtime_expires += TICK_NSEC;
4162 /* global deadline is ahead, expiration has passed */
4163 cfs_rq->runtime_remaining = 0;
4167 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4169 /* dock delta_exec before expiring quota (as it could span periods) */
4170 cfs_rq->runtime_remaining -= delta_exec;
4171 expire_cfs_rq_runtime(cfs_rq);
4173 if (likely(cfs_rq->runtime_remaining > 0))
4177 * if we're unable to extend our runtime we resched so that the active
4178 * hierarchy can be throttled
4180 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4181 resched_curr(rq_of(cfs_rq));
4184 static __always_inline
4185 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4187 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4190 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4193 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4195 return cfs_bandwidth_used() && cfs_rq->throttled;
4198 /* check whether cfs_rq, or any parent, is throttled */
4199 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4201 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4205 * Ensure that neither of the group entities corresponding to src_cpu or
4206 * dest_cpu are members of a throttled hierarchy when performing group
4207 * load-balance operations.
4209 static inline int throttled_lb_pair(struct task_group *tg,
4210 int src_cpu, int dest_cpu)
4212 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4214 src_cfs_rq = tg->cfs_rq[src_cpu];
4215 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4217 return throttled_hierarchy(src_cfs_rq) ||
4218 throttled_hierarchy(dest_cfs_rq);
4221 /* updated child weight may affect parent so we have to do this bottom up */
4222 static int tg_unthrottle_up(struct task_group *tg, void *data)
4224 struct rq *rq = data;
4225 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4227 cfs_rq->throttle_count--;
4228 if (!cfs_rq->throttle_count) {
4229 /* adjust cfs_rq_clock_task() */
4230 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4231 cfs_rq->throttled_clock_task;
4237 static int tg_throttle_down(struct task_group *tg, void *data)
4239 struct rq *rq = data;
4240 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4242 /* group is entering throttled state, stop time */
4243 if (!cfs_rq->throttle_count)
4244 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4245 cfs_rq->throttle_count++;
4250 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4252 struct rq *rq = rq_of(cfs_rq);
4253 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4254 struct sched_entity *se;
4255 long task_delta, dequeue = 1;
4258 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4260 /* freeze hierarchy runnable averages while throttled */
4262 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4265 task_delta = cfs_rq->h_nr_running;
4266 for_each_sched_entity(se) {
4267 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4268 /* throttled entity or throttle-on-deactivate */
4273 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4274 qcfs_rq->h_nr_running -= task_delta;
4276 if (qcfs_rq->load.weight)
4281 sub_nr_running(rq, task_delta);
4283 cfs_rq->throttled = 1;
4284 cfs_rq->throttled_clock = rq_clock(rq);
4285 raw_spin_lock(&cfs_b->lock);
4286 empty = list_empty(&cfs_b->throttled_cfs_rq);
4289 * Add to the _head_ of the list, so that an already-started
4290 * distribute_cfs_runtime will not see us
4292 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4295 * If we're the first throttled task, make sure the bandwidth
4299 start_cfs_bandwidth(cfs_b);
4301 raw_spin_unlock(&cfs_b->lock);
4304 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4306 struct rq *rq = rq_of(cfs_rq);
4307 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4308 struct sched_entity *se;
4312 se = cfs_rq->tg->se[cpu_of(rq)];
4314 cfs_rq->throttled = 0;
4316 update_rq_clock(rq);
4318 raw_spin_lock(&cfs_b->lock);
4319 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4320 list_del_rcu(&cfs_rq->throttled_list);
4321 raw_spin_unlock(&cfs_b->lock);
4323 /* update hierarchical throttle state */
4324 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4326 if (!cfs_rq->load.weight)
4329 task_delta = cfs_rq->h_nr_running;
4330 for_each_sched_entity(se) {
4334 cfs_rq = cfs_rq_of(se);
4336 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4337 cfs_rq->h_nr_running += task_delta;
4339 if (cfs_rq_throttled(cfs_rq))
4344 add_nr_running(rq, task_delta);
4346 /* determine whether we need to wake up potentially idle cpu */
4347 if (rq->curr == rq->idle && rq->cfs.nr_running)
4351 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4352 u64 remaining, u64 expires)
4354 struct cfs_rq *cfs_rq;
4356 u64 starting_runtime = remaining;
4359 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4361 struct rq *rq = rq_of(cfs_rq);
4365 if (!cfs_rq_throttled(cfs_rq))
4368 runtime = -cfs_rq->runtime_remaining + 1;
4369 if (runtime > remaining)
4370 runtime = remaining;
4371 remaining -= runtime;
4373 cfs_rq->runtime_remaining += runtime;
4374 cfs_rq->runtime_expires = expires;
4376 /* we check whether we're throttled above */
4377 if (cfs_rq->runtime_remaining > 0)
4378 unthrottle_cfs_rq(cfs_rq);
4388 return starting_runtime - remaining;
4392 * Responsible for refilling a task_group's bandwidth and unthrottling its
4393 * cfs_rqs as appropriate. If there has been no activity within the last
4394 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4395 * used to track this state.
4397 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4399 u64 runtime, runtime_expires;
4402 /* no need to continue the timer with no bandwidth constraint */
4403 if (cfs_b->quota == RUNTIME_INF)
4404 goto out_deactivate;
4406 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4407 cfs_b->nr_periods += overrun;
4410 * idle depends on !throttled (for the case of a large deficit), and if
4411 * we're going inactive then everything else can be deferred
4413 if (cfs_b->idle && !throttled)
4414 goto out_deactivate;
4416 __refill_cfs_bandwidth_runtime(cfs_b);
4419 /* mark as potentially idle for the upcoming period */
4424 /* account preceding periods in which throttling occurred */
4425 cfs_b->nr_throttled += overrun;
4427 runtime_expires = cfs_b->runtime_expires;
4430 * This check is repeated as we are holding onto the new bandwidth while
4431 * we unthrottle. This can potentially race with an unthrottled group
4432 * trying to acquire new bandwidth from the global pool. This can result
4433 * in us over-using our runtime if it is all used during this loop, but
4434 * only by limited amounts in that extreme case.
4436 while (throttled && cfs_b->runtime > 0) {
4437 runtime = cfs_b->runtime;
4438 raw_spin_unlock(&cfs_b->lock);
4439 /* we can't nest cfs_b->lock while distributing bandwidth */
4440 runtime = distribute_cfs_runtime(cfs_b, runtime,
4442 raw_spin_lock(&cfs_b->lock);
4444 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4446 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4450 * While we are ensured activity in the period following an
4451 * unthrottle, this also covers the case in which the new bandwidth is
4452 * insufficient to cover the existing bandwidth deficit. (Forcing the
4453 * timer to remain active while there are any throttled entities.)
4463 /* a cfs_rq won't donate quota below this amount */
4464 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4465 /* minimum remaining period time to redistribute slack quota */
4466 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4467 /* how long we wait to gather additional slack before distributing */
4468 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4471 * Are we near the end of the current quota period?
4473 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4474 * hrtimer base being cleared by hrtimer_start. In the case of
4475 * migrate_hrtimers, base is never cleared, so we are fine.
4477 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4479 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4482 /* if the call-back is running a quota refresh is already occurring */
4483 if (hrtimer_callback_running(refresh_timer))
4486 /* is a quota refresh about to occur? */
4487 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4488 if (remaining < min_expire)
4494 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4496 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4498 /* if there's a quota refresh soon don't bother with slack */
4499 if (runtime_refresh_within(cfs_b, min_left))
4502 hrtimer_start(&cfs_b->slack_timer,
4503 ns_to_ktime(cfs_bandwidth_slack_period),
4507 /* we know any runtime found here is valid as update_curr() precedes return */
4508 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4510 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4511 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4513 if (slack_runtime <= 0)
4516 raw_spin_lock(&cfs_b->lock);
4517 if (cfs_b->quota != RUNTIME_INF &&
4518 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4519 cfs_b->runtime += slack_runtime;
4521 /* we are under rq->lock, defer unthrottling using a timer */
4522 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4523 !list_empty(&cfs_b->throttled_cfs_rq))
4524 start_cfs_slack_bandwidth(cfs_b);
4526 raw_spin_unlock(&cfs_b->lock);
4528 /* even if it's not valid for return we don't want to try again */
4529 cfs_rq->runtime_remaining -= slack_runtime;
4532 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4534 if (!cfs_bandwidth_used())
4537 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4540 __return_cfs_rq_runtime(cfs_rq);
4544 * This is done with a timer (instead of inline with bandwidth return) since
4545 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4547 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4549 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4552 /* confirm we're still not at a refresh boundary */
4553 raw_spin_lock(&cfs_b->lock);
4554 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4555 raw_spin_unlock(&cfs_b->lock);
4559 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4560 runtime = cfs_b->runtime;
4562 expires = cfs_b->runtime_expires;
4563 raw_spin_unlock(&cfs_b->lock);
4568 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4570 raw_spin_lock(&cfs_b->lock);
4571 if (expires == cfs_b->runtime_expires)
4572 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4573 raw_spin_unlock(&cfs_b->lock);
4577 * When a group wakes up we want to make sure that its quota is not already
4578 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4579 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4581 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4583 if (!cfs_bandwidth_used())
4586 /* an active group must be handled by the update_curr()->put() path */
4587 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4590 /* ensure the group is not already throttled */
4591 if (cfs_rq_throttled(cfs_rq))
4594 /* update runtime allocation */
4595 account_cfs_rq_runtime(cfs_rq, 0);
4596 if (cfs_rq->runtime_remaining <= 0)
4597 throttle_cfs_rq(cfs_rq);
4600 static void sync_throttle(struct task_group *tg, int cpu)
4602 struct cfs_rq *pcfs_rq, *cfs_rq;
4604 if (!cfs_bandwidth_used())
4610 cfs_rq = tg->cfs_rq[cpu];
4611 pcfs_rq = tg->parent->cfs_rq[cpu];
4613 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4614 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4617 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4618 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4620 if (!cfs_bandwidth_used())
4623 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4627 * it's possible for a throttled entity to be forced into a running
4628 * state (e.g. set_curr_task), in this case we're finished.
4630 if (cfs_rq_throttled(cfs_rq))
4633 throttle_cfs_rq(cfs_rq);
4637 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4639 struct cfs_bandwidth *cfs_b =
4640 container_of(timer, struct cfs_bandwidth, slack_timer);
4642 do_sched_cfs_slack_timer(cfs_b);
4644 return HRTIMER_NORESTART;
4647 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4649 struct cfs_bandwidth *cfs_b =
4650 container_of(timer, struct cfs_bandwidth, period_timer);
4654 raw_spin_lock(&cfs_b->lock);
4656 overrun = hrtimer_forward_now(timer, cfs_b->period);
4660 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4663 cfs_b->period_active = 0;
4664 raw_spin_unlock(&cfs_b->lock);
4666 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4669 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4671 raw_spin_lock_init(&cfs_b->lock);
4673 cfs_b->quota = RUNTIME_INF;
4674 cfs_b->period = ns_to_ktime(default_cfs_period());
4676 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4677 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4678 cfs_b->period_timer.function = sched_cfs_period_timer;
4679 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4680 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4683 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4685 cfs_rq->runtime_enabled = 0;
4686 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4689 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4691 lockdep_assert_held(&cfs_b->lock);
4693 if (!cfs_b->period_active) {
4694 cfs_b->period_active = 1;
4695 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4696 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4700 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4702 /* init_cfs_bandwidth() was not called */
4703 if (!cfs_b->throttled_cfs_rq.next)
4706 hrtimer_cancel(&cfs_b->period_timer);
4707 hrtimer_cancel(&cfs_b->slack_timer);
4711 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4713 * The race is harmless, since modifying bandwidth settings of unhooked group
4714 * bits doesn't do much.
4717 /* cpu online calback */
4718 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4720 struct task_group *tg;
4722 lockdep_assert_held(&rq->lock);
4725 list_for_each_entry_rcu(tg, &task_groups, list) {
4726 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4727 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4729 raw_spin_lock(&cfs_b->lock);
4730 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4731 raw_spin_unlock(&cfs_b->lock);
4736 /* cpu offline callback */
4737 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4739 struct task_group *tg;
4741 lockdep_assert_held(&rq->lock);
4744 list_for_each_entry_rcu(tg, &task_groups, list) {
4745 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4747 if (!cfs_rq->runtime_enabled)
4751 * clock_task is not advancing so we just need to make sure
4752 * there's some valid quota amount
4754 cfs_rq->runtime_remaining = 1;
4756 * Offline rq is schedulable till cpu is completely disabled
4757 * in take_cpu_down(), so we prevent new cfs throttling here.
4759 cfs_rq->runtime_enabled = 0;
4761 if (cfs_rq_throttled(cfs_rq))
4762 unthrottle_cfs_rq(cfs_rq);
4767 #else /* CONFIG_CFS_BANDWIDTH */
4768 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4770 return rq_clock_task(rq_of(cfs_rq));
4773 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4774 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4775 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4776 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4777 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4779 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4784 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4789 static inline int throttled_lb_pair(struct task_group *tg,
4790 int src_cpu, int dest_cpu)
4795 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4797 #ifdef CONFIG_FAIR_GROUP_SCHED
4798 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4801 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4805 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4806 static inline void update_runtime_enabled(struct rq *rq) {}
4807 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4809 #endif /* CONFIG_CFS_BANDWIDTH */
4811 /**************************************************
4812 * CFS operations on tasks:
4815 #ifdef CONFIG_SCHED_HRTICK
4816 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4818 struct sched_entity *se = &p->se;
4819 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4821 SCHED_WARN_ON(task_rq(p) != rq);
4823 if (rq->cfs.h_nr_running > 1) {
4824 u64 slice = sched_slice(cfs_rq, se);
4825 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4826 s64 delta = slice - ran;
4833 hrtick_start(rq, delta);
4838 * called from enqueue/dequeue and updates the hrtick when the
4839 * current task is from our class and nr_running is low enough
4842 static void hrtick_update(struct rq *rq)
4844 struct task_struct *curr = rq->curr;
4846 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4849 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4850 hrtick_start_fair(rq, curr);
4852 #else /* !CONFIG_SCHED_HRTICK */
4854 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4858 static inline void hrtick_update(struct rq *rq)
4864 * The enqueue_task method is called before nr_running is
4865 * increased. Here we update the fair scheduling stats and
4866 * then put the task into the rbtree:
4869 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4871 struct cfs_rq *cfs_rq;
4872 struct sched_entity *se = &p->se;
4875 * If in_iowait is set, the code below may not trigger any cpufreq
4876 * utilization updates, so do it here explicitly with the IOWAIT flag
4880 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
4882 for_each_sched_entity(se) {
4885 cfs_rq = cfs_rq_of(se);
4886 enqueue_entity(cfs_rq, se, flags);
4889 * end evaluation on encountering a throttled cfs_rq
4891 * note: in the case of encountering a throttled cfs_rq we will
4892 * post the final h_nr_running increment below.
4894 if (cfs_rq_throttled(cfs_rq))
4896 cfs_rq->h_nr_running++;
4898 flags = ENQUEUE_WAKEUP;
4901 for_each_sched_entity(se) {
4902 cfs_rq = cfs_rq_of(se);
4903 cfs_rq->h_nr_running++;
4905 if (cfs_rq_throttled(cfs_rq))
4908 update_load_avg(se, UPDATE_TG);
4909 update_cfs_shares(se);
4913 add_nr_running(rq, 1);
4918 static void set_next_buddy(struct sched_entity *se);
4921 * The dequeue_task method is called before nr_running is
4922 * decreased. We remove the task from the rbtree and
4923 * update the fair scheduling stats:
4925 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4927 struct cfs_rq *cfs_rq;
4928 struct sched_entity *se = &p->se;
4929 int task_sleep = flags & DEQUEUE_SLEEP;
4931 for_each_sched_entity(se) {
4932 cfs_rq = cfs_rq_of(se);
4933 dequeue_entity(cfs_rq, se, flags);
4936 * end evaluation on encountering a throttled cfs_rq
4938 * note: in the case of encountering a throttled cfs_rq we will
4939 * post the final h_nr_running decrement below.
4941 if (cfs_rq_throttled(cfs_rq))
4943 cfs_rq->h_nr_running--;
4945 /* Don't dequeue parent if it has other entities besides us */
4946 if (cfs_rq->load.weight) {
4947 /* Avoid re-evaluating load for this entity: */
4948 se = parent_entity(se);
4950 * Bias pick_next to pick a task from this cfs_rq, as
4951 * p is sleeping when it is within its sched_slice.
4953 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4957 flags |= DEQUEUE_SLEEP;
4960 for_each_sched_entity(se) {
4961 cfs_rq = cfs_rq_of(se);
4962 cfs_rq->h_nr_running--;
4964 if (cfs_rq_throttled(cfs_rq))
4967 update_load_avg(se, UPDATE_TG);
4968 update_cfs_shares(se);
4972 sub_nr_running(rq, 1);
4979 /* Working cpumask for: load_balance, load_balance_newidle. */
4980 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4981 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4983 #ifdef CONFIG_NO_HZ_COMMON
4985 * per rq 'load' arrray crap; XXX kill this.
4989 * The exact cpuload calculated at every tick would be:
4991 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4993 * If a cpu misses updates for n ticks (as it was idle) and update gets
4994 * called on the n+1-th tick when cpu may be busy, then we have:
4996 * load_n = (1 - 1/2^i)^n * load_0
4997 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4999 * decay_load_missed() below does efficient calculation of
5001 * load' = (1 - 1/2^i)^n * load
5003 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5004 * This allows us to precompute the above in said factors, thereby allowing the
5005 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5006 * fixed_power_int())
5008 * The calculation is approximated on a 128 point scale.
5010 #define DEGRADE_SHIFT 7
5012 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5013 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5014 { 0, 0, 0, 0, 0, 0, 0, 0 },
5015 { 64, 32, 8, 0, 0, 0, 0, 0 },
5016 { 96, 72, 40, 12, 1, 0, 0, 0 },
5017 { 112, 98, 75, 43, 15, 1, 0, 0 },
5018 { 120, 112, 98, 76, 45, 16, 2, 0 }
5022 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5023 * would be when CPU is idle and so we just decay the old load without
5024 * adding any new load.
5026 static unsigned long
5027 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5031 if (!missed_updates)
5034 if (missed_updates >= degrade_zero_ticks[idx])
5038 return load >> missed_updates;
5040 while (missed_updates) {
5041 if (missed_updates % 2)
5042 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5044 missed_updates >>= 1;
5049 #endif /* CONFIG_NO_HZ_COMMON */
5052 * __cpu_load_update - update the rq->cpu_load[] statistics
5053 * @this_rq: The rq to update statistics for
5054 * @this_load: The current load
5055 * @pending_updates: The number of missed updates
5057 * Update rq->cpu_load[] statistics. This function is usually called every
5058 * scheduler tick (TICK_NSEC).
5060 * This function computes a decaying average:
5062 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5064 * Because of NOHZ it might not get called on every tick which gives need for
5065 * the @pending_updates argument.
5067 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5068 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5069 * = A * (A * load[i]_n-2 + B) + B
5070 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5071 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5072 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5073 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5074 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5076 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5077 * any change in load would have resulted in the tick being turned back on.
5079 * For regular NOHZ, this reduces to:
5081 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5083 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5086 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5087 unsigned long pending_updates)
5089 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5092 this_rq->nr_load_updates++;
5094 /* Update our load: */
5095 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5096 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5097 unsigned long old_load, new_load;
5099 /* scale is effectively 1 << i now, and >> i divides by scale */
5101 old_load = this_rq->cpu_load[i];
5102 #ifdef CONFIG_NO_HZ_COMMON
5103 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5104 if (tickless_load) {
5105 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5107 * old_load can never be a negative value because a
5108 * decayed tickless_load cannot be greater than the
5109 * original tickless_load.
5111 old_load += tickless_load;
5114 new_load = this_load;
5116 * Round up the averaging division if load is increasing. This
5117 * prevents us from getting stuck on 9 if the load is 10, for
5120 if (new_load > old_load)
5121 new_load += scale - 1;
5123 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5126 sched_avg_update(this_rq);
5129 /* Used instead of source_load when we know the type == 0 */
5130 static unsigned long weighted_cpuload(const int cpu)
5132 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
5135 #ifdef CONFIG_NO_HZ_COMMON
5137 * There is no sane way to deal with nohz on smp when using jiffies because the
5138 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5139 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5141 * Therefore we need to avoid the delta approach from the regular tick when
5142 * possible since that would seriously skew the load calculation. This is why we
5143 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5144 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5145 * loop exit, nohz_idle_balance, nohz full exit...)
5147 * This means we might still be one tick off for nohz periods.
5150 static void cpu_load_update_nohz(struct rq *this_rq,
5151 unsigned long curr_jiffies,
5154 unsigned long pending_updates;
5156 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5157 if (pending_updates) {
5158 this_rq->last_load_update_tick = curr_jiffies;
5160 * In the regular NOHZ case, we were idle, this means load 0.
5161 * In the NOHZ_FULL case, we were non-idle, we should consider
5162 * its weighted load.
5164 cpu_load_update(this_rq, load, pending_updates);
5169 * Called from nohz_idle_balance() to update the load ratings before doing the
5172 static void cpu_load_update_idle(struct rq *this_rq)
5175 * bail if there's load or we're actually up-to-date.
5177 if (weighted_cpuload(cpu_of(this_rq)))
5180 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5184 * Record CPU load on nohz entry so we know the tickless load to account
5185 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5186 * than other cpu_load[idx] but it should be fine as cpu_load readers
5187 * shouldn't rely into synchronized cpu_load[*] updates.
5189 void cpu_load_update_nohz_start(void)
5191 struct rq *this_rq = this_rq();
5194 * This is all lockless but should be fine. If weighted_cpuload changes
5195 * concurrently we'll exit nohz. And cpu_load write can race with
5196 * cpu_load_update_idle() but both updater would be writing the same.
5198 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
5202 * Account the tickless load in the end of a nohz frame.
5204 void cpu_load_update_nohz_stop(void)
5206 unsigned long curr_jiffies = READ_ONCE(jiffies);
5207 struct rq *this_rq = this_rq();
5211 if (curr_jiffies == this_rq->last_load_update_tick)
5214 load = weighted_cpuload(cpu_of(this_rq));
5215 rq_lock(this_rq, &rf);
5216 update_rq_clock(this_rq);
5217 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5218 rq_unlock(this_rq, &rf);
5220 #else /* !CONFIG_NO_HZ_COMMON */
5221 static inline void cpu_load_update_nohz(struct rq *this_rq,
5222 unsigned long curr_jiffies,
5223 unsigned long load) { }
5224 #endif /* CONFIG_NO_HZ_COMMON */
5226 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5228 #ifdef CONFIG_NO_HZ_COMMON
5229 /* See the mess around cpu_load_update_nohz(). */
5230 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5232 cpu_load_update(this_rq, load, 1);
5236 * Called from scheduler_tick()
5238 void cpu_load_update_active(struct rq *this_rq)
5240 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5242 if (tick_nohz_tick_stopped())
5243 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5245 cpu_load_update_periodic(this_rq, load);
5249 * Return a low guess at the load of a migration-source cpu weighted
5250 * according to the scheduling class and "nice" value.
5252 * We want to under-estimate the load of migration sources, to
5253 * balance conservatively.
5255 static unsigned long source_load(int cpu, int type)
5257 struct rq *rq = cpu_rq(cpu);
5258 unsigned long total = weighted_cpuload(cpu);
5260 if (type == 0 || !sched_feat(LB_BIAS))
5263 return min(rq->cpu_load[type-1], total);
5267 * Return a high guess at the load of a migration-target cpu weighted
5268 * according to the scheduling class and "nice" value.
5270 static unsigned long target_load(int cpu, int type)
5272 struct rq *rq = cpu_rq(cpu);
5273 unsigned long total = weighted_cpuload(cpu);
5275 if (type == 0 || !sched_feat(LB_BIAS))
5278 return max(rq->cpu_load[type-1], total);
5281 static unsigned long capacity_of(int cpu)
5283 return cpu_rq(cpu)->cpu_capacity;
5286 static unsigned long capacity_orig_of(int cpu)
5288 return cpu_rq(cpu)->cpu_capacity_orig;
5291 static unsigned long cpu_avg_load_per_task(int cpu)
5293 struct rq *rq = cpu_rq(cpu);
5294 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5295 unsigned long load_avg = weighted_cpuload(cpu);
5298 return load_avg / nr_running;
5303 static void record_wakee(struct task_struct *p)
5306 * Only decay a single time; tasks that have less then 1 wakeup per
5307 * jiffy will not have built up many flips.
5309 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5310 current->wakee_flips >>= 1;
5311 current->wakee_flip_decay_ts = jiffies;
5314 if (current->last_wakee != p) {
5315 current->last_wakee = p;
5316 current->wakee_flips++;
5321 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5323 * A waker of many should wake a different task than the one last awakened
5324 * at a frequency roughly N times higher than one of its wakees.
5326 * In order to determine whether we should let the load spread vs consolidating
5327 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5328 * partner, and a factor of lls_size higher frequency in the other.
5330 * With both conditions met, we can be relatively sure that the relationship is
5331 * non-monogamous, with partner count exceeding socket size.
5333 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5334 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5337 static int wake_wide(struct task_struct *p)
5339 unsigned int master = current->wakee_flips;
5340 unsigned int slave = p->wakee_flips;
5341 int factor = this_cpu_read(sd_llc_size);
5344 swap(master, slave);
5345 if (slave < factor || master < slave * factor)
5350 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5351 int prev_cpu, int sync)
5353 int this_cpu = smp_processor_id();
5354 bool affine = false;
5357 * Common case: CPUs are in the same socket, and select_idle_sibling()
5358 * will do its thing regardless of what we return:
5360 if (cpus_share_cache(prev_cpu, this_cpu))
5363 affine = numa_wake_affine(sd, p, this_cpu, prev_cpu, sync);
5365 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5367 schedstat_inc(sd->ttwu_move_affine);
5368 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5374 static inline int task_util(struct task_struct *p);
5375 static int cpu_util_wake(int cpu, struct task_struct *p);
5377 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5379 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5383 * find_idlest_group finds and returns the least busy CPU group within the
5386 static struct sched_group *
5387 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5388 int this_cpu, int sd_flag)
5390 struct sched_group *idlest = NULL, *group = sd->groups;
5391 struct sched_group *most_spare_sg = NULL;
5392 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5393 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5394 unsigned long most_spare = 0, this_spare = 0;
5395 int load_idx = sd->forkexec_idx;
5396 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5397 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5398 (sd->imbalance_pct-100) / 100;
5400 if (sd_flag & SD_BALANCE_WAKE)
5401 load_idx = sd->wake_idx;
5404 unsigned long load, avg_load, runnable_load;
5405 unsigned long spare_cap, max_spare_cap;
5409 /* Skip over this group if it has no CPUs allowed */
5410 if (!cpumask_intersects(sched_group_span(group),
5414 local_group = cpumask_test_cpu(this_cpu,
5415 sched_group_span(group));
5418 * Tally up the load of all CPUs in the group and find
5419 * the group containing the CPU with most spare capacity.
5425 for_each_cpu(i, sched_group_span(group)) {
5426 /* Bias balancing toward cpus of our domain */
5428 load = source_load(i, load_idx);
5430 load = target_load(i, load_idx);
5432 runnable_load += load;
5434 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5436 spare_cap = capacity_spare_wake(i, p);
5438 if (spare_cap > max_spare_cap)
5439 max_spare_cap = spare_cap;
5442 /* Adjust by relative CPU capacity of the group */
5443 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5444 group->sgc->capacity;
5445 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5446 group->sgc->capacity;
5449 this_runnable_load = runnable_load;
5450 this_avg_load = avg_load;
5451 this_spare = max_spare_cap;
5453 if (min_runnable_load > (runnable_load + imbalance)) {
5455 * The runnable load is significantly smaller
5456 * so we can pick this new cpu
5458 min_runnable_load = runnable_load;
5459 min_avg_load = avg_load;
5461 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5462 (100*min_avg_load > imbalance_scale*avg_load)) {
5464 * The runnable loads are close so take the
5465 * blocked load into account through avg_load.
5467 min_avg_load = avg_load;
5471 if (most_spare < max_spare_cap) {
5472 most_spare = max_spare_cap;
5473 most_spare_sg = group;
5476 } while (group = group->next, group != sd->groups);
5479 * The cross-over point between using spare capacity or least load
5480 * is too conservative for high utilization tasks on partially
5481 * utilized systems if we require spare_capacity > task_util(p),
5482 * so we allow for some task stuffing by using
5483 * spare_capacity > task_util(p)/2.
5485 * Spare capacity can't be used for fork because the utilization has
5486 * not been set yet, we must first select a rq to compute the initial
5489 if (sd_flag & SD_BALANCE_FORK)
5492 if (this_spare > task_util(p) / 2 &&
5493 imbalance_scale*this_spare > 100*most_spare)
5496 if (most_spare > task_util(p) / 2)
5497 return most_spare_sg;
5503 if (min_runnable_load > (this_runnable_load + imbalance))
5506 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5507 (100*this_avg_load < imbalance_scale*min_avg_load))
5514 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5517 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5519 unsigned long load, min_load = ULONG_MAX;
5520 unsigned int min_exit_latency = UINT_MAX;
5521 u64 latest_idle_timestamp = 0;
5522 int least_loaded_cpu = this_cpu;
5523 int shallowest_idle_cpu = -1;
5526 /* Check if we have any choice: */
5527 if (group->group_weight == 1)
5528 return cpumask_first(sched_group_span(group));
5530 /* Traverse only the allowed CPUs */
5531 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5533 struct rq *rq = cpu_rq(i);
5534 struct cpuidle_state *idle = idle_get_state(rq);
5535 if (idle && idle->exit_latency < min_exit_latency) {
5537 * We give priority to a CPU whose idle state
5538 * has the smallest exit latency irrespective
5539 * of any idle timestamp.
5541 min_exit_latency = idle->exit_latency;
5542 latest_idle_timestamp = rq->idle_stamp;
5543 shallowest_idle_cpu = i;
5544 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5545 rq->idle_stamp > latest_idle_timestamp) {
5547 * If equal or no active idle state, then
5548 * the most recently idled CPU might have
5551 latest_idle_timestamp = rq->idle_stamp;
5552 shallowest_idle_cpu = i;
5554 } else if (shallowest_idle_cpu == -1) {
5555 load = weighted_cpuload(i);
5556 if (load < min_load || (load == min_load && i == this_cpu)) {
5558 least_loaded_cpu = i;
5563 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5566 #ifdef CONFIG_SCHED_SMT
5568 static inline void set_idle_cores(int cpu, int val)
5570 struct sched_domain_shared *sds;
5572 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5574 WRITE_ONCE(sds->has_idle_cores, val);
5577 static inline bool test_idle_cores(int cpu, bool def)
5579 struct sched_domain_shared *sds;
5581 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5583 return READ_ONCE(sds->has_idle_cores);
5589 * Scans the local SMT mask to see if the entire core is idle, and records this
5590 * information in sd_llc_shared->has_idle_cores.
5592 * Since SMT siblings share all cache levels, inspecting this limited remote
5593 * state should be fairly cheap.
5595 void __update_idle_core(struct rq *rq)
5597 int core = cpu_of(rq);
5601 if (test_idle_cores(core, true))
5604 for_each_cpu(cpu, cpu_smt_mask(core)) {
5612 set_idle_cores(core, 1);
5618 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5619 * there are no idle cores left in the system; tracked through
5620 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5622 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5624 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5627 if (!static_branch_likely(&sched_smt_present))
5630 if (!test_idle_cores(target, false))
5633 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5635 for_each_cpu_wrap(core, cpus, target) {
5638 for_each_cpu(cpu, cpu_smt_mask(core)) {
5639 cpumask_clear_cpu(cpu, cpus);
5649 * Failed to find an idle core; stop looking for one.
5651 set_idle_cores(target, 0);
5657 * Scan the local SMT mask for idle CPUs.
5659 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5663 if (!static_branch_likely(&sched_smt_present))
5666 for_each_cpu(cpu, cpu_smt_mask(target)) {
5667 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5676 #else /* CONFIG_SCHED_SMT */
5678 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5683 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5688 #endif /* CONFIG_SCHED_SMT */
5691 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5692 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5693 * average idle time for this rq (as found in rq->avg_idle).
5695 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5697 struct sched_domain *this_sd;
5698 u64 avg_cost, avg_idle;
5701 int cpu, nr = INT_MAX;
5703 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5708 * Due to large variance we need a large fuzz factor; hackbench in
5709 * particularly is sensitive here.
5711 avg_idle = this_rq()->avg_idle / 512;
5712 avg_cost = this_sd->avg_scan_cost + 1;
5714 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5717 if (sched_feat(SIS_PROP)) {
5718 u64 span_avg = sd->span_weight * avg_idle;
5719 if (span_avg > 4*avg_cost)
5720 nr = div_u64(span_avg, avg_cost);
5725 time = local_clock();
5727 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5730 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5736 time = local_clock() - time;
5737 cost = this_sd->avg_scan_cost;
5738 delta = (s64)(time - cost) / 8;
5739 this_sd->avg_scan_cost += delta;
5745 * Try and locate an idle core/thread in the LLC cache domain.
5747 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5749 struct sched_domain *sd;
5752 if (idle_cpu(target))
5756 * If the previous cpu is cache affine and idle, don't be stupid.
5758 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5761 sd = rcu_dereference(per_cpu(sd_llc, target));
5765 i = select_idle_core(p, sd, target);
5766 if ((unsigned)i < nr_cpumask_bits)
5769 i = select_idle_cpu(p, sd, target);
5770 if ((unsigned)i < nr_cpumask_bits)
5773 i = select_idle_smt(p, sd, target);
5774 if ((unsigned)i < nr_cpumask_bits)
5781 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5782 * tasks. The unit of the return value must be the one of capacity so we can
5783 * compare the utilization with the capacity of the CPU that is available for
5784 * CFS task (ie cpu_capacity).
5786 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5787 * recent utilization of currently non-runnable tasks on a CPU. It represents
5788 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5789 * capacity_orig is the cpu_capacity available at the highest frequency
5790 * (arch_scale_freq_capacity()).
5791 * The utilization of a CPU converges towards a sum equal to or less than the
5792 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5793 * the running time on this CPU scaled by capacity_curr.
5795 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5796 * higher than capacity_orig because of unfortunate rounding in
5797 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5798 * the average stabilizes with the new running time. We need to check that the
5799 * utilization stays within the range of [0..capacity_orig] and cap it if
5800 * necessary. Without utilization capping, a group could be seen as overloaded
5801 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5802 * available capacity. We allow utilization to overshoot capacity_curr (but not
5803 * capacity_orig) as it useful for predicting the capacity required after task
5804 * migrations (scheduler-driven DVFS).
5806 static int cpu_util(int cpu)
5808 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5809 unsigned long capacity = capacity_orig_of(cpu);
5811 return (util >= capacity) ? capacity : util;
5814 static inline int task_util(struct task_struct *p)
5816 return p->se.avg.util_avg;
5820 * cpu_util_wake: Compute cpu utilization with any contributions from
5821 * the waking task p removed.
5823 static int cpu_util_wake(int cpu, struct task_struct *p)
5825 unsigned long util, capacity;
5827 /* Task has no contribution or is new */
5828 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5829 return cpu_util(cpu);
5831 capacity = capacity_orig_of(cpu);
5832 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5834 return (util >= capacity) ? capacity : util;
5838 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5839 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5841 * In that case WAKE_AFFINE doesn't make sense and we'll let
5842 * BALANCE_WAKE sort things out.
5844 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5846 long min_cap, max_cap;
5848 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5849 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5851 /* Minimum capacity is close to max, no need to abort wake_affine */
5852 if (max_cap - min_cap < max_cap >> 3)
5855 /* Bring task utilization in sync with prev_cpu */
5856 sync_entity_load_avg(&p->se);
5858 return min_cap * 1024 < task_util(p) * capacity_margin;
5862 * select_task_rq_fair: Select target runqueue for the waking task in domains
5863 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5864 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5866 * Balances load by selecting the idlest cpu in the idlest group, or under
5867 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5869 * Returns the target cpu number.
5871 * preempt must be disabled.
5874 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5876 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5877 int cpu = smp_processor_id();
5878 int new_cpu = prev_cpu;
5879 int want_affine = 0;
5880 int sync = wake_flags & WF_SYNC;
5882 if (sd_flag & SD_BALANCE_WAKE) {
5884 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5885 && cpumask_test_cpu(cpu, &p->cpus_allowed);
5889 for_each_domain(cpu, tmp) {
5890 if (!(tmp->flags & SD_LOAD_BALANCE))
5894 * If both cpu and prev_cpu are part of this domain,
5895 * cpu is a valid SD_WAKE_AFFINE target.
5897 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5898 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5903 if (tmp->flags & sd_flag)
5905 else if (!want_affine)
5910 sd = NULL; /* Prefer wake_affine over balance flags */
5911 if (cpu == prev_cpu)
5914 if (wake_affine(affine_sd, p, prev_cpu, sync))
5920 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5921 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5924 struct sched_group *group;
5927 if (!(sd->flags & sd_flag)) {
5932 group = find_idlest_group(sd, p, cpu, sd_flag);
5938 new_cpu = find_idlest_cpu(group, p, cpu);
5939 if (new_cpu == -1 || new_cpu == cpu) {
5940 /* Now try balancing at a lower domain level of cpu */
5945 /* Now try balancing at a lower domain level of new_cpu */
5947 weight = sd->span_weight;
5949 for_each_domain(cpu, tmp) {
5950 if (weight <= tmp->span_weight)
5952 if (tmp->flags & sd_flag)
5955 /* while loop will break here if sd == NULL */
5963 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5964 * cfs_rq_of(p) references at time of call are still valid and identify the
5965 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5967 static void migrate_task_rq_fair(struct task_struct *p)
5970 * As blocked tasks retain absolute vruntime the migration needs to
5971 * deal with this by subtracting the old and adding the new
5972 * min_vruntime -- the latter is done by enqueue_entity() when placing
5973 * the task on the new runqueue.
5975 if (p->state == TASK_WAKING) {
5976 struct sched_entity *se = &p->se;
5977 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5980 #ifndef CONFIG_64BIT
5981 u64 min_vruntime_copy;
5984 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5986 min_vruntime = cfs_rq->min_vruntime;
5987 } while (min_vruntime != min_vruntime_copy);
5989 min_vruntime = cfs_rq->min_vruntime;
5992 se->vruntime -= min_vruntime;
5996 * We are supposed to update the task to "current" time, then its up to date
5997 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5998 * what current time is, so simply throw away the out-of-date time. This
5999 * will result in the wakee task is less decayed, but giving the wakee more
6000 * load sounds not bad.
6002 remove_entity_load_avg(&p->se);
6004 /* Tell new CPU we are migrated */
6005 p->se.avg.last_update_time = 0;
6007 /* We have migrated, no longer consider this task hot */
6008 p->se.exec_start = 0;
6011 static void task_dead_fair(struct task_struct *p)
6013 remove_entity_load_avg(&p->se);
6015 #endif /* CONFIG_SMP */
6017 static unsigned long
6018 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6020 unsigned long gran = sysctl_sched_wakeup_granularity;
6023 * Since its curr running now, convert the gran from real-time
6024 * to virtual-time in his units.
6026 * By using 'se' instead of 'curr' we penalize light tasks, so
6027 * they get preempted easier. That is, if 'se' < 'curr' then
6028 * the resulting gran will be larger, therefore penalizing the
6029 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6030 * be smaller, again penalizing the lighter task.
6032 * This is especially important for buddies when the leftmost
6033 * task is higher priority than the buddy.
6035 return calc_delta_fair(gran, se);
6039 * Should 'se' preempt 'curr'.
6053 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6055 s64 gran, vdiff = curr->vruntime - se->vruntime;
6060 gran = wakeup_gran(curr, se);
6067 static void set_last_buddy(struct sched_entity *se)
6069 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6072 for_each_sched_entity(se) {
6073 if (SCHED_WARN_ON(!se->on_rq))
6075 cfs_rq_of(se)->last = se;
6079 static void set_next_buddy(struct sched_entity *se)
6081 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6084 for_each_sched_entity(se) {
6085 if (SCHED_WARN_ON(!se->on_rq))
6087 cfs_rq_of(se)->next = se;
6091 static void set_skip_buddy(struct sched_entity *se)
6093 for_each_sched_entity(se)
6094 cfs_rq_of(se)->skip = se;
6098 * Preempt the current task with a newly woken task if needed:
6100 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6102 struct task_struct *curr = rq->curr;
6103 struct sched_entity *se = &curr->se, *pse = &p->se;
6104 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6105 int scale = cfs_rq->nr_running >= sched_nr_latency;
6106 int next_buddy_marked = 0;
6108 if (unlikely(se == pse))
6112 * This is possible from callers such as attach_tasks(), in which we
6113 * unconditionally check_prempt_curr() after an enqueue (which may have
6114 * lead to a throttle). This both saves work and prevents false
6115 * next-buddy nomination below.
6117 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6120 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6121 set_next_buddy(pse);
6122 next_buddy_marked = 1;
6126 * We can come here with TIF_NEED_RESCHED already set from new task
6129 * Note: this also catches the edge-case of curr being in a throttled
6130 * group (e.g. via set_curr_task), since update_curr() (in the
6131 * enqueue of curr) will have resulted in resched being set. This
6132 * prevents us from potentially nominating it as a false LAST_BUDDY
6135 if (test_tsk_need_resched(curr))
6138 /* Idle tasks are by definition preempted by non-idle tasks. */
6139 if (unlikely(curr->policy == SCHED_IDLE) &&
6140 likely(p->policy != SCHED_IDLE))
6144 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6145 * is driven by the tick):
6147 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6150 find_matching_se(&se, &pse);
6151 update_curr(cfs_rq_of(se));
6153 if (wakeup_preempt_entity(se, pse) == 1) {
6155 * Bias pick_next to pick the sched entity that is
6156 * triggering this preemption.
6158 if (!next_buddy_marked)
6159 set_next_buddy(pse);
6168 * Only set the backward buddy when the current task is still
6169 * on the rq. This can happen when a wakeup gets interleaved
6170 * with schedule on the ->pre_schedule() or idle_balance()
6171 * point, either of which can * drop the rq lock.
6173 * Also, during early boot the idle thread is in the fair class,
6174 * for obvious reasons its a bad idea to schedule back to it.
6176 if (unlikely(!se->on_rq || curr == rq->idle))
6179 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6183 static struct task_struct *
6184 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6186 struct cfs_rq *cfs_rq = &rq->cfs;
6187 struct sched_entity *se;
6188 struct task_struct *p;
6192 #ifdef CONFIG_FAIR_GROUP_SCHED
6193 if (!cfs_rq->nr_running)
6196 if (prev->sched_class != &fair_sched_class)
6200 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6201 * likely that a next task is from the same cgroup as the current.
6203 * Therefore attempt to avoid putting and setting the entire cgroup
6204 * hierarchy, only change the part that actually changes.
6208 struct sched_entity *curr = cfs_rq->curr;
6211 * Since we got here without doing put_prev_entity() we also
6212 * have to consider cfs_rq->curr. If it is still a runnable
6213 * entity, update_curr() will update its vruntime, otherwise
6214 * forget we've ever seen it.
6218 update_curr(cfs_rq);
6223 * This call to check_cfs_rq_runtime() will do the
6224 * throttle and dequeue its entity in the parent(s).
6225 * Therefore the 'simple' nr_running test will indeed
6228 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6232 se = pick_next_entity(cfs_rq, curr);
6233 cfs_rq = group_cfs_rq(se);
6239 * Since we haven't yet done put_prev_entity and if the selected task
6240 * is a different task than we started out with, try and touch the
6241 * least amount of cfs_rqs.
6244 struct sched_entity *pse = &prev->se;
6246 while (!(cfs_rq = is_same_group(se, pse))) {
6247 int se_depth = se->depth;
6248 int pse_depth = pse->depth;
6250 if (se_depth <= pse_depth) {
6251 put_prev_entity(cfs_rq_of(pse), pse);
6252 pse = parent_entity(pse);
6254 if (se_depth >= pse_depth) {
6255 set_next_entity(cfs_rq_of(se), se);
6256 se = parent_entity(se);
6260 put_prev_entity(cfs_rq, pse);
6261 set_next_entity(cfs_rq, se);
6264 if (hrtick_enabled(rq))
6265 hrtick_start_fair(rq, p);
6272 if (!cfs_rq->nr_running)
6275 put_prev_task(rq, prev);
6278 se = pick_next_entity(cfs_rq, NULL);
6279 set_next_entity(cfs_rq, se);
6280 cfs_rq = group_cfs_rq(se);
6285 if (hrtick_enabled(rq))
6286 hrtick_start_fair(rq, p);
6291 new_tasks = idle_balance(rq, rf);
6294 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6295 * possible for any higher priority task to appear. In that case we
6296 * must re-start the pick_next_entity() loop.
6308 * Account for a descheduled task:
6310 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6312 struct sched_entity *se = &prev->se;
6313 struct cfs_rq *cfs_rq;
6315 for_each_sched_entity(se) {
6316 cfs_rq = cfs_rq_of(se);
6317 put_prev_entity(cfs_rq, se);
6322 * sched_yield() is very simple
6324 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6326 static void yield_task_fair(struct rq *rq)
6328 struct task_struct *curr = rq->curr;
6329 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6330 struct sched_entity *se = &curr->se;
6333 * Are we the only task in the tree?
6335 if (unlikely(rq->nr_running == 1))
6338 clear_buddies(cfs_rq, se);
6340 if (curr->policy != SCHED_BATCH) {
6341 update_rq_clock(rq);
6343 * Update run-time statistics of the 'current'.
6345 update_curr(cfs_rq);
6347 * Tell update_rq_clock() that we've just updated,
6348 * so we don't do microscopic update in schedule()
6349 * and double the fastpath cost.
6351 rq_clock_skip_update(rq, true);
6357 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6359 struct sched_entity *se = &p->se;
6361 /* throttled hierarchies are not runnable */
6362 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6365 /* Tell the scheduler that we'd really like pse to run next. */
6368 yield_task_fair(rq);
6374 /**************************************************
6375 * Fair scheduling class load-balancing methods.
6379 * The purpose of load-balancing is to achieve the same basic fairness the
6380 * per-cpu scheduler provides, namely provide a proportional amount of compute
6381 * time to each task. This is expressed in the following equation:
6383 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6385 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6386 * W_i,0 is defined as:
6388 * W_i,0 = \Sum_j w_i,j (2)
6390 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6391 * is derived from the nice value as per sched_prio_to_weight[].
6393 * The weight average is an exponential decay average of the instantaneous
6396 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6398 * C_i is the compute capacity of cpu i, typically it is the
6399 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6400 * can also include other factors [XXX].
6402 * To achieve this balance we define a measure of imbalance which follows
6403 * directly from (1):
6405 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6407 * We them move tasks around to minimize the imbalance. In the continuous
6408 * function space it is obvious this converges, in the discrete case we get
6409 * a few fun cases generally called infeasible weight scenarios.
6412 * - infeasible weights;
6413 * - local vs global optima in the discrete case. ]
6418 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6419 * for all i,j solution, we create a tree of cpus that follows the hardware
6420 * topology where each level pairs two lower groups (or better). This results
6421 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6422 * tree to only the first of the previous level and we decrease the frequency
6423 * of load-balance at each level inv. proportional to the number of cpus in
6429 * \Sum { --- * --- * 2^i } = O(n) (5)
6431 * `- size of each group
6432 * | | `- number of cpus doing load-balance
6434 * `- sum over all levels
6436 * Coupled with a limit on how many tasks we can migrate every balance pass,
6437 * this makes (5) the runtime complexity of the balancer.
6439 * An important property here is that each CPU is still (indirectly) connected
6440 * to every other cpu in at most O(log n) steps:
6442 * The adjacency matrix of the resulting graph is given by:
6445 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6448 * And you'll find that:
6450 * A^(log_2 n)_i,j != 0 for all i,j (7)
6452 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6453 * The task movement gives a factor of O(m), giving a convergence complexity
6456 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6461 * In order to avoid CPUs going idle while there's still work to do, new idle
6462 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6463 * tree itself instead of relying on other CPUs to bring it work.
6465 * This adds some complexity to both (5) and (8) but it reduces the total idle
6473 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6476 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6481 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6483 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6485 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6488 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6489 * rewrite all of this once again.]
6492 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6494 enum fbq_type { regular, remote, all };
6496 #define LBF_ALL_PINNED 0x01
6497 #define LBF_NEED_BREAK 0x02
6498 #define LBF_DST_PINNED 0x04
6499 #define LBF_SOME_PINNED 0x08
6502 struct sched_domain *sd;
6510 struct cpumask *dst_grpmask;
6512 enum cpu_idle_type idle;
6514 /* The set of CPUs under consideration for load-balancing */
6515 struct cpumask *cpus;
6520 unsigned int loop_break;
6521 unsigned int loop_max;
6523 enum fbq_type fbq_type;
6524 struct list_head tasks;
6528 * Is this task likely cache-hot:
6530 static int task_hot(struct task_struct *p, struct lb_env *env)
6534 lockdep_assert_held(&env->src_rq->lock);
6536 if (p->sched_class != &fair_sched_class)
6539 if (unlikely(p->policy == SCHED_IDLE))
6543 * Buddy candidates are cache hot:
6545 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6546 (&p->se == cfs_rq_of(&p->se)->next ||
6547 &p->se == cfs_rq_of(&p->se)->last))
6550 if (sysctl_sched_migration_cost == -1)
6552 if (sysctl_sched_migration_cost == 0)
6555 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6557 return delta < (s64)sysctl_sched_migration_cost;
6560 #ifdef CONFIG_NUMA_BALANCING
6562 * Returns 1, if task migration degrades locality
6563 * Returns 0, if task migration improves locality i.e migration preferred.
6564 * Returns -1, if task migration is not affected by locality.
6566 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6568 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6569 unsigned long src_faults, dst_faults;
6570 int src_nid, dst_nid;
6572 if (!static_branch_likely(&sched_numa_balancing))
6575 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6578 src_nid = cpu_to_node(env->src_cpu);
6579 dst_nid = cpu_to_node(env->dst_cpu);
6581 if (src_nid == dst_nid)
6584 /* Migrating away from the preferred node is always bad. */
6585 if (src_nid == p->numa_preferred_nid) {
6586 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6592 /* Encourage migration to the preferred node. */
6593 if (dst_nid == p->numa_preferred_nid)
6596 /* Leaving a core idle is often worse than degrading locality. */
6597 if (env->idle != CPU_NOT_IDLE)
6601 src_faults = group_faults(p, src_nid);
6602 dst_faults = group_faults(p, dst_nid);
6604 src_faults = task_faults(p, src_nid);
6605 dst_faults = task_faults(p, dst_nid);
6608 return dst_faults < src_faults;
6612 static inline int migrate_degrades_locality(struct task_struct *p,
6620 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6623 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6627 lockdep_assert_held(&env->src_rq->lock);
6630 * We do not migrate tasks that are:
6631 * 1) throttled_lb_pair, or
6632 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6633 * 3) running (obviously), or
6634 * 4) are cache-hot on their current CPU.
6636 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6639 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6642 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6644 env->flags |= LBF_SOME_PINNED;
6647 * Remember if this task can be migrated to any other cpu in
6648 * our sched_group. We may want to revisit it if we couldn't
6649 * meet load balance goals by pulling other tasks on src_cpu.
6651 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6652 * already computed one in current iteration.
6654 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
6657 /* Prevent to re-select dst_cpu via env's cpus */
6658 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6659 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6660 env->flags |= LBF_DST_PINNED;
6661 env->new_dst_cpu = cpu;
6669 /* Record that we found atleast one task that could run on dst_cpu */
6670 env->flags &= ~LBF_ALL_PINNED;
6672 if (task_running(env->src_rq, p)) {
6673 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6678 * Aggressive migration if:
6679 * 1) destination numa is preferred
6680 * 2) task is cache cold, or
6681 * 3) too many balance attempts have failed.
6683 tsk_cache_hot = migrate_degrades_locality(p, env);
6684 if (tsk_cache_hot == -1)
6685 tsk_cache_hot = task_hot(p, env);
6687 if (tsk_cache_hot <= 0 ||
6688 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6689 if (tsk_cache_hot == 1) {
6690 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6691 schedstat_inc(p->se.statistics.nr_forced_migrations);
6696 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6701 * detach_task() -- detach the task for the migration specified in env
6703 static void detach_task(struct task_struct *p, struct lb_env *env)
6705 lockdep_assert_held(&env->src_rq->lock);
6707 p->on_rq = TASK_ON_RQ_MIGRATING;
6708 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6709 set_task_cpu(p, env->dst_cpu);
6713 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6714 * part of active balancing operations within "domain".
6716 * Returns a task if successful and NULL otherwise.
6718 static struct task_struct *detach_one_task(struct lb_env *env)
6720 struct task_struct *p, *n;
6722 lockdep_assert_held(&env->src_rq->lock);
6724 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6725 if (!can_migrate_task(p, env))
6728 detach_task(p, env);
6731 * Right now, this is only the second place where
6732 * lb_gained[env->idle] is updated (other is detach_tasks)
6733 * so we can safely collect stats here rather than
6734 * inside detach_tasks().
6736 schedstat_inc(env->sd->lb_gained[env->idle]);
6742 static const unsigned int sched_nr_migrate_break = 32;
6745 * detach_tasks() -- tries to detach up to imbalance weighted load from
6746 * busiest_rq, as part of a balancing operation within domain "sd".
6748 * Returns number of detached tasks if successful and 0 otherwise.
6750 static int detach_tasks(struct lb_env *env)
6752 struct list_head *tasks = &env->src_rq->cfs_tasks;
6753 struct task_struct *p;
6757 lockdep_assert_held(&env->src_rq->lock);
6759 if (env->imbalance <= 0)
6762 while (!list_empty(tasks)) {
6764 * We don't want to steal all, otherwise we may be treated likewise,
6765 * which could at worst lead to a livelock crash.
6767 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6770 p = list_first_entry(tasks, struct task_struct, se.group_node);
6773 /* We've more or less seen every task there is, call it quits */
6774 if (env->loop > env->loop_max)
6777 /* take a breather every nr_migrate tasks */
6778 if (env->loop > env->loop_break) {
6779 env->loop_break += sched_nr_migrate_break;
6780 env->flags |= LBF_NEED_BREAK;
6784 if (!can_migrate_task(p, env))
6787 load = task_h_load(p);
6789 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6792 if ((load / 2) > env->imbalance)
6795 detach_task(p, env);
6796 list_add(&p->se.group_node, &env->tasks);
6799 env->imbalance -= load;
6801 #ifdef CONFIG_PREEMPT
6803 * NEWIDLE balancing is a source of latency, so preemptible
6804 * kernels will stop after the first task is detached to minimize
6805 * the critical section.
6807 if (env->idle == CPU_NEWLY_IDLE)
6812 * We only want to steal up to the prescribed amount of
6815 if (env->imbalance <= 0)
6820 list_move_tail(&p->se.group_node, tasks);
6824 * Right now, this is one of only two places we collect this stat
6825 * so we can safely collect detach_one_task() stats here rather
6826 * than inside detach_one_task().
6828 schedstat_add(env->sd->lb_gained[env->idle], detached);
6834 * attach_task() -- attach the task detached by detach_task() to its new rq.
6836 static void attach_task(struct rq *rq, struct task_struct *p)
6838 lockdep_assert_held(&rq->lock);
6840 BUG_ON(task_rq(p) != rq);
6841 activate_task(rq, p, ENQUEUE_NOCLOCK);
6842 p->on_rq = TASK_ON_RQ_QUEUED;
6843 check_preempt_curr(rq, p, 0);
6847 * attach_one_task() -- attaches the task returned from detach_one_task() to
6850 static void attach_one_task(struct rq *rq, struct task_struct *p)
6855 update_rq_clock(rq);
6861 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6864 static void attach_tasks(struct lb_env *env)
6866 struct list_head *tasks = &env->tasks;
6867 struct task_struct *p;
6870 rq_lock(env->dst_rq, &rf);
6871 update_rq_clock(env->dst_rq);
6873 while (!list_empty(tasks)) {
6874 p = list_first_entry(tasks, struct task_struct, se.group_node);
6875 list_del_init(&p->se.group_node);
6877 attach_task(env->dst_rq, p);
6880 rq_unlock(env->dst_rq, &rf);
6883 #ifdef CONFIG_FAIR_GROUP_SCHED
6885 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
6887 if (cfs_rq->load.weight)
6890 if (cfs_rq->avg.load_sum)
6893 if (cfs_rq->avg.util_sum)
6896 if (cfs_rq->runnable_load_sum)
6902 static void update_blocked_averages(int cpu)
6904 struct rq *rq = cpu_rq(cpu);
6905 struct cfs_rq *cfs_rq, *pos;
6908 rq_lock_irqsave(rq, &rf);
6909 update_rq_clock(rq);
6912 * Iterates the task_group tree in a bottom up fashion, see
6913 * list_add_leaf_cfs_rq() for details.
6915 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
6916 struct sched_entity *se;
6918 /* throttled entities do not contribute to load */
6919 if (throttled_hierarchy(cfs_rq))
6922 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6923 update_tg_load_avg(cfs_rq, 0);
6925 /* Propagate pending load changes to the parent, if any: */
6926 se = cfs_rq->tg->se[cpu];
6927 if (se && !skip_blocked_update(se))
6928 update_load_avg(se, 0);
6931 * There can be a lot of idle CPU cgroups. Don't let fully
6932 * decayed cfs_rqs linger on the list.
6934 if (cfs_rq_is_decayed(cfs_rq))
6935 list_del_leaf_cfs_rq(cfs_rq);
6937 rq_unlock_irqrestore(rq, &rf);
6941 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6942 * This needs to be done in a top-down fashion because the load of a child
6943 * group is a fraction of its parents load.
6945 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6947 struct rq *rq = rq_of(cfs_rq);
6948 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6949 unsigned long now = jiffies;
6952 if (cfs_rq->last_h_load_update == now)
6955 cfs_rq->h_load_next = NULL;
6956 for_each_sched_entity(se) {
6957 cfs_rq = cfs_rq_of(se);
6958 cfs_rq->h_load_next = se;
6959 if (cfs_rq->last_h_load_update == now)
6964 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6965 cfs_rq->last_h_load_update = now;
6968 while ((se = cfs_rq->h_load_next) != NULL) {
6969 load = cfs_rq->h_load;
6970 load = div64_ul(load * se->avg.load_avg,
6971 cfs_rq_load_avg(cfs_rq) + 1);
6972 cfs_rq = group_cfs_rq(se);
6973 cfs_rq->h_load = load;
6974 cfs_rq->last_h_load_update = now;
6978 static unsigned long task_h_load(struct task_struct *p)
6980 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6982 update_cfs_rq_h_load(cfs_rq);
6983 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6984 cfs_rq_load_avg(cfs_rq) + 1);
6987 static inline void update_blocked_averages(int cpu)
6989 struct rq *rq = cpu_rq(cpu);
6990 struct cfs_rq *cfs_rq = &rq->cfs;
6993 rq_lock_irqsave(rq, &rf);
6994 update_rq_clock(rq);
6995 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6996 rq_unlock_irqrestore(rq, &rf);
6999 static unsigned long task_h_load(struct task_struct *p)
7001 return p->se.avg.load_avg;
7005 /********** Helpers for find_busiest_group ************************/
7014 * sg_lb_stats - stats of a sched_group required for load_balancing
7016 struct sg_lb_stats {
7017 unsigned long avg_load; /*Avg load across the CPUs of the group */
7018 unsigned long group_load; /* Total load over the CPUs of the group */
7019 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7020 unsigned long load_per_task;
7021 unsigned long group_capacity;
7022 unsigned long group_util; /* Total utilization of the group */
7023 unsigned int sum_nr_running; /* Nr tasks running in the group */
7024 unsigned int idle_cpus;
7025 unsigned int group_weight;
7026 enum group_type group_type;
7027 int group_no_capacity;
7028 #ifdef CONFIG_NUMA_BALANCING
7029 unsigned int nr_numa_running;
7030 unsigned int nr_preferred_running;
7035 * sd_lb_stats - Structure to store the statistics of a sched_domain
7036 * during load balancing.
7038 struct sd_lb_stats {
7039 struct sched_group *busiest; /* Busiest group in this sd */
7040 struct sched_group *local; /* Local group in this sd */
7041 unsigned long total_load; /* Total load of all groups in sd */
7042 unsigned long total_capacity; /* Total capacity of all groups in sd */
7043 unsigned long avg_load; /* Average load across all groups in sd */
7045 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7046 struct sg_lb_stats local_stat; /* Statistics of the local group */
7049 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7052 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7053 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7054 * We must however clear busiest_stat::avg_load because
7055 * update_sd_pick_busiest() reads this before assignment.
7057 *sds = (struct sd_lb_stats){
7061 .total_capacity = 0UL,
7064 .sum_nr_running = 0,
7065 .group_type = group_other,
7071 * get_sd_load_idx - Obtain the load index for a given sched domain.
7072 * @sd: The sched_domain whose load_idx is to be obtained.
7073 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7075 * Return: The load index.
7077 static inline int get_sd_load_idx(struct sched_domain *sd,
7078 enum cpu_idle_type idle)
7084 load_idx = sd->busy_idx;
7087 case CPU_NEWLY_IDLE:
7088 load_idx = sd->newidle_idx;
7091 load_idx = sd->idle_idx;
7098 static unsigned long scale_rt_capacity(int cpu)
7100 struct rq *rq = cpu_rq(cpu);
7101 u64 total, used, age_stamp, avg;
7105 * Since we're reading these variables without serialization make sure
7106 * we read them once before doing sanity checks on them.
7108 age_stamp = READ_ONCE(rq->age_stamp);
7109 avg = READ_ONCE(rq->rt_avg);
7110 delta = __rq_clock_broken(rq) - age_stamp;
7112 if (unlikely(delta < 0))
7115 total = sched_avg_period() + delta;
7117 used = div_u64(avg, total);
7119 if (likely(used < SCHED_CAPACITY_SCALE))
7120 return SCHED_CAPACITY_SCALE - used;
7125 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7127 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7128 struct sched_group *sdg = sd->groups;
7130 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7132 capacity *= scale_rt_capacity(cpu);
7133 capacity >>= SCHED_CAPACITY_SHIFT;
7138 cpu_rq(cpu)->cpu_capacity = capacity;
7139 sdg->sgc->capacity = capacity;
7140 sdg->sgc->min_capacity = capacity;
7143 void update_group_capacity(struct sched_domain *sd, int cpu)
7145 struct sched_domain *child = sd->child;
7146 struct sched_group *group, *sdg = sd->groups;
7147 unsigned long capacity, min_capacity;
7148 unsigned long interval;
7150 interval = msecs_to_jiffies(sd->balance_interval);
7151 interval = clamp(interval, 1UL, max_load_balance_interval);
7152 sdg->sgc->next_update = jiffies + interval;
7155 update_cpu_capacity(sd, cpu);
7160 min_capacity = ULONG_MAX;
7162 if (child->flags & SD_OVERLAP) {
7164 * SD_OVERLAP domains cannot assume that child groups
7165 * span the current group.
7168 for_each_cpu(cpu, sched_group_span(sdg)) {
7169 struct sched_group_capacity *sgc;
7170 struct rq *rq = cpu_rq(cpu);
7173 * build_sched_domains() -> init_sched_groups_capacity()
7174 * gets here before we've attached the domains to the
7177 * Use capacity_of(), which is set irrespective of domains
7178 * in update_cpu_capacity().
7180 * This avoids capacity from being 0 and
7181 * causing divide-by-zero issues on boot.
7183 if (unlikely(!rq->sd)) {
7184 capacity += capacity_of(cpu);
7186 sgc = rq->sd->groups->sgc;
7187 capacity += sgc->capacity;
7190 min_capacity = min(capacity, min_capacity);
7194 * !SD_OVERLAP domains can assume that child groups
7195 * span the current group.
7198 group = child->groups;
7200 struct sched_group_capacity *sgc = group->sgc;
7202 capacity += sgc->capacity;
7203 min_capacity = min(sgc->min_capacity, min_capacity);
7204 group = group->next;
7205 } while (group != child->groups);
7208 sdg->sgc->capacity = capacity;
7209 sdg->sgc->min_capacity = min_capacity;
7213 * Check whether the capacity of the rq has been noticeably reduced by side
7214 * activity. The imbalance_pct is used for the threshold.
7215 * Return true is the capacity is reduced
7218 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7220 return ((rq->cpu_capacity * sd->imbalance_pct) <
7221 (rq->cpu_capacity_orig * 100));
7225 * Group imbalance indicates (and tries to solve) the problem where balancing
7226 * groups is inadequate due to ->cpus_allowed constraints.
7228 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7229 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7232 * { 0 1 2 3 } { 4 5 6 7 }
7235 * If we were to balance group-wise we'd place two tasks in the first group and
7236 * two tasks in the second group. Clearly this is undesired as it will overload
7237 * cpu 3 and leave one of the cpus in the second group unused.
7239 * The current solution to this issue is detecting the skew in the first group
7240 * by noticing the lower domain failed to reach balance and had difficulty
7241 * moving tasks due to affinity constraints.
7243 * When this is so detected; this group becomes a candidate for busiest; see
7244 * update_sd_pick_busiest(). And calculate_imbalance() and
7245 * find_busiest_group() avoid some of the usual balance conditions to allow it
7246 * to create an effective group imbalance.
7248 * This is a somewhat tricky proposition since the next run might not find the
7249 * group imbalance and decide the groups need to be balanced again. A most
7250 * subtle and fragile situation.
7253 static inline int sg_imbalanced(struct sched_group *group)
7255 return group->sgc->imbalance;
7259 * group_has_capacity returns true if the group has spare capacity that could
7260 * be used by some tasks.
7261 * We consider that a group has spare capacity if the * number of task is
7262 * smaller than the number of CPUs or if the utilization is lower than the
7263 * available capacity for CFS tasks.
7264 * For the latter, we use a threshold to stabilize the state, to take into
7265 * account the variance of the tasks' load and to return true if the available
7266 * capacity in meaningful for the load balancer.
7267 * As an example, an available capacity of 1% can appear but it doesn't make
7268 * any benefit for the load balance.
7271 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7273 if (sgs->sum_nr_running < sgs->group_weight)
7276 if ((sgs->group_capacity * 100) >
7277 (sgs->group_util * env->sd->imbalance_pct))
7284 * group_is_overloaded returns true if the group has more tasks than it can
7286 * group_is_overloaded is not equals to !group_has_capacity because a group
7287 * with the exact right number of tasks, has no more spare capacity but is not
7288 * overloaded so both group_has_capacity and group_is_overloaded return
7292 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7294 if (sgs->sum_nr_running <= sgs->group_weight)
7297 if ((sgs->group_capacity * 100) <
7298 (sgs->group_util * env->sd->imbalance_pct))
7305 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7306 * per-CPU capacity than sched_group ref.
7309 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7311 return sg->sgc->min_capacity * capacity_margin <
7312 ref->sgc->min_capacity * 1024;
7316 group_type group_classify(struct sched_group *group,
7317 struct sg_lb_stats *sgs)
7319 if (sgs->group_no_capacity)
7320 return group_overloaded;
7322 if (sg_imbalanced(group))
7323 return group_imbalanced;
7329 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7330 * @env: The load balancing environment.
7331 * @group: sched_group whose statistics are to be updated.
7332 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7333 * @local_group: Does group contain this_cpu.
7334 * @sgs: variable to hold the statistics for this group.
7335 * @overload: Indicate more than one runnable task for any CPU.
7337 static inline void update_sg_lb_stats(struct lb_env *env,
7338 struct sched_group *group, int load_idx,
7339 int local_group, struct sg_lb_stats *sgs,
7345 memset(sgs, 0, sizeof(*sgs));
7347 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7348 struct rq *rq = cpu_rq(i);
7350 /* Bias balancing toward cpus of our domain */
7352 load = target_load(i, load_idx);
7354 load = source_load(i, load_idx);
7356 sgs->group_load += load;
7357 sgs->group_util += cpu_util(i);
7358 sgs->sum_nr_running += rq->cfs.h_nr_running;
7360 nr_running = rq->nr_running;
7364 #ifdef CONFIG_NUMA_BALANCING
7365 sgs->nr_numa_running += rq->nr_numa_running;
7366 sgs->nr_preferred_running += rq->nr_preferred_running;
7368 sgs->sum_weighted_load += weighted_cpuload(i);
7370 * No need to call idle_cpu() if nr_running is not 0
7372 if (!nr_running && idle_cpu(i))
7376 /* Adjust by relative CPU capacity of the group */
7377 sgs->group_capacity = group->sgc->capacity;
7378 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7380 if (sgs->sum_nr_running)
7381 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7383 sgs->group_weight = group->group_weight;
7385 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7386 sgs->group_type = group_classify(group, sgs);
7390 * update_sd_pick_busiest - return 1 on busiest group
7391 * @env: The load balancing environment.
7392 * @sds: sched_domain statistics
7393 * @sg: sched_group candidate to be checked for being the busiest
7394 * @sgs: sched_group statistics
7396 * Determine if @sg is a busier group than the previously selected
7399 * Return: %true if @sg is a busier group than the previously selected
7400 * busiest group. %false otherwise.
7402 static bool update_sd_pick_busiest(struct lb_env *env,
7403 struct sd_lb_stats *sds,
7404 struct sched_group *sg,
7405 struct sg_lb_stats *sgs)
7407 struct sg_lb_stats *busiest = &sds->busiest_stat;
7409 if (sgs->group_type > busiest->group_type)
7412 if (sgs->group_type < busiest->group_type)
7415 if (sgs->avg_load <= busiest->avg_load)
7418 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7422 * Candidate sg has no more than one task per CPU and
7423 * has higher per-CPU capacity. Migrating tasks to less
7424 * capable CPUs may harm throughput. Maximize throughput,
7425 * power/energy consequences are not considered.
7427 if (sgs->sum_nr_running <= sgs->group_weight &&
7428 group_smaller_cpu_capacity(sds->local, sg))
7432 /* This is the busiest node in its class. */
7433 if (!(env->sd->flags & SD_ASYM_PACKING))
7436 /* No ASYM_PACKING if target cpu is already busy */
7437 if (env->idle == CPU_NOT_IDLE)
7440 * ASYM_PACKING needs to move all the work to the highest
7441 * prority CPUs in the group, therefore mark all groups
7442 * of lower priority than ourself as busy.
7444 if (sgs->sum_nr_running &&
7445 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7449 /* Prefer to move from lowest priority cpu's work */
7450 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7451 sg->asym_prefer_cpu))
7458 #ifdef CONFIG_NUMA_BALANCING
7459 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7461 if (sgs->sum_nr_running > sgs->nr_numa_running)
7463 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7468 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7470 if (rq->nr_running > rq->nr_numa_running)
7472 if (rq->nr_running > rq->nr_preferred_running)
7477 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7482 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7486 #endif /* CONFIG_NUMA_BALANCING */
7489 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7490 * @env: The load balancing environment.
7491 * @sds: variable to hold the statistics for this sched_domain.
7493 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7495 struct sched_domain *child = env->sd->child;
7496 struct sched_group *sg = env->sd->groups;
7497 struct sg_lb_stats *local = &sds->local_stat;
7498 struct sg_lb_stats tmp_sgs;
7499 int load_idx, prefer_sibling = 0;
7500 bool overload = false;
7502 if (child && child->flags & SD_PREFER_SIBLING)
7505 load_idx = get_sd_load_idx(env->sd, env->idle);
7508 struct sg_lb_stats *sgs = &tmp_sgs;
7511 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7516 if (env->idle != CPU_NEWLY_IDLE ||
7517 time_after_eq(jiffies, sg->sgc->next_update))
7518 update_group_capacity(env->sd, env->dst_cpu);
7521 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7528 * In case the child domain prefers tasks go to siblings
7529 * first, lower the sg capacity so that we'll try
7530 * and move all the excess tasks away. We lower the capacity
7531 * of a group only if the local group has the capacity to fit
7532 * these excess tasks. The extra check prevents the case where
7533 * you always pull from the heaviest group when it is already
7534 * under-utilized (possible with a large weight task outweighs
7535 * the tasks on the system).
7537 if (prefer_sibling && sds->local &&
7538 group_has_capacity(env, local) &&
7539 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7540 sgs->group_no_capacity = 1;
7541 sgs->group_type = group_classify(sg, sgs);
7544 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7546 sds->busiest_stat = *sgs;
7550 /* Now, start updating sd_lb_stats */
7551 sds->total_load += sgs->group_load;
7552 sds->total_capacity += sgs->group_capacity;
7555 } while (sg != env->sd->groups);
7557 if (env->sd->flags & SD_NUMA)
7558 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7560 if (!env->sd->parent) {
7561 /* update overload indicator if we are at root domain */
7562 if (env->dst_rq->rd->overload != overload)
7563 env->dst_rq->rd->overload = overload;
7569 * check_asym_packing - Check to see if the group is packed into the
7572 * This is primarily intended to used at the sibling level. Some
7573 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7574 * case of POWER7, it can move to lower SMT modes only when higher
7575 * threads are idle. When in lower SMT modes, the threads will
7576 * perform better since they share less core resources. Hence when we
7577 * have idle threads, we want them to be the higher ones.
7579 * This packing function is run on idle threads. It checks to see if
7580 * the busiest CPU in this domain (core in the P7 case) has a higher
7581 * CPU number than the packing function is being run on. Here we are
7582 * assuming lower CPU number will be equivalent to lower a SMT thread
7585 * Return: 1 when packing is required and a task should be moved to
7586 * this CPU. The amount of the imbalance is returned in *imbalance.
7588 * @env: The load balancing environment.
7589 * @sds: Statistics of the sched_domain which is to be packed
7591 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7595 if (!(env->sd->flags & SD_ASYM_PACKING))
7598 if (env->idle == CPU_NOT_IDLE)
7604 busiest_cpu = sds->busiest->asym_prefer_cpu;
7605 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7608 env->imbalance = DIV_ROUND_CLOSEST(
7609 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7610 SCHED_CAPACITY_SCALE);
7616 * fix_small_imbalance - Calculate the minor imbalance that exists
7617 * amongst the groups of a sched_domain, during
7619 * @env: The load balancing environment.
7620 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7623 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7625 unsigned long tmp, capa_now = 0, capa_move = 0;
7626 unsigned int imbn = 2;
7627 unsigned long scaled_busy_load_per_task;
7628 struct sg_lb_stats *local, *busiest;
7630 local = &sds->local_stat;
7631 busiest = &sds->busiest_stat;
7633 if (!local->sum_nr_running)
7634 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7635 else if (busiest->load_per_task > local->load_per_task)
7638 scaled_busy_load_per_task =
7639 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7640 busiest->group_capacity;
7642 if (busiest->avg_load + scaled_busy_load_per_task >=
7643 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7644 env->imbalance = busiest->load_per_task;
7649 * OK, we don't have enough imbalance to justify moving tasks,
7650 * however we may be able to increase total CPU capacity used by
7654 capa_now += busiest->group_capacity *
7655 min(busiest->load_per_task, busiest->avg_load);
7656 capa_now += local->group_capacity *
7657 min(local->load_per_task, local->avg_load);
7658 capa_now /= SCHED_CAPACITY_SCALE;
7660 /* Amount of load we'd subtract */
7661 if (busiest->avg_load > scaled_busy_load_per_task) {
7662 capa_move += busiest->group_capacity *
7663 min(busiest->load_per_task,
7664 busiest->avg_load - scaled_busy_load_per_task);
7667 /* Amount of load we'd add */
7668 if (busiest->avg_load * busiest->group_capacity <
7669 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7670 tmp = (busiest->avg_load * busiest->group_capacity) /
7671 local->group_capacity;
7673 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7674 local->group_capacity;
7676 capa_move += local->group_capacity *
7677 min(local->load_per_task, local->avg_load + tmp);
7678 capa_move /= SCHED_CAPACITY_SCALE;
7680 /* Move if we gain throughput */
7681 if (capa_move > capa_now)
7682 env->imbalance = busiest->load_per_task;
7686 * calculate_imbalance - Calculate the amount of imbalance present within the
7687 * groups of a given sched_domain during load balance.
7688 * @env: load balance environment
7689 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7691 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7693 unsigned long max_pull, load_above_capacity = ~0UL;
7694 struct sg_lb_stats *local, *busiest;
7696 local = &sds->local_stat;
7697 busiest = &sds->busiest_stat;
7699 if (busiest->group_type == group_imbalanced) {
7701 * In the group_imb case we cannot rely on group-wide averages
7702 * to ensure cpu-load equilibrium, look at wider averages. XXX
7704 busiest->load_per_task =
7705 min(busiest->load_per_task, sds->avg_load);
7709 * Avg load of busiest sg can be less and avg load of local sg can
7710 * be greater than avg load across all sgs of sd because avg load
7711 * factors in sg capacity and sgs with smaller group_type are
7712 * skipped when updating the busiest sg:
7714 if (busiest->avg_load <= sds->avg_load ||
7715 local->avg_load >= sds->avg_load) {
7717 return fix_small_imbalance(env, sds);
7721 * If there aren't any idle cpus, avoid creating some.
7723 if (busiest->group_type == group_overloaded &&
7724 local->group_type == group_overloaded) {
7725 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7726 if (load_above_capacity > busiest->group_capacity) {
7727 load_above_capacity -= busiest->group_capacity;
7728 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7729 load_above_capacity /= busiest->group_capacity;
7731 load_above_capacity = ~0UL;
7735 * We're trying to get all the cpus to the average_load, so we don't
7736 * want to push ourselves above the average load, nor do we wish to
7737 * reduce the max loaded cpu below the average load. At the same time,
7738 * we also don't want to reduce the group load below the group
7739 * capacity. Thus we look for the minimum possible imbalance.
7741 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7743 /* How much load to actually move to equalise the imbalance */
7744 env->imbalance = min(
7745 max_pull * busiest->group_capacity,
7746 (sds->avg_load - local->avg_load) * local->group_capacity
7747 ) / SCHED_CAPACITY_SCALE;
7750 * if *imbalance is less than the average load per runnable task
7751 * there is no guarantee that any tasks will be moved so we'll have
7752 * a think about bumping its value to force at least one task to be
7755 if (env->imbalance < busiest->load_per_task)
7756 return fix_small_imbalance(env, sds);
7759 /******* find_busiest_group() helpers end here *********************/
7762 * find_busiest_group - Returns the busiest group within the sched_domain
7763 * if there is an imbalance.
7765 * Also calculates the amount of weighted load which should be moved
7766 * to restore balance.
7768 * @env: The load balancing environment.
7770 * Return: - The busiest group if imbalance exists.
7772 static struct sched_group *find_busiest_group(struct lb_env *env)
7774 struct sg_lb_stats *local, *busiest;
7775 struct sd_lb_stats sds;
7777 init_sd_lb_stats(&sds);
7780 * Compute the various statistics relavent for load balancing at
7783 update_sd_lb_stats(env, &sds);
7784 local = &sds.local_stat;
7785 busiest = &sds.busiest_stat;
7787 /* ASYM feature bypasses nice load balance check */
7788 if (check_asym_packing(env, &sds))
7791 /* There is no busy sibling group to pull tasks from */
7792 if (!sds.busiest || busiest->sum_nr_running == 0)
7795 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7796 / sds.total_capacity;
7799 * If the busiest group is imbalanced the below checks don't
7800 * work because they assume all things are equal, which typically
7801 * isn't true due to cpus_allowed constraints and the like.
7803 if (busiest->group_type == group_imbalanced)
7806 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7807 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7808 busiest->group_no_capacity)
7812 * If the local group is busier than the selected busiest group
7813 * don't try and pull any tasks.
7815 if (local->avg_load >= busiest->avg_load)
7819 * Don't pull any tasks if this group is already above the domain
7822 if (local->avg_load >= sds.avg_load)
7825 if (env->idle == CPU_IDLE) {
7827 * This cpu is idle. If the busiest group is not overloaded
7828 * and there is no imbalance between this and busiest group
7829 * wrt idle cpus, it is balanced. The imbalance becomes
7830 * significant if the diff is greater than 1 otherwise we
7831 * might end up to just move the imbalance on another group
7833 if ((busiest->group_type != group_overloaded) &&
7834 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7838 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7839 * imbalance_pct to be conservative.
7841 if (100 * busiest->avg_load <=
7842 env->sd->imbalance_pct * local->avg_load)
7847 /* Looks like there is an imbalance. Compute it */
7848 calculate_imbalance(env, &sds);
7857 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7859 static struct rq *find_busiest_queue(struct lb_env *env,
7860 struct sched_group *group)
7862 struct rq *busiest = NULL, *rq;
7863 unsigned long busiest_load = 0, busiest_capacity = 1;
7866 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7867 unsigned long capacity, wl;
7871 rt = fbq_classify_rq(rq);
7874 * We classify groups/runqueues into three groups:
7875 * - regular: there are !numa tasks
7876 * - remote: there are numa tasks that run on the 'wrong' node
7877 * - all: there is no distinction
7879 * In order to avoid migrating ideally placed numa tasks,
7880 * ignore those when there's better options.
7882 * If we ignore the actual busiest queue to migrate another
7883 * task, the next balance pass can still reduce the busiest
7884 * queue by moving tasks around inside the node.
7886 * If we cannot move enough load due to this classification
7887 * the next pass will adjust the group classification and
7888 * allow migration of more tasks.
7890 * Both cases only affect the total convergence complexity.
7892 if (rt > env->fbq_type)
7895 capacity = capacity_of(i);
7897 wl = weighted_cpuload(i);
7900 * When comparing with imbalance, use weighted_cpuload()
7901 * which is not scaled with the cpu capacity.
7904 if (rq->nr_running == 1 && wl > env->imbalance &&
7905 !check_cpu_capacity(rq, env->sd))
7909 * For the load comparisons with the other cpu's, consider
7910 * the weighted_cpuload() scaled with the cpu capacity, so
7911 * that the load can be moved away from the cpu that is
7912 * potentially running at a lower capacity.
7914 * Thus we're looking for max(wl_i / capacity_i), crosswise
7915 * multiplication to rid ourselves of the division works out
7916 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7917 * our previous maximum.
7919 if (wl * busiest_capacity > busiest_load * capacity) {
7921 busiest_capacity = capacity;
7930 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7931 * so long as it is large enough.
7933 #define MAX_PINNED_INTERVAL 512
7935 static int need_active_balance(struct lb_env *env)
7937 struct sched_domain *sd = env->sd;
7939 if (env->idle == CPU_NEWLY_IDLE) {
7942 * ASYM_PACKING needs to force migrate tasks from busy but
7943 * lower priority CPUs in order to pack all tasks in the
7944 * highest priority CPUs.
7946 if ((sd->flags & SD_ASYM_PACKING) &&
7947 sched_asym_prefer(env->dst_cpu, env->src_cpu))
7952 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7953 * It's worth migrating the task if the src_cpu's capacity is reduced
7954 * because of other sched_class or IRQs if more capacity stays
7955 * available on dst_cpu.
7957 if ((env->idle != CPU_NOT_IDLE) &&
7958 (env->src_rq->cfs.h_nr_running == 1)) {
7959 if ((check_cpu_capacity(env->src_rq, sd)) &&
7960 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7964 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7967 static int active_load_balance_cpu_stop(void *data);
7969 static int should_we_balance(struct lb_env *env)
7971 struct sched_group *sg = env->sd->groups;
7972 int cpu, balance_cpu = -1;
7975 * In the newly idle case, we will allow all the cpu's
7976 * to do the newly idle load balance.
7978 if (env->idle == CPU_NEWLY_IDLE)
7981 /* Try to find first idle cpu */
7982 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
7990 if (balance_cpu == -1)
7991 balance_cpu = group_balance_cpu(sg);
7994 * First idle cpu or the first cpu(busiest) in this sched group
7995 * is eligible for doing load balancing at this and above domains.
7997 return balance_cpu == env->dst_cpu;
8001 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8002 * tasks if there is an imbalance.
8004 static int load_balance(int this_cpu, struct rq *this_rq,
8005 struct sched_domain *sd, enum cpu_idle_type idle,
8006 int *continue_balancing)
8008 int ld_moved, cur_ld_moved, active_balance = 0;
8009 struct sched_domain *sd_parent = sd->parent;
8010 struct sched_group *group;
8013 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8015 struct lb_env env = {
8017 .dst_cpu = this_cpu,
8019 .dst_grpmask = sched_group_span(sd->groups),
8021 .loop_break = sched_nr_migrate_break,
8024 .tasks = LIST_HEAD_INIT(env.tasks),
8027 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8029 schedstat_inc(sd->lb_count[idle]);
8032 if (!should_we_balance(&env)) {
8033 *continue_balancing = 0;
8037 group = find_busiest_group(&env);
8039 schedstat_inc(sd->lb_nobusyg[idle]);
8043 busiest = find_busiest_queue(&env, group);
8045 schedstat_inc(sd->lb_nobusyq[idle]);
8049 BUG_ON(busiest == env.dst_rq);
8051 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8053 env.src_cpu = busiest->cpu;
8054 env.src_rq = busiest;
8057 if (busiest->nr_running > 1) {
8059 * Attempt to move tasks. If find_busiest_group has found
8060 * an imbalance but busiest->nr_running <= 1, the group is
8061 * still unbalanced. ld_moved simply stays zero, so it is
8062 * correctly treated as an imbalance.
8064 env.flags |= LBF_ALL_PINNED;
8065 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8068 rq_lock_irqsave(busiest, &rf);
8069 update_rq_clock(busiest);
8072 * cur_ld_moved - load moved in current iteration
8073 * ld_moved - cumulative load moved across iterations
8075 cur_ld_moved = detach_tasks(&env);
8078 * We've detached some tasks from busiest_rq. Every
8079 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8080 * unlock busiest->lock, and we are able to be sure
8081 * that nobody can manipulate the tasks in parallel.
8082 * See task_rq_lock() family for the details.
8085 rq_unlock(busiest, &rf);
8089 ld_moved += cur_ld_moved;
8092 local_irq_restore(rf.flags);
8094 if (env.flags & LBF_NEED_BREAK) {
8095 env.flags &= ~LBF_NEED_BREAK;
8100 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8101 * us and move them to an alternate dst_cpu in our sched_group
8102 * where they can run. The upper limit on how many times we
8103 * iterate on same src_cpu is dependent on number of cpus in our
8106 * This changes load balance semantics a bit on who can move
8107 * load to a given_cpu. In addition to the given_cpu itself
8108 * (or a ilb_cpu acting on its behalf where given_cpu is
8109 * nohz-idle), we now have balance_cpu in a position to move
8110 * load to given_cpu. In rare situations, this may cause
8111 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8112 * _independently_ and at _same_ time to move some load to
8113 * given_cpu) causing exceess load to be moved to given_cpu.
8114 * This however should not happen so much in practice and
8115 * moreover subsequent load balance cycles should correct the
8116 * excess load moved.
8118 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8120 /* Prevent to re-select dst_cpu via env's cpus */
8121 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8123 env.dst_rq = cpu_rq(env.new_dst_cpu);
8124 env.dst_cpu = env.new_dst_cpu;
8125 env.flags &= ~LBF_DST_PINNED;
8127 env.loop_break = sched_nr_migrate_break;
8130 * Go back to "more_balance" rather than "redo" since we
8131 * need to continue with same src_cpu.
8137 * We failed to reach balance because of affinity.
8140 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8142 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8143 *group_imbalance = 1;
8146 /* All tasks on this runqueue were pinned by CPU affinity */
8147 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8148 cpumask_clear_cpu(cpu_of(busiest), cpus);
8150 * Attempting to continue load balancing at the current
8151 * sched_domain level only makes sense if there are
8152 * active CPUs remaining as possible busiest CPUs to
8153 * pull load from which are not contained within the
8154 * destination group that is receiving any migrated
8157 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8159 env.loop_break = sched_nr_migrate_break;
8162 goto out_all_pinned;
8167 schedstat_inc(sd->lb_failed[idle]);
8169 * Increment the failure counter only on periodic balance.
8170 * We do not want newidle balance, which can be very
8171 * frequent, pollute the failure counter causing
8172 * excessive cache_hot migrations and active balances.
8174 if (idle != CPU_NEWLY_IDLE)
8175 sd->nr_balance_failed++;
8177 if (need_active_balance(&env)) {
8178 unsigned long flags;
8180 raw_spin_lock_irqsave(&busiest->lock, flags);
8182 /* don't kick the active_load_balance_cpu_stop,
8183 * if the curr task on busiest cpu can't be
8186 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8187 raw_spin_unlock_irqrestore(&busiest->lock,
8189 env.flags |= LBF_ALL_PINNED;
8190 goto out_one_pinned;
8194 * ->active_balance synchronizes accesses to
8195 * ->active_balance_work. Once set, it's cleared
8196 * only after active load balance is finished.
8198 if (!busiest->active_balance) {
8199 busiest->active_balance = 1;
8200 busiest->push_cpu = this_cpu;
8203 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8205 if (active_balance) {
8206 stop_one_cpu_nowait(cpu_of(busiest),
8207 active_load_balance_cpu_stop, busiest,
8208 &busiest->active_balance_work);
8211 /* We've kicked active balancing, force task migration. */
8212 sd->nr_balance_failed = sd->cache_nice_tries+1;
8215 sd->nr_balance_failed = 0;
8217 if (likely(!active_balance)) {
8218 /* We were unbalanced, so reset the balancing interval */
8219 sd->balance_interval = sd->min_interval;
8222 * If we've begun active balancing, start to back off. This
8223 * case may not be covered by the all_pinned logic if there
8224 * is only 1 task on the busy runqueue (because we don't call
8227 if (sd->balance_interval < sd->max_interval)
8228 sd->balance_interval *= 2;
8235 * We reach balance although we may have faced some affinity
8236 * constraints. Clear the imbalance flag if it was set.
8239 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8241 if (*group_imbalance)
8242 *group_imbalance = 0;
8247 * We reach balance because all tasks are pinned at this level so
8248 * we can't migrate them. Let the imbalance flag set so parent level
8249 * can try to migrate them.
8251 schedstat_inc(sd->lb_balanced[idle]);
8253 sd->nr_balance_failed = 0;
8256 /* tune up the balancing interval */
8257 if (((env.flags & LBF_ALL_PINNED) &&
8258 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8259 (sd->balance_interval < sd->max_interval))
8260 sd->balance_interval *= 2;
8267 static inline unsigned long
8268 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8270 unsigned long interval = sd->balance_interval;
8273 interval *= sd->busy_factor;
8275 /* scale ms to jiffies */
8276 interval = msecs_to_jiffies(interval);
8277 interval = clamp(interval, 1UL, max_load_balance_interval);
8283 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8285 unsigned long interval, next;
8287 /* used by idle balance, so cpu_busy = 0 */
8288 interval = get_sd_balance_interval(sd, 0);
8289 next = sd->last_balance + interval;
8291 if (time_after(*next_balance, next))
8292 *next_balance = next;
8296 * idle_balance is called by schedule() if this_cpu is about to become
8297 * idle. Attempts to pull tasks from other CPUs.
8299 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8301 unsigned long next_balance = jiffies + HZ;
8302 int this_cpu = this_rq->cpu;
8303 struct sched_domain *sd;
8304 int pulled_task = 0;
8308 * We must set idle_stamp _before_ calling idle_balance(), such that we
8309 * measure the duration of idle_balance() as idle time.
8311 this_rq->idle_stamp = rq_clock(this_rq);
8314 * This is OK, because current is on_cpu, which avoids it being picked
8315 * for load-balance and preemption/IRQs are still disabled avoiding
8316 * further scheduler activity on it and we're being very careful to
8317 * re-start the picking loop.
8319 rq_unpin_lock(this_rq, rf);
8321 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8322 !this_rq->rd->overload) {
8324 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8326 update_next_balance(sd, &next_balance);
8332 raw_spin_unlock(&this_rq->lock);
8334 update_blocked_averages(this_cpu);
8336 for_each_domain(this_cpu, sd) {
8337 int continue_balancing = 1;
8338 u64 t0, domain_cost;
8340 if (!(sd->flags & SD_LOAD_BALANCE))
8343 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8344 update_next_balance(sd, &next_balance);
8348 if (sd->flags & SD_BALANCE_NEWIDLE) {
8349 t0 = sched_clock_cpu(this_cpu);
8351 pulled_task = load_balance(this_cpu, this_rq,
8353 &continue_balancing);
8355 domain_cost = sched_clock_cpu(this_cpu) - t0;
8356 if (domain_cost > sd->max_newidle_lb_cost)
8357 sd->max_newidle_lb_cost = domain_cost;
8359 curr_cost += domain_cost;
8362 update_next_balance(sd, &next_balance);
8365 * Stop searching for tasks to pull if there are
8366 * now runnable tasks on this rq.
8368 if (pulled_task || this_rq->nr_running > 0)
8373 raw_spin_lock(&this_rq->lock);
8375 if (curr_cost > this_rq->max_idle_balance_cost)
8376 this_rq->max_idle_balance_cost = curr_cost;
8379 * While browsing the domains, we released the rq lock, a task could
8380 * have been enqueued in the meantime. Since we're not going idle,
8381 * pretend we pulled a task.
8383 if (this_rq->cfs.h_nr_running && !pulled_task)
8387 /* Move the next balance forward */
8388 if (time_after(this_rq->next_balance, next_balance))
8389 this_rq->next_balance = next_balance;
8391 /* Is there a task of a high priority class? */
8392 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8396 this_rq->idle_stamp = 0;
8398 rq_repin_lock(this_rq, rf);
8404 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8405 * running tasks off the busiest CPU onto idle CPUs. It requires at
8406 * least 1 task to be running on each physical CPU where possible, and
8407 * avoids physical / logical imbalances.
8409 static int active_load_balance_cpu_stop(void *data)
8411 struct rq *busiest_rq = data;
8412 int busiest_cpu = cpu_of(busiest_rq);
8413 int target_cpu = busiest_rq->push_cpu;
8414 struct rq *target_rq = cpu_rq(target_cpu);
8415 struct sched_domain *sd;
8416 struct task_struct *p = NULL;
8419 rq_lock_irq(busiest_rq, &rf);
8421 /* make sure the requested cpu hasn't gone down in the meantime */
8422 if (unlikely(busiest_cpu != smp_processor_id() ||
8423 !busiest_rq->active_balance))
8426 /* Is there any task to move? */
8427 if (busiest_rq->nr_running <= 1)
8431 * This condition is "impossible", if it occurs
8432 * we need to fix it. Originally reported by
8433 * Bjorn Helgaas on a 128-cpu setup.
8435 BUG_ON(busiest_rq == target_rq);
8437 /* Search for an sd spanning us and the target CPU. */
8439 for_each_domain(target_cpu, sd) {
8440 if ((sd->flags & SD_LOAD_BALANCE) &&
8441 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8446 struct lb_env env = {
8448 .dst_cpu = target_cpu,
8449 .dst_rq = target_rq,
8450 .src_cpu = busiest_rq->cpu,
8451 .src_rq = busiest_rq,
8454 * can_migrate_task() doesn't need to compute new_dst_cpu
8455 * for active balancing. Since we have CPU_IDLE, but no
8456 * @dst_grpmask we need to make that test go away with lying
8459 .flags = LBF_DST_PINNED,
8462 schedstat_inc(sd->alb_count);
8463 update_rq_clock(busiest_rq);
8465 p = detach_one_task(&env);
8467 schedstat_inc(sd->alb_pushed);
8468 /* Active balancing done, reset the failure counter. */
8469 sd->nr_balance_failed = 0;
8471 schedstat_inc(sd->alb_failed);
8476 busiest_rq->active_balance = 0;
8477 rq_unlock(busiest_rq, &rf);
8480 attach_one_task(target_rq, p);
8487 static inline int on_null_domain(struct rq *rq)
8489 return unlikely(!rcu_dereference_sched(rq->sd));
8492 #ifdef CONFIG_NO_HZ_COMMON
8494 * idle load balancing details
8495 * - When one of the busy CPUs notice that there may be an idle rebalancing
8496 * needed, they will kick the idle load balancer, which then does idle
8497 * load balancing for all the idle CPUs.
8500 cpumask_var_t idle_cpus_mask;
8502 unsigned long next_balance; /* in jiffy units */
8503 } nohz ____cacheline_aligned;
8505 static inline int find_new_ilb(void)
8507 int ilb = cpumask_first(nohz.idle_cpus_mask);
8509 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8516 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8517 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8518 * CPU (if there is one).
8520 static void nohz_balancer_kick(void)
8524 nohz.next_balance++;
8526 ilb_cpu = find_new_ilb();
8528 if (ilb_cpu >= nr_cpu_ids)
8531 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8534 * Use smp_send_reschedule() instead of resched_cpu().
8535 * This way we generate a sched IPI on the target cpu which
8536 * is idle. And the softirq performing nohz idle load balance
8537 * will be run before returning from the IPI.
8539 smp_send_reschedule(ilb_cpu);
8543 void nohz_balance_exit_idle(unsigned int cpu)
8545 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8547 * Completely isolated CPUs don't ever set, so we must test.
8549 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8550 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8551 atomic_dec(&nohz.nr_cpus);
8553 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8557 static inline void set_cpu_sd_state_busy(void)
8559 struct sched_domain *sd;
8560 int cpu = smp_processor_id();
8563 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8565 if (!sd || !sd->nohz_idle)
8569 atomic_inc(&sd->shared->nr_busy_cpus);
8574 void set_cpu_sd_state_idle(void)
8576 struct sched_domain *sd;
8577 int cpu = smp_processor_id();
8580 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8582 if (!sd || sd->nohz_idle)
8586 atomic_dec(&sd->shared->nr_busy_cpus);
8592 * This routine will record that the cpu is going idle with tick stopped.
8593 * This info will be used in performing idle load balancing in the future.
8595 void nohz_balance_enter_idle(int cpu)
8598 * If this cpu is going down, then nothing needs to be done.
8600 if (!cpu_active(cpu))
8603 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8604 if (!is_housekeeping_cpu(cpu))
8607 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8611 * If we're a completely isolated CPU, we don't play.
8613 if (on_null_domain(cpu_rq(cpu)))
8616 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8617 atomic_inc(&nohz.nr_cpus);
8618 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8622 static DEFINE_SPINLOCK(balancing);
8625 * Scale the max load_balance interval with the number of CPUs in the system.
8626 * This trades load-balance latency on larger machines for less cross talk.
8628 void update_max_interval(void)
8630 max_load_balance_interval = HZ*num_online_cpus()/10;
8634 * It checks each scheduling domain to see if it is due to be balanced,
8635 * and initiates a balancing operation if so.
8637 * Balancing parameters are set up in init_sched_domains.
8639 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8641 int continue_balancing = 1;
8643 unsigned long interval;
8644 struct sched_domain *sd;
8645 /* Earliest time when we have to do rebalance again */
8646 unsigned long next_balance = jiffies + 60*HZ;
8647 int update_next_balance = 0;
8648 int need_serialize, need_decay = 0;
8651 update_blocked_averages(cpu);
8654 for_each_domain(cpu, sd) {
8656 * Decay the newidle max times here because this is a regular
8657 * visit to all the domains. Decay ~1% per second.
8659 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8660 sd->max_newidle_lb_cost =
8661 (sd->max_newidle_lb_cost * 253) / 256;
8662 sd->next_decay_max_lb_cost = jiffies + HZ;
8665 max_cost += sd->max_newidle_lb_cost;
8667 if (!(sd->flags & SD_LOAD_BALANCE))
8671 * Stop the load balance at this level. There is another
8672 * CPU in our sched group which is doing load balancing more
8675 if (!continue_balancing) {
8681 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8683 need_serialize = sd->flags & SD_SERIALIZE;
8684 if (need_serialize) {
8685 if (!spin_trylock(&balancing))
8689 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8690 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8692 * The LBF_DST_PINNED logic could have changed
8693 * env->dst_cpu, so we can't know our idle
8694 * state even if we migrated tasks. Update it.
8696 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8698 sd->last_balance = jiffies;
8699 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8702 spin_unlock(&balancing);
8704 if (time_after(next_balance, sd->last_balance + interval)) {
8705 next_balance = sd->last_balance + interval;
8706 update_next_balance = 1;
8711 * Ensure the rq-wide value also decays but keep it at a
8712 * reasonable floor to avoid funnies with rq->avg_idle.
8714 rq->max_idle_balance_cost =
8715 max((u64)sysctl_sched_migration_cost, max_cost);
8720 * next_balance will be updated only when there is a need.
8721 * When the cpu is attached to null domain for ex, it will not be
8724 if (likely(update_next_balance)) {
8725 rq->next_balance = next_balance;
8727 #ifdef CONFIG_NO_HZ_COMMON
8729 * If this CPU has been elected to perform the nohz idle
8730 * balance. Other idle CPUs have already rebalanced with
8731 * nohz_idle_balance() and nohz.next_balance has been
8732 * updated accordingly. This CPU is now running the idle load
8733 * balance for itself and we need to update the
8734 * nohz.next_balance accordingly.
8736 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8737 nohz.next_balance = rq->next_balance;
8742 #ifdef CONFIG_NO_HZ_COMMON
8744 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8745 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8747 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8749 int this_cpu = this_rq->cpu;
8752 /* Earliest time when we have to do rebalance again */
8753 unsigned long next_balance = jiffies + 60*HZ;
8754 int update_next_balance = 0;
8756 if (idle != CPU_IDLE ||
8757 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8760 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8761 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8765 * If this cpu gets work to do, stop the load balancing
8766 * work being done for other cpus. Next load
8767 * balancing owner will pick it up.
8772 rq = cpu_rq(balance_cpu);
8775 * If time for next balance is due,
8778 if (time_after_eq(jiffies, rq->next_balance)) {
8781 rq_lock_irq(rq, &rf);
8782 update_rq_clock(rq);
8783 cpu_load_update_idle(rq);
8784 rq_unlock_irq(rq, &rf);
8786 rebalance_domains(rq, CPU_IDLE);
8789 if (time_after(next_balance, rq->next_balance)) {
8790 next_balance = rq->next_balance;
8791 update_next_balance = 1;
8796 * next_balance will be updated only when there is a need.
8797 * When the CPU is attached to null domain for ex, it will not be
8800 if (likely(update_next_balance))
8801 nohz.next_balance = next_balance;
8803 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8807 * Current heuristic for kicking the idle load balancer in the presence
8808 * of an idle cpu in the system.
8809 * - This rq has more than one task.
8810 * - This rq has at least one CFS task and the capacity of the CPU is
8811 * significantly reduced because of RT tasks or IRQs.
8812 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8813 * multiple busy cpu.
8814 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8815 * domain span are idle.
8817 static inline bool nohz_kick_needed(struct rq *rq)
8819 unsigned long now = jiffies;
8820 struct sched_domain_shared *sds;
8821 struct sched_domain *sd;
8822 int nr_busy, i, cpu = rq->cpu;
8825 if (unlikely(rq->idle_balance))
8829 * We may be recently in ticked or tickless idle mode. At the first
8830 * busy tick after returning from idle, we will update the busy stats.
8832 set_cpu_sd_state_busy();
8833 nohz_balance_exit_idle(cpu);
8836 * None are in tickless mode and hence no need for NOHZ idle load
8839 if (likely(!atomic_read(&nohz.nr_cpus)))
8842 if (time_before(now, nohz.next_balance))
8845 if (rq->nr_running >= 2)
8849 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8852 * XXX: write a coherent comment on why we do this.
8853 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8855 nr_busy = atomic_read(&sds->nr_busy_cpus);
8863 sd = rcu_dereference(rq->sd);
8865 if ((rq->cfs.h_nr_running >= 1) &&
8866 check_cpu_capacity(rq, sd)) {
8872 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8874 for_each_cpu(i, sched_domain_span(sd)) {
8876 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
8879 if (sched_asym_prefer(i, cpu)) {
8890 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8894 * run_rebalance_domains is triggered when needed from the scheduler tick.
8895 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8897 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
8899 struct rq *this_rq = this_rq();
8900 enum cpu_idle_type idle = this_rq->idle_balance ?
8901 CPU_IDLE : CPU_NOT_IDLE;
8904 * If this cpu has a pending nohz_balance_kick, then do the
8905 * balancing on behalf of the other idle cpus whose ticks are
8906 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8907 * give the idle cpus a chance to load balance. Else we may
8908 * load balance only within the local sched_domain hierarchy
8909 * and abort nohz_idle_balance altogether if we pull some load.
8911 nohz_idle_balance(this_rq, idle);
8912 rebalance_domains(this_rq, idle);
8916 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8918 void trigger_load_balance(struct rq *rq)
8920 /* Don't need to rebalance while attached to NULL domain */
8921 if (unlikely(on_null_domain(rq)))
8924 if (time_after_eq(jiffies, rq->next_balance))
8925 raise_softirq(SCHED_SOFTIRQ);
8926 #ifdef CONFIG_NO_HZ_COMMON
8927 if (nohz_kick_needed(rq))
8928 nohz_balancer_kick();
8932 static void rq_online_fair(struct rq *rq)
8936 update_runtime_enabled(rq);
8939 static void rq_offline_fair(struct rq *rq)
8943 /* Ensure any throttled groups are reachable by pick_next_task */
8944 unthrottle_offline_cfs_rqs(rq);
8947 #endif /* CONFIG_SMP */
8950 * scheduler tick hitting a task of our scheduling class:
8952 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8954 struct cfs_rq *cfs_rq;
8955 struct sched_entity *se = &curr->se;
8957 for_each_sched_entity(se) {
8958 cfs_rq = cfs_rq_of(se);
8959 entity_tick(cfs_rq, se, queued);
8962 if (static_branch_unlikely(&sched_numa_balancing))
8963 task_tick_numa(rq, curr);
8967 * called on fork with the child task as argument from the parent's context
8968 * - child not yet on the tasklist
8969 * - preemption disabled
8971 static void task_fork_fair(struct task_struct *p)
8973 struct cfs_rq *cfs_rq;
8974 struct sched_entity *se = &p->se, *curr;
8975 struct rq *rq = this_rq();
8979 update_rq_clock(rq);
8981 cfs_rq = task_cfs_rq(current);
8982 curr = cfs_rq->curr;
8984 update_curr(cfs_rq);
8985 se->vruntime = curr->vruntime;
8987 place_entity(cfs_rq, se, 1);
8989 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8991 * Upon rescheduling, sched_class::put_prev_task() will place
8992 * 'current' within the tree based on its new key value.
8994 swap(curr->vruntime, se->vruntime);
8998 se->vruntime -= cfs_rq->min_vruntime;
9003 * Priority of the task has changed. Check to see if we preempt
9007 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9009 if (!task_on_rq_queued(p))
9013 * Reschedule if we are currently running on this runqueue and
9014 * our priority decreased, or if we are not currently running on
9015 * this runqueue and our priority is higher than the current's
9017 if (rq->curr == p) {
9018 if (p->prio > oldprio)
9021 check_preempt_curr(rq, p, 0);
9024 static inline bool vruntime_normalized(struct task_struct *p)
9026 struct sched_entity *se = &p->se;
9029 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9030 * the dequeue_entity(.flags=0) will already have normalized the
9037 * When !on_rq, vruntime of the task has usually NOT been normalized.
9038 * But there are some cases where it has already been normalized:
9040 * - A forked child which is waiting for being woken up by
9041 * wake_up_new_task().
9042 * - A task which has been woken up by try_to_wake_up() and
9043 * waiting for actually being woken up by sched_ttwu_pending().
9045 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9051 #ifdef CONFIG_FAIR_GROUP_SCHED
9053 * Propagate the changes of the sched_entity across the tg tree to make it
9054 * visible to the root
9056 static void propagate_entity_cfs_rq(struct sched_entity *se)
9058 struct cfs_rq *cfs_rq;
9060 /* Start to propagate at parent */
9063 for_each_sched_entity(se) {
9064 cfs_rq = cfs_rq_of(se);
9066 if (cfs_rq_throttled(cfs_rq))
9069 update_load_avg(se, UPDATE_TG);
9073 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9076 static void detach_entity_cfs_rq(struct sched_entity *se)
9078 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9080 /* Catch up with the cfs_rq and remove our load when we leave */
9081 update_load_avg(se, 0);
9082 detach_entity_load_avg(cfs_rq, se);
9083 update_tg_load_avg(cfs_rq, false);
9084 propagate_entity_cfs_rq(se);
9087 static void attach_entity_cfs_rq(struct sched_entity *se)
9089 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9091 #ifdef CONFIG_FAIR_GROUP_SCHED
9093 * Since the real-depth could have been changed (only FAIR
9094 * class maintain depth value), reset depth properly.
9096 se->depth = se->parent ? se->parent->depth + 1 : 0;
9099 /* Synchronize entity with its cfs_rq */
9100 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9101 attach_entity_load_avg(cfs_rq, se);
9102 update_tg_load_avg(cfs_rq, false);
9103 propagate_entity_cfs_rq(se);
9106 static void detach_task_cfs_rq(struct task_struct *p)
9108 struct sched_entity *se = &p->se;
9109 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9111 if (!vruntime_normalized(p)) {
9113 * Fix up our vruntime so that the current sleep doesn't
9114 * cause 'unlimited' sleep bonus.
9116 place_entity(cfs_rq, se, 0);
9117 se->vruntime -= cfs_rq->min_vruntime;
9120 detach_entity_cfs_rq(se);
9123 static void attach_task_cfs_rq(struct task_struct *p)
9125 struct sched_entity *se = &p->se;
9126 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9128 attach_entity_cfs_rq(se);
9130 if (!vruntime_normalized(p))
9131 se->vruntime += cfs_rq->min_vruntime;
9134 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9136 detach_task_cfs_rq(p);
9139 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9141 attach_task_cfs_rq(p);
9143 if (task_on_rq_queued(p)) {
9145 * We were most likely switched from sched_rt, so
9146 * kick off the schedule if running, otherwise just see
9147 * if we can still preempt the current task.
9152 check_preempt_curr(rq, p, 0);
9156 /* Account for a task changing its policy or group.
9158 * This routine is mostly called to set cfs_rq->curr field when a task
9159 * migrates between groups/classes.
9161 static void set_curr_task_fair(struct rq *rq)
9163 struct sched_entity *se = &rq->curr->se;
9165 for_each_sched_entity(se) {
9166 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9168 set_next_entity(cfs_rq, se);
9169 /* ensure bandwidth has been allocated on our new cfs_rq */
9170 account_cfs_rq_runtime(cfs_rq, 0);
9174 void init_cfs_rq(struct cfs_rq *cfs_rq)
9176 cfs_rq->tasks_timeline = RB_ROOT;
9177 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9178 #ifndef CONFIG_64BIT
9179 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9182 #ifdef CONFIG_FAIR_GROUP_SCHED
9183 cfs_rq->propagate_avg = 0;
9185 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9186 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9191 static void task_set_group_fair(struct task_struct *p)
9193 struct sched_entity *se = &p->se;
9195 set_task_rq(p, task_cpu(p));
9196 se->depth = se->parent ? se->parent->depth + 1 : 0;
9199 static void task_move_group_fair(struct task_struct *p)
9201 detach_task_cfs_rq(p);
9202 set_task_rq(p, task_cpu(p));
9205 /* Tell se's cfs_rq has been changed -- migrated */
9206 p->se.avg.last_update_time = 0;
9208 attach_task_cfs_rq(p);
9211 static void task_change_group_fair(struct task_struct *p, int type)
9214 case TASK_SET_GROUP:
9215 task_set_group_fair(p);
9218 case TASK_MOVE_GROUP:
9219 task_move_group_fair(p);
9224 void free_fair_sched_group(struct task_group *tg)
9228 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9230 for_each_possible_cpu(i) {
9232 kfree(tg->cfs_rq[i]);
9241 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9243 struct sched_entity *se;
9244 struct cfs_rq *cfs_rq;
9247 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9250 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9254 tg->shares = NICE_0_LOAD;
9256 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9258 for_each_possible_cpu(i) {
9259 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9260 GFP_KERNEL, cpu_to_node(i));
9264 se = kzalloc_node(sizeof(struct sched_entity),
9265 GFP_KERNEL, cpu_to_node(i));
9269 init_cfs_rq(cfs_rq);
9270 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9271 init_entity_runnable_average(se);
9282 void online_fair_sched_group(struct task_group *tg)
9284 struct sched_entity *se;
9288 for_each_possible_cpu(i) {
9292 raw_spin_lock_irq(&rq->lock);
9293 update_rq_clock(rq);
9294 attach_entity_cfs_rq(se);
9295 sync_throttle(tg, i);
9296 raw_spin_unlock_irq(&rq->lock);
9300 void unregister_fair_sched_group(struct task_group *tg)
9302 unsigned long flags;
9306 for_each_possible_cpu(cpu) {
9308 remove_entity_load_avg(tg->se[cpu]);
9311 * Only empty task groups can be destroyed; so we can speculatively
9312 * check on_list without danger of it being re-added.
9314 if (!tg->cfs_rq[cpu]->on_list)
9319 raw_spin_lock_irqsave(&rq->lock, flags);
9320 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9321 raw_spin_unlock_irqrestore(&rq->lock, flags);
9325 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9326 struct sched_entity *se, int cpu,
9327 struct sched_entity *parent)
9329 struct rq *rq = cpu_rq(cpu);
9333 init_cfs_rq_runtime(cfs_rq);
9335 tg->cfs_rq[cpu] = cfs_rq;
9338 /* se could be NULL for root_task_group */
9343 se->cfs_rq = &rq->cfs;
9346 se->cfs_rq = parent->my_q;
9347 se->depth = parent->depth + 1;
9351 /* guarantee group entities always have weight */
9352 update_load_set(&se->load, NICE_0_LOAD);
9353 se->parent = parent;
9356 static DEFINE_MUTEX(shares_mutex);
9358 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9363 * We can't change the weight of the root cgroup.
9368 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9370 mutex_lock(&shares_mutex);
9371 if (tg->shares == shares)
9374 tg->shares = shares;
9375 for_each_possible_cpu(i) {
9376 struct rq *rq = cpu_rq(i);
9377 struct sched_entity *se = tg->se[i];
9380 /* Propagate contribution to hierarchy */
9381 rq_lock_irqsave(rq, &rf);
9382 update_rq_clock(rq);
9383 for_each_sched_entity(se) {
9384 update_load_avg(se, UPDATE_TG);
9385 update_cfs_shares(se);
9387 rq_unlock_irqrestore(rq, &rf);
9391 mutex_unlock(&shares_mutex);
9394 #else /* CONFIG_FAIR_GROUP_SCHED */
9396 void free_fair_sched_group(struct task_group *tg) { }
9398 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9403 void online_fair_sched_group(struct task_group *tg) { }
9405 void unregister_fair_sched_group(struct task_group *tg) { }
9407 #endif /* CONFIG_FAIR_GROUP_SCHED */
9410 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9412 struct sched_entity *se = &task->se;
9413 unsigned int rr_interval = 0;
9416 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9419 if (rq->cfs.load.weight)
9420 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9426 * All the scheduling class methods:
9428 const struct sched_class fair_sched_class = {
9429 .next = &idle_sched_class,
9430 .enqueue_task = enqueue_task_fair,
9431 .dequeue_task = dequeue_task_fair,
9432 .yield_task = yield_task_fair,
9433 .yield_to_task = yield_to_task_fair,
9435 .check_preempt_curr = check_preempt_wakeup,
9437 .pick_next_task = pick_next_task_fair,
9438 .put_prev_task = put_prev_task_fair,
9441 .select_task_rq = select_task_rq_fair,
9442 .migrate_task_rq = migrate_task_rq_fair,
9444 .rq_online = rq_online_fair,
9445 .rq_offline = rq_offline_fair,
9447 .task_dead = task_dead_fair,
9448 .set_cpus_allowed = set_cpus_allowed_common,
9451 .set_curr_task = set_curr_task_fair,
9452 .task_tick = task_tick_fair,
9453 .task_fork = task_fork_fair,
9455 .prio_changed = prio_changed_fair,
9456 .switched_from = switched_from_fair,
9457 .switched_to = switched_to_fair,
9459 .get_rr_interval = get_rr_interval_fair,
9461 .update_curr = update_curr_fair,
9463 #ifdef CONFIG_FAIR_GROUP_SCHED
9464 .task_change_group = task_change_group_fair,
9468 #ifdef CONFIG_SCHED_DEBUG
9469 void print_cfs_stats(struct seq_file *m, int cpu)
9471 struct cfs_rq *cfs_rq, *pos;
9474 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9475 print_cfs_rq(m, cpu, cfs_rq);
9479 #ifdef CONFIG_NUMA_BALANCING
9480 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9483 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9485 for_each_online_node(node) {
9486 if (p->numa_faults) {
9487 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9488 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9490 if (p->numa_group) {
9491 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9492 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9494 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9497 #endif /* CONFIG_NUMA_BALANCING */
9498 #endif /* CONFIG_SCHED_DEBUG */
9500 __init void init_sched_fair_class(void)
9503 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9505 #ifdef CONFIG_NO_HZ_COMMON
9506 nohz.next_balance = jiffies;
9507 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);