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/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
33 #include <linux/module.h>
35 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
56 unsigned int sysctl_sched_sync_hint_enable = 1;
57 unsigned int sysctl_sched_initial_task_util = 0;
58 unsigned int sysctl_sched_cstate_aware = 1;
60 #ifdef CONFIG_SCHED_WALT
61 unsigned int sysctl_sched_use_walt_cpu_util = 1;
62 unsigned int sysctl_sched_use_walt_task_util = 1;
63 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
67 * The initial- and re-scaling of tunables is configurable
68 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
72 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
73 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75 enum sched_tunable_scaling sysctl_sched_tunable_scaling
76 = SCHED_TUNABLESCALING_LOG;
79 * Minimal preemption granularity for CPU-bound tasks:
80 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_min_granularity = 750000ULL;
83 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
86 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
88 static unsigned int sched_nr_latency = 8;
91 * After fork, child runs first. If set to 0 (default) then
92 * parent will (try to) run first.
94 unsigned int sysctl_sched_child_runs_first __read_mostly;
97 * SCHED_OTHER wake-up granularity.
98 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
100 * This option delays the preemption effects of decoupled workloads
101 * and reduces their over-scheduling. Synchronous workloads will still
102 * have immediate wakeup/sleep latencies.
104 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
105 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
107 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
110 * The exponential sliding window over which load is averaged for shares
114 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
116 #ifdef CONFIG_CFS_BANDWIDTH
118 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
119 * each time a cfs_rq requests quota.
121 * Note: in the case that the slice exceeds the runtime remaining (either due
122 * to consumption or the quota being specified to be smaller than the slice)
123 * we will always only issue the remaining available time.
125 * default: 5 msec, units: microseconds
127 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 * The margin used when comparing utilization with CPU capacity:
132 * util * margin < capacity * 1024
134 unsigned int capacity_margin = 1280; /* ~20% */
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
172 case SCHED_TUNABLESCALING_LINEAR:
175 case SCHED_TUNABLESCALING_LOG:
177 factor = 1 + ilog2(cpus);
184 static void update_sysctl(void)
186 unsigned int factor = get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity);
191 SET_SYSCTL(sched_latency);
192 SET_SYSCTL(sched_wakeup_granularity);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
208 if (likely(lw->inv_weight))
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
235 u64 fact = scale_load_down(weight);
236 int shift = WMULT_SHIFT;
238 __update_inv_weight(lw);
240 if (unlikely(fact >> 32)) {
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
255 return mul_u64_u32_shr(delta_exec, fact, shift);
259 const struct sched_class fair_sched_class;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct *task_of(struct sched_entity *se)
278 #ifdef CONFIG_SCHED_DEBUG
279 WARN_ON_ONCE(!entity_is_task(se));
281 return container_of(se, struct task_struct, se);
284 /* Walk up scheduling entities hierarchy */
285 #define for_each_sched_entity(se) \
286 for (; se; se = se->parent)
288 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
293 /* runqueue on which this entity is (to be) queued */
294 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
299 /* runqueue "owned" by this group */
300 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
305 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
307 if (!cfs_rq->on_list) {
308 struct rq *rq = rq_of(cfs_rq);
309 int cpu = cpu_of(rq);
311 * Ensure we either appear before our parent (if already
312 * enqueued) or force our parent to appear after us when it is
313 * enqueued. The fact that we always enqueue bottom-up
314 * reduces this to two cases and a special case for the root
315 * cfs_rq. Furthermore, it also means that we will always reset
316 * tmp_alone_branch either when the branch is connected
317 * to a tree or when we reach the beg of the tree
319 if (cfs_rq->tg->parent &&
320 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
322 * If parent is already on the list, we add the child
323 * just before. Thanks to circular linked property of
324 * the list, this means to put the child at the tail
325 * of the list that starts by parent.
327 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
328 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
330 * The branch is now connected to its tree so we can
331 * reset tmp_alone_branch to the beginning of the
334 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
335 } else if (!cfs_rq->tg->parent) {
337 * cfs rq without parent should be put
338 * at the tail of the list.
340 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
341 &rq->leaf_cfs_rq_list);
343 * We have reach the beg of a tree so we can reset
344 * tmp_alone_branch to the beginning of the list.
346 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
349 * The parent has not already been added so we want to
350 * make sure that it will be put after us.
351 * tmp_alone_branch points to the beg of the branch
352 * where we will add parent.
354 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
355 rq->tmp_alone_branch);
357 * update tmp_alone_branch to points to the new beg
360 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
367 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
369 if (cfs_rq->on_list) {
370 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
375 /* Iterate thr' all leaf cfs_rq's on a runqueue */
376 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
377 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
379 /* Do the two (enqueued) entities belong to the same group ? */
380 static inline struct cfs_rq *
381 is_same_group(struct sched_entity *se, struct sched_entity *pse)
383 if (se->cfs_rq == pse->cfs_rq)
389 static inline struct sched_entity *parent_entity(struct sched_entity *se)
395 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
397 int se_depth, pse_depth;
400 * preemption test can be made between sibling entities who are in the
401 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
402 * both tasks until we find their ancestors who are siblings of common
406 /* First walk up until both entities are at same depth */
407 se_depth = (*se)->depth;
408 pse_depth = (*pse)->depth;
410 while (se_depth > pse_depth) {
412 *se = parent_entity(*se);
415 while (pse_depth > se_depth) {
417 *pse = parent_entity(*pse);
420 while (!is_same_group(*se, *pse)) {
421 *se = parent_entity(*se);
422 *pse = parent_entity(*pse);
426 #else /* !CONFIG_FAIR_GROUP_SCHED */
428 static inline struct task_struct *task_of(struct sched_entity *se)
430 return container_of(se, struct task_struct, se);
433 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
435 return container_of(cfs_rq, struct rq, cfs);
438 #define entity_is_task(se) 1
440 #define for_each_sched_entity(se) \
441 for (; se; se = NULL)
443 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
445 return &task_rq(p)->cfs;
448 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
450 struct task_struct *p = task_of(se);
451 struct rq *rq = task_rq(p);
456 /* runqueue "owned" by this group */
457 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
462 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
466 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
470 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
471 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
473 static inline struct sched_entity *parent_entity(struct sched_entity *se)
479 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
483 #endif /* CONFIG_FAIR_GROUP_SCHED */
485 static __always_inline
486 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
488 /**************************************************************
489 * Scheduling class tree data structure manipulation methods:
492 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
494 s64 delta = (s64)(vruntime - max_vruntime);
496 max_vruntime = vruntime;
501 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
503 s64 delta = (s64)(vruntime - min_vruntime);
505 min_vruntime = vruntime;
510 static inline int entity_before(struct sched_entity *a,
511 struct sched_entity *b)
513 return (s64)(a->vruntime - b->vruntime) < 0;
516 static void update_min_vruntime(struct cfs_rq *cfs_rq)
518 u64 vruntime = cfs_rq->min_vruntime;
521 vruntime = cfs_rq->curr->vruntime;
523 if (cfs_rq->rb_leftmost) {
524 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
529 vruntime = se->vruntime;
531 vruntime = min_vruntime(vruntime, se->vruntime);
534 /* ensure we never gain time by being placed backwards. */
535 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
538 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
543 * Enqueue an entity into the rb-tree:
545 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
548 struct rb_node *parent = NULL;
549 struct sched_entity *entry;
553 * Find the right place in the rbtree:
557 entry = rb_entry(parent, struct sched_entity, run_node);
559 * We dont care about collisions. Nodes with
560 * the same key stay together.
562 if (entity_before(se, entry)) {
563 link = &parent->rb_left;
565 link = &parent->rb_right;
571 * Maintain a cache of leftmost tree entries (it is frequently
575 cfs_rq->rb_leftmost = &se->run_node;
577 rb_link_node(&se->run_node, parent, link);
578 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
581 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
583 if (cfs_rq->rb_leftmost == &se->run_node) {
584 struct rb_node *next_node;
586 next_node = rb_next(&se->run_node);
587 cfs_rq->rb_leftmost = next_node;
590 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
593 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
595 struct rb_node *left = cfs_rq->rb_leftmost;
600 return rb_entry(left, struct sched_entity, run_node);
603 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
605 struct rb_node *next = rb_next(&se->run_node);
610 return rb_entry(next, struct sched_entity, run_node);
613 #ifdef CONFIG_SCHED_DEBUG
614 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
616 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
621 return rb_entry(last, struct sched_entity, run_node);
624 /**************************************************************
625 * Scheduling class statistics methods:
628 int sched_proc_update_handler(struct ctl_table *table, int write,
629 void __user *buffer, size_t *lenp,
632 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
633 unsigned int factor = get_update_sysctl_factor();
638 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
639 sysctl_sched_min_granularity);
641 #define WRT_SYSCTL(name) \
642 (normalized_sysctl_##name = sysctl_##name / (factor))
643 WRT_SYSCTL(sched_min_granularity);
644 WRT_SYSCTL(sched_latency);
645 WRT_SYSCTL(sched_wakeup_granularity);
655 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
657 if (unlikely(se->load.weight != NICE_0_LOAD))
658 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
664 * The idea is to set a period in which each task runs once.
666 * When there are too many tasks (sched_nr_latency) we have to stretch
667 * this period because otherwise the slices get too small.
669 * p = (nr <= nl) ? l : l*nr/nl
671 static u64 __sched_period(unsigned long nr_running)
673 if (unlikely(nr_running > sched_nr_latency))
674 return nr_running * sysctl_sched_min_granularity;
676 return sysctl_sched_latency;
680 * We calculate the wall-time slice from the period by taking a part
681 * proportional to the weight.
685 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
689 for_each_sched_entity(se) {
690 struct load_weight *load;
691 struct load_weight lw;
693 cfs_rq = cfs_rq_of(se);
694 load = &cfs_rq->load;
696 if (unlikely(!se->on_rq)) {
699 update_load_add(&lw, se->load.weight);
702 slice = __calc_delta(slice, se->load.weight, load);
708 * We calculate the vruntime slice of a to-be-inserted task.
712 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
714 return calc_delta_fair(sched_slice(cfs_rq, se), se);
718 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
719 static unsigned long task_h_load(struct task_struct *p);
722 * We choose a half-life close to 1 scheduling period.
723 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
724 * dependent on this value.
726 #define LOAD_AVG_PERIOD 32
727 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
728 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
730 /* Give new sched_entity start runnable values to heavy its load in infant time */
731 void init_entity_runnable_average(struct sched_entity *se)
733 struct sched_avg *sa = &se->avg;
735 sa->last_update_time = 0;
737 * sched_avg's period_contrib should be strictly less then 1024, so
738 * we give it 1023 to make sure it is almost a period (1024us), and
739 * will definitely be update (after enqueue).
741 sa->period_contrib = 1023;
742 sa->load_avg = scale_load_down(se->load.weight);
743 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
745 * In previous Android versions, we used to have:
746 * sa->util_avg = sched_freq() ?
747 * sysctl_sched_initial_task_util :
748 * scale_load_down(SCHED_LOAD_SCALE);
749 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
750 * However, that functionality has been moved to enqueue.
751 * It is unclear if we should restore this in enqueue.
754 * At this point, util_avg won't be used in select_task_rq_fair anyway
758 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
761 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
762 static void attach_entity_cfs_rq(struct sched_entity *se);
765 * With new tasks being created, their initial util_avgs are extrapolated
766 * based on the cfs_rq's current util_avg:
768 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
770 * However, in many cases, the above util_avg does not give a desired
771 * value. Moreover, the sum of the util_avgs may be divergent, such
772 * as when the series is a harmonic series.
774 * To solve this problem, we also cap the util_avg of successive tasks to
775 * only 1/2 of the left utilization budget:
777 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
779 * where n denotes the nth task.
781 * For example, a simplest series from the beginning would be like:
783 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
784 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
786 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
787 * if util_avg > util_avg_cap.
789 void post_init_entity_util_avg(struct sched_entity *se)
791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
792 struct sched_avg *sa = &se->avg;
793 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
796 if (cfs_rq->avg.util_avg != 0) {
797 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
798 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
800 if (sa->util_avg > cap)
806 * If we wish to restore tuning via setting initial util,
807 * this is where we should do it.
809 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
812 if (entity_is_task(se)) {
813 struct task_struct *p = task_of(se);
814 if (p->sched_class != &fair_sched_class) {
816 * For !fair tasks do:
818 update_cfs_rq_load_avg(now, cfs_rq, false);
819 attach_entity_load_avg(cfs_rq, se);
820 switched_from_fair(rq, p);
822 * such that the next switched_to_fair() has the
825 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
830 attach_entity_cfs_rq(se);
833 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
834 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
836 void init_entity_runnable_average(struct sched_entity *se)
839 void post_init_entity_util_avg(struct sched_entity *se)
842 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
845 #endif /* CONFIG_SMP */
848 * Update the current task's runtime statistics.
850 static void update_curr(struct cfs_rq *cfs_rq)
852 struct sched_entity *curr = cfs_rq->curr;
853 u64 now = rq_clock_task(rq_of(cfs_rq));
859 delta_exec = now - curr->exec_start;
860 if (unlikely((s64)delta_exec <= 0))
863 curr->exec_start = now;
865 schedstat_set(curr->statistics.exec_max,
866 max(delta_exec, curr->statistics.exec_max));
868 curr->sum_exec_runtime += delta_exec;
869 schedstat_add(cfs_rq, exec_clock, delta_exec);
871 curr->vruntime += calc_delta_fair(delta_exec, curr);
872 update_min_vruntime(cfs_rq);
874 if (entity_is_task(curr)) {
875 struct task_struct *curtask = task_of(curr);
877 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
878 cpuacct_charge(curtask, delta_exec);
879 account_group_exec_runtime(curtask, delta_exec);
882 account_cfs_rq_runtime(cfs_rq, delta_exec);
885 static void update_curr_fair(struct rq *rq)
887 update_curr(cfs_rq_of(&rq->curr->se));
891 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
893 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
897 * Task is being enqueued - update stats:
899 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
902 * Are we enqueueing a waiting task? (for current tasks
903 * a dequeue/enqueue event is a NOP)
905 if (se != cfs_rq->curr)
906 update_stats_wait_start(cfs_rq, se);
910 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
913 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
914 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
915 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
916 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
917 #ifdef CONFIG_SCHEDSTATS
918 if (entity_is_task(se)) {
919 trace_sched_stat_wait(task_of(se),
920 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
923 schedstat_set(se->statistics.wait_start, 0);
927 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
930 * Mark the end of the wait period if dequeueing a
933 if (se != cfs_rq->curr)
934 update_stats_wait_end(cfs_rq, se);
938 * We are picking a new current task - update its stats:
941 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
944 * We are starting a new run period:
946 se->exec_start = rq_clock_task(rq_of(cfs_rq));
949 /**************************************************
950 * Scheduling class queueing methods:
953 #ifdef CONFIG_NUMA_BALANCING
955 * Approximate time to scan a full NUMA task in ms. The task scan period is
956 * calculated based on the tasks virtual memory size and
957 * numa_balancing_scan_size.
959 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
960 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
962 /* Portion of address space to scan in MB */
963 unsigned int sysctl_numa_balancing_scan_size = 256;
965 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
966 unsigned int sysctl_numa_balancing_scan_delay = 1000;
968 static unsigned int task_nr_scan_windows(struct task_struct *p)
970 unsigned long rss = 0;
971 unsigned long nr_scan_pages;
974 * Calculations based on RSS as non-present and empty pages are skipped
975 * by the PTE scanner and NUMA hinting faults should be trapped based
978 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
979 rss = get_mm_rss(p->mm);
983 rss = round_up(rss, nr_scan_pages);
984 return rss / nr_scan_pages;
987 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
988 #define MAX_SCAN_WINDOW 2560
990 static unsigned int task_scan_min(struct task_struct *p)
992 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
993 unsigned int scan, floor;
994 unsigned int windows = 1;
996 if (scan_size < MAX_SCAN_WINDOW)
997 windows = MAX_SCAN_WINDOW / scan_size;
998 floor = 1000 / windows;
1000 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1001 return max_t(unsigned int, floor, scan);
1004 static unsigned int task_scan_max(struct task_struct *p)
1006 unsigned int smin = task_scan_min(p);
1009 /* Watch for min being lower than max due to floor calculations */
1010 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1011 return max(smin, smax);
1014 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1016 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1017 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1020 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1022 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1023 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1029 spinlock_t lock; /* nr_tasks, tasks */
1033 struct rcu_head rcu;
1034 nodemask_t active_nodes;
1035 unsigned long total_faults;
1037 * Faults_cpu is used to decide whether memory should move
1038 * towards the CPU. As a consequence, these stats are weighted
1039 * more by CPU use than by memory faults.
1041 unsigned long *faults_cpu;
1042 unsigned long faults[0];
1045 /* Shared or private faults. */
1046 #define NR_NUMA_HINT_FAULT_TYPES 2
1048 /* Memory and CPU locality */
1049 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1051 /* Averaged statistics, and temporary buffers. */
1052 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1054 pid_t task_numa_group_id(struct task_struct *p)
1056 return p->numa_group ? p->numa_group->gid : 0;
1060 * The averaged statistics, shared & private, memory & cpu,
1061 * occupy the first half of the array. The second half of the
1062 * array is for current counters, which are averaged into the
1063 * first set by task_numa_placement.
1065 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1067 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1070 static inline unsigned long task_faults(struct task_struct *p, int nid)
1072 if (!p->numa_faults)
1075 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1076 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1079 static inline unsigned long group_faults(struct task_struct *p, int nid)
1084 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1085 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1088 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1090 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1091 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1094 /* Handle placement on systems where not all nodes are directly connected. */
1095 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1096 int maxdist, bool task)
1098 unsigned long score = 0;
1102 * All nodes are directly connected, and the same distance
1103 * from each other. No need for fancy placement algorithms.
1105 if (sched_numa_topology_type == NUMA_DIRECT)
1109 * This code is called for each node, introducing N^2 complexity,
1110 * which should be ok given the number of nodes rarely exceeds 8.
1112 for_each_online_node(node) {
1113 unsigned long faults;
1114 int dist = node_distance(nid, node);
1117 * The furthest away nodes in the system are not interesting
1118 * for placement; nid was already counted.
1120 if (dist == sched_max_numa_distance || node == nid)
1124 * On systems with a backplane NUMA topology, compare groups
1125 * of nodes, and move tasks towards the group with the most
1126 * memory accesses. When comparing two nodes at distance
1127 * "hoplimit", only nodes closer by than "hoplimit" are part
1128 * of each group. Skip other nodes.
1130 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1134 /* Add up the faults from nearby nodes. */
1136 faults = task_faults(p, node);
1138 faults = group_faults(p, node);
1141 * On systems with a glueless mesh NUMA topology, there are
1142 * no fixed "groups of nodes". Instead, nodes that are not
1143 * directly connected bounce traffic through intermediate
1144 * nodes; a numa_group can occupy any set of nodes.
1145 * The further away a node is, the less the faults count.
1146 * This seems to result in good task placement.
1148 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1149 faults *= (sched_max_numa_distance - dist);
1150 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1160 * These return the fraction of accesses done by a particular task, or
1161 * task group, on a particular numa node. The group weight is given a
1162 * larger multiplier, in order to group tasks together that are almost
1163 * evenly spread out between numa nodes.
1165 static inline unsigned long task_weight(struct task_struct *p, int nid,
1168 unsigned long faults, total_faults;
1170 if (!p->numa_faults)
1173 total_faults = p->total_numa_faults;
1178 faults = task_faults(p, nid);
1179 faults += score_nearby_nodes(p, nid, dist, true);
1181 return 1000 * faults / total_faults;
1184 static inline unsigned long group_weight(struct task_struct *p, int nid,
1187 unsigned long faults, total_faults;
1192 total_faults = p->numa_group->total_faults;
1197 faults = group_faults(p, nid);
1198 faults += score_nearby_nodes(p, nid, dist, false);
1200 return 1000 * faults / total_faults;
1203 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1204 int src_nid, int dst_cpu)
1206 struct numa_group *ng = p->numa_group;
1207 int dst_nid = cpu_to_node(dst_cpu);
1208 int last_cpupid, this_cpupid;
1210 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1213 * Multi-stage node selection is used in conjunction with a periodic
1214 * migration fault to build a temporal task<->page relation. By using
1215 * a two-stage filter we remove short/unlikely relations.
1217 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1218 * a task's usage of a particular page (n_p) per total usage of this
1219 * page (n_t) (in a given time-span) to a probability.
1221 * Our periodic faults will sample this probability and getting the
1222 * same result twice in a row, given these samples are fully
1223 * independent, is then given by P(n)^2, provided our sample period
1224 * is sufficiently short compared to the usage pattern.
1226 * This quadric squishes small probabilities, making it less likely we
1227 * act on an unlikely task<->page relation.
1229 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1230 if (!cpupid_pid_unset(last_cpupid) &&
1231 cpupid_to_nid(last_cpupid) != dst_nid)
1234 /* Always allow migrate on private faults */
1235 if (cpupid_match_pid(p, last_cpupid))
1238 /* A shared fault, but p->numa_group has not been set up yet. */
1243 * Do not migrate if the destination is not a node that
1244 * is actively used by this numa group.
1246 if (!node_isset(dst_nid, ng->active_nodes))
1250 * Source is a node that is not actively used by this
1251 * numa group, while the destination is. Migrate.
1253 if (!node_isset(src_nid, ng->active_nodes))
1257 * Both source and destination are nodes in active
1258 * use by this numa group. Maximize memory bandwidth
1259 * by migrating from more heavily used groups, to less
1260 * heavily used ones, spreading the load around.
1261 * Use a 1/4 hysteresis to avoid spurious page movement.
1263 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1266 static unsigned long weighted_cpuload(const int cpu);
1267 static unsigned long source_load(int cpu, int type);
1268 static unsigned long target_load(int cpu, int type);
1269 static unsigned long capacity_of(int cpu);
1270 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1272 /* Cached statistics for all CPUs within a node */
1274 unsigned long nr_running;
1277 /* Total compute capacity of CPUs on a node */
1278 unsigned long compute_capacity;
1280 /* Approximate capacity in terms of runnable tasks on a node */
1281 unsigned long task_capacity;
1282 int has_free_capacity;
1286 * XXX borrowed from update_sg_lb_stats
1288 static void update_numa_stats(struct numa_stats *ns, int nid)
1290 int smt, cpu, cpus = 0;
1291 unsigned long capacity;
1293 memset(ns, 0, sizeof(*ns));
1294 for_each_cpu(cpu, cpumask_of_node(nid)) {
1295 struct rq *rq = cpu_rq(cpu);
1297 ns->nr_running += rq->nr_running;
1298 ns->load += weighted_cpuload(cpu);
1299 ns->compute_capacity += capacity_of(cpu);
1305 * If we raced with hotplug and there are no CPUs left in our mask
1306 * the @ns structure is NULL'ed and task_numa_compare() will
1307 * not find this node attractive.
1309 * We'll either bail at !has_free_capacity, or we'll detect a huge
1310 * imbalance and bail there.
1315 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1316 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1317 capacity = cpus / smt; /* cores */
1319 ns->task_capacity = min_t(unsigned, capacity,
1320 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1321 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1324 struct task_numa_env {
1325 struct task_struct *p;
1327 int src_cpu, src_nid;
1328 int dst_cpu, dst_nid;
1330 struct numa_stats src_stats, dst_stats;
1335 struct task_struct *best_task;
1340 static void task_numa_assign(struct task_numa_env *env,
1341 struct task_struct *p, long imp)
1344 put_task_struct(env->best_task);
1347 env->best_imp = imp;
1348 env->best_cpu = env->dst_cpu;
1351 static bool load_too_imbalanced(long src_load, long dst_load,
1352 struct task_numa_env *env)
1355 long orig_src_load, orig_dst_load;
1356 long src_capacity, dst_capacity;
1359 * The load is corrected for the CPU capacity available on each node.
1362 * ------------ vs ---------
1363 * src_capacity dst_capacity
1365 src_capacity = env->src_stats.compute_capacity;
1366 dst_capacity = env->dst_stats.compute_capacity;
1368 /* We care about the slope of the imbalance, not the direction. */
1369 if (dst_load < src_load)
1370 swap(dst_load, src_load);
1372 /* Is the difference below the threshold? */
1373 imb = dst_load * src_capacity * 100 -
1374 src_load * dst_capacity * env->imbalance_pct;
1379 * The imbalance is above the allowed threshold.
1380 * Compare it with the old imbalance.
1382 orig_src_load = env->src_stats.load;
1383 orig_dst_load = env->dst_stats.load;
1385 if (orig_dst_load < orig_src_load)
1386 swap(orig_dst_load, orig_src_load);
1388 old_imb = orig_dst_load * src_capacity * 100 -
1389 orig_src_load * dst_capacity * env->imbalance_pct;
1391 /* Would this change make things worse? */
1392 return (imb > old_imb);
1396 * This checks if the overall compute and NUMA accesses of the system would
1397 * be improved if the source tasks was migrated to the target dst_cpu taking
1398 * into account that it might be best if task running on the dst_cpu should
1399 * be exchanged with the source task
1401 static void task_numa_compare(struct task_numa_env *env,
1402 long taskimp, long groupimp)
1404 struct rq *src_rq = cpu_rq(env->src_cpu);
1405 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1406 struct task_struct *cur;
1407 long src_load, dst_load;
1409 long imp = env->p->numa_group ? groupimp : taskimp;
1411 int dist = env->dist;
1412 bool assigned = false;
1416 raw_spin_lock_irq(&dst_rq->lock);
1419 * No need to move the exiting task or idle task.
1421 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1425 * The task_struct must be protected here to protect the
1426 * p->numa_faults access in the task_weight since the
1427 * numa_faults could already be freed in the following path:
1428 * finish_task_switch()
1429 * --> put_task_struct()
1430 * --> __put_task_struct()
1431 * --> task_numa_free()
1433 get_task_struct(cur);
1436 raw_spin_unlock_irq(&dst_rq->lock);
1439 * Because we have preemption enabled we can get migrated around and
1440 * end try selecting ourselves (current == env->p) as a swap candidate.
1446 * "imp" is the fault differential for the source task between the
1447 * source and destination node. Calculate the total differential for
1448 * the source task and potential destination task. The more negative
1449 * the value is, the more rmeote accesses that would be expected to
1450 * be incurred if the tasks were swapped.
1453 /* Skip this swap candidate if cannot move to the source cpu */
1454 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1458 * If dst and source tasks are in the same NUMA group, or not
1459 * in any group then look only at task weights.
1461 if (cur->numa_group == env->p->numa_group) {
1462 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1463 task_weight(cur, env->dst_nid, dist);
1465 * Add some hysteresis to prevent swapping the
1466 * tasks within a group over tiny differences.
1468 if (cur->numa_group)
1472 * Compare the group weights. If a task is all by
1473 * itself (not part of a group), use the task weight
1476 if (cur->numa_group)
1477 imp += group_weight(cur, env->src_nid, dist) -
1478 group_weight(cur, env->dst_nid, dist);
1480 imp += task_weight(cur, env->src_nid, dist) -
1481 task_weight(cur, env->dst_nid, dist);
1485 if (imp <= env->best_imp && moveimp <= env->best_imp)
1489 /* Is there capacity at our destination? */
1490 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1491 !env->dst_stats.has_free_capacity)
1497 /* Balance doesn't matter much if we're running a task per cpu */
1498 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1499 dst_rq->nr_running == 1)
1503 * In the overloaded case, try and keep the load balanced.
1506 load = task_h_load(env->p);
1507 dst_load = env->dst_stats.load + load;
1508 src_load = env->src_stats.load - load;
1510 if (moveimp > imp && moveimp > env->best_imp) {
1512 * If the improvement from just moving env->p direction is
1513 * better than swapping tasks around, check if a move is
1514 * possible. Store a slightly smaller score than moveimp,
1515 * so an actually idle CPU will win.
1517 if (!load_too_imbalanced(src_load, dst_load, env)) {
1519 put_task_struct(cur);
1525 if (imp <= env->best_imp)
1529 load = task_h_load(cur);
1534 if (load_too_imbalanced(src_load, dst_load, env))
1538 * One idle CPU per node is evaluated for a task numa move.
1539 * Call select_idle_sibling to maybe find a better one.
1542 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1547 task_numa_assign(env, cur, imp);
1551 * The dst_rq->curr isn't assigned. The protection for task_struct is
1554 if (cur && !assigned)
1555 put_task_struct(cur);
1558 static void task_numa_find_cpu(struct task_numa_env *env,
1559 long taskimp, long groupimp)
1563 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1564 /* Skip this CPU if the source task cannot migrate */
1565 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1569 task_numa_compare(env, taskimp, groupimp);
1573 /* Only move tasks to a NUMA node less busy than the current node. */
1574 static bool numa_has_capacity(struct task_numa_env *env)
1576 struct numa_stats *src = &env->src_stats;
1577 struct numa_stats *dst = &env->dst_stats;
1579 if (src->has_free_capacity && !dst->has_free_capacity)
1583 * Only consider a task move if the source has a higher load
1584 * than the destination, corrected for CPU capacity on each node.
1586 * src->load dst->load
1587 * --------------------- vs ---------------------
1588 * src->compute_capacity dst->compute_capacity
1590 if (src->load * dst->compute_capacity * env->imbalance_pct >
1592 dst->load * src->compute_capacity * 100)
1598 static int task_numa_migrate(struct task_struct *p)
1600 struct task_numa_env env = {
1603 .src_cpu = task_cpu(p),
1604 .src_nid = task_node(p),
1606 .imbalance_pct = 112,
1612 struct sched_domain *sd;
1613 unsigned long taskweight, groupweight;
1615 long taskimp, groupimp;
1618 * Pick the lowest SD_NUMA domain, as that would have the smallest
1619 * imbalance and would be the first to start moving tasks about.
1621 * And we want to avoid any moving of tasks about, as that would create
1622 * random movement of tasks -- counter the numa conditions we're trying
1626 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1628 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1632 * Cpusets can break the scheduler domain tree into smaller
1633 * balance domains, some of which do not cross NUMA boundaries.
1634 * Tasks that are "trapped" in such domains cannot be migrated
1635 * elsewhere, so there is no point in (re)trying.
1637 if (unlikely(!sd)) {
1638 p->numa_preferred_nid = task_node(p);
1642 env.dst_nid = p->numa_preferred_nid;
1643 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1644 taskweight = task_weight(p, env.src_nid, dist);
1645 groupweight = group_weight(p, env.src_nid, dist);
1646 update_numa_stats(&env.src_stats, env.src_nid);
1647 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1648 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1649 update_numa_stats(&env.dst_stats, env.dst_nid);
1651 /* Try to find a spot on the preferred nid. */
1652 if (numa_has_capacity(&env))
1653 task_numa_find_cpu(&env, taskimp, groupimp);
1656 * Look at other nodes in these cases:
1657 * - there is no space available on the preferred_nid
1658 * - the task is part of a numa_group that is interleaved across
1659 * multiple NUMA nodes; in order to better consolidate the group,
1660 * we need to check other locations.
1662 if (env.best_cpu == -1 || (p->numa_group &&
1663 nodes_weight(p->numa_group->active_nodes) > 1)) {
1664 for_each_online_node(nid) {
1665 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1668 dist = node_distance(env.src_nid, env.dst_nid);
1669 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1671 taskweight = task_weight(p, env.src_nid, dist);
1672 groupweight = group_weight(p, env.src_nid, dist);
1675 /* Only consider nodes where both task and groups benefit */
1676 taskimp = task_weight(p, nid, dist) - taskweight;
1677 groupimp = group_weight(p, nid, dist) - groupweight;
1678 if (taskimp < 0 && groupimp < 0)
1683 update_numa_stats(&env.dst_stats, env.dst_nid);
1684 if (numa_has_capacity(&env))
1685 task_numa_find_cpu(&env, taskimp, groupimp);
1690 * If the task is part of a workload that spans multiple NUMA nodes,
1691 * and is migrating into one of the workload's active nodes, remember
1692 * this node as the task's preferred numa node, so the workload can
1694 * A task that migrated to a second choice node will be better off
1695 * trying for a better one later. Do not set the preferred node here.
1697 if (p->numa_group) {
1698 if (env.best_cpu == -1)
1703 if (node_isset(nid, p->numa_group->active_nodes))
1704 sched_setnuma(p, env.dst_nid);
1707 /* No better CPU than the current one was found. */
1708 if (env.best_cpu == -1)
1712 * Reset the scan period if the task is being rescheduled on an
1713 * alternative node to recheck if the tasks is now properly placed.
1715 p->numa_scan_period = task_scan_min(p);
1717 if (env.best_task == NULL) {
1718 ret = migrate_task_to(p, env.best_cpu);
1720 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1724 ret = migrate_swap(p, env.best_task);
1726 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1727 put_task_struct(env.best_task);
1731 /* Attempt to migrate a task to a CPU on the preferred node. */
1732 static void numa_migrate_preferred(struct task_struct *p)
1734 unsigned long interval = HZ;
1736 /* This task has no NUMA fault statistics yet */
1737 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1740 /* Periodically retry migrating the task to the preferred node */
1741 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1742 p->numa_migrate_retry = jiffies + interval;
1744 /* Success if task is already running on preferred CPU */
1745 if (task_node(p) == p->numa_preferred_nid)
1748 /* Otherwise, try migrate to a CPU on the preferred node */
1749 task_numa_migrate(p);
1753 * Find the nodes on which the workload is actively running. We do this by
1754 * tracking the nodes from which NUMA hinting faults are triggered. This can
1755 * be different from the set of nodes where the workload's memory is currently
1758 * The bitmask is used to make smarter decisions on when to do NUMA page
1759 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1760 * are added when they cause over 6/16 of the maximum number of faults, but
1761 * only removed when they drop below 3/16.
1763 static void update_numa_active_node_mask(struct numa_group *numa_group)
1765 unsigned long faults, max_faults = 0;
1768 for_each_online_node(nid) {
1769 faults = group_faults_cpu(numa_group, nid);
1770 if (faults > max_faults)
1771 max_faults = faults;
1774 for_each_online_node(nid) {
1775 faults = group_faults_cpu(numa_group, nid);
1776 if (!node_isset(nid, numa_group->active_nodes)) {
1777 if (faults > max_faults * 6 / 16)
1778 node_set(nid, numa_group->active_nodes);
1779 } else if (faults < max_faults * 3 / 16)
1780 node_clear(nid, numa_group->active_nodes);
1785 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1786 * increments. The more local the fault statistics are, the higher the scan
1787 * period will be for the next scan window. If local/(local+remote) ratio is
1788 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1789 * the scan period will decrease. Aim for 70% local accesses.
1791 #define NUMA_PERIOD_SLOTS 10
1792 #define NUMA_PERIOD_THRESHOLD 7
1795 * Increase the scan period (slow down scanning) if the majority of
1796 * our memory is already on our local node, or if the majority of
1797 * the page accesses are shared with other processes.
1798 * Otherwise, decrease the scan period.
1800 static void update_task_scan_period(struct task_struct *p,
1801 unsigned long shared, unsigned long private)
1803 unsigned int period_slot;
1807 unsigned long remote = p->numa_faults_locality[0];
1808 unsigned long local = p->numa_faults_locality[1];
1811 * If there were no record hinting faults then either the task is
1812 * completely idle or all activity is areas that are not of interest
1813 * to automatic numa balancing. Related to that, if there were failed
1814 * migration then it implies we are migrating too quickly or the local
1815 * node is overloaded. In either case, scan slower
1817 if (local + shared == 0 || p->numa_faults_locality[2]) {
1818 p->numa_scan_period = min(p->numa_scan_period_max,
1819 p->numa_scan_period << 1);
1821 p->mm->numa_next_scan = jiffies +
1822 msecs_to_jiffies(p->numa_scan_period);
1828 * Prepare to scale scan period relative to the current period.
1829 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1830 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1831 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1833 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1834 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1835 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1836 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1839 diff = slot * period_slot;
1841 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1844 * Scale scan rate increases based on sharing. There is an
1845 * inverse relationship between the degree of sharing and
1846 * the adjustment made to the scanning period. Broadly
1847 * speaking the intent is that there is little point
1848 * scanning faster if shared accesses dominate as it may
1849 * simply bounce migrations uselessly
1851 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1852 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1855 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1856 task_scan_min(p), task_scan_max(p));
1857 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1861 * Get the fraction of time the task has been running since the last
1862 * NUMA placement cycle. The scheduler keeps similar statistics, but
1863 * decays those on a 32ms period, which is orders of magnitude off
1864 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1865 * stats only if the task is so new there are no NUMA statistics yet.
1867 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1869 u64 runtime, delta, now;
1870 /* Use the start of this time slice to avoid calculations. */
1871 now = p->se.exec_start;
1872 runtime = p->se.sum_exec_runtime;
1874 if (p->last_task_numa_placement) {
1875 delta = runtime - p->last_sum_exec_runtime;
1876 *period = now - p->last_task_numa_placement;
1878 delta = p->se.avg.load_sum / p->se.load.weight;
1879 *period = LOAD_AVG_MAX;
1882 p->last_sum_exec_runtime = runtime;
1883 p->last_task_numa_placement = now;
1889 * Determine the preferred nid for a task in a numa_group. This needs to
1890 * be done in a way that produces consistent results with group_weight,
1891 * otherwise workloads might not converge.
1893 static int preferred_group_nid(struct task_struct *p, int nid)
1898 /* Direct connections between all NUMA nodes. */
1899 if (sched_numa_topology_type == NUMA_DIRECT)
1903 * On a system with glueless mesh NUMA topology, group_weight
1904 * scores nodes according to the number of NUMA hinting faults on
1905 * both the node itself, and on nearby nodes.
1907 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1908 unsigned long score, max_score = 0;
1909 int node, max_node = nid;
1911 dist = sched_max_numa_distance;
1913 for_each_online_node(node) {
1914 score = group_weight(p, node, dist);
1915 if (score > max_score) {
1924 * Finding the preferred nid in a system with NUMA backplane
1925 * interconnect topology is more involved. The goal is to locate
1926 * tasks from numa_groups near each other in the system, and
1927 * untangle workloads from different sides of the system. This requires
1928 * searching down the hierarchy of node groups, recursively searching
1929 * inside the highest scoring group of nodes. The nodemask tricks
1930 * keep the complexity of the search down.
1932 nodes = node_online_map;
1933 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1934 unsigned long max_faults = 0;
1935 nodemask_t max_group = NODE_MASK_NONE;
1938 /* Are there nodes at this distance from each other? */
1939 if (!find_numa_distance(dist))
1942 for_each_node_mask(a, nodes) {
1943 unsigned long faults = 0;
1944 nodemask_t this_group;
1945 nodes_clear(this_group);
1947 /* Sum group's NUMA faults; includes a==b case. */
1948 for_each_node_mask(b, nodes) {
1949 if (node_distance(a, b) < dist) {
1950 faults += group_faults(p, b);
1951 node_set(b, this_group);
1952 node_clear(b, nodes);
1956 /* Remember the top group. */
1957 if (faults > max_faults) {
1958 max_faults = faults;
1959 max_group = this_group;
1961 * subtle: at the smallest distance there is
1962 * just one node left in each "group", the
1963 * winner is the preferred nid.
1968 /* Next round, evaluate the nodes within max_group. */
1976 static void task_numa_placement(struct task_struct *p)
1978 int seq, nid, max_nid = -1, max_group_nid = -1;
1979 unsigned long max_faults = 0, max_group_faults = 0;
1980 unsigned long fault_types[2] = { 0, 0 };
1981 unsigned long total_faults;
1982 u64 runtime, period;
1983 spinlock_t *group_lock = NULL;
1986 * The p->mm->numa_scan_seq field gets updated without
1987 * exclusive access. Use READ_ONCE() here to ensure
1988 * that the field is read in a single access:
1990 seq = READ_ONCE(p->mm->numa_scan_seq);
1991 if (p->numa_scan_seq == seq)
1993 p->numa_scan_seq = seq;
1994 p->numa_scan_period_max = task_scan_max(p);
1996 total_faults = p->numa_faults_locality[0] +
1997 p->numa_faults_locality[1];
1998 runtime = numa_get_avg_runtime(p, &period);
2000 /* If the task is part of a group prevent parallel updates to group stats */
2001 if (p->numa_group) {
2002 group_lock = &p->numa_group->lock;
2003 spin_lock_irq(group_lock);
2006 /* Find the node with the highest number of faults */
2007 for_each_online_node(nid) {
2008 /* Keep track of the offsets in numa_faults array */
2009 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2010 unsigned long faults = 0, group_faults = 0;
2013 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2014 long diff, f_diff, f_weight;
2016 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2017 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2018 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2019 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2021 /* Decay existing window, copy faults since last scan */
2022 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2023 fault_types[priv] += p->numa_faults[membuf_idx];
2024 p->numa_faults[membuf_idx] = 0;
2027 * Normalize the faults_from, so all tasks in a group
2028 * count according to CPU use, instead of by the raw
2029 * number of faults. Tasks with little runtime have
2030 * little over-all impact on throughput, and thus their
2031 * faults are less important.
2033 f_weight = div64_u64(runtime << 16, period + 1);
2034 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2036 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2037 p->numa_faults[cpubuf_idx] = 0;
2039 p->numa_faults[mem_idx] += diff;
2040 p->numa_faults[cpu_idx] += f_diff;
2041 faults += p->numa_faults[mem_idx];
2042 p->total_numa_faults += diff;
2043 if (p->numa_group) {
2045 * safe because we can only change our own group
2047 * mem_idx represents the offset for a given
2048 * nid and priv in a specific region because it
2049 * is at the beginning of the numa_faults array.
2051 p->numa_group->faults[mem_idx] += diff;
2052 p->numa_group->faults_cpu[mem_idx] += f_diff;
2053 p->numa_group->total_faults += diff;
2054 group_faults += p->numa_group->faults[mem_idx];
2058 if (faults > max_faults) {
2059 max_faults = faults;
2063 if (group_faults > max_group_faults) {
2064 max_group_faults = group_faults;
2065 max_group_nid = nid;
2069 update_task_scan_period(p, fault_types[0], fault_types[1]);
2071 if (p->numa_group) {
2072 update_numa_active_node_mask(p->numa_group);
2073 spin_unlock_irq(group_lock);
2074 max_nid = preferred_group_nid(p, max_group_nid);
2078 /* Set the new preferred node */
2079 if (max_nid != p->numa_preferred_nid)
2080 sched_setnuma(p, max_nid);
2082 if (task_node(p) != p->numa_preferred_nid)
2083 numa_migrate_preferred(p);
2087 static inline int get_numa_group(struct numa_group *grp)
2089 return atomic_inc_not_zero(&grp->refcount);
2092 static inline void put_numa_group(struct numa_group *grp)
2094 if (atomic_dec_and_test(&grp->refcount))
2095 kfree_rcu(grp, rcu);
2098 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2101 struct numa_group *grp, *my_grp;
2102 struct task_struct *tsk;
2104 int cpu = cpupid_to_cpu(cpupid);
2107 if (unlikely(!p->numa_group)) {
2108 unsigned int size = sizeof(struct numa_group) +
2109 4*nr_node_ids*sizeof(unsigned long);
2111 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2115 atomic_set(&grp->refcount, 1);
2116 spin_lock_init(&grp->lock);
2118 /* Second half of the array tracks nids where faults happen */
2119 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2122 node_set(task_node(current), grp->active_nodes);
2124 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2125 grp->faults[i] = p->numa_faults[i];
2127 grp->total_faults = p->total_numa_faults;
2130 rcu_assign_pointer(p->numa_group, grp);
2134 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2136 if (!cpupid_match_pid(tsk, cpupid))
2139 grp = rcu_dereference(tsk->numa_group);
2143 my_grp = p->numa_group;
2148 * Only join the other group if its bigger; if we're the bigger group,
2149 * the other task will join us.
2151 if (my_grp->nr_tasks > grp->nr_tasks)
2155 * Tie-break on the grp address.
2157 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2160 /* Always join threads in the same process. */
2161 if (tsk->mm == current->mm)
2164 /* Simple filter to avoid false positives due to PID collisions */
2165 if (flags & TNF_SHARED)
2168 /* Update priv based on whether false sharing was detected */
2171 if (join && !get_numa_group(grp))
2179 BUG_ON(irqs_disabled());
2180 double_lock_irq(&my_grp->lock, &grp->lock);
2182 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2183 my_grp->faults[i] -= p->numa_faults[i];
2184 grp->faults[i] += p->numa_faults[i];
2186 my_grp->total_faults -= p->total_numa_faults;
2187 grp->total_faults += p->total_numa_faults;
2192 spin_unlock(&my_grp->lock);
2193 spin_unlock_irq(&grp->lock);
2195 rcu_assign_pointer(p->numa_group, grp);
2197 put_numa_group(my_grp);
2205 void task_numa_free(struct task_struct *p)
2207 struct numa_group *grp = p->numa_group;
2208 void *numa_faults = p->numa_faults;
2209 unsigned long flags;
2213 spin_lock_irqsave(&grp->lock, flags);
2214 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2215 grp->faults[i] -= p->numa_faults[i];
2216 grp->total_faults -= p->total_numa_faults;
2219 spin_unlock_irqrestore(&grp->lock, flags);
2220 RCU_INIT_POINTER(p->numa_group, NULL);
2221 put_numa_group(grp);
2224 p->numa_faults = NULL;
2229 * Got a PROT_NONE fault for a page on @node.
2231 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2233 struct task_struct *p = current;
2234 bool migrated = flags & TNF_MIGRATED;
2235 int cpu_node = task_node(current);
2236 int local = !!(flags & TNF_FAULT_LOCAL);
2239 if (!static_branch_likely(&sched_numa_balancing))
2242 /* for example, ksmd faulting in a user's mm */
2246 /* Allocate buffer to track faults on a per-node basis */
2247 if (unlikely(!p->numa_faults)) {
2248 int size = sizeof(*p->numa_faults) *
2249 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2251 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2252 if (!p->numa_faults)
2255 p->total_numa_faults = 0;
2256 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2260 * First accesses are treated as private, otherwise consider accesses
2261 * to be private if the accessing pid has not changed
2263 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2266 priv = cpupid_match_pid(p, last_cpupid);
2267 if (!priv && !(flags & TNF_NO_GROUP))
2268 task_numa_group(p, last_cpupid, flags, &priv);
2272 * If a workload spans multiple NUMA nodes, a shared fault that
2273 * occurs wholly within the set of nodes that the workload is
2274 * actively using should be counted as local. This allows the
2275 * scan rate to slow down when a workload has settled down.
2277 if (!priv && !local && p->numa_group &&
2278 node_isset(cpu_node, p->numa_group->active_nodes) &&
2279 node_isset(mem_node, p->numa_group->active_nodes))
2282 task_numa_placement(p);
2285 * Retry task to preferred node migration periodically, in case it
2286 * case it previously failed, or the scheduler moved us.
2288 if (time_after(jiffies, p->numa_migrate_retry))
2289 numa_migrate_preferred(p);
2292 p->numa_pages_migrated += pages;
2293 if (flags & TNF_MIGRATE_FAIL)
2294 p->numa_faults_locality[2] += pages;
2296 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2297 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2298 p->numa_faults_locality[local] += pages;
2301 static void reset_ptenuma_scan(struct task_struct *p)
2304 * We only did a read acquisition of the mmap sem, so
2305 * p->mm->numa_scan_seq is written to without exclusive access
2306 * and the update is not guaranteed to be atomic. That's not
2307 * much of an issue though, since this is just used for
2308 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2309 * expensive, to avoid any form of compiler optimizations:
2311 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2312 p->mm->numa_scan_offset = 0;
2316 * The expensive part of numa migration is done from task_work context.
2317 * Triggered from task_tick_numa().
2319 void task_numa_work(struct callback_head *work)
2321 unsigned long migrate, next_scan, now = jiffies;
2322 struct task_struct *p = current;
2323 struct mm_struct *mm = p->mm;
2324 struct vm_area_struct *vma;
2325 unsigned long start, end;
2326 unsigned long nr_pte_updates = 0;
2327 long pages, virtpages;
2329 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2331 work->next = work; /* protect against double add */
2333 * Who cares about NUMA placement when they're dying.
2335 * NOTE: make sure not to dereference p->mm before this check,
2336 * exit_task_work() happens _after_ exit_mm() so we could be called
2337 * without p->mm even though we still had it when we enqueued this
2340 if (p->flags & PF_EXITING)
2343 if (!mm->numa_next_scan) {
2344 mm->numa_next_scan = now +
2345 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2349 * Enforce maximal scan/migration frequency..
2351 migrate = mm->numa_next_scan;
2352 if (time_before(now, migrate))
2355 if (p->numa_scan_period == 0) {
2356 p->numa_scan_period_max = task_scan_max(p);
2357 p->numa_scan_period = task_scan_min(p);
2360 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2361 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2365 * Delay this task enough that another task of this mm will likely win
2366 * the next time around.
2368 p->node_stamp += 2 * TICK_NSEC;
2370 start = mm->numa_scan_offset;
2371 pages = sysctl_numa_balancing_scan_size;
2372 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2373 virtpages = pages * 8; /* Scan up to this much virtual space */
2378 down_read(&mm->mmap_sem);
2379 vma = find_vma(mm, start);
2381 reset_ptenuma_scan(p);
2385 for (; vma; vma = vma->vm_next) {
2386 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2387 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2392 * Shared library pages mapped by multiple processes are not
2393 * migrated as it is expected they are cache replicated. Avoid
2394 * hinting faults in read-only file-backed mappings or the vdso
2395 * as migrating the pages will be of marginal benefit.
2398 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2402 * Skip inaccessible VMAs to avoid any confusion between
2403 * PROT_NONE and NUMA hinting ptes
2405 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2409 start = max(start, vma->vm_start);
2410 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2411 end = min(end, vma->vm_end);
2412 nr_pte_updates = change_prot_numa(vma, start, end);
2415 * Try to scan sysctl_numa_balancing_size worth of
2416 * hpages that have at least one present PTE that
2417 * is not already pte-numa. If the VMA contains
2418 * areas that are unused or already full of prot_numa
2419 * PTEs, scan up to virtpages, to skip through those
2423 pages -= (end - start) >> PAGE_SHIFT;
2424 virtpages -= (end - start) >> PAGE_SHIFT;
2427 if (pages <= 0 || virtpages <= 0)
2431 } while (end != vma->vm_end);
2436 * It is possible to reach the end of the VMA list but the last few
2437 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2438 * would find the !migratable VMA on the next scan but not reset the
2439 * scanner to the start so check it now.
2442 mm->numa_scan_offset = start;
2444 reset_ptenuma_scan(p);
2445 up_read(&mm->mmap_sem);
2449 * Drive the periodic memory faults..
2451 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2453 struct callback_head *work = &curr->numa_work;
2457 * We don't care about NUMA placement if we don't have memory.
2459 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2463 * Using runtime rather than walltime has the dual advantage that
2464 * we (mostly) drive the selection from busy threads and that the
2465 * task needs to have done some actual work before we bother with
2468 now = curr->se.sum_exec_runtime;
2469 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2471 if (now > curr->node_stamp + period) {
2472 if (!curr->node_stamp)
2473 curr->numa_scan_period = task_scan_min(curr);
2474 curr->node_stamp += period;
2476 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2477 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2478 task_work_add(curr, work, true);
2483 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2487 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2491 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2494 #endif /* CONFIG_NUMA_BALANCING */
2497 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2499 update_load_add(&cfs_rq->load, se->load.weight);
2500 if (!parent_entity(se))
2501 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2503 if (entity_is_task(se)) {
2504 struct rq *rq = rq_of(cfs_rq);
2506 account_numa_enqueue(rq, task_of(se));
2507 list_add(&se->group_node, &rq->cfs_tasks);
2510 cfs_rq->nr_running++;
2514 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2516 update_load_sub(&cfs_rq->load, se->load.weight);
2517 if (!parent_entity(se))
2518 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2519 if (entity_is_task(se)) {
2520 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2521 list_del_init(&se->group_node);
2523 cfs_rq->nr_running--;
2526 #ifdef CONFIG_FAIR_GROUP_SCHED
2528 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2533 * Use this CPU's real-time load instead of the last load contribution
2534 * as the updating of the contribution is delayed, and we will use the
2535 * the real-time load to calc the share. See update_tg_load_avg().
2537 tg_weight = atomic_long_read(&tg->load_avg);
2538 tg_weight -= cfs_rq->tg_load_avg_contrib;
2539 tg_weight += cfs_rq->load.weight;
2544 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2546 long tg_weight, load, shares;
2548 tg_weight = calc_tg_weight(tg, cfs_rq);
2549 load = cfs_rq->load.weight;
2551 shares = (tg->shares * load);
2553 shares /= tg_weight;
2555 if (shares < MIN_SHARES)
2556 shares = MIN_SHARES;
2557 if (shares > tg->shares)
2558 shares = tg->shares;
2562 # else /* CONFIG_SMP */
2563 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2567 # endif /* CONFIG_SMP */
2568 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2569 unsigned long weight)
2572 /* commit outstanding execution time */
2573 if (cfs_rq->curr == se)
2574 update_curr(cfs_rq);
2575 account_entity_dequeue(cfs_rq, se);
2578 update_load_set(&se->load, weight);
2581 account_entity_enqueue(cfs_rq, se);
2584 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2586 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2588 struct task_group *tg;
2589 struct sched_entity *se;
2593 se = tg->se[cpu_of(rq_of(cfs_rq))];
2594 if (!se || throttled_hierarchy(cfs_rq))
2597 if (likely(se->load.weight == tg->shares))
2600 shares = calc_cfs_shares(cfs_rq, tg);
2602 reweight_entity(cfs_rq_of(se), se, shares);
2604 #else /* CONFIG_FAIR_GROUP_SCHED */
2605 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2608 #endif /* CONFIG_FAIR_GROUP_SCHED */
2611 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2612 static const u32 runnable_avg_yN_inv[] = {
2613 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2614 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2615 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2616 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2617 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2618 0x85aac367, 0x82cd8698,
2622 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2623 * over-estimates when re-combining.
2625 static const u32 runnable_avg_yN_sum[] = {
2626 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2627 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2628 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2633 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2635 static __always_inline u64 decay_load(u64 val, u64 n)
2637 unsigned int local_n;
2641 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2644 /* after bounds checking we can collapse to 32-bit */
2648 * As y^PERIOD = 1/2, we can combine
2649 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2650 * With a look-up table which covers y^n (n<PERIOD)
2652 * To achieve constant time decay_load.
2654 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2655 val >>= local_n / LOAD_AVG_PERIOD;
2656 local_n %= LOAD_AVG_PERIOD;
2659 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2664 * For updates fully spanning n periods, the contribution to runnable
2665 * average will be: \Sum 1024*y^n
2667 * We can compute this reasonably efficiently by combining:
2668 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2670 static u32 __compute_runnable_contrib(u64 n)
2674 if (likely(n <= LOAD_AVG_PERIOD))
2675 return runnable_avg_yN_sum[n];
2676 else if (unlikely(n >= LOAD_AVG_MAX_N))
2677 return LOAD_AVG_MAX;
2679 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2681 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2682 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2684 n -= LOAD_AVG_PERIOD;
2685 } while (n > LOAD_AVG_PERIOD);
2687 contrib = decay_load(contrib, n);
2688 return contrib + runnable_avg_yN_sum[n];
2691 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2692 #error "load tracking assumes 2^10 as unit"
2695 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2698 * We can represent the historical contribution to runnable average as the
2699 * coefficients of a geometric series. To do this we sub-divide our runnable
2700 * history into segments of approximately 1ms (1024us); label the segment that
2701 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2703 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2705 * (now) (~1ms ago) (~2ms ago)
2707 * Let u_i denote the fraction of p_i that the entity was runnable.
2709 * We then designate the fractions u_i as our co-efficients, yielding the
2710 * following representation of historical load:
2711 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2713 * We choose y based on the with of a reasonably scheduling period, fixing:
2716 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2717 * approximately half as much as the contribution to load within the last ms
2720 * When a period "rolls over" and we have new u_0`, multiplying the previous
2721 * sum again by y is sufficient to update:
2722 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2723 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2725 static __always_inline int
2726 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2727 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2729 u64 delta, scaled_delta, periods;
2731 unsigned int delta_w, scaled_delta_w, decayed = 0;
2732 unsigned long scale_freq, scale_cpu;
2734 delta = now - sa->last_update_time;
2736 * This should only happen when time goes backwards, which it
2737 * unfortunately does during sched clock init when we swap over to TSC.
2739 if ((s64)delta < 0) {
2740 sa->last_update_time = now;
2745 * Use 1024ns as the unit of measurement since it's a reasonable
2746 * approximation of 1us and fast to compute.
2751 sa->last_update_time = now;
2753 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2754 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2755 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2757 /* delta_w is the amount already accumulated against our next period */
2758 delta_w = sa->period_contrib;
2759 if (delta + delta_w >= 1024) {
2762 /* how much left for next period will start over, we don't know yet */
2763 sa->period_contrib = 0;
2766 * Now that we know we're crossing a period boundary, figure
2767 * out how much from delta we need to complete the current
2768 * period and accrue it.
2770 delta_w = 1024 - delta_w;
2771 scaled_delta_w = cap_scale(delta_w, scale_freq);
2773 sa->load_sum += weight * scaled_delta_w;
2775 cfs_rq->runnable_load_sum +=
2776 weight * scaled_delta_w;
2780 sa->util_sum += scaled_delta_w * scale_cpu;
2784 /* Figure out how many additional periods this update spans */
2785 periods = delta / 1024;
2788 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2790 cfs_rq->runnable_load_sum =
2791 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2793 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2795 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2796 contrib = __compute_runnable_contrib(periods);
2797 contrib = cap_scale(contrib, scale_freq);
2799 sa->load_sum += weight * contrib;
2801 cfs_rq->runnable_load_sum += weight * contrib;
2804 sa->util_sum += contrib * scale_cpu;
2807 /* Remainder of delta accrued against u_0` */
2808 scaled_delta = cap_scale(delta, scale_freq);
2810 sa->load_sum += weight * scaled_delta;
2812 cfs_rq->runnable_load_sum += weight * scaled_delta;
2815 sa->util_sum += scaled_delta * scale_cpu;
2817 sa->period_contrib += delta;
2820 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2822 cfs_rq->runnable_load_avg =
2823 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2825 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2831 #ifdef CONFIG_FAIR_GROUP_SCHED
2833 * update_tg_load_avg - update the tg's load avg
2834 * @cfs_rq: the cfs_rq whose avg changed
2835 * @force: update regardless of how small the difference
2837 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2838 * However, because tg->load_avg is a global value there are performance
2841 * In order to avoid having to look at the other cfs_rq's, we use a
2842 * differential update where we store the last value we propagated. This in
2843 * turn allows skipping updates if the differential is 'small'.
2845 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2846 * done) and effective_load() (which is not done because it is too costly).
2848 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2850 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2852 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2853 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2854 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2858 #else /* CONFIG_FAIR_GROUP_SCHED */
2859 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2860 #endif /* CONFIG_FAIR_GROUP_SCHED */
2862 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2864 if (&this_rq()->cfs == cfs_rq) {
2866 * There are a few boundary cases this might miss but it should
2867 * get called often enough that that should (hopefully) not be
2868 * a real problem -- added to that it only calls on the local
2869 * CPU, so if we enqueue remotely we'll miss an update, but
2870 * the next tick/schedule should update.
2872 * It will not get called when we go idle, because the idle
2873 * thread is a different class (!fair), nor will the utilization
2874 * number include things like RT tasks.
2876 * As is, the util number is not freq-invariant (we'd have to
2877 * implement arch_scale_freq_capacity() for that).
2881 cpufreq_update_util(rq_of(cfs_rq), 0);
2885 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2888 * Unsigned subtract and clamp on underflow.
2890 * Explicitly do a load-store to ensure the intermediate value never hits
2891 * memory. This allows lockless observations without ever seeing the negative
2894 #define sub_positive(_ptr, _val) do { \
2895 typeof(_ptr) ptr = (_ptr); \
2896 typeof(*ptr) val = (_val); \
2897 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2901 WRITE_ONCE(*ptr, res); \
2905 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2906 * @now: current time, as per cfs_rq_clock_task()
2907 * @cfs_rq: cfs_rq to update
2908 * @update_freq: should we call cfs_rq_util_change() or will the call do so
2910 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2911 * avg. The immediate corollary is that all (fair) tasks must be attached, see
2912 * post_init_entity_util_avg().
2914 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2916 * Returns true if the load decayed or we removed load.
2918 * Since both these conditions indicate a changed cfs_rq->avg.load we should
2919 * call update_tg_load_avg() when this function returns true.
2922 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2924 struct sched_avg *sa = &cfs_rq->avg;
2925 int decayed, removed = 0, removed_util = 0;
2927 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2928 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2929 sub_positive(&sa->load_avg, r);
2930 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2934 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2935 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2936 sub_positive(&sa->util_avg, r);
2937 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2941 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2942 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2944 #ifndef CONFIG_64BIT
2946 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2949 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2950 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2951 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2953 if (update_freq && (decayed || removed_util))
2954 cfs_rq_util_change(cfs_rq);
2956 return decayed || removed;
2960 * Optional action to be done while updating the load average
2962 #define UPDATE_TG 0x1
2963 #define SKIP_AGE_LOAD 0x2
2965 /* Update task and its cfs_rq load average */
2966 static inline void update_load_avg(struct sched_entity *se, int flags)
2968 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2969 u64 now = cfs_rq_clock_task(cfs_rq);
2970 int cpu = cpu_of(rq_of(cfs_rq));
2973 * Track task load average for carrying it to new CPU after migrated, and
2974 * track group sched_entity load average for task_h_load calc in migration
2976 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
2977 __update_load_avg(now, cpu, &se->avg,
2978 se->on_rq * scale_load_down(se->load.weight),
2979 cfs_rq->curr == se, NULL);
2982 if (update_cfs_rq_load_avg(now, cfs_rq, true) && (flags & UPDATE_TG))
2983 update_tg_load_avg(cfs_rq, 0);
2985 if (entity_is_task(se))
2986 trace_sched_load_avg_task(task_of(se), &se->avg);
2990 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2991 * @cfs_rq: cfs_rq to attach to
2992 * @se: sched_entity to attach
2994 * Must call update_cfs_rq_load_avg() before this, since we rely on
2995 * cfs_rq->avg.last_update_time being current.
2997 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2999 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3000 cfs_rq->avg.load_avg += se->avg.load_avg;
3001 cfs_rq->avg.load_sum += se->avg.load_sum;
3002 cfs_rq->avg.util_avg += se->avg.util_avg;
3003 cfs_rq->avg.util_sum += se->avg.util_sum;
3005 cfs_rq_util_change(cfs_rq);
3009 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3010 * @cfs_rq: cfs_rq to detach from
3011 * @se: sched_entity to detach
3013 * Must call update_cfs_rq_load_avg() before this, since we rely on
3014 * cfs_rq->avg.last_update_time being current.
3016 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3019 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3020 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3021 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3022 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3024 cfs_rq_util_change(cfs_rq);
3027 /* Add the load generated by se into cfs_rq's load average */
3029 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3031 struct sched_avg *sa = &se->avg;
3033 cfs_rq->runnable_load_avg += sa->load_avg;
3034 cfs_rq->runnable_load_sum += sa->load_sum;
3036 if (!sa->last_update_time) {
3037 attach_entity_load_avg(cfs_rq, se);
3038 update_tg_load_avg(cfs_rq, 0);
3042 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3044 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3046 cfs_rq->runnable_load_avg =
3047 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3048 cfs_rq->runnable_load_sum =
3049 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3052 #ifndef CONFIG_64BIT
3053 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3055 u64 last_update_time_copy;
3056 u64 last_update_time;
3059 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3061 last_update_time = cfs_rq->avg.last_update_time;
3062 } while (last_update_time != last_update_time_copy);
3064 return last_update_time;
3067 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3069 return cfs_rq->avg.last_update_time;
3074 * Synchronize entity load avg of dequeued entity without locking
3077 void sync_entity_load_avg(struct sched_entity *se)
3079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3080 u64 last_update_time;
3082 last_update_time = cfs_rq_last_update_time(cfs_rq);
3083 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3087 * Task first catches up with cfs_rq, and then subtract
3088 * itself from the cfs_rq (task must be off the queue now).
3090 void remove_entity_load_avg(struct sched_entity *se)
3092 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3095 * Newly created task or never used group entity should not be removed
3096 * from its (source) cfs_rq
3098 if (se->avg.last_update_time == 0)
3101 sync_entity_load_avg(se);
3102 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3103 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3107 * Update the rq's load with the elapsed running time before entering
3108 * idle. if the last scheduled task is not a CFS task, idle_enter will
3109 * be the only way to update the runnable statistic.
3111 void idle_enter_fair(struct rq *this_rq)
3116 * Update the rq's load with the elapsed idle time before a task is
3117 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3118 * be the only way to update the runnable statistic.
3120 void idle_exit_fair(struct rq *this_rq)
3124 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3126 return cfs_rq->runnable_load_avg;
3129 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3131 return cfs_rq->avg.load_avg;
3134 static int idle_balance(struct rq *this_rq);
3136 #else /* CONFIG_SMP */
3139 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3144 #define UPDATE_TG 0x0
3145 #define SKIP_AGE_LOAD 0x0
3147 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3149 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3151 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3152 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3155 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3157 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3159 static inline int idle_balance(struct rq *rq)
3164 #endif /* CONFIG_SMP */
3166 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3168 #ifdef CONFIG_SCHEDSTATS
3169 struct task_struct *tsk = NULL;
3171 if (entity_is_task(se))
3174 if (se->statistics.sleep_start) {
3175 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3180 if (unlikely(delta > se->statistics.sleep_max))
3181 se->statistics.sleep_max = delta;
3183 se->statistics.sleep_start = 0;
3184 se->statistics.sum_sleep_runtime += delta;
3187 account_scheduler_latency(tsk, delta >> 10, 1);
3188 trace_sched_stat_sleep(tsk, delta);
3191 if (se->statistics.block_start) {
3192 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3197 if (unlikely(delta > se->statistics.block_max))
3198 se->statistics.block_max = delta;
3200 se->statistics.block_start = 0;
3201 se->statistics.sum_sleep_runtime += delta;
3204 if (tsk->in_iowait) {
3205 se->statistics.iowait_sum += delta;
3206 se->statistics.iowait_count++;
3207 trace_sched_stat_iowait(tsk, delta);
3210 trace_sched_stat_blocked(tsk, delta);
3211 trace_sched_blocked_reason(tsk);
3214 * Blocking time is in units of nanosecs, so shift by
3215 * 20 to get a milliseconds-range estimation of the
3216 * amount of time that the task spent sleeping:
3218 if (unlikely(prof_on == SLEEP_PROFILING)) {
3219 profile_hits(SLEEP_PROFILING,
3220 (void *)get_wchan(tsk),
3223 account_scheduler_latency(tsk, delta >> 10, 0);
3229 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3231 #ifdef CONFIG_SCHED_DEBUG
3232 s64 d = se->vruntime - cfs_rq->min_vruntime;
3237 if (d > 3*sysctl_sched_latency)
3238 schedstat_inc(cfs_rq, nr_spread_over);
3243 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3245 u64 vruntime = cfs_rq->min_vruntime;
3248 * The 'current' period is already promised to the current tasks,
3249 * however the extra weight of the new task will slow them down a
3250 * little, place the new task so that it fits in the slot that
3251 * stays open at the end.
3253 if (initial && sched_feat(START_DEBIT))
3254 vruntime += sched_vslice(cfs_rq, se);
3256 /* sleeps up to a single latency don't count. */
3258 unsigned long thresh = sysctl_sched_latency;
3261 * Halve their sleep time's effect, to allow
3262 * for a gentler effect of sleepers:
3264 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3270 /* ensure we never gain time by being placed backwards. */
3271 se->vruntime = max_vruntime(se->vruntime, vruntime);
3274 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3277 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3280 * Update the normalized vruntime before updating min_vruntime
3281 * through calling update_curr().
3283 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3284 se->vruntime += cfs_rq->min_vruntime;
3287 * Update run-time statistics of the 'current'.
3289 update_curr(cfs_rq);
3290 update_load_avg(se, UPDATE_TG);
3291 enqueue_entity_load_avg(cfs_rq, se);
3292 account_entity_enqueue(cfs_rq, se);
3293 update_cfs_shares(cfs_rq);
3295 if (flags & ENQUEUE_WAKEUP) {
3296 place_entity(cfs_rq, se, 0);
3297 enqueue_sleeper(cfs_rq, se);
3300 update_stats_enqueue(cfs_rq, se);
3301 check_spread(cfs_rq, se);
3302 if (se != cfs_rq->curr)
3303 __enqueue_entity(cfs_rq, se);
3306 if (cfs_rq->nr_running == 1) {
3307 list_add_leaf_cfs_rq(cfs_rq);
3308 check_enqueue_throttle(cfs_rq);
3312 static void __clear_buddies_last(struct sched_entity *se)
3314 for_each_sched_entity(se) {
3315 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3316 if (cfs_rq->last != se)
3319 cfs_rq->last = NULL;
3323 static void __clear_buddies_next(struct sched_entity *se)
3325 for_each_sched_entity(se) {
3326 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3327 if (cfs_rq->next != se)
3330 cfs_rq->next = NULL;
3334 static void __clear_buddies_skip(struct sched_entity *se)
3336 for_each_sched_entity(se) {
3337 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3338 if (cfs_rq->skip != se)
3341 cfs_rq->skip = NULL;
3345 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3347 if (cfs_rq->last == se)
3348 __clear_buddies_last(se);
3350 if (cfs_rq->next == se)
3351 __clear_buddies_next(se);
3353 if (cfs_rq->skip == se)
3354 __clear_buddies_skip(se);
3357 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3360 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3363 * Update run-time statistics of the 'current'.
3365 update_curr(cfs_rq);
3366 update_load_avg(se, UPDATE_TG);
3367 dequeue_entity_load_avg(cfs_rq, se);
3369 update_stats_dequeue(cfs_rq, se);
3370 if (flags & DEQUEUE_SLEEP) {
3371 #ifdef CONFIG_SCHEDSTATS
3372 if (entity_is_task(se)) {
3373 struct task_struct *tsk = task_of(se);
3375 if (tsk->state & TASK_INTERRUPTIBLE)
3376 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3377 if (tsk->state & TASK_UNINTERRUPTIBLE)
3378 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3383 clear_buddies(cfs_rq, se);
3385 if (se != cfs_rq->curr)
3386 __dequeue_entity(cfs_rq, se);
3388 account_entity_dequeue(cfs_rq, se);
3391 * Normalize the entity after updating the min_vruntime because the
3392 * update can refer to the ->curr item and we need to reflect this
3393 * movement in our normalized position.
3395 if (!(flags & DEQUEUE_SLEEP))
3396 se->vruntime -= cfs_rq->min_vruntime;
3398 /* return excess runtime on last dequeue */
3399 return_cfs_rq_runtime(cfs_rq);
3401 update_min_vruntime(cfs_rq);
3402 update_cfs_shares(cfs_rq);
3406 * Preempt the current task with a newly woken task if needed:
3409 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3411 unsigned long ideal_runtime, delta_exec;
3412 struct sched_entity *se;
3415 ideal_runtime = sched_slice(cfs_rq, curr);
3416 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3417 if (delta_exec > ideal_runtime) {
3418 resched_curr(rq_of(cfs_rq));
3420 * The current task ran long enough, ensure it doesn't get
3421 * re-elected due to buddy favours.
3423 clear_buddies(cfs_rq, curr);
3428 * Ensure that a task that missed wakeup preemption by a
3429 * narrow margin doesn't have to wait for a full slice.
3430 * This also mitigates buddy induced latencies under load.
3432 if (delta_exec < sysctl_sched_min_granularity)
3435 se = __pick_first_entity(cfs_rq);
3436 delta = curr->vruntime - se->vruntime;
3441 if (delta > ideal_runtime)
3442 resched_curr(rq_of(cfs_rq));
3446 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3448 /* 'current' is not kept within the tree. */
3451 * Any task has to be enqueued before it get to execute on
3452 * a CPU. So account for the time it spent waiting on the
3455 update_stats_wait_end(cfs_rq, se);
3456 __dequeue_entity(cfs_rq, se);
3457 update_load_avg(se, UPDATE_TG);
3460 update_stats_curr_start(cfs_rq, se);
3462 #ifdef CONFIG_SCHEDSTATS
3464 * Track our maximum slice length, if the CPU's load is at
3465 * least twice that of our own weight (i.e. dont track it
3466 * when there are only lesser-weight tasks around):
3468 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3469 se->statistics.slice_max = max(se->statistics.slice_max,
3470 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3473 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3477 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3480 * Pick the next process, keeping these things in mind, in this order:
3481 * 1) keep things fair between processes/task groups
3482 * 2) pick the "next" process, since someone really wants that to run
3483 * 3) pick the "last" process, for cache locality
3484 * 4) do not run the "skip" process, if something else is available
3486 static struct sched_entity *
3487 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3489 struct sched_entity *left = __pick_first_entity(cfs_rq);
3490 struct sched_entity *se;
3493 * If curr is set we have to see if its left of the leftmost entity
3494 * still in the tree, provided there was anything in the tree at all.
3496 if (!left || (curr && entity_before(curr, left)))
3499 se = left; /* ideally we run the leftmost entity */
3502 * Avoid running the skip buddy, if running something else can
3503 * be done without getting too unfair.
3505 if (cfs_rq->skip == se) {
3506 struct sched_entity *second;
3509 second = __pick_first_entity(cfs_rq);
3511 second = __pick_next_entity(se);
3512 if (!second || (curr && entity_before(curr, second)))
3516 if (second && wakeup_preempt_entity(second, left) < 1)
3521 * Prefer last buddy, try to return the CPU to a preempted task.
3523 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3527 * Someone really wants this to run. If it's not unfair, run it.
3529 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3532 clear_buddies(cfs_rq, se);
3537 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3539 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3542 * If still on the runqueue then deactivate_task()
3543 * was not called and update_curr() has to be done:
3546 update_curr(cfs_rq);
3548 /* throttle cfs_rqs exceeding runtime */
3549 check_cfs_rq_runtime(cfs_rq);
3551 check_spread(cfs_rq, prev);
3553 update_stats_wait_start(cfs_rq, prev);
3554 /* Put 'current' back into the tree. */
3555 __enqueue_entity(cfs_rq, prev);
3556 /* in !on_rq case, update occurred at dequeue */
3557 update_load_avg(prev, 0);
3559 cfs_rq->curr = NULL;
3563 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3566 * Update run-time statistics of the 'current'.
3568 update_curr(cfs_rq);
3571 * Ensure that runnable average is periodically updated.
3573 update_load_avg(curr, UPDATE_TG);
3574 update_cfs_shares(cfs_rq);
3576 #ifdef CONFIG_SCHED_HRTICK
3578 * queued ticks are scheduled to match the slice, so don't bother
3579 * validating it and just reschedule.
3582 resched_curr(rq_of(cfs_rq));
3586 * don't let the period tick interfere with the hrtick preemption
3588 if (!sched_feat(DOUBLE_TICK) &&
3589 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3593 if (cfs_rq->nr_running > 1)
3594 check_preempt_tick(cfs_rq, curr);
3598 /**************************************************
3599 * CFS bandwidth control machinery
3602 #ifdef CONFIG_CFS_BANDWIDTH
3604 #ifdef HAVE_JUMP_LABEL
3605 static struct static_key __cfs_bandwidth_used;
3607 static inline bool cfs_bandwidth_used(void)
3609 return static_key_false(&__cfs_bandwidth_used);
3612 void cfs_bandwidth_usage_inc(void)
3614 static_key_slow_inc(&__cfs_bandwidth_used);
3617 void cfs_bandwidth_usage_dec(void)
3619 static_key_slow_dec(&__cfs_bandwidth_used);
3621 #else /* HAVE_JUMP_LABEL */
3622 static bool cfs_bandwidth_used(void)
3627 void cfs_bandwidth_usage_inc(void) {}
3628 void cfs_bandwidth_usage_dec(void) {}
3629 #endif /* HAVE_JUMP_LABEL */
3632 * default period for cfs group bandwidth.
3633 * default: 0.1s, units: nanoseconds
3635 static inline u64 default_cfs_period(void)
3637 return 100000000ULL;
3640 static inline u64 sched_cfs_bandwidth_slice(void)
3642 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3646 * Replenish runtime according to assigned quota and update expiration time.
3647 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3648 * additional synchronization around rq->lock.
3650 * requires cfs_b->lock
3652 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3656 if (cfs_b->quota == RUNTIME_INF)
3659 now = sched_clock_cpu(smp_processor_id());
3660 cfs_b->runtime = cfs_b->quota;
3661 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3664 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3666 return &tg->cfs_bandwidth;
3669 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3670 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3672 if (unlikely(cfs_rq->throttle_count))
3673 return cfs_rq->throttled_clock_task;
3675 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3678 /* returns 0 on failure to allocate runtime */
3679 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3681 struct task_group *tg = cfs_rq->tg;
3682 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3683 u64 amount = 0, min_amount, expires;
3685 /* note: this is a positive sum as runtime_remaining <= 0 */
3686 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3688 raw_spin_lock(&cfs_b->lock);
3689 if (cfs_b->quota == RUNTIME_INF)
3690 amount = min_amount;
3692 start_cfs_bandwidth(cfs_b);
3694 if (cfs_b->runtime > 0) {
3695 amount = min(cfs_b->runtime, min_amount);
3696 cfs_b->runtime -= amount;
3700 expires = cfs_b->runtime_expires;
3701 raw_spin_unlock(&cfs_b->lock);
3703 cfs_rq->runtime_remaining += amount;
3705 * we may have advanced our local expiration to account for allowed
3706 * spread between our sched_clock and the one on which runtime was
3709 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3710 cfs_rq->runtime_expires = expires;
3712 return cfs_rq->runtime_remaining > 0;
3716 * Note: This depends on the synchronization provided by sched_clock and the
3717 * fact that rq->clock snapshots this value.
3719 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3721 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3723 /* if the deadline is ahead of our clock, nothing to do */
3724 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3727 if (cfs_rq->runtime_remaining < 0)
3731 * If the local deadline has passed we have to consider the
3732 * possibility that our sched_clock is 'fast' and the global deadline
3733 * has not truly expired.
3735 * Fortunately we can check determine whether this the case by checking
3736 * whether the global deadline has advanced. It is valid to compare
3737 * cfs_b->runtime_expires without any locks since we only care about
3738 * exact equality, so a partial write will still work.
3741 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3742 /* extend local deadline, drift is bounded above by 2 ticks */
3743 cfs_rq->runtime_expires += TICK_NSEC;
3745 /* global deadline is ahead, expiration has passed */
3746 cfs_rq->runtime_remaining = 0;
3750 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3752 /* dock delta_exec before expiring quota (as it could span periods) */
3753 cfs_rq->runtime_remaining -= delta_exec;
3754 expire_cfs_rq_runtime(cfs_rq);
3756 if (likely(cfs_rq->runtime_remaining > 0))
3760 * if we're unable to extend our runtime we resched so that the active
3761 * hierarchy can be throttled
3763 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3764 resched_curr(rq_of(cfs_rq));
3767 static __always_inline
3768 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3770 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3773 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3776 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3778 return cfs_bandwidth_used() && cfs_rq->throttled;
3781 /* check whether cfs_rq, or any parent, is throttled */
3782 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3784 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3788 * Ensure that neither of the group entities corresponding to src_cpu or
3789 * dest_cpu are members of a throttled hierarchy when performing group
3790 * load-balance operations.
3792 static inline int throttled_lb_pair(struct task_group *tg,
3793 int src_cpu, int dest_cpu)
3795 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3797 src_cfs_rq = tg->cfs_rq[src_cpu];
3798 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3800 return throttled_hierarchy(src_cfs_rq) ||
3801 throttled_hierarchy(dest_cfs_rq);
3804 /* updated child weight may affect parent so we have to do this bottom up */
3805 static int tg_unthrottle_up(struct task_group *tg, void *data)
3807 struct rq *rq = data;
3808 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3810 cfs_rq->throttle_count--;
3812 if (!cfs_rq->throttle_count) {
3813 /* adjust cfs_rq_clock_task() */
3814 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3815 cfs_rq->throttled_clock_task;
3822 static int tg_throttle_down(struct task_group *tg, void *data)
3824 struct rq *rq = data;
3825 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3827 /* group is entering throttled state, stop time */
3828 if (!cfs_rq->throttle_count)
3829 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3830 cfs_rq->throttle_count++;
3835 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3837 struct rq *rq = rq_of(cfs_rq);
3838 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3839 struct sched_entity *se;
3840 long task_delta, dequeue = 1;
3843 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3845 /* freeze hierarchy runnable averages while throttled */
3847 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3850 task_delta = cfs_rq->h_nr_running;
3851 for_each_sched_entity(se) {
3852 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3853 /* throttled entity or throttle-on-deactivate */
3858 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3859 qcfs_rq->h_nr_running -= task_delta;
3861 if (qcfs_rq->load.weight)
3866 sub_nr_running(rq, task_delta);
3868 cfs_rq->throttled = 1;
3869 cfs_rq->throttled_clock = rq_clock(rq);
3870 raw_spin_lock(&cfs_b->lock);
3871 empty = list_empty(&cfs_b->throttled_cfs_rq);
3874 * Add to the _head_ of the list, so that an already-started
3875 * distribute_cfs_runtime will not see us
3877 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3880 * If we're the first throttled task, make sure the bandwidth
3884 start_cfs_bandwidth(cfs_b);
3886 raw_spin_unlock(&cfs_b->lock);
3889 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3891 struct rq *rq = rq_of(cfs_rq);
3892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3893 struct sched_entity *se;
3897 se = cfs_rq->tg->se[cpu_of(rq)];
3899 cfs_rq->throttled = 0;
3901 update_rq_clock(rq);
3903 raw_spin_lock(&cfs_b->lock);
3904 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3905 list_del_rcu(&cfs_rq->throttled_list);
3906 raw_spin_unlock(&cfs_b->lock);
3908 /* update hierarchical throttle state */
3909 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3911 if (!cfs_rq->load.weight)
3914 task_delta = cfs_rq->h_nr_running;
3915 for_each_sched_entity(se) {
3919 cfs_rq = cfs_rq_of(se);
3921 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3922 cfs_rq->h_nr_running += task_delta;
3924 if (cfs_rq_throttled(cfs_rq))
3929 add_nr_running(rq, task_delta);
3931 /* determine whether we need to wake up potentially idle cpu */
3932 if (rq->curr == rq->idle && rq->cfs.nr_running)
3936 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3937 u64 remaining, u64 expires)
3939 struct cfs_rq *cfs_rq;
3941 u64 starting_runtime = remaining;
3944 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3946 struct rq *rq = rq_of(cfs_rq);
3948 raw_spin_lock(&rq->lock);
3949 if (!cfs_rq_throttled(cfs_rq))
3952 runtime = -cfs_rq->runtime_remaining + 1;
3953 if (runtime > remaining)
3954 runtime = remaining;
3955 remaining -= runtime;
3957 cfs_rq->runtime_remaining += runtime;
3958 cfs_rq->runtime_expires = expires;
3960 /* we check whether we're throttled above */
3961 if (cfs_rq->runtime_remaining > 0)
3962 unthrottle_cfs_rq(cfs_rq);
3965 raw_spin_unlock(&rq->lock);
3972 return starting_runtime - remaining;
3976 * Responsible for refilling a task_group's bandwidth and unthrottling its
3977 * cfs_rqs as appropriate. If there has been no activity within the last
3978 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3979 * used to track this state.
3981 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3983 u64 runtime, runtime_expires;
3986 /* no need to continue the timer with no bandwidth constraint */
3987 if (cfs_b->quota == RUNTIME_INF)
3988 goto out_deactivate;
3990 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3991 cfs_b->nr_periods += overrun;
3994 * idle depends on !throttled (for the case of a large deficit), and if
3995 * we're going inactive then everything else can be deferred
3997 if (cfs_b->idle && !throttled)
3998 goto out_deactivate;
4000 __refill_cfs_bandwidth_runtime(cfs_b);
4003 /* mark as potentially idle for the upcoming period */
4008 /* account preceding periods in which throttling occurred */
4009 cfs_b->nr_throttled += overrun;
4011 runtime_expires = cfs_b->runtime_expires;
4014 * This check is repeated as we are holding onto the new bandwidth while
4015 * we unthrottle. This can potentially race with an unthrottled group
4016 * trying to acquire new bandwidth from the global pool. This can result
4017 * in us over-using our runtime if it is all used during this loop, but
4018 * only by limited amounts in that extreme case.
4020 while (throttled && cfs_b->runtime > 0) {
4021 runtime = cfs_b->runtime;
4022 raw_spin_unlock(&cfs_b->lock);
4023 /* we can't nest cfs_b->lock while distributing bandwidth */
4024 runtime = distribute_cfs_runtime(cfs_b, runtime,
4026 raw_spin_lock(&cfs_b->lock);
4028 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4030 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4034 * While we are ensured activity in the period following an
4035 * unthrottle, this also covers the case in which the new bandwidth is
4036 * insufficient to cover the existing bandwidth deficit. (Forcing the
4037 * timer to remain active while there are any throttled entities.)
4047 /* a cfs_rq won't donate quota below this amount */
4048 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4049 /* minimum remaining period time to redistribute slack quota */
4050 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4051 /* how long we wait to gather additional slack before distributing */
4052 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4055 * Are we near the end of the current quota period?
4057 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4058 * hrtimer base being cleared by hrtimer_start. In the case of
4059 * migrate_hrtimers, base is never cleared, so we are fine.
4061 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4063 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4066 /* if the call-back is running a quota refresh is already occurring */
4067 if (hrtimer_callback_running(refresh_timer))
4070 /* is a quota refresh about to occur? */
4071 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4072 if (remaining < min_expire)
4078 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4080 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4082 /* if there's a quota refresh soon don't bother with slack */
4083 if (runtime_refresh_within(cfs_b, min_left))
4086 hrtimer_start(&cfs_b->slack_timer,
4087 ns_to_ktime(cfs_bandwidth_slack_period),
4091 /* we know any runtime found here is valid as update_curr() precedes return */
4092 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4094 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4095 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4097 if (slack_runtime <= 0)
4100 raw_spin_lock(&cfs_b->lock);
4101 if (cfs_b->quota != RUNTIME_INF &&
4102 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4103 cfs_b->runtime += slack_runtime;
4105 /* we are under rq->lock, defer unthrottling using a timer */
4106 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4107 !list_empty(&cfs_b->throttled_cfs_rq))
4108 start_cfs_slack_bandwidth(cfs_b);
4110 raw_spin_unlock(&cfs_b->lock);
4112 /* even if it's not valid for return we don't want to try again */
4113 cfs_rq->runtime_remaining -= slack_runtime;
4116 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4118 if (!cfs_bandwidth_used())
4121 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4124 __return_cfs_rq_runtime(cfs_rq);
4128 * This is done with a timer (instead of inline with bandwidth return) since
4129 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4131 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4133 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4136 /* confirm we're still not at a refresh boundary */
4137 raw_spin_lock(&cfs_b->lock);
4138 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4139 raw_spin_unlock(&cfs_b->lock);
4143 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4144 runtime = cfs_b->runtime;
4146 expires = cfs_b->runtime_expires;
4147 raw_spin_unlock(&cfs_b->lock);
4152 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4154 raw_spin_lock(&cfs_b->lock);
4155 if (expires == cfs_b->runtime_expires)
4156 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4157 raw_spin_unlock(&cfs_b->lock);
4161 * When a group wakes up we want to make sure that its quota is not already
4162 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4163 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4165 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4167 if (!cfs_bandwidth_used())
4170 /* Synchronize hierarchical throttle counter: */
4171 if (unlikely(!cfs_rq->throttle_uptodate)) {
4172 struct rq *rq = rq_of(cfs_rq);
4173 struct cfs_rq *pcfs_rq;
4174 struct task_group *tg;
4176 cfs_rq->throttle_uptodate = 1;
4178 /* Get closest up-to-date node, because leaves go first: */
4179 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4180 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4181 if (pcfs_rq->throttle_uptodate)
4185 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4186 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4190 /* an active group must be handled by the update_curr()->put() path */
4191 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4194 /* ensure the group is not already throttled */
4195 if (cfs_rq_throttled(cfs_rq))
4198 /* update runtime allocation */
4199 account_cfs_rq_runtime(cfs_rq, 0);
4200 if (cfs_rq->runtime_remaining <= 0)
4201 throttle_cfs_rq(cfs_rq);
4204 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4205 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4207 if (!cfs_bandwidth_used())
4210 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4214 * it's possible for a throttled entity to be forced into a running
4215 * state (e.g. set_curr_task), in this case we're finished.
4217 if (cfs_rq_throttled(cfs_rq))
4220 throttle_cfs_rq(cfs_rq);
4224 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4226 struct cfs_bandwidth *cfs_b =
4227 container_of(timer, struct cfs_bandwidth, slack_timer);
4229 do_sched_cfs_slack_timer(cfs_b);
4231 return HRTIMER_NORESTART;
4234 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4236 struct cfs_bandwidth *cfs_b =
4237 container_of(timer, struct cfs_bandwidth, period_timer);
4241 raw_spin_lock(&cfs_b->lock);
4243 overrun = hrtimer_forward_now(timer, cfs_b->period);
4247 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4250 cfs_b->period_active = 0;
4251 raw_spin_unlock(&cfs_b->lock);
4253 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4256 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4258 raw_spin_lock_init(&cfs_b->lock);
4260 cfs_b->quota = RUNTIME_INF;
4261 cfs_b->period = ns_to_ktime(default_cfs_period());
4263 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4264 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4265 cfs_b->period_timer.function = sched_cfs_period_timer;
4266 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4267 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4270 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4272 cfs_rq->runtime_enabled = 0;
4273 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4276 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4278 lockdep_assert_held(&cfs_b->lock);
4280 if (!cfs_b->period_active) {
4281 cfs_b->period_active = 1;
4282 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4283 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4287 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4289 /* init_cfs_bandwidth() was not called */
4290 if (!cfs_b->throttled_cfs_rq.next)
4293 hrtimer_cancel(&cfs_b->period_timer);
4294 hrtimer_cancel(&cfs_b->slack_timer);
4297 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4299 struct cfs_rq *cfs_rq;
4301 for_each_leaf_cfs_rq(rq, cfs_rq) {
4302 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4304 raw_spin_lock(&cfs_b->lock);
4305 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4306 raw_spin_unlock(&cfs_b->lock);
4310 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4312 struct cfs_rq *cfs_rq;
4314 for_each_leaf_cfs_rq(rq, cfs_rq) {
4315 if (!cfs_rq->runtime_enabled)
4319 * clock_task is not advancing so we just need to make sure
4320 * there's some valid quota amount
4322 cfs_rq->runtime_remaining = 1;
4324 * Offline rq is schedulable till cpu is completely disabled
4325 * in take_cpu_down(), so we prevent new cfs throttling here.
4327 cfs_rq->runtime_enabled = 0;
4329 if (cfs_rq_throttled(cfs_rq))
4330 unthrottle_cfs_rq(cfs_rq);
4334 #else /* CONFIG_CFS_BANDWIDTH */
4335 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4337 return rq_clock_task(rq_of(cfs_rq));
4340 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4341 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4342 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4343 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4345 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4350 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4355 static inline int throttled_lb_pair(struct task_group *tg,
4356 int src_cpu, int dest_cpu)
4361 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4363 #ifdef CONFIG_FAIR_GROUP_SCHED
4364 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4367 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4371 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4372 static inline void update_runtime_enabled(struct rq *rq) {}
4373 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4375 #endif /* CONFIG_CFS_BANDWIDTH */
4377 /**************************************************
4378 * CFS operations on tasks:
4381 #ifdef CONFIG_SCHED_HRTICK
4382 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4384 struct sched_entity *se = &p->se;
4385 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4387 WARN_ON(task_rq(p) != rq);
4389 if (cfs_rq->nr_running > 1) {
4390 u64 slice = sched_slice(cfs_rq, se);
4391 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4392 s64 delta = slice - ran;
4399 hrtick_start(rq, delta);
4404 * called from enqueue/dequeue and updates the hrtick when the
4405 * current task is from our class and nr_running is low enough
4408 static void hrtick_update(struct rq *rq)
4410 struct task_struct *curr = rq->curr;
4412 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4415 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4416 hrtick_start_fair(rq, curr);
4418 #else /* !CONFIG_SCHED_HRTICK */
4420 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4424 static inline void hrtick_update(struct rq *rq)
4430 static bool cpu_overutilized(int cpu);
4431 unsigned long boosted_cpu_util(int cpu);
4433 #define boosted_cpu_util(cpu) cpu_util(cpu)
4437 static void update_capacity_of(int cpu)
4439 unsigned long req_cap;
4444 /* Convert scale-invariant capacity to cpu. */
4445 req_cap = boosted_cpu_util(cpu);
4446 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4447 set_cfs_cpu_capacity(cpu, true, req_cap);
4452 * The enqueue_task method is called before nr_running is
4453 * increased. Here we update the fair scheduling stats and
4454 * then put the task into the rbtree:
4457 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4459 struct cfs_rq *cfs_rq;
4460 struct sched_entity *se = &p->se;
4462 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4463 int task_wakeup = flags & ENQUEUE_WAKEUP;
4467 * If in_iowait is set, the code below may not trigger any cpufreq
4468 * utilization updates, so do it here explicitly with the IOWAIT flag
4472 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4474 for_each_sched_entity(se) {
4477 cfs_rq = cfs_rq_of(se);
4478 enqueue_entity(cfs_rq, se, flags);
4481 * end evaluation on encountering a throttled cfs_rq
4483 * note: in the case of encountering a throttled cfs_rq we will
4484 * post the final h_nr_running increment below.
4486 if (cfs_rq_throttled(cfs_rq))
4488 cfs_rq->h_nr_running++;
4489 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4491 flags = ENQUEUE_WAKEUP;
4494 for_each_sched_entity(se) {
4495 cfs_rq = cfs_rq_of(se);
4496 cfs_rq->h_nr_running++;
4497 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4499 if (cfs_rq_throttled(cfs_rq))
4502 update_load_avg(se, UPDATE_TG);
4503 update_cfs_shares(cfs_rq);
4507 add_nr_running(rq, 1);
4512 * Update SchedTune accounting.
4514 * We do it before updating the CPU capacity to ensure the
4515 * boost value of the current task is accounted for in the
4516 * selection of the OPP.
4518 * We do it also in the case where we enqueue a throttled task;
4519 * we could argue that a throttled task should not boost a CPU,
4521 * a) properly implementing CPU boosting considering throttled
4522 * tasks will increase a lot the complexity of the solution
4523 * b) it's not easy to quantify the benefits introduced by
4524 * such a more complex solution.
4525 * Thus, for the time being we go for the simple solution and boost
4526 * also for throttled RQs.
4528 schedtune_enqueue_task(p, cpu_of(rq));
4531 walt_inc_cumulative_runnable_avg(rq, p);
4532 if (!task_new && !rq->rd->overutilized &&
4533 cpu_overutilized(rq->cpu)) {
4534 rq->rd->overutilized = true;
4535 trace_sched_overutilized(true);
4539 * We want to potentially trigger a freq switch
4540 * request only for tasks that are waking up; this is
4541 * because we get here also during load balancing, but
4542 * in these cases it seems wise to trigger as single
4543 * request after load balancing is done.
4545 if (task_new || task_wakeup)
4546 update_capacity_of(cpu_of(rq));
4549 #endif /* CONFIG_SMP */
4553 static void set_next_buddy(struct sched_entity *se);
4556 * The dequeue_task method is called before nr_running is
4557 * decreased. We remove the task from the rbtree and
4558 * update the fair scheduling stats:
4560 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4562 struct cfs_rq *cfs_rq;
4563 struct sched_entity *se = &p->se;
4564 int task_sleep = flags & DEQUEUE_SLEEP;
4566 for_each_sched_entity(se) {
4567 cfs_rq = cfs_rq_of(se);
4568 dequeue_entity(cfs_rq, se, flags);
4571 * end evaluation on encountering a throttled cfs_rq
4573 * note: in the case of encountering a throttled cfs_rq we will
4574 * post the final h_nr_running decrement below.
4576 if (cfs_rq_throttled(cfs_rq))
4578 cfs_rq->h_nr_running--;
4579 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4581 /* Don't dequeue parent if it has other entities besides us */
4582 if (cfs_rq->load.weight) {
4583 /* Avoid re-evaluating load for this entity: */
4584 se = parent_entity(se);
4586 * Bias pick_next to pick a task from this cfs_rq, as
4587 * p is sleeping when it is within its sched_slice.
4589 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4593 flags |= DEQUEUE_SLEEP;
4596 for_each_sched_entity(se) {
4597 cfs_rq = cfs_rq_of(se);
4598 cfs_rq->h_nr_running--;
4599 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4601 if (cfs_rq_throttled(cfs_rq))
4604 update_load_avg(se, UPDATE_TG);
4605 update_cfs_shares(cfs_rq);
4609 sub_nr_running(rq, 1);
4614 * Update SchedTune accounting
4616 * We do it before updating the CPU capacity to ensure the
4617 * boost value of the current task is accounted for in the
4618 * selection of the OPP.
4620 schedtune_dequeue_task(p, cpu_of(rq));
4623 walt_dec_cumulative_runnable_avg(rq, p);
4626 * We want to potentially trigger a freq switch
4627 * request only for tasks that are going to sleep;
4628 * this is because we get here also during load
4629 * balancing, but in these cases it seems wise to
4630 * trigger as single request after load balancing is
4634 if (rq->cfs.nr_running)
4635 update_capacity_of(cpu_of(rq));
4636 else if (sched_freq())
4637 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4641 #endif /* CONFIG_SMP */
4649 * per rq 'load' arrray crap; XXX kill this.
4653 * The exact cpuload at various idx values, calculated at every tick would be
4654 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4656 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4657 * on nth tick when cpu may be busy, then we have:
4658 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4659 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4661 * decay_load_missed() below does efficient calculation of
4662 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4663 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4665 * The calculation is approximated on a 128 point scale.
4666 * degrade_zero_ticks is the number of ticks after which load at any
4667 * particular idx is approximated to be zero.
4668 * degrade_factor is a precomputed table, a row for each load idx.
4669 * Each column corresponds to degradation factor for a power of two ticks,
4670 * based on 128 point scale.
4672 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4673 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4675 * With this power of 2 load factors, we can degrade the load n times
4676 * by looking at 1 bits in n and doing as many mult/shift instead of
4677 * n mult/shifts needed by the exact degradation.
4679 #define DEGRADE_SHIFT 7
4680 static const unsigned char
4681 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4682 static const unsigned char
4683 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4684 {0, 0, 0, 0, 0, 0, 0, 0},
4685 {64, 32, 8, 0, 0, 0, 0, 0},
4686 {96, 72, 40, 12, 1, 0, 0},
4687 {112, 98, 75, 43, 15, 1, 0},
4688 {120, 112, 98, 76, 45, 16, 2} };
4691 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4692 * would be when CPU is idle and so we just decay the old load without
4693 * adding any new load.
4695 static unsigned long
4696 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4700 if (!missed_updates)
4703 if (missed_updates >= degrade_zero_ticks[idx])
4707 return load >> missed_updates;
4709 while (missed_updates) {
4710 if (missed_updates % 2)
4711 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4713 missed_updates >>= 1;
4720 * Update rq->cpu_load[] statistics. This function is usually called every
4721 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4722 * every tick. We fix it up based on jiffies.
4724 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4725 unsigned long pending_updates)
4729 this_rq->nr_load_updates++;
4731 /* Update our load: */
4732 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4733 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4734 unsigned long old_load, new_load;
4736 /* scale is effectively 1 << i now, and >> i divides by scale */
4738 old_load = this_rq->cpu_load[i];
4739 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4740 new_load = this_load;
4742 * Round up the averaging division if load is increasing. This
4743 * prevents us from getting stuck on 9 if the load is 10, for
4746 if (new_load > old_load)
4747 new_load += scale - 1;
4749 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4752 sched_avg_update(this_rq);
4755 /* Used instead of source_load when we know the type == 0 */
4756 static unsigned long weighted_cpuload(const int cpu)
4758 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4761 #ifdef CONFIG_NO_HZ_COMMON
4763 * There is no sane way to deal with nohz on smp when using jiffies because the
4764 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4765 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4767 * Therefore we cannot use the delta approach from the regular tick since that
4768 * would seriously skew the load calculation. However we'll make do for those
4769 * updates happening while idle (nohz_idle_balance) or coming out of idle
4770 * (tick_nohz_idle_exit).
4772 * This means we might still be one tick off for nohz periods.
4776 * Called from nohz_idle_balance() to update the load ratings before doing the
4779 static void update_idle_cpu_load(struct rq *this_rq)
4781 unsigned long curr_jiffies = READ_ONCE(jiffies);
4782 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4783 unsigned long pending_updates;
4786 * bail if there's load or we're actually up-to-date.
4788 if (load || curr_jiffies == this_rq->last_load_update_tick)
4791 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4792 this_rq->last_load_update_tick = curr_jiffies;
4794 __update_cpu_load(this_rq, load, pending_updates);
4798 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4800 void update_cpu_load_nohz(void)
4802 struct rq *this_rq = this_rq();
4803 unsigned long curr_jiffies = READ_ONCE(jiffies);
4804 unsigned long pending_updates;
4806 if (curr_jiffies == this_rq->last_load_update_tick)
4809 raw_spin_lock(&this_rq->lock);
4810 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4811 if (pending_updates) {
4812 this_rq->last_load_update_tick = curr_jiffies;
4814 * We were idle, this means load 0, the current load might be
4815 * !0 due to remote wakeups and the sort.
4817 __update_cpu_load(this_rq, 0, pending_updates);
4819 raw_spin_unlock(&this_rq->lock);
4821 #endif /* CONFIG_NO_HZ */
4824 * Called from scheduler_tick()
4826 void update_cpu_load_active(struct rq *this_rq)
4828 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4830 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4832 this_rq->last_load_update_tick = jiffies;
4833 __update_cpu_load(this_rq, load, 1);
4837 * Return a low guess at the load of a migration-source cpu weighted
4838 * according to the scheduling class and "nice" value.
4840 * We want to under-estimate the load of migration sources, to
4841 * balance conservatively.
4843 static unsigned long source_load(int cpu, int type)
4845 struct rq *rq = cpu_rq(cpu);
4846 unsigned long total = weighted_cpuload(cpu);
4848 if (type == 0 || !sched_feat(LB_BIAS))
4851 return min(rq->cpu_load[type-1], total);
4855 * Return a high guess at the load of a migration-target cpu weighted
4856 * according to the scheduling class and "nice" value.
4858 static unsigned long target_load(int cpu, int type)
4860 struct rq *rq = cpu_rq(cpu);
4861 unsigned long total = weighted_cpuload(cpu);
4863 if (type == 0 || !sched_feat(LB_BIAS))
4866 return max(rq->cpu_load[type-1], total);
4870 static unsigned long cpu_avg_load_per_task(int cpu)
4872 struct rq *rq = cpu_rq(cpu);
4873 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4874 unsigned long load_avg = weighted_cpuload(cpu);
4877 return load_avg / nr_running;
4882 static void record_wakee(struct task_struct *p)
4885 * Rough decay (wiping) for cost saving, don't worry
4886 * about the boundary, really active task won't care
4889 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4890 current->wakee_flips >>= 1;
4891 current->wakee_flip_decay_ts = jiffies;
4894 if (current->last_wakee != p) {
4895 current->last_wakee = p;
4896 current->wakee_flips++;
4900 static void task_waking_fair(struct task_struct *p)
4902 struct sched_entity *se = &p->se;
4903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4906 #ifndef CONFIG_64BIT
4907 u64 min_vruntime_copy;
4910 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4912 min_vruntime = cfs_rq->min_vruntime;
4913 } while (min_vruntime != min_vruntime_copy);
4915 min_vruntime = cfs_rq->min_vruntime;
4918 se->vruntime -= min_vruntime;
4922 #ifdef CONFIG_FAIR_GROUP_SCHED
4924 * effective_load() calculates the load change as seen from the root_task_group
4926 * Adding load to a group doesn't make a group heavier, but can cause movement
4927 * of group shares between cpus. Assuming the shares were perfectly aligned one
4928 * can calculate the shift in shares.
4930 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4931 * on this @cpu and results in a total addition (subtraction) of @wg to the
4932 * total group weight.
4934 * Given a runqueue weight distribution (rw_i) we can compute a shares
4935 * distribution (s_i) using:
4937 * s_i = rw_i / \Sum rw_j (1)
4939 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4940 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4941 * shares distribution (s_i):
4943 * rw_i = { 2, 4, 1, 0 }
4944 * s_i = { 2/7, 4/7, 1/7, 0 }
4946 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4947 * task used to run on and the CPU the waker is running on), we need to
4948 * compute the effect of waking a task on either CPU and, in case of a sync
4949 * wakeup, compute the effect of the current task going to sleep.
4951 * So for a change of @wl to the local @cpu with an overall group weight change
4952 * of @wl we can compute the new shares distribution (s'_i) using:
4954 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4956 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4957 * differences in waking a task to CPU 0. The additional task changes the
4958 * weight and shares distributions like:
4960 * rw'_i = { 3, 4, 1, 0 }
4961 * s'_i = { 3/8, 4/8, 1/8, 0 }
4963 * We can then compute the difference in effective weight by using:
4965 * dw_i = S * (s'_i - s_i) (3)
4967 * Where 'S' is the group weight as seen by its parent.
4969 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4970 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4971 * 4/7) times the weight of the group.
4973 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4975 struct sched_entity *se = tg->se[cpu];
4977 if (!tg->parent) /* the trivial, non-cgroup case */
4980 for_each_sched_entity(se) {
4981 struct cfs_rq *cfs_rq = se->my_q;
4982 long W, w = cfs_rq_load_avg(cfs_rq);
4987 * W = @wg + \Sum rw_j
4989 W = wg + atomic_long_read(&tg->load_avg);
4991 /* Ensure \Sum rw_j >= rw_i */
4992 W -= cfs_rq->tg_load_avg_contrib;
5001 * wl = S * s'_i; see (2)
5004 wl = (w * (long)tg->shares) / W;
5009 * Per the above, wl is the new se->load.weight value; since
5010 * those are clipped to [MIN_SHARES, ...) do so now. See
5011 * calc_cfs_shares().
5013 if (wl < MIN_SHARES)
5017 * wl = dw_i = S * (s'_i - s_i); see (3)
5019 wl -= se->avg.load_avg;
5022 * Recursively apply this logic to all parent groups to compute
5023 * the final effective load change on the root group. Since
5024 * only the @tg group gets extra weight, all parent groups can
5025 * only redistribute existing shares. @wl is the shift in shares
5026 * resulting from this level per the above.
5035 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5043 * Returns the current capacity of cpu after applying both
5044 * cpu and freq scaling.
5046 unsigned long capacity_curr_of(int cpu)
5048 return cpu_rq(cpu)->cpu_capacity_orig *
5049 arch_scale_freq_capacity(NULL, cpu)
5050 >> SCHED_CAPACITY_SHIFT;
5053 static inline bool energy_aware(void)
5055 return sched_feat(ENERGY_AWARE);
5059 struct sched_group *sg_top;
5060 struct sched_group *sg_cap;
5067 struct task_struct *task;
5082 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5083 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5084 * energy calculations. Using the scale-invariant util returned by
5085 * cpu_util() and approximating scale-invariant util by:
5087 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5089 * the normalized util can be found using the specific capacity.
5091 * capacity = capacity_orig * curr_freq/max_freq
5093 * norm_util = running_time/time ~ util/capacity
5095 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5097 int util = __cpu_util(cpu, delta);
5099 if (util >= capacity)
5100 return SCHED_CAPACITY_SCALE;
5102 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5105 static int calc_util_delta(struct energy_env *eenv, int cpu)
5107 if (cpu == eenv->src_cpu)
5108 return -eenv->util_delta;
5109 if (cpu == eenv->dst_cpu)
5110 return eenv->util_delta;
5115 unsigned long group_max_util(struct energy_env *eenv)
5118 unsigned long max_util = 0;
5120 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5121 delta = calc_util_delta(eenv, i);
5122 max_util = max(max_util, __cpu_util(i, delta));
5129 * group_norm_util() returns the approximated group util relative to it's
5130 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5131 * energy calculations. Since task executions may or may not overlap in time in
5132 * the group the true normalized util is between max(cpu_norm_util(i)) and
5133 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5134 * latter is used as the estimate as it leads to a more pessimistic energy
5135 * estimate (more busy).
5138 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5141 unsigned long util_sum = 0;
5142 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5144 for_each_cpu(i, sched_group_cpus(sg)) {
5145 delta = calc_util_delta(eenv, i);
5146 util_sum += __cpu_norm_util(i, capacity, delta);
5149 if (util_sum > SCHED_CAPACITY_SCALE)
5150 return SCHED_CAPACITY_SCALE;
5154 static int find_new_capacity(struct energy_env *eenv,
5155 const struct sched_group_energy * const sge)
5158 unsigned long util = group_max_util(eenv);
5160 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5161 if (sge->cap_states[idx].cap >= util)
5165 eenv->cap_idx = idx;
5170 static int group_idle_state(struct sched_group *sg)
5172 int i, state = INT_MAX;
5174 /* Find the shallowest idle state in the sched group. */
5175 for_each_cpu(i, sched_group_cpus(sg))
5176 state = min(state, idle_get_state_idx(cpu_rq(i)));
5178 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5185 * sched_group_energy(): Computes the absolute energy consumption of cpus
5186 * belonging to the sched_group including shared resources shared only by
5187 * members of the group. Iterates over all cpus in the hierarchy below the
5188 * sched_group starting from the bottom working it's way up before going to
5189 * the next cpu until all cpus are covered at all levels. The current
5190 * implementation is likely to gather the same util statistics multiple times.
5191 * This can probably be done in a faster but more complex way.
5192 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5194 static int sched_group_energy(struct energy_env *eenv)
5196 struct sched_domain *sd;
5197 int cpu, total_energy = 0;
5198 struct cpumask visit_cpus;
5199 struct sched_group *sg;
5201 WARN_ON(!eenv->sg_top->sge);
5203 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5205 while (!cpumask_empty(&visit_cpus)) {
5206 struct sched_group *sg_shared_cap = NULL;
5208 cpu = cpumask_first(&visit_cpus);
5211 * Is the group utilization affected by cpus outside this
5214 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5218 * We most probably raced with hotplug; returning a
5219 * wrong energy estimation is better than entering an
5225 sg_shared_cap = sd->parent->groups;
5227 for_each_domain(cpu, sd) {
5230 /* Has this sched_domain already been visited? */
5231 if (sd->child && group_first_cpu(sg) != cpu)
5235 unsigned long group_util;
5236 int sg_busy_energy, sg_idle_energy;
5237 int cap_idx, idle_idx;
5239 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5240 eenv->sg_cap = sg_shared_cap;
5244 cap_idx = find_new_capacity(eenv, sg->sge);
5246 if (sg->group_weight == 1) {
5247 /* Remove capacity of src CPU (before task move) */
5248 if (eenv->util_delta == 0 &&
5249 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5250 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5251 eenv->cap.delta -= eenv->cap.before;
5253 /* Add capacity of dst CPU (after task move) */
5254 if (eenv->util_delta != 0 &&
5255 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5256 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5257 eenv->cap.delta += eenv->cap.after;
5261 idle_idx = group_idle_state(sg);
5262 group_util = group_norm_util(eenv, sg);
5263 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5264 >> SCHED_CAPACITY_SHIFT;
5265 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5266 * sg->sge->idle_states[idle_idx].power)
5267 >> SCHED_CAPACITY_SHIFT;
5269 total_energy += sg_busy_energy + sg_idle_energy;
5272 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5274 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5277 } while (sg = sg->next, sg != sd->groups);
5280 cpumask_clear_cpu(cpu, &visit_cpus);
5284 eenv->energy = total_energy;
5288 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5290 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5294 * energy_diff(): Estimate the energy impact of changing the utilization
5295 * distribution. eenv specifies the change: utilisation amount, source, and
5296 * destination cpu. Source or destination cpu may be -1 in which case the
5297 * utilization is removed from or added to the system (e.g. task wake-up). If
5298 * both are specified, the utilization is migrated.
5300 static inline int __energy_diff(struct energy_env *eenv)
5302 struct sched_domain *sd;
5303 struct sched_group *sg;
5304 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5307 struct energy_env eenv_before = {
5309 .src_cpu = eenv->src_cpu,
5310 .dst_cpu = eenv->dst_cpu,
5311 .nrg = { 0, 0, 0, 0},
5315 if (eenv->src_cpu == eenv->dst_cpu)
5318 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5319 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5322 return 0; /* Error */
5327 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5328 eenv_before.sg_top = eenv->sg_top = sg;
5330 if (sched_group_energy(&eenv_before))
5331 return 0; /* Invalid result abort */
5332 energy_before += eenv_before.energy;
5334 /* Keep track of SRC cpu (before) capacity */
5335 eenv->cap.before = eenv_before.cap.before;
5336 eenv->cap.delta = eenv_before.cap.delta;
5338 if (sched_group_energy(eenv))
5339 return 0; /* Invalid result abort */
5340 energy_after += eenv->energy;
5342 } while (sg = sg->next, sg != sd->groups);
5344 eenv->nrg.before = energy_before;
5345 eenv->nrg.after = energy_after;
5346 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5349 trace_sched_energy_diff(eenv->task,
5350 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5351 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5352 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5353 eenv->nrg.delta, eenv->payoff);
5356 * Dead-zone margin preventing too many migrations.
5359 margin = eenv->nrg.before >> 6; /* ~1.56% */
5361 diff = eenv->nrg.after - eenv->nrg.before;
5363 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5365 return eenv->nrg.diff;
5368 #ifdef CONFIG_SCHED_TUNE
5370 struct target_nrg schedtune_target_nrg;
5373 * System energy normalization
5374 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5375 * corresponding to the specified energy variation.
5378 normalize_energy(int energy_diff)
5381 #ifdef CONFIG_SCHED_DEBUG
5384 /* Check for boundaries */
5385 max_delta = schedtune_target_nrg.max_power;
5386 max_delta -= schedtune_target_nrg.min_power;
5387 WARN_ON(abs(energy_diff) >= max_delta);
5390 /* Do scaling using positive numbers to increase the range */
5391 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5393 /* Scale by energy magnitude */
5394 normalized_nrg <<= SCHED_LOAD_SHIFT;
5396 /* Normalize on max energy for target platform */
5397 normalized_nrg = reciprocal_divide(
5398 normalized_nrg, schedtune_target_nrg.rdiv);
5400 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5404 energy_diff(struct energy_env *eenv)
5406 int boost = schedtune_task_boost(eenv->task);
5409 /* Conpute "absolute" energy diff */
5410 __energy_diff(eenv);
5412 /* Return energy diff when boost margin is 0 */
5414 return eenv->nrg.diff;
5416 /* Compute normalized energy diff */
5417 nrg_delta = normalize_energy(eenv->nrg.diff);
5418 eenv->nrg.delta = nrg_delta;
5420 eenv->payoff = schedtune_accept_deltas(
5426 * When SchedTune is enabled, the energy_diff() function will return
5427 * the computed energy payoff value. Since the energy_diff() return
5428 * value is expected to be negative by its callers, this evaluation
5429 * function return a negative value each time the evaluation return a
5430 * positive payoff, which is the condition for the acceptance of
5431 * a scheduling decision
5433 return -eenv->payoff;
5435 #else /* CONFIG_SCHED_TUNE */
5436 #define energy_diff(eenv) __energy_diff(eenv)
5440 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5441 * A waker of many should wake a different task than the one last awakened
5442 * at a frequency roughly N times higher than one of its wakees. In order
5443 * to determine whether we should let the load spread vs consolodating to
5444 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5445 * partner, and a factor of lls_size higher frequency in the other. With
5446 * both conditions met, we can be relatively sure that the relationship is
5447 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5448 * being client/server, worker/dispatcher, interrupt source or whatever is
5449 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5451 static int wake_wide(struct task_struct *p)
5453 unsigned int master = current->wakee_flips;
5454 unsigned int slave = p->wakee_flips;
5455 int factor = this_cpu_read(sd_llc_size);
5458 swap(master, slave);
5459 if (slave < factor || master < slave * factor)
5464 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5465 int prev_cpu, int sync)
5467 s64 this_load, load;
5468 s64 this_eff_load, prev_eff_load;
5470 struct task_group *tg;
5471 unsigned long weight;
5475 this_cpu = smp_processor_id();
5476 load = source_load(prev_cpu, idx);
5477 this_load = target_load(this_cpu, idx);
5480 * If sync wakeup then subtract the (maximum possible)
5481 * effect of the currently running task from the load
5482 * of the current CPU:
5485 tg = task_group(current);
5486 weight = current->se.avg.load_avg;
5488 this_load += effective_load(tg, this_cpu, -weight, -weight);
5489 load += effective_load(tg, prev_cpu, 0, -weight);
5493 weight = p->se.avg.load_avg;
5496 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5497 * due to the sync cause above having dropped this_load to 0, we'll
5498 * always have an imbalance, but there's really nothing you can do
5499 * about that, so that's good too.
5501 * Otherwise check if either cpus are near enough in load to allow this
5502 * task to be woken on this_cpu.
5504 this_eff_load = 100;
5505 this_eff_load *= capacity_of(prev_cpu);
5507 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5508 prev_eff_load *= capacity_of(this_cpu);
5510 if (this_load > 0) {
5511 this_eff_load *= this_load +
5512 effective_load(tg, this_cpu, weight, weight);
5514 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5517 balanced = this_eff_load <= prev_eff_load;
5519 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5524 schedstat_inc(sd, ttwu_move_affine);
5525 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5530 static inline unsigned long task_util(struct task_struct *p)
5532 #ifdef CONFIG_SCHED_WALT
5533 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5534 unsigned long demand = p->ravg.demand;
5535 return (demand << 10) / walt_ravg_window;
5538 return p->se.avg.util_avg;
5541 static inline unsigned long boosted_task_util(struct task_struct *task);
5543 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5545 unsigned long capacity = capacity_of(cpu);
5547 util += boosted_task_util(p);
5549 return (capacity * 1024) > (util * capacity_margin);
5552 static inline bool task_fits_max(struct task_struct *p, int cpu)
5554 unsigned long capacity = capacity_of(cpu);
5555 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5557 if (capacity == max_capacity)
5560 if (capacity * capacity_margin > max_capacity * 1024)
5563 return __task_fits(p, cpu, 0);
5566 static bool cpu_overutilized(int cpu)
5568 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5571 #ifdef CONFIG_SCHED_TUNE
5574 schedtune_margin(unsigned long signal, long boost)
5576 long long margin = 0;
5579 * Signal proportional compensation (SPC)
5581 * The Boost (B) value is used to compute a Margin (M) which is
5582 * proportional to the complement of the original Signal (S):
5583 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5584 * M = B * S, if B is negative
5585 * The obtained M could be used by the caller to "boost" S.
5588 margin = SCHED_LOAD_SCALE - signal;
5591 margin = -signal * boost;
5593 * Fast integer division by constant:
5594 * Constant : (C) = 100
5595 * Precision : 0.1% (P) = 0.1
5596 * Reference : C * 100 / P (R) = 100000
5599 * Shift bits : ceil(log(R,2)) (S) = 17
5600 * Mult const : round(2^S/C) (M) = 1311
5613 schedtune_cpu_margin(unsigned long util, int cpu)
5615 int boost = schedtune_cpu_boost(cpu);
5620 return schedtune_margin(util, boost);
5624 schedtune_task_margin(struct task_struct *task)
5626 int boost = schedtune_task_boost(task);
5633 util = task_util(task);
5634 margin = schedtune_margin(util, boost);
5639 #else /* CONFIG_SCHED_TUNE */
5642 schedtune_cpu_margin(unsigned long util, int cpu)
5648 schedtune_task_margin(struct task_struct *task)
5653 #endif /* CONFIG_SCHED_TUNE */
5656 boosted_cpu_util(int cpu)
5658 unsigned long util = cpu_util(cpu);
5659 long margin = schedtune_cpu_margin(util, cpu);
5661 trace_sched_boost_cpu(cpu, util, margin);
5663 return util + margin;
5666 static inline unsigned long
5667 boosted_task_util(struct task_struct *task)
5669 unsigned long util = task_util(task);
5670 long margin = schedtune_task_margin(task);
5672 trace_sched_boost_task(task, util, margin);
5674 return util + margin;
5677 static int cpu_util_wake(int cpu, struct task_struct *p);
5679 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5681 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5685 * find_idlest_group finds and returns the least busy CPU group within the
5688 static struct sched_group *
5689 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5690 int this_cpu, int sd_flag)
5692 struct sched_group *idlest = NULL, *group = sd->groups;
5693 struct sched_group *most_spare_sg = NULL;
5694 unsigned long min_load = ULONG_MAX, this_load = 0;
5695 unsigned long most_spare = 0, this_spare = 0;
5696 int load_idx = sd->forkexec_idx;
5697 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5699 if (sd_flag & SD_BALANCE_WAKE)
5700 load_idx = sd->wake_idx;
5703 unsigned long load, avg_load, spare_cap, max_spare_cap;
5707 /* Skip over this group if it has no CPUs allowed */
5708 if (!cpumask_intersects(sched_group_cpus(group),
5709 tsk_cpus_allowed(p)))
5712 local_group = cpumask_test_cpu(this_cpu,
5713 sched_group_cpus(group));
5716 * Tally up the load of all CPUs in the group and find
5717 * the group containing the CPU with most spare capacity.
5722 for_each_cpu(i, sched_group_cpus(group)) {
5723 /* Bias balancing toward cpus of our domain */
5725 load = source_load(i, load_idx);
5727 load = target_load(i, load_idx);
5731 spare_cap = capacity_spare_wake(i, p);
5733 if (spare_cap > max_spare_cap)
5734 max_spare_cap = spare_cap;
5737 /* Adjust by relative CPU capacity of the group */
5738 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5741 this_load = avg_load;
5742 this_spare = max_spare_cap;
5744 if (avg_load < min_load) {
5745 min_load = avg_load;
5749 if (most_spare < max_spare_cap) {
5750 most_spare = max_spare_cap;
5751 most_spare_sg = group;
5754 } while (group = group->next, group != sd->groups);
5757 * The cross-over point between using spare capacity or least load
5758 * is too conservative for high utilization tasks on partially
5759 * utilized systems if we require spare_capacity > task_util(p),
5760 * so we allow for some task stuffing by using
5761 * spare_capacity > task_util(p)/2.
5763 if (this_spare > task_util(p) / 2 &&
5764 imbalance*this_spare > 100*most_spare)
5766 else if (most_spare > task_util(p) / 2)
5767 return most_spare_sg;
5769 if (!idlest || 100*this_load < imbalance*min_load)
5775 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5778 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5780 unsigned long load, min_load = ULONG_MAX;
5781 unsigned int min_exit_latency = UINT_MAX;
5782 u64 latest_idle_timestamp = 0;
5783 int least_loaded_cpu = this_cpu;
5784 int shallowest_idle_cpu = -1;
5787 /* Check if we have any choice: */
5788 if (group->group_weight == 1)
5789 return cpumask_first(sched_group_cpus(group));
5791 /* Traverse only the allowed CPUs */
5792 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5794 struct rq *rq = cpu_rq(i);
5795 struct cpuidle_state *idle = idle_get_state(rq);
5796 if (idle && idle->exit_latency < min_exit_latency) {
5798 * We give priority to a CPU whose idle state
5799 * has the smallest exit latency irrespective
5800 * of any idle timestamp.
5802 min_exit_latency = idle->exit_latency;
5803 latest_idle_timestamp = rq->idle_stamp;
5804 shallowest_idle_cpu = i;
5805 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5806 rq->idle_stamp > latest_idle_timestamp) {
5808 * If equal or no active idle state, then
5809 * the most recently idled CPU might have
5812 latest_idle_timestamp = rq->idle_stamp;
5813 shallowest_idle_cpu = i;
5815 } else if (shallowest_idle_cpu == -1) {
5816 load = weighted_cpuload(i);
5817 if (load < min_load || (load == min_load && i == this_cpu)) {
5819 least_loaded_cpu = i;
5824 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5828 * Try and locate an idle CPU in the sched_domain.
5830 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5832 struct sched_domain *sd;
5833 struct sched_group *sg;
5834 int best_idle_cpu = -1;
5835 int best_idle_cstate = INT_MAX;
5836 unsigned long best_idle_capacity = ULONG_MAX;
5838 if (!sysctl_sched_cstate_aware) {
5839 if (idle_cpu(target))
5843 * If the prevous cpu is cache affine and idle, don't be stupid.
5845 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5850 * Otherwise, iterate the domains and find an elegible idle cpu.
5852 sd = rcu_dereference(per_cpu(sd_llc, target));
5853 for_each_lower_domain(sd) {
5857 if (!cpumask_intersects(sched_group_cpus(sg),
5858 tsk_cpus_allowed(p)))
5861 if (sysctl_sched_cstate_aware) {
5862 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5863 int idle_idx = idle_get_state_idx(cpu_rq(i));
5864 unsigned long new_usage = boosted_task_util(p);
5865 unsigned long capacity_orig = capacity_orig_of(i);
5867 if (new_usage > capacity_orig || !idle_cpu(i))
5870 if (i == target && new_usage <= capacity_curr_of(target))
5873 if (idle_idx < best_idle_cstate &&
5874 capacity_orig <= best_idle_capacity) {
5876 best_idle_cstate = idle_idx;
5877 best_idle_capacity = capacity_orig;
5881 for_each_cpu(i, sched_group_cpus(sg)) {
5882 if (i == target || !idle_cpu(i))
5886 target = cpumask_first_and(sched_group_cpus(sg),
5887 tsk_cpus_allowed(p));
5892 } while (sg != sd->groups);
5895 if (best_idle_cpu >= 0)
5896 target = best_idle_cpu;
5902 static int start_cpu(bool boosted)
5904 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5906 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5907 "sched RCU must be held");
5909 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5912 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5914 int target_cpu = -1;
5915 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5916 unsigned long backup_capacity = ULONG_MAX;
5917 int best_idle_cpu = -1;
5918 int best_idle_cstate = INT_MAX;
5919 int backup_cpu = -1;
5920 unsigned long min_util = boosted_task_util(p);
5921 struct sched_domain *sd;
5922 struct sched_group *sg;
5923 int cpu = start_cpu(boosted);
5928 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5938 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5939 unsigned long cur_capacity, new_util;
5945 * p's blocked utilization is still accounted for on prev_cpu
5946 * so prev_cpu will receive a negative bias due to the double
5947 * accounting. However, the blocked utilization may be zero.
5949 new_util = cpu_util(i) + task_util(p);
5952 * Ensure minimum capacity to grant the required boost.
5953 * The target CPU can be already at a capacity level higher
5954 * than the one required to boost the task.
5956 new_util = max(min_util, new_util);
5958 if (new_util > capacity_orig_of(i))
5961 #ifdef CONFIG_SCHED_WALT
5962 if (walt_cpu_high_irqload(i))
5967 * Unconditionally favoring tasks that prefer idle cpus to
5970 if (idle_cpu(i) && prefer_idle)
5973 cur_capacity = capacity_curr_of(i);
5975 if (new_util < cur_capacity) {
5976 if (cpu_rq(i)->nr_running) {
5978 * Find a target cpu with the lowest/highest
5979 * utilization if prefer_idle/!prefer_idle.
5981 if ((prefer_idle && target_util > new_util) ||
5982 (!prefer_idle && target_util < new_util)) {
5983 target_util = new_util;
5986 } else if (!prefer_idle) {
5987 int idle_idx = idle_get_state_idx(cpu_rq(i));
5989 if (best_idle_cpu < 0 ||
5990 (sysctl_sched_cstate_aware &&
5991 best_idle_cstate > idle_idx)) {
5992 best_idle_cstate = idle_idx;
5996 } else if (backup_capacity > cur_capacity) {
5997 /* Find a backup cpu with least capacity. */
5998 backup_capacity = cur_capacity;
6002 } while (sg = sg->next, sg != sd->groups);
6005 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
6011 * cpu_util_wake: Compute cpu utilization with any contributions from
6012 * the waking task p removed.
6014 static int cpu_util_wake(int cpu, struct task_struct *p)
6016 unsigned long util, capacity;
6018 /* Task has no contribution or is new */
6019 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6020 return cpu_util(cpu);
6022 capacity = capacity_orig_of(cpu);
6023 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6025 return (util >= capacity) ? capacity : util;
6029 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6030 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6032 * In that case WAKE_AFFINE doesn't make sense and we'll let
6033 * BALANCE_WAKE sort things out.
6035 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6037 long min_cap, max_cap;
6039 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6040 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6042 /* Minimum capacity is close to max, no need to abort wake_affine */
6043 if (max_cap - min_cap < max_cap >> 3)
6046 /* Bring task utilization in sync with prev_cpu */
6047 sync_entity_load_avg(&p->se);
6049 return min_cap * 1024 < task_util(p) * capacity_margin;
6052 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6054 struct sched_domain *sd;
6055 int target_cpu = prev_cpu, tmp_target;
6056 bool boosted, prefer_idle;
6058 if (sysctl_sched_sync_hint_enable && sync) {
6059 int cpu = smp_processor_id();
6061 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
6066 #ifdef CONFIG_CGROUP_SCHEDTUNE
6067 boosted = schedtune_task_boost(p) > 0;
6068 prefer_idle = schedtune_prefer_idle(p) > 0;
6070 boosted = get_sysctl_sched_cfs_boost() > 0;
6074 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6075 /* Find a cpu with sufficient capacity */
6076 tmp_target = find_best_target(p, boosted, prefer_idle);
6080 if (tmp_target >= 0) {
6081 target_cpu = tmp_target;
6082 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
6086 if (target_cpu != prev_cpu) {
6087 struct energy_env eenv = {
6088 .util_delta = task_util(p),
6089 .src_cpu = prev_cpu,
6090 .dst_cpu = target_cpu,
6094 /* Not enough spare capacity on previous cpu */
6095 if (cpu_overutilized(prev_cpu))
6098 if (energy_diff(&eenv) >= 0)
6099 target_cpu = prev_cpu;
6108 * select_task_rq_fair: Select target runqueue for the waking task in domains
6109 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6110 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6112 * Balances load by selecting the idlest cpu in the idlest group, or under
6113 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6115 * Returns the target cpu number.
6117 * preempt must be disabled.
6120 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6122 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6123 int cpu = smp_processor_id();
6124 int new_cpu = prev_cpu;
6125 int want_affine = 0;
6126 int sync = wake_flags & WF_SYNC;
6128 if (sd_flag & SD_BALANCE_WAKE)
6129 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6130 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6132 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6133 return select_energy_cpu_brute(p, prev_cpu, sync);
6136 for_each_domain(cpu, tmp) {
6137 if (!(tmp->flags & SD_LOAD_BALANCE))
6141 * If both cpu and prev_cpu are part of this domain,
6142 * cpu is a valid SD_WAKE_AFFINE target.
6144 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6145 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6150 if (tmp->flags & sd_flag)
6152 else if (!want_affine)
6157 sd = NULL; /* Prefer wake_affine over balance flags */
6158 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6163 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6164 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6167 struct sched_group *group;
6170 if (!(sd->flags & sd_flag)) {
6175 group = find_idlest_group(sd, p, cpu, sd_flag);
6181 new_cpu = find_idlest_cpu(group, p, cpu);
6182 if (new_cpu == -1 || new_cpu == cpu) {
6183 /* Now try balancing at a lower domain level of cpu */
6188 /* Now try balancing at a lower domain level of new_cpu */
6190 weight = sd->span_weight;
6192 for_each_domain(cpu, tmp) {
6193 if (weight <= tmp->span_weight)
6195 if (tmp->flags & sd_flag)
6198 /* while loop will break here if sd == NULL */
6206 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6207 * cfs_rq_of(p) references at time of call are still valid and identify the
6208 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6209 * other assumptions, including the state of rq->lock, should be made.
6211 static void migrate_task_rq_fair(struct task_struct *p)
6214 * We are supposed to update the task to "current" time, then its up to date
6215 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6216 * what current time is, so simply throw away the out-of-date time. This
6217 * will result in the wakee task is less decayed, but giving the wakee more
6218 * load sounds not bad.
6220 remove_entity_load_avg(&p->se);
6222 /* Tell new CPU we are migrated */
6223 p->se.avg.last_update_time = 0;
6225 /* We have migrated, no longer consider this task hot */
6226 p->se.exec_start = 0;
6229 static void task_dead_fair(struct task_struct *p)
6231 remove_entity_load_avg(&p->se);
6234 #define task_fits_max(p, cpu) true
6235 #endif /* CONFIG_SMP */
6237 static unsigned long
6238 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6240 unsigned long gran = sysctl_sched_wakeup_granularity;
6243 * Since its curr running now, convert the gran from real-time
6244 * to virtual-time in his units.
6246 * By using 'se' instead of 'curr' we penalize light tasks, so
6247 * they get preempted easier. That is, if 'se' < 'curr' then
6248 * the resulting gran will be larger, therefore penalizing the
6249 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6250 * be smaller, again penalizing the lighter task.
6252 * This is especially important for buddies when the leftmost
6253 * task is higher priority than the buddy.
6255 return calc_delta_fair(gran, se);
6259 * Should 'se' preempt 'curr'.
6273 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6275 s64 gran, vdiff = curr->vruntime - se->vruntime;
6280 gran = wakeup_gran(curr, se);
6287 static void set_last_buddy(struct sched_entity *se)
6289 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6292 for_each_sched_entity(se)
6293 cfs_rq_of(se)->last = se;
6296 static void set_next_buddy(struct sched_entity *se)
6298 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6301 for_each_sched_entity(se)
6302 cfs_rq_of(se)->next = se;
6305 static void set_skip_buddy(struct sched_entity *se)
6307 for_each_sched_entity(se)
6308 cfs_rq_of(se)->skip = se;
6312 * Preempt the current task with a newly woken task if needed:
6314 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6316 struct task_struct *curr = rq->curr;
6317 struct sched_entity *se = &curr->se, *pse = &p->se;
6318 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6319 int scale = cfs_rq->nr_running >= sched_nr_latency;
6320 int next_buddy_marked = 0;
6322 if (unlikely(se == pse))
6326 * This is possible from callers such as attach_tasks(), in which we
6327 * unconditionally check_prempt_curr() after an enqueue (which may have
6328 * lead to a throttle). This both saves work and prevents false
6329 * next-buddy nomination below.
6331 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6334 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6335 set_next_buddy(pse);
6336 next_buddy_marked = 1;
6340 * We can come here with TIF_NEED_RESCHED already set from new task
6343 * Note: this also catches the edge-case of curr being in a throttled
6344 * group (e.g. via set_curr_task), since update_curr() (in the
6345 * enqueue of curr) will have resulted in resched being set. This
6346 * prevents us from potentially nominating it as a false LAST_BUDDY
6349 if (test_tsk_need_resched(curr))
6352 /* Idle tasks are by definition preempted by non-idle tasks. */
6353 if (unlikely(curr->policy == SCHED_IDLE) &&
6354 likely(p->policy != SCHED_IDLE))
6358 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6359 * is driven by the tick):
6361 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6364 find_matching_se(&se, &pse);
6365 update_curr(cfs_rq_of(se));
6367 if (wakeup_preempt_entity(se, pse) == 1) {
6369 * Bias pick_next to pick the sched entity that is
6370 * triggering this preemption.
6372 if (!next_buddy_marked)
6373 set_next_buddy(pse);
6382 * Only set the backward buddy when the current task is still
6383 * on the rq. This can happen when a wakeup gets interleaved
6384 * with schedule on the ->pre_schedule() or idle_balance()
6385 * point, either of which can * drop the rq lock.
6387 * Also, during early boot the idle thread is in the fair class,
6388 * for obvious reasons its a bad idea to schedule back to it.
6390 if (unlikely(!se->on_rq || curr == rq->idle))
6393 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6397 static struct task_struct *
6398 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6400 struct cfs_rq *cfs_rq = &rq->cfs;
6401 struct sched_entity *se;
6402 struct task_struct *p;
6406 #ifdef CONFIG_FAIR_GROUP_SCHED
6407 if (!cfs_rq->nr_running)
6410 if (prev->sched_class != &fair_sched_class)
6414 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6415 * likely that a next task is from the same cgroup as the current.
6417 * Therefore attempt to avoid putting and setting the entire cgroup
6418 * hierarchy, only change the part that actually changes.
6422 struct sched_entity *curr = cfs_rq->curr;
6425 * Since we got here without doing put_prev_entity() we also
6426 * have to consider cfs_rq->curr. If it is still a runnable
6427 * entity, update_curr() will update its vruntime, otherwise
6428 * forget we've ever seen it.
6432 update_curr(cfs_rq);
6437 * This call to check_cfs_rq_runtime() will do the
6438 * throttle and dequeue its entity in the parent(s).
6439 * Therefore the 'simple' nr_running test will indeed
6442 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6446 se = pick_next_entity(cfs_rq, curr);
6447 cfs_rq = group_cfs_rq(se);
6453 * Since we haven't yet done put_prev_entity and if the selected task
6454 * is a different task than we started out with, try and touch the
6455 * least amount of cfs_rqs.
6458 struct sched_entity *pse = &prev->se;
6460 while (!(cfs_rq = is_same_group(se, pse))) {
6461 int se_depth = se->depth;
6462 int pse_depth = pse->depth;
6464 if (se_depth <= pse_depth) {
6465 put_prev_entity(cfs_rq_of(pse), pse);
6466 pse = parent_entity(pse);
6468 if (se_depth >= pse_depth) {
6469 set_next_entity(cfs_rq_of(se), se);
6470 se = parent_entity(se);
6474 put_prev_entity(cfs_rq, pse);
6475 set_next_entity(cfs_rq, se);
6478 if (hrtick_enabled(rq))
6479 hrtick_start_fair(rq, p);
6481 rq->misfit_task = !task_fits_max(p, rq->cpu);
6488 if (!cfs_rq->nr_running)
6491 put_prev_task(rq, prev);
6494 se = pick_next_entity(cfs_rq, NULL);
6495 set_next_entity(cfs_rq, se);
6496 cfs_rq = group_cfs_rq(se);
6501 if (hrtick_enabled(rq))
6502 hrtick_start_fair(rq, p);
6504 rq->misfit_task = !task_fits_max(p, rq->cpu);
6509 rq->misfit_task = 0;
6511 * This is OK, because current is on_cpu, which avoids it being picked
6512 * for load-balance and preemption/IRQs are still disabled avoiding
6513 * further scheduler activity on it and we're being very careful to
6514 * re-start the picking loop.
6516 lockdep_unpin_lock(&rq->lock);
6517 new_tasks = idle_balance(rq);
6518 lockdep_pin_lock(&rq->lock);
6520 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6521 * possible for any higher priority task to appear. In that case we
6522 * must re-start the pick_next_entity() loop.
6534 * Account for a descheduled task:
6536 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6538 struct sched_entity *se = &prev->se;
6539 struct cfs_rq *cfs_rq;
6541 for_each_sched_entity(se) {
6542 cfs_rq = cfs_rq_of(se);
6543 put_prev_entity(cfs_rq, se);
6548 * sched_yield() is very simple
6550 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6552 static void yield_task_fair(struct rq *rq)
6554 struct task_struct *curr = rq->curr;
6555 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6556 struct sched_entity *se = &curr->se;
6559 * Are we the only task in the tree?
6561 if (unlikely(rq->nr_running == 1))
6564 clear_buddies(cfs_rq, se);
6566 if (curr->policy != SCHED_BATCH) {
6567 update_rq_clock(rq);
6569 * Update run-time statistics of the 'current'.
6571 update_curr(cfs_rq);
6573 * Tell update_rq_clock() that we've just updated,
6574 * so we don't do microscopic update in schedule()
6575 * and double the fastpath cost.
6577 rq_clock_skip_update(rq, true);
6583 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6585 struct sched_entity *se = &p->se;
6587 /* throttled hierarchies are not runnable */
6588 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6591 /* Tell the scheduler that we'd really like pse to run next. */
6594 yield_task_fair(rq);
6600 /**************************************************
6601 * Fair scheduling class load-balancing methods.
6605 * The purpose of load-balancing is to achieve the same basic fairness the
6606 * per-cpu scheduler provides, namely provide a proportional amount of compute
6607 * time to each task. This is expressed in the following equation:
6609 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6611 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6612 * W_i,0 is defined as:
6614 * W_i,0 = \Sum_j w_i,j (2)
6616 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6617 * is derived from the nice value as per prio_to_weight[].
6619 * The weight average is an exponential decay average of the instantaneous
6622 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6624 * C_i is the compute capacity of cpu i, typically it is the
6625 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6626 * can also include other factors [XXX].
6628 * To achieve this balance we define a measure of imbalance which follows
6629 * directly from (1):
6631 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6633 * We them move tasks around to minimize the imbalance. In the continuous
6634 * function space it is obvious this converges, in the discrete case we get
6635 * a few fun cases generally called infeasible weight scenarios.
6638 * - infeasible weights;
6639 * - local vs global optima in the discrete case. ]
6644 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6645 * for all i,j solution, we create a tree of cpus that follows the hardware
6646 * topology where each level pairs two lower groups (or better). This results
6647 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6648 * tree to only the first of the previous level and we decrease the frequency
6649 * of load-balance at each level inv. proportional to the number of cpus in
6655 * \Sum { --- * --- * 2^i } = O(n) (5)
6657 * `- size of each group
6658 * | | `- number of cpus doing load-balance
6660 * `- sum over all levels
6662 * Coupled with a limit on how many tasks we can migrate every balance pass,
6663 * this makes (5) the runtime complexity of the balancer.
6665 * An important property here is that each CPU is still (indirectly) connected
6666 * to every other cpu in at most O(log n) steps:
6668 * The adjacency matrix of the resulting graph is given by:
6671 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6674 * And you'll find that:
6676 * A^(log_2 n)_i,j != 0 for all i,j (7)
6678 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6679 * The task movement gives a factor of O(m), giving a convergence complexity
6682 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6687 * In order to avoid CPUs going idle while there's still work to do, new idle
6688 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6689 * tree itself instead of relying on other CPUs to bring it work.
6691 * This adds some complexity to both (5) and (8) but it reduces the total idle
6699 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6702 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6707 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6709 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6711 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6714 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6715 * rewrite all of this once again.]
6718 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6720 enum fbq_type { regular, remote, all };
6729 #define LBF_ALL_PINNED 0x01
6730 #define LBF_NEED_BREAK 0x02
6731 #define LBF_DST_PINNED 0x04
6732 #define LBF_SOME_PINNED 0x08
6735 struct sched_domain *sd;
6743 struct cpumask *dst_grpmask;
6745 enum cpu_idle_type idle;
6747 unsigned int src_grp_nr_running;
6748 /* The set of CPUs under consideration for load-balancing */
6749 struct cpumask *cpus;
6754 unsigned int loop_break;
6755 unsigned int loop_max;
6757 enum fbq_type fbq_type;
6758 enum group_type busiest_group_type;
6759 struct list_head tasks;
6763 * Is this task likely cache-hot:
6765 static int task_hot(struct task_struct *p, struct lb_env *env)
6769 lockdep_assert_held(&env->src_rq->lock);
6771 if (p->sched_class != &fair_sched_class)
6774 if (unlikely(p->policy == SCHED_IDLE))
6778 * Buddy candidates are cache hot:
6780 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6781 (&p->se == cfs_rq_of(&p->se)->next ||
6782 &p->se == cfs_rq_of(&p->se)->last))
6785 if (sysctl_sched_migration_cost == -1)
6787 if (sysctl_sched_migration_cost == 0)
6790 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6792 return delta < (s64)sysctl_sched_migration_cost;
6795 #ifdef CONFIG_NUMA_BALANCING
6797 * Returns 1, if task migration degrades locality
6798 * Returns 0, if task migration improves locality i.e migration preferred.
6799 * Returns -1, if task migration is not affected by locality.
6801 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6803 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6804 unsigned long src_faults, dst_faults;
6805 int src_nid, dst_nid;
6807 if (!static_branch_likely(&sched_numa_balancing))
6810 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6813 src_nid = cpu_to_node(env->src_cpu);
6814 dst_nid = cpu_to_node(env->dst_cpu);
6816 if (src_nid == dst_nid)
6819 /* Migrating away from the preferred node is always bad. */
6820 if (src_nid == p->numa_preferred_nid) {
6821 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6827 /* Encourage migration to the preferred node. */
6828 if (dst_nid == p->numa_preferred_nid)
6832 src_faults = group_faults(p, src_nid);
6833 dst_faults = group_faults(p, dst_nid);
6835 src_faults = task_faults(p, src_nid);
6836 dst_faults = task_faults(p, dst_nid);
6839 return dst_faults < src_faults;
6843 static inline int migrate_degrades_locality(struct task_struct *p,
6851 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6854 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6858 lockdep_assert_held(&env->src_rq->lock);
6861 * We do not migrate tasks that are:
6862 * 1) throttled_lb_pair, or
6863 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6864 * 3) running (obviously), or
6865 * 4) are cache-hot on their current CPU.
6867 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6870 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6873 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6875 env->flags |= LBF_SOME_PINNED;
6878 * Remember if this task can be migrated to any other cpu in
6879 * our sched_group. We may want to revisit it if we couldn't
6880 * meet load balance goals by pulling other tasks on src_cpu.
6882 * Also avoid computing new_dst_cpu if we have already computed
6883 * one in current iteration.
6885 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6888 /* Prevent to re-select dst_cpu via env's cpus */
6889 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6890 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6891 env->flags |= LBF_DST_PINNED;
6892 env->new_dst_cpu = cpu;
6900 /* Record that we found atleast one task that could run on dst_cpu */
6901 env->flags &= ~LBF_ALL_PINNED;
6903 if (task_running(env->src_rq, p)) {
6904 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6909 * Aggressive migration if:
6910 * 1) destination numa is preferred
6911 * 2) task is cache cold, or
6912 * 3) too many balance attempts have failed.
6914 tsk_cache_hot = migrate_degrades_locality(p, env);
6915 if (tsk_cache_hot == -1)
6916 tsk_cache_hot = task_hot(p, env);
6918 if (tsk_cache_hot <= 0 ||
6919 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6920 if (tsk_cache_hot == 1) {
6921 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6922 schedstat_inc(p, se.statistics.nr_forced_migrations);
6927 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6932 * detach_task() -- detach the task for the migration specified in env
6934 static void detach_task(struct task_struct *p, struct lb_env *env)
6936 lockdep_assert_held(&env->src_rq->lock);
6938 deactivate_task(env->src_rq, p, 0);
6939 p->on_rq = TASK_ON_RQ_MIGRATING;
6940 double_lock_balance(env->src_rq, env->dst_rq);
6941 set_task_cpu(p, env->dst_cpu);
6942 double_unlock_balance(env->src_rq, env->dst_rq);
6946 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6947 * part of active balancing operations within "domain".
6949 * Returns a task if successful and NULL otherwise.
6951 static struct task_struct *detach_one_task(struct lb_env *env)
6953 struct task_struct *p, *n;
6955 lockdep_assert_held(&env->src_rq->lock);
6957 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6958 if (!can_migrate_task(p, env))
6961 detach_task(p, env);
6964 * Right now, this is only the second place where
6965 * lb_gained[env->idle] is updated (other is detach_tasks)
6966 * so we can safely collect stats here rather than
6967 * inside detach_tasks().
6969 schedstat_inc(env->sd, lb_gained[env->idle]);
6975 static const unsigned int sched_nr_migrate_break = 32;
6978 * detach_tasks() -- tries to detach up to imbalance weighted load from
6979 * busiest_rq, as part of a balancing operation within domain "sd".
6981 * Returns number of detached tasks if successful and 0 otherwise.
6983 static int detach_tasks(struct lb_env *env)
6985 struct list_head *tasks = &env->src_rq->cfs_tasks;
6986 struct task_struct *p;
6990 lockdep_assert_held(&env->src_rq->lock);
6992 if (env->imbalance <= 0)
6995 while (!list_empty(tasks)) {
6997 * We don't want to steal all, otherwise we may be treated likewise,
6998 * which could at worst lead to a livelock crash.
7000 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7003 p = list_first_entry(tasks, struct task_struct, se.group_node);
7006 /* We've more or less seen every task there is, call it quits */
7007 if (env->loop > env->loop_max)
7010 /* take a breather every nr_migrate tasks */
7011 if (env->loop > env->loop_break) {
7012 env->loop_break += sched_nr_migrate_break;
7013 env->flags |= LBF_NEED_BREAK;
7017 if (!can_migrate_task(p, env))
7020 load = task_h_load(p);
7022 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7025 if ((load / 2) > env->imbalance)
7028 detach_task(p, env);
7029 list_add(&p->se.group_node, &env->tasks);
7032 env->imbalance -= load;
7034 #ifdef CONFIG_PREEMPT
7036 * NEWIDLE balancing is a source of latency, so preemptible
7037 * kernels will stop after the first task is detached to minimize
7038 * the critical section.
7040 if (env->idle == CPU_NEWLY_IDLE)
7045 * We only want to steal up to the prescribed amount of
7048 if (env->imbalance <= 0)
7053 list_move_tail(&p->se.group_node, tasks);
7057 * Right now, this is one of only two places we collect this stat
7058 * so we can safely collect detach_one_task() stats here rather
7059 * than inside detach_one_task().
7061 schedstat_add(env->sd, lb_gained[env->idle], detached);
7067 * attach_task() -- attach the task detached by detach_task() to its new rq.
7069 static void attach_task(struct rq *rq, struct task_struct *p)
7071 lockdep_assert_held(&rq->lock);
7073 BUG_ON(task_rq(p) != rq);
7074 p->on_rq = TASK_ON_RQ_QUEUED;
7075 activate_task(rq, p, 0);
7076 check_preempt_curr(rq, p, 0);
7080 * attach_one_task() -- attaches the task returned from detach_one_task() to
7083 static void attach_one_task(struct rq *rq, struct task_struct *p)
7085 raw_spin_lock(&rq->lock);
7088 * We want to potentially raise target_cpu's OPP.
7090 update_capacity_of(cpu_of(rq));
7091 raw_spin_unlock(&rq->lock);
7095 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7098 static void attach_tasks(struct lb_env *env)
7100 struct list_head *tasks = &env->tasks;
7101 struct task_struct *p;
7103 raw_spin_lock(&env->dst_rq->lock);
7105 while (!list_empty(tasks)) {
7106 p = list_first_entry(tasks, struct task_struct, se.group_node);
7107 list_del_init(&p->se.group_node);
7109 attach_task(env->dst_rq, p);
7113 * We want to potentially raise env.dst_cpu's OPP.
7115 update_capacity_of(env->dst_cpu);
7117 raw_spin_unlock(&env->dst_rq->lock);
7120 #ifdef CONFIG_FAIR_GROUP_SCHED
7121 static void update_blocked_averages(int cpu)
7123 struct rq *rq = cpu_rq(cpu);
7124 struct cfs_rq *cfs_rq;
7125 unsigned long flags;
7127 raw_spin_lock_irqsave(&rq->lock, flags);
7128 update_rq_clock(rq);
7131 * Iterates the task_group tree in a bottom up fashion, see
7132 * list_add_leaf_cfs_rq() for details.
7134 for_each_leaf_cfs_rq(rq, cfs_rq) {
7135 /* throttled entities do not contribute to load */
7136 if (throttled_hierarchy(cfs_rq))
7139 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7141 update_tg_load_avg(cfs_rq, 0);
7143 raw_spin_unlock_irqrestore(&rq->lock, flags);
7147 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7148 * This needs to be done in a top-down fashion because the load of a child
7149 * group is a fraction of its parents load.
7151 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7153 struct rq *rq = rq_of(cfs_rq);
7154 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7155 unsigned long now = jiffies;
7158 if (cfs_rq->last_h_load_update == now)
7161 cfs_rq->h_load_next = NULL;
7162 for_each_sched_entity(se) {
7163 cfs_rq = cfs_rq_of(se);
7164 cfs_rq->h_load_next = se;
7165 if (cfs_rq->last_h_load_update == now)
7170 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7171 cfs_rq->last_h_load_update = now;
7174 while ((se = cfs_rq->h_load_next) != NULL) {
7175 load = cfs_rq->h_load;
7176 load = div64_ul(load * se->avg.load_avg,
7177 cfs_rq_load_avg(cfs_rq) + 1);
7178 cfs_rq = group_cfs_rq(se);
7179 cfs_rq->h_load = load;
7180 cfs_rq->last_h_load_update = now;
7184 static unsigned long task_h_load(struct task_struct *p)
7186 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7188 update_cfs_rq_h_load(cfs_rq);
7189 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7190 cfs_rq_load_avg(cfs_rq) + 1);
7193 static inline void update_blocked_averages(int cpu)
7195 struct rq *rq = cpu_rq(cpu);
7196 struct cfs_rq *cfs_rq = &rq->cfs;
7197 unsigned long flags;
7199 raw_spin_lock_irqsave(&rq->lock, flags);
7200 update_rq_clock(rq);
7201 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7202 raw_spin_unlock_irqrestore(&rq->lock, flags);
7205 static unsigned long task_h_load(struct task_struct *p)
7207 return p->se.avg.load_avg;
7211 /********** Helpers for find_busiest_group ************************/
7214 * sg_lb_stats - stats of a sched_group required for load_balancing
7216 struct sg_lb_stats {
7217 unsigned long avg_load; /*Avg load across the CPUs of the group */
7218 unsigned long group_load; /* Total load over the CPUs of the group */
7219 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7220 unsigned long load_per_task;
7221 unsigned long group_capacity;
7222 unsigned long group_util; /* Total utilization of the group */
7223 unsigned int sum_nr_running; /* Nr tasks running in the group */
7224 unsigned int idle_cpus;
7225 unsigned int group_weight;
7226 enum group_type group_type;
7227 int group_no_capacity;
7228 int group_misfit_task; /* A cpu has a task too big for its capacity */
7229 #ifdef CONFIG_NUMA_BALANCING
7230 unsigned int nr_numa_running;
7231 unsigned int nr_preferred_running;
7236 * sd_lb_stats - Structure to store the statistics of a sched_domain
7237 * during load balancing.
7239 struct sd_lb_stats {
7240 struct sched_group *busiest; /* Busiest group in this sd */
7241 struct sched_group *local; /* Local group in this sd */
7242 unsigned long total_load; /* Total load of all groups in sd */
7243 unsigned long total_capacity; /* Total capacity of all groups in sd */
7244 unsigned long avg_load; /* Average load across all groups in sd */
7246 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7247 struct sg_lb_stats local_stat; /* Statistics of the local group */
7250 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7253 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7254 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7255 * We must however clear busiest_stat::avg_load because
7256 * update_sd_pick_busiest() reads this before assignment.
7258 *sds = (struct sd_lb_stats){
7262 .total_capacity = 0UL,
7265 .sum_nr_running = 0,
7266 .group_type = group_other,
7272 * get_sd_load_idx - Obtain the load index for a given sched domain.
7273 * @sd: The sched_domain whose load_idx is to be obtained.
7274 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7276 * Return: The load index.
7278 static inline int get_sd_load_idx(struct sched_domain *sd,
7279 enum cpu_idle_type idle)
7285 load_idx = sd->busy_idx;
7288 case CPU_NEWLY_IDLE:
7289 load_idx = sd->newidle_idx;
7292 load_idx = sd->idle_idx;
7299 static unsigned long scale_rt_capacity(int cpu)
7301 struct rq *rq = cpu_rq(cpu);
7302 u64 total, used, age_stamp, avg;
7306 * Since we're reading these variables without serialization make sure
7307 * we read them once before doing sanity checks on them.
7309 age_stamp = READ_ONCE(rq->age_stamp);
7310 avg = READ_ONCE(rq->rt_avg);
7311 delta = __rq_clock_broken(rq) - age_stamp;
7313 if (unlikely(delta < 0))
7316 total = sched_avg_period() + delta;
7318 used = div_u64(avg, total);
7321 * deadline bandwidth is defined at system level so we must
7322 * weight this bandwidth with the max capacity of the system.
7323 * As a reminder, avg_bw is 20bits width and
7324 * scale_cpu_capacity is 10 bits width
7326 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7328 if (likely(used < SCHED_CAPACITY_SCALE))
7329 return SCHED_CAPACITY_SCALE - used;
7334 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7336 raw_spin_lock_init(&mcc->lock);
7341 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7343 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7344 struct sched_group *sdg = sd->groups;
7345 struct max_cpu_capacity *mcc;
7346 unsigned long max_capacity;
7348 unsigned long flags;
7350 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7352 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7354 raw_spin_lock_irqsave(&mcc->lock, flags);
7355 max_capacity = mcc->val;
7356 max_cap_cpu = mcc->cpu;
7358 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7359 (max_capacity < capacity)) {
7360 mcc->val = capacity;
7362 #ifdef CONFIG_SCHED_DEBUG
7363 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7364 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7369 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7371 skip_unlock: __attribute__ ((unused));
7372 capacity *= scale_rt_capacity(cpu);
7373 capacity >>= SCHED_CAPACITY_SHIFT;
7378 cpu_rq(cpu)->cpu_capacity = capacity;
7379 sdg->sgc->capacity = capacity;
7380 sdg->sgc->max_capacity = capacity;
7381 sdg->sgc->min_capacity = capacity;
7384 void update_group_capacity(struct sched_domain *sd, int cpu)
7386 struct sched_domain *child = sd->child;
7387 struct sched_group *group, *sdg = sd->groups;
7388 unsigned long capacity, max_capacity, min_capacity;
7389 unsigned long interval;
7391 interval = msecs_to_jiffies(sd->balance_interval);
7392 interval = clamp(interval, 1UL, max_load_balance_interval);
7393 sdg->sgc->next_update = jiffies + interval;
7396 update_cpu_capacity(sd, cpu);
7402 min_capacity = ULONG_MAX;
7404 if (child->flags & SD_OVERLAP) {
7406 * SD_OVERLAP domains cannot assume that child groups
7407 * span the current group.
7410 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7411 struct sched_group_capacity *sgc;
7412 struct rq *rq = cpu_rq(cpu);
7415 * build_sched_domains() -> init_sched_groups_capacity()
7416 * gets here before we've attached the domains to the
7419 * Use capacity_of(), which is set irrespective of domains
7420 * in update_cpu_capacity().
7422 * This avoids capacity from being 0 and
7423 * causing divide-by-zero issues on boot.
7425 if (unlikely(!rq->sd)) {
7426 capacity += capacity_of(cpu);
7428 sgc = rq->sd->groups->sgc;
7429 capacity += sgc->capacity;
7432 max_capacity = max(capacity, max_capacity);
7433 min_capacity = min(capacity, min_capacity);
7437 * !SD_OVERLAP domains can assume that child groups
7438 * span the current group.
7441 group = child->groups;
7443 struct sched_group_capacity *sgc = group->sgc;
7445 capacity += sgc->capacity;
7446 max_capacity = max(sgc->max_capacity, max_capacity);
7447 min_capacity = min(sgc->min_capacity, min_capacity);
7448 group = group->next;
7449 } while (group != child->groups);
7452 sdg->sgc->capacity = capacity;
7453 sdg->sgc->max_capacity = max_capacity;
7454 sdg->sgc->min_capacity = min_capacity;
7458 * Check whether the capacity of the rq has been noticeably reduced by side
7459 * activity. The imbalance_pct is used for the threshold.
7460 * Return true is the capacity is reduced
7463 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7465 return ((rq->cpu_capacity * sd->imbalance_pct) <
7466 (rq->cpu_capacity_orig * 100));
7470 * Group imbalance indicates (and tries to solve) the problem where balancing
7471 * groups is inadequate due to tsk_cpus_allowed() constraints.
7473 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7474 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7477 * { 0 1 2 3 } { 4 5 6 7 }
7480 * If we were to balance group-wise we'd place two tasks in the first group and
7481 * two tasks in the second group. Clearly this is undesired as it will overload
7482 * cpu 3 and leave one of the cpus in the second group unused.
7484 * The current solution to this issue is detecting the skew in the first group
7485 * by noticing the lower domain failed to reach balance and had difficulty
7486 * moving tasks due to affinity constraints.
7488 * When this is so detected; this group becomes a candidate for busiest; see
7489 * update_sd_pick_busiest(). And calculate_imbalance() and
7490 * find_busiest_group() avoid some of the usual balance conditions to allow it
7491 * to create an effective group imbalance.
7493 * This is a somewhat tricky proposition since the next run might not find the
7494 * group imbalance and decide the groups need to be balanced again. A most
7495 * subtle and fragile situation.
7498 static inline int sg_imbalanced(struct sched_group *group)
7500 return group->sgc->imbalance;
7504 * group_has_capacity returns true if the group has spare capacity that could
7505 * be used by some tasks.
7506 * We consider that a group has spare capacity if the * number of task is
7507 * smaller than the number of CPUs or if the utilization is lower than the
7508 * available capacity for CFS tasks.
7509 * For the latter, we use a threshold to stabilize the state, to take into
7510 * account the variance of the tasks' load and to return true if the available
7511 * capacity in meaningful for the load balancer.
7512 * As an example, an available capacity of 1% can appear but it doesn't make
7513 * any benefit for the load balance.
7516 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7518 if (sgs->sum_nr_running < sgs->group_weight)
7521 if ((sgs->group_capacity * 100) >
7522 (sgs->group_util * env->sd->imbalance_pct))
7529 * group_is_overloaded returns true if the group has more tasks than it can
7531 * group_is_overloaded is not equals to !group_has_capacity because a group
7532 * with the exact right number of tasks, has no more spare capacity but is not
7533 * overloaded so both group_has_capacity and group_is_overloaded return
7537 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7539 if (sgs->sum_nr_running <= sgs->group_weight)
7542 if ((sgs->group_capacity * 100) <
7543 (sgs->group_util * env->sd->imbalance_pct))
7551 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7552 * per-cpu capacity than sched_group ref.
7555 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7557 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7558 ref->sgc->max_capacity;
7562 group_type group_classify(struct sched_group *group,
7563 struct sg_lb_stats *sgs)
7565 if (sgs->group_no_capacity)
7566 return group_overloaded;
7568 if (sg_imbalanced(group))
7569 return group_imbalanced;
7571 if (sgs->group_misfit_task)
7572 return group_misfit_task;
7578 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7579 * @env: The load balancing environment.
7580 * @group: sched_group whose statistics are to be updated.
7581 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7582 * @local_group: Does group contain this_cpu.
7583 * @sgs: variable to hold the statistics for this group.
7584 * @overload: Indicate more than one runnable task for any CPU.
7585 * @overutilized: Indicate overutilization for any CPU.
7587 static inline void update_sg_lb_stats(struct lb_env *env,
7588 struct sched_group *group, int load_idx,
7589 int local_group, struct sg_lb_stats *sgs,
7590 bool *overload, bool *overutilized)
7595 memset(sgs, 0, sizeof(*sgs));
7597 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7598 struct rq *rq = cpu_rq(i);
7600 /* Bias balancing toward cpus of our domain */
7602 load = target_load(i, load_idx);
7604 load = source_load(i, load_idx);
7606 sgs->group_load += load;
7607 sgs->group_util += cpu_util(i);
7608 sgs->sum_nr_running += rq->cfs.h_nr_running;
7610 nr_running = rq->nr_running;
7614 #ifdef CONFIG_NUMA_BALANCING
7615 sgs->nr_numa_running += rq->nr_numa_running;
7616 sgs->nr_preferred_running += rq->nr_preferred_running;
7618 sgs->sum_weighted_load += weighted_cpuload(i);
7620 * No need to call idle_cpu() if nr_running is not 0
7622 if (!nr_running && idle_cpu(i))
7625 if (cpu_overutilized(i)) {
7626 *overutilized = true;
7627 if (!sgs->group_misfit_task && rq->misfit_task)
7628 sgs->group_misfit_task = capacity_of(i);
7632 /* Adjust by relative CPU capacity of the group */
7633 sgs->group_capacity = group->sgc->capacity;
7634 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7636 if (sgs->sum_nr_running)
7637 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7639 sgs->group_weight = group->group_weight;
7641 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7642 sgs->group_type = group_classify(group, sgs);
7646 * update_sd_pick_busiest - return 1 on busiest group
7647 * @env: The load balancing environment.
7648 * @sds: sched_domain statistics
7649 * @sg: sched_group candidate to be checked for being the busiest
7650 * @sgs: sched_group statistics
7652 * Determine if @sg is a busier group than the previously selected
7655 * Return: %true if @sg is a busier group than the previously selected
7656 * busiest group. %false otherwise.
7658 static bool update_sd_pick_busiest(struct lb_env *env,
7659 struct sd_lb_stats *sds,
7660 struct sched_group *sg,
7661 struct sg_lb_stats *sgs)
7663 struct sg_lb_stats *busiest = &sds->busiest_stat;
7665 if (sgs->group_type > busiest->group_type)
7668 if (sgs->group_type < busiest->group_type)
7672 * Candidate sg doesn't face any serious load-balance problems
7673 * so don't pick it if the local sg is already filled up.
7675 if (sgs->group_type == group_other &&
7676 !group_has_capacity(env, &sds->local_stat))
7679 if (sgs->avg_load <= busiest->avg_load)
7682 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7686 * Candidate sg has no more than one task per CPU and
7687 * has higher per-CPU capacity. Migrating tasks to less
7688 * capable CPUs may harm throughput. Maximize throughput,
7689 * power/energy consequences are not considered.
7691 if (sgs->sum_nr_running <= sgs->group_weight &&
7692 group_smaller_cpu_capacity(sds->local, sg))
7696 /* This is the busiest node in its class. */
7697 if (!(env->sd->flags & SD_ASYM_PACKING))
7701 * ASYM_PACKING needs to move all the work to the lowest
7702 * numbered CPUs in the group, therefore mark all groups
7703 * higher than ourself as busy.
7705 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7709 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7716 #ifdef CONFIG_NUMA_BALANCING
7717 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7719 if (sgs->sum_nr_running > sgs->nr_numa_running)
7721 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7726 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7728 if (rq->nr_running > rq->nr_numa_running)
7730 if (rq->nr_running > rq->nr_preferred_running)
7735 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7740 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7744 #endif /* CONFIG_NUMA_BALANCING */
7747 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7748 * @env: The load balancing environment.
7749 * @sds: variable to hold the statistics for this sched_domain.
7751 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7753 struct sched_domain *child = env->sd->child;
7754 struct sched_group *sg = env->sd->groups;
7755 struct sg_lb_stats tmp_sgs;
7756 int load_idx, prefer_sibling = 0;
7757 bool overload = false, overutilized = false;
7759 if (child && child->flags & SD_PREFER_SIBLING)
7762 load_idx = get_sd_load_idx(env->sd, env->idle);
7765 struct sg_lb_stats *sgs = &tmp_sgs;
7768 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7771 sgs = &sds->local_stat;
7773 if (env->idle != CPU_NEWLY_IDLE ||
7774 time_after_eq(jiffies, sg->sgc->next_update))
7775 update_group_capacity(env->sd, env->dst_cpu);
7778 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7779 &overload, &overutilized);
7785 * In case the child domain prefers tasks go to siblings
7786 * first, lower the sg capacity so that we'll try
7787 * and move all the excess tasks away. We lower the capacity
7788 * of a group only if the local group has the capacity to fit
7789 * these excess tasks. The extra check prevents the case where
7790 * you always pull from the heaviest group when it is already
7791 * under-utilized (possible with a large weight task outweighs
7792 * the tasks on the system).
7794 if (prefer_sibling && sds->local &&
7795 group_has_capacity(env, &sds->local_stat) &&
7796 (sgs->sum_nr_running > 1)) {
7797 sgs->group_no_capacity = 1;
7798 sgs->group_type = group_classify(sg, sgs);
7802 * Ignore task groups with misfit tasks if local group has no
7803 * capacity or if per-cpu capacity isn't higher.
7805 if (sgs->group_type == group_misfit_task &&
7806 (!group_has_capacity(env, &sds->local_stat) ||
7807 !group_smaller_cpu_capacity(sg, sds->local)))
7808 sgs->group_type = group_other;
7810 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7812 sds->busiest_stat = *sgs;
7816 /* Now, start updating sd_lb_stats */
7817 sds->total_load += sgs->group_load;
7818 sds->total_capacity += sgs->group_capacity;
7821 } while (sg != env->sd->groups);
7823 if (env->sd->flags & SD_NUMA)
7824 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7826 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7828 if (!env->sd->parent) {
7829 /* update overload indicator if we are at root domain */
7830 if (env->dst_rq->rd->overload != overload)
7831 env->dst_rq->rd->overload = overload;
7833 /* Update over-utilization (tipping point, U >= 0) indicator */
7834 if (env->dst_rq->rd->overutilized != overutilized) {
7835 env->dst_rq->rd->overutilized = overutilized;
7836 trace_sched_overutilized(overutilized);
7839 if (!env->dst_rq->rd->overutilized && overutilized) {
7840 env->dst_rq->rd->overutilized = true;
7841 trace_sched_overutilized(true);
7848 * check_asym_packing - Check to see if the group is packed into the
7851 * This is primarily intended to used at the sibling level. Some
7852 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7853 * case of POWER7, it can move to lower SMT modes only when higher
7854 * threads are idle. When in lower SMT modes, the threads will
7855 * perform better since they share less core resources. Hence when we
7856 * have idle threads, we want them to be the higher ones.
7858 * This packing function is run on idle threads. It checks to see if
7859 * the busiest CPU in this domain (core in the P7 case) has a higher
7860 * CPU number than the packing function is being run on. Here we are
7861 * assuming lower CPU number will be equivalent to lower a SMT thread
7864 * Return: 1 when packing is required and a task should be moved to
7865 * this CPU. The amount of the imbalance is returned in *imbalance.
7867 * @env: The load balancing environment.
7868 * @sds: Statistics of the sched_domain which is to be packed
7870 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7874 if (!(env->sd->flags & SD_ASYM_PACKING))
7880 busiest_cpu = group_first_cpu(sds->busiest);
7881 if (env->dst_cpu > busiest_cpu)
7884 env->imbalance = DIV_ROUND_CLOSEST(
7885 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7886 SCHED_CAPACITY_SCALE);
7892 * fix_small_imbalance - Calculate the minor imbalance that exists
7893 * amongst the groups of a sched_domain, during
7895 * @env: The load balancing environment.
7896 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7899 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7901 unsigned long tmp, capa_now = 0, capa_move = 0;
7902 unsigned int imbn = 2;
7903 unsigned long scaled_busy_load_per_task;
7904 struct sg_lb_stats *local, *busiest;
7906 local = &sds->local_stat;
7907 busiest = &sds->busiest_stat;
7909 if (!local->sum_nr_running)
7910 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7911 else if (busiest->load_per_task > local->load_per_task)
7914 scaled_busy_load_per_task =
7915 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7916 busiest->group_capacity;
7918 if (busiest->avg_load + scaled_busy_load_per_task >=
7919 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7920 env->imbalance = busiest->load_per_task;
7925 * OK, we don't have enough imbalance to justify moving tasks,
7926 * however we may be able to increase total CPU capacity used by
7930 capa_now += busiest->group_capacity *
7931 min(busiest->load_per_task, busiest->avg_load);
7932 capa_now += local->group_capacity *
7933 min(local->load_per_task, local->avg_load);
7934 capa_now /= SCHED_CAPACITY_SCALE;
7936 /* Amount of load we'd subtract */
7937 if (busiest->avg_load > scaled_busy_load_per_task) {
7938 capa_move += busiest->group_capacity *
7939 min(busiest->load_per_task,
7940 busiest->avg_load - scaled_busy_load_per_task);
7943 /* Amount of load we'd add */
7944 if (busiest->avg_load * busiest->group_capacity <
7945 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7946 tmp = (busiest->avg_load * busiest->group_capacity) /
7947 local->group_capacity;
7949 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7950 local->group_capacity;
7952 capa_move += local->group_capacity *
7953 min(local->load_per_task, local->avg_load + tmp);
7954 capa_move /= SCHED_CAPACITY_SCALE;
7956 /* Move if we gain throughput */
7957 if (capa_move > capa_now)
7958 env->imbalance = busiest->load_per_task;
7962 * calculate_imbalance - Calculate the amount of imbalance present within the
7963 * groups of a given sched_domain during load balance.
7964 * @env: load balance environment
7965 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7967 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7969 unsigned long max_pull, load_above_capacity = ~0UL;
7970 struct sg_lb_stats *local, *busiest;
7972 local = &sds->local_stat;
7973 busiest = &sds->busiest_stat;
7975 if (busiest->group_type == group_imbalanced) {
7977 * In the group_imb case we cannot rely on group-wide averages
7978 * to ensure cpu-load equilibrium, look at wider averages. XXX
7980 busiest->load_per_task =
7981 min(busiest->load_per_task, sds->avg_load);
7985 * In the presence of smp nice balancing, certain scenarios can have
7986 * max load less than avg load(as we skip the groups at or below
7987 * its cpu_capacity, while calculating max_load..)
7989 if (busiest->avg_load <= sds->avg_load ||
7990 local->avg_load >= sds->avg_load) {
7991 /* Misfitting tasks should be migrated in any case */
7992 if (busiest->group_type == group_misfit_task) {
7993 env->imbalance = busiest->group_misfit_task;
7998 * Busiest group is overloaded, local is not, use the spare
7999 * cycles to maximize throughput
8001 if (busiest->group_type == group_overloaded &&
8002 local->group_type <= group_misfit_task) {
8003 env->imbalance = busiest->load_per_task;
8008 return fix_small_imbalance(env, sds);
8012 * If there aren't any idle cpus, avoid creating some.
8014 if (busiest->group_type == group_overloaded &&
8015 local->group_type == group_overloaded) {
8016 load_above_capacity = busiest->sum_nr_running *
8018 if (load_above_capacity > busiest->group_capacity)
8019 load_above_capacity -= busiest->group_capacity;
8021 load_above_capacity = ~0UL;
8025 * We're trying to get all the cpus to the average_load, so we don't
8026 * want to push ourselves above the average load, nor do we wish to
8027 * reduce the max loaded cpu below the average load. At the same time,
8028 * we also don't want to reduce the group load below the group capacity
8029 * (so that we can implement power-savings policies etc). Thus we look
8030 * for the minimum possible imbalance.
8032 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8034 /* How much load to actually move to equalise the imbalance */
8035 env->imbalance = min(
8036 max_pull * busiest->group_capacity,
8037 (sds->avg_load - local->avg_load) * local->group_capacity
8038 ) / SCHED_CAPACITY_SCALE;
8040 /* Boost imbalance to allow misfit task to be balanced. */
8041 if (busiest->group_type == group_misfit_task)
8042 env->imbalance = max_t(long, env->imbalance,
8043 busiest->group_misfit_task);
8046 * if *imbalance is less than the average load per runnable task
8047 * there is no guarantee that any tasks will be moved so we'll have
8048 * a think about bumping its value to force at least one task to be
8051 if (env->imbalance < busiest->load_per_task)
8052 return fix_small_imbalance(env, sds);
8055 /******* find_busiest_group() helpers end here *********************/
8058 * find_busiest_group - Returns the busiest group within the sched_domain
8059 * if there is an imbalance. If there isn't an imbalance, and
8060 * the user has opted for power-savings, it returns a group whose
8061 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8062 * such a group exists.
8064 * Also calculates the amount of weighted load which should be moved
8065 * to restore balance.
8067 * @env: The load balancing environment.
8069 * Return: - The busiest group if imbalance exists.
8070 * - If no imbalance and user has opted for power-savings balance,
8071 * return the least loaded group whose CPUs can be
8072 * put to idle by rebalancing its tasks onto our group.
8074 static struct sched_group *find_busiest_group(struct lb_env *env)
8076 struct sg_lb_stats *local, *busiest;
8077 struct sd_lb_stats sds;
8079 init_sd_lb_stats(&sds);
8082 * Compute the various statistics relavent for load balancing at
8085 update_sd_lb_stats(env, &sds);
8087 if (energy_aware() && !env->dst_rq->rd->overutilized)
8090 local = &sds.local_stat;
8091 busiest = &sds.busiest_stat;
8093 /* ASYM feature bypasses nice load balance check */
8094 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8095 check_asym_packing(env, &sds))
8098 /* There is no busy sibling group to pull tasks from */
8099 if (!sds.busiest || busiest->sum_nr_running == 0)
8102 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8103 / sds.total_capacity;
8106 * If the busiest group is imbalanced the below checks don't
8107 * work because they assume all things are equal, which typically
8108 * isn't true due to cpus_allowed constraints and the like.
8110 if (busiest->group_type == group_imbalanced)
8113 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8114 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8115 busiest->group_no_capacity)
8118 /* Misfitting tasks should be dealt with regardless of the avg load */
8119 if (busiest->group_type == group_misfit_task) {
8124 * If the local group is busier than the selected busiest group
8125 * don't try and pull any tasks.
8127 if (local->avg_load >= busiest->avg_load)
8131 * Don't pull any tasks if this group is already above the domain
8134 if (local->avg_load >= sds.avg_load)
8137 if (env->idle == CPU_IDLE) {
8139 * This cpu is idle. If the busiest group is not overloaded
8140 * and there is no imbalance between this and busiest group
8141 * wrt idle cpus, it is balanced. The imbalance becomes
8142 * significant if the diff is greater than 1 otherwise we
8143 * might end up to just move the imbalance on another group
8145 if ((busiest->group_type != group_overloaded) &&
8146 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8147 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8151 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8152 * imbalance_pct to be conservative.
8154 if (100 * busiest->avg_load <=
8155 env->sd->imbalance_pct * local->avg_load)
8160 env->busiest_group_type = busiest->group_type;
8161 /* Looks like there is an imbalance. Compute it */
8162 calculate_imbalance(env, &sds);
8171 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8173 static struct rq *find_busiest_queue(struct lb_env *env,
8174 struct sched_group *group)
8176 struct rq *busiest = NULL, *rq;
8177 unsigned long busiest_load = 0, busiest_capacity = 1;
8180 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8181 unsigned long capacity, wl;
8185 rt = fbq_classify_rq(rq);
8188 * We classify groups/runqueues into three groups:
8189 * - regular: there are !numa tasks
8190 * - remote: there are numa tasks that run on the 'wrong' node
8191 * - all: there is no distinction
8193 * In order to avoid migrating ideally placed numa tasks,
8194 * ignore those when there's better options.
8196 * If we ignore the actual busiest queue to migrate another
8197 * task, the next balance pass can still reduce the busiest
8198 * queue by moving tasks around inside the node.
8200 * If we cannot move enough load due to this classification
8201 * the next pass will adjust the group classification and
8202 * allow migration of more tasks.
8204 * Both cases only affect the total convergence complexity.
8206 if (rt > env->fbq_type)
8209 capacity = capacity_of(i);
8211 wl = weighted_cpuload(i);
8214 * When comparing with imbalance, use weighted_cpuload()
8215 * which is not scaled with the cpu capacity.
8218 if (rq->nr_running == 1 && wl > env->imbalance &&
8219 !check_cpu_capacity(rq, env->sd) &&
8220 env->busiest_group_type != group_misfit_task)
8224 * For the load comparisons with the other cpu's, consider
8225 * the weighted_cpuload() scaled with the cpu capacity, so
8226 * that the load can be moved away from the cpu that is
8227 * potentially running at a lower capacity.
8229 * Thus we're looking for max(wl_i / capacity_i), crosswise
8230 * multiplication to rid ourselves of the division works out
8231 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8232 * our previous maximum.
8234 if (wl * busiest_capacity > busiest_load * capacity) {
8236 busiest_capacity = capacity;
8245 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8246 * so long as it is large enough.
8248 #define MAX_PINNED_INTERVAL 512
8250 /* Working cpumask for load_balance and load_balance_newidle. */
8251 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8253 static int need_active_balance(struct lb_env *env)
8255 struct sched_domain *sd = env->sd;
8257 if (env->idle == CPU_NEWLY_IDLE) {
8260 * ASYM_PACKING needs to force migrate tasks from busy but
8261 * higher numbered CPUs in order to pack all tasks in the
8262 * lowest numbered CPUs.
8264 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8269 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8270 * It's worth migrating the task if the src_cpu's capacity is reduced
8271 * because of other sched_class or IRQs if more capacity stays
8272 * available on dst_cpu.
8274 if ((env->idle != CPU_NOT_IDLE) &&
8275 (env->src_rq->cfs.h_nr_running == 1)) {
8276 if ((check_cpu_capacity(env->src_rq, sd)) &&
8277 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8281 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8282 env->src_rq->cfs.h_nr_running == 1 &&
8283 cpu_overutilized(env->src_cpu) &&
8284 !cpu_overutilized(env->dst_cpu)) {
8288 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8291 static int active_load_balance_cpu_stop(void *data);
8293 static int should_we_balance(struct lb_env *env)
8295 struct sched_group *sg = env->sd->groups;
8296 struct cpumask *sg_cpus, *sg_mask;
8297 int cpu, balance_cpu = -1;
8300 * In the newly idle case, we will allow all the cpu's
8301 * to do the newly idle load balance.
8303 if (env->idle == CPU_NEWLY_IDLE)
8306 sg_cpus = sched_group_cpus(sg);
8307 sg_mask = sched_group_mask(sg);
8308 /* Try to find first idle cpu */
8309 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8310 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8317 if (balance_cpu == -1)
8318 balance_cpu = group_balance_cpu(sg);
8321 * First idle cpu or the first cpu(busiest) in this sched group
8322 * is eligible for doing load balancing at this and above domains.
8324 return balance_cpu == env->dst_cpu;
8328 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8329 * tasks if there is an imbalance.
8331 static int load_balance(int this_cpu, struct rq *this_rq,
8332 struct sched_domain *sd, enum cpu_idle_type idle,
8333 int *continue_balancing)
8335 int ld_moved, cur_ld_moved, active_balance = 0;
8336 struct sched_domain *sd_parent = sd->parent;
8337 struct sched_group *group;
8339 unsigned long flags;
8340 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8342 struct lb_env env = {
8344 .dst_cpu = this_cpu,
8346 .dst_grpmask = sched_group_cpus(sd->groups),
8348 .loop_break = sched_nr_migrate_break,
8351 .tasks = LIST_HEAD_INIT(env.tasks),
8355 * For NEWLY_IDLE load_balancing, we don't need to consider
8356 * other cpus in our group
8358 if (idle == CPU_NEWLY_IDLE)
8359 env.dst_grpmask = NULL;
8361 cpumask_copy(cpus, cpu_active_mask);
8363 schedstat_inc(sd, lb_count[idle]);
8366 if (!should_we_balance(&env)) {
8367 *continue_balancing = 0;
8371 group = find_busiest_group(&env);
8373 schedstat_inc(sd, lb_nobusyg[idle]);
8377 busiest = find_busiest_queue(&env, group);
8379 schedstat_inc(sd, lb_nobusyq[idle]);
8383 BUG_ON(busiest == env.dst_rq);
8385 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8387 env.src_cpu = busiest->cpu;
8388 env.src_rq = busiest;
8391 if (busiest->nr_running > 1) {
8393 * Attempt to move tasks. If find_busiest_group has found
8394 * an imbalance but busiest->nr_running <= 1, the group is
8395 * still unbalanced. ld_moved simply stays zero, so it is
8396 * correctly treated as an imbalance.
8398 env.flags |= LBF_ALL_PINNED;
8399 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8402 raw_spin_lock_irqsave(&busiest->lock, flags);
8405 * cur_ld_moved - load moved in current iteration
8406 * ld_moved - cumulative load moved across iterations
8408 cur_ld_moved = detach_tasks(&env);
8410 * We want to potentially lower env.src_cpu's OPP.
8413 update_capacity_of(env.src_cpu);
8416 * We've detached some tasks from busiest_rq. Every
8417 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8418 * unlock busiest->lock, and we are able to be sure
8419 * that nobody can manipulate the tasks in parallel.
8420 * See task_rq_lock() family for the details.
8423 raw_spin_unlock(&busiest->lock);
8427 ld_moved += cur_ld_moved;
8430 local_irq_restore(flags);
8432 if (env.flags & LBF_NEED_BREAK) {
8433 env.flags &= ~LBF_NEED_BREAK;
8438 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8439 * us and move them to an alternate dst_cpu in our sched_group
8440 * where they can run. The upper limit on how many times we
8441 * iterate on same src_cpu is dependent on number of cpus in our
8444 * This changes load balance semantics a bit on who can move
8445 * load to a given_cpu. In addition to the given_cpu itself
8446 * (or a ilb_cpu acting on its behalf where given_cpu is
8447 * nohz-idle), we now have balance_cpu in a position to move
8448 * load to given_cpu. In rare situations, this may cause
8449 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8450 * _independently_ and at _same_ time to move some load to
8451 * given_cpu) causing exceess load to be moved to given_cpu.
8452 * This however should not happen so much in practice and
8453 * moreover subsequent load balance cycles should correct the
8454 * excess load moved.
8456 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8458 /* Prevent to re-select dst_cpu via env's cpus */
8459 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8461 env.dst_rq = cpu_rq(env.new_dst_cpu);
8462 env.dst_cpu = env.new_dst_cpu;
8463 env.flags &= ~LBF_DST_PINNED;
8465 env.loop_break = sched_nr_migrate_break;
8468 * Go back to "more_balance" rather than "redo" since we
8469 * need to continue with same src_cpu.
8475 * We failed to reach balance because of affinity.
8478 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8480 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8481 *group_imbalance = 1;
8484 /* All tasks on this runqueue were pinned by CPU affinity */
8485 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8486 cpumask_clear_cpu(cpu_of(busiest), cpus);
8487 if (!cpumask_empty(cpus)) {
8489 env.loop_break = sched_nr_migrate_break;
8492 goto out_all_pinned;
8497 schedstat_inc(sd, lb_failed[idle]);
8499 * Increment the failure counter only on periodic balance.
8500 * We do not want newidle balance, which can be very
8501 * frequent, pollute the failure counter causing
8502 * excessive cache_hot migrations and active balances.
8504 if (idle != CPU_NEWLY_IDLE)
8505 if (env.src_grp_nr_running > 1)
8506 sd->nr_balance_failed++;
8508 if (need_active_balance(&env)) {
8509 raw_spin_lock_irqsave(&busiest->lock, flags);
8511 /* don't kick the active_load_balance_cpu_stop,
8512 * if the curr task on busiest cpu can't be
8515 if (!cpumask_test_cpu(this_cpu,
8516 tsk_cpus_allowed(busiest->curr))) {
8517 raw_spin_unlock_irqrestore(&busiest->lock,
8519 env.flags |= LBF_ALL_PINNED;
8520 goto out_one_pinned;
8524 * ->active_balance synchronizes accesses to
8525 * ->active_balance_work. Once set, it's cleared
8526 * only after active load balance is finished.
8528 if (!busiest->active_balance) {
8529 busiest->active_balance = 1;
8530 busiest->push_cpu = this_cpu;
8533 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8535 if (active_balance) {
8536 stop_one_cpu_nowait(cpu_of(busiest),
8537 active_load_balance_cpu_stop, busiest,
8538 &busiest->active_balance_work);
8542 * We've kicked active balancing, reset the failure
8545 sd->nr_balance_failed = sd->cache_nice_tries+1;
8548 sd->nr_balance_failed = 0;
8550 if (likely(!active_balance)) {
8551 /* We were unbalanced, so reset the balancing interval */
8552 sd->balance_interval = sd->min_interval;
8555 * If we've begun active balancing, start to back off. This
8556 * case may not be covered by the all_pinned logic if there
8557 * is only 1 task on the busy runqueue (because we don't call
8560 if (sd->balance_interval < sd->max_interval)
8561 sd->balance_interval *= 2;
8568 * We reach balance although we may have faced some affinity
8569 * constraints. Clear the imbalance flag if it was set.
8572 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8574 if (*group_imbalance)
8575 *group_imbalance = 0;
8580 * We reach balance because all tasks are pinned at this level so
8581 * we can't migrate them. Let the imbalance flag set so parent level
8582 * can try to migrate them.
8584 schedstat_inc(sd, lb_balanced[idle]);
8586 sd->nr_balance_failed = 0;
8589 /* tune up the balancing interval */
8590 if (((env.flags & LBF_ALL_PINNED) &&
8591 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8592 (sd->balance_interval < sd->max_interval))
8593 sd->balance_interval *= 2;
8600 static inline unsigned long
8601 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8603 unsigned long interval = sd->balance_interval;
8606 interval *= sd->busy_factor;
8608 /* scale ms to jiffies */
8609 interval = msecs_to_jiffies(interval);
8610 interval = clamp(interval, 1UL, max_load_balance_interval);
8616 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8618 unsigned long interval, next;
8620 interval = get_sd_balance_interval(sd, cpu_busy);
8621 next = sd->last_balance + interval;
8623 if (time_after(*next_balance, next))
8624 *next_balance = next;
8628 * idle_balance is called by schedule() if this_cpu is about to become
8629 * idle. Attempts to pull tasks from other CPUs.
8631 static int idle_balance(struct rq *this_rq)
8633 unsigned long next_balance = jiffies + HZ;
8634 int this_cpu = this_rq->cpu;
8635 struct sched_domain *sd;
8636 int pulled_task = 0;
8638 long removed_util=0;
8640 idle_enter_fair(this_rq);
8643 * We must set idle_stamp _before_ calling idle_balance(), such that we
8644 * measure the duration of idle_balance() as idle time.
8646 this_rq->idle_stamp = rq_clock(this_rq);
8648 if (!energy_aware() &&
8649 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8650 !this_rq->rd->overload)) {
8652 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8654 update_next_balance(sd, 0, &next_balance);
8660 raw_spin_unlock(&this_rq->lock);
8663 * If removed_util_avg is !0 we most probably migrated some task away
8664 * from this_cpu. In this case we might be willing to trigger an OPP
8665 * update, but we want to do so if we don't find anybody else to pull
8666 * here (we will trigger an OPP update with the pulled task's enqueue
8669 * Record removed_util before calling update_blocked_averages, and use
8670 * it below (before returning) to see if an OPP update is required.
8672 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8673 update_blocked_averages(this_cpu);
8675 for_each_domain(this_cpu, sd) {
8676 int continue_balancing = 1;
8677 u64 t0, domain_cost;
8679 if (!(sd->flags & SD_LOAD_BALANCE))
8682 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8683 update_next_balance(sd, 0, &next_balance);
8687 if (sd->flags & SD_BALANCE_NEWIDLE) {
8688 t0 = sched_clock_cpu(this_cpu);
8690 pulled_task = load_balance(this_cpu, this_rq,
8692 &continue_balancing);
8694 domain_cost = sched_clock_cpu(this_cpu) - t0;
8695 if (domain_cost > sd->max_newidle_lb_cost)
8696 sd->max_newidle_lb_cost = domain_cost;
8698 curr_cost += domain_cost;
8701 update_next_balance(sd, 0, &next_balance);
8704 * Stop searching for tasks to pull if there are
8705 * now runnable tasks on this rq.
8707 if (pulled_task || this_rq->nr_running > 0)
8712 raw_spin_lock(&this_rq->lock);
8714 if (curr_cost > this_rq->max_idle_balance_cost)
8715 this_rq->max_idle_balance_cost = curr_cost;
8718 * While browsing the domains, we released the rq lock, a task could
8719 * have been enqueued in the meantime. Since we're not going idle,
8720 * pretend we pulled a task.
8722 if (this_rq->cfs.h_nr_running && !pulled_task)
8726 /* Move the next balance forward */
8727 if (time_after(this_rq->next_balance, next_balance))
8728 this_rq->next_balance = next_balance;
8730 /* Is there a task of a high priority class? */
8731 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8735 idle_exit_fair(this_rq);
8736 this_rq->idle_stamp = 0;
8737 } else if (removed_util) {
8739 * No task pulled and someone has been migrated away.
8740 * Good case to trigger an OPP update.
8742 update_capacity_of(this_cpu);
8749 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8750 * running tasks off the busiest CPU onto idle CPUs. It requires at
8751 * least 1 task to be running on each physical CPU where possible, and
8752 * avoids physical / logical imbalances.
8754 static int active_load_balance_cpu_stop(void *data)
8756 struct rq *busiest_rq = data;
8757 int busiest_cpu = cpu_of(busiest_rq);
8758 int target_cpu = busiest_rq->push_cpu;
8759 struct rq *target_rq = cpu_rq(target_cpu);
8760 struct sched_domain *sd;
8761 struct task_struct *p = NULL;
8763 raw_spin_lock_irq(&busiest_rq->lock);
8765 /* make sure the requested cpu hasn't gone down in the meantime */
8766 if (unlikely(busiest_cpu != smp_processor_id() ||
8767 !busiest_rq->active_balance))
8770 /* Is there any task to move? */
8771 if (busiest_rq->nr_running <= 1)
8775 * This condition is "impossible", if it occurs
8776 * we need to fix it. Originally reported by
8777 * Bjorn Helgaas on a 128-cpu setup.
8779 BUG_ON(busiest_rq == target_rq);
8781 /* Search for an sd spanning us and the target CPU. */
8783 for_each_domain(target_cpu, sd) {
8784 if ((sd->flags & SD_LOAD_BALANCE) &&
8785 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8790 struct lb_env env = {
8792 .dst_cpu = target_cpu,
8793 .dst_rq = target_rq,
8794 .src_cpu = busiest_rq->cpu,
8795 .src_rq = busiest_rq,
8799 schedstat_inc(sd, alb_count);
8801 p = detach_one_task(&env);
8803 schedstat_inc(sd, alb_pushed);
8805 * We want to potentially lower env.src_cpu's OPP.
8807 update_capacity_of(env.src_cpu);
8810 schedstat_inc(sd, alb_failed);
8814 busiest_rq->active_balance = 0;
8815 raw_spin_unlock(&busiest_rq->lock);
8818 attach_one_task(target_rq, p);
8825 static inline int on_null_domain(struct rq *rq)
8827 return unlikely(!rcu_dereference_sched(rq->sd));
8830 #ifdef CONFIG_NO_HZ_COMMON
8832 * idle load balancing details
8833 * - When one of the busy CPUs notice that there may be an idle rebalancing
8834 * needed, they will kick the idle load balancer, which then does idle
8835 * load balancing for all the idle CPUs.
8838 cpumask_var_t idle_cpus_mask;
8840 unsigned long next_balance; /* in jiffy units */
8841 } nohz ____cacheline_aligned;
8843 static inline int find_new_ilb(void)
8845 int ilb = cpumask_first(nohz.idle_cpus_mask);
8847 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8854 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8855 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8856 * CPU (if there is one).
8858 static void nohz_balancer_kick(void)
8862 nohz.next_balance++;
8864 ilb_cpu = find_new_ilb();
8866 if (ilb_cpu >= nr_cpu_ids)
8869 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8872 * Use smp_send_reschedule() instead of resched_cpu().
8873 * This way we generate a sched IPI on the target cpu which
8874 * is idle. And the softirq performing nohz idle load balance
8875 * will be run before returning from the IPI.
8877 smp_send_reschedule(ilb_cpu);
8881 static inline void nohz_balance_exit_idle(int cpu)
8883 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8885 * Completely isolated CPUs don't ever set, so we must test.
8887 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8888 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8889 atomic_dec(&nohz.nr_cpus);
8891 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8895 static inline void set_cpu_sd_state_busy(void)
8897 struct sched_domain *sd;
8898 int cpu = smp_processor_id();
8901 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8903 if (!sd || !sd->nohz_idle)
8907 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8912 void set_cpu_sd_state_idle(void)
8914 struct sched_domain *sd;
8915 int cpu = smp_processor_id();
8918 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8920 if (!sd || sd->nohz_idle)
8924 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8930 * This routine will record that the cpu is going idle with tick stopped.
8931 * This info will be used in performing idle load balancing in the future.
8933 void nohz_balance_enter_idle(int cpu)
8936 * If this cpu is going down, then nothing needs to be done.
8938 if (!cpu_active(cpu))
8941 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8945 * If we're a completely isolated CPU, we don't play.
8947 if (on_null_domain(cpu_rq(cpu)))
8950 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8951 atomic_inc(&nohz.nr_cpus);
8952 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8955 static int sched_ilb_notifier(struct notifier_block *nfb,
8956 unsigned long action, void *hcpu)
8958 switch (action & ~CPU_TASKS_FROZEN) {
8960 nohz_balance_exit_idle(smp_processor_id());
8968 static DEFINE_SPINLOCK(balancing);
8971 * Scale the max load_balance interval with the number of CPUs in the system.
8972 * This trades load-balance latency on larger machines for less cross talk.
8974 void update_max_interval(void)
8976 max_load_balance_interval = HZ*num_online_cpus()/10;
8980 * It checks each scheduling domain to see if it is due to be balanced,
8981 * and initiates a balancing operation if so.
8983 * Balancing parameters are set up in init_sched_domains.
8985 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8987 int continue_balancing = 1;
8989 unsigned long interval;
8990 struct sched_domain *sd;
8991 /* Earliest time when we have to do rebalance again */
8992 unsigned long next_balance = jiffies + 60*HZ;
8993 int update_next_balance = 0;
8994 int need_serialize, need_decay = 0;
8997 update_blocked_averages(cpu);
9000 for_each_domain(cpu, sd) {
9002 * Decay the newidle max times here because this is a regular
9003 * visit to all the domains. Decay ~1% per second.
9005 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9006 sd->max_newidle_lb_cost =
9007 (sd->max_newidle_lb_cost * 253) / 256;
9008 sd->next_decay_max_lb_cost = jiffies + HZ;
9011 max_cost += sd->max_newidle_lb_cost;
9013 if (!(sd->flags & SD_LOAD_BALANCE))
9017 * Stop the load balance at this level. There is another
9018 * CPU in our sched group which is doing load balancing more
9021 if (!continue_balancing) {
9027 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9029 need_serialize = sd->flags & SD_SERIALIZE;
9030 if (need_serialize) {
9031 if (!spin_trylock(&balancing))
9035 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9036 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9038 * The LBF_DST_PINNED logic could have changed
9039 * env->dst_cpu, so we can't know our idle
9040 * state even if we migrated tasks. Update it.
9042 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9044 sd->last_balance = jiffies;
9045 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9048 spin_unlock(&balancing);
9050 if (time_after(next_balance, sd->last_balance + interval)) {
9051 next_balance = sd->last_balance + interval;
9052 update_next_balance = 1;
9057 * Ensure the rq-wide value also decays but keep it at a
9058 * reasonable floor to avoid funnies with rq->avg_idle.
9060 rq->max_idle_balance_cost =
9061 max((u64)sysctl_sched_migration_cost, max_cost);
9066 * next_balance will be updated only when there is a need.
9067 * When the cpu is attached to null domain for ex, it will not be
9070 if (likely(update_next_balance)) {
9071 rq->next_balance = next_balance;
9073 #ifdef CONFIG_NO_HZ_COMMON
9075 * If this CPU has been elected to perform the nohz idle
9076 * balance. Other idle CPUs have already rebalanced with
9077 * nohz_idle_balance() and nohz.next_balance has been
9078 * updated accordingly. This CPU is now running the idle load
9079 * balance for itself and we need to update the
9080 * nohz.next_balance accordingly.
9082 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9083 nohz.next_balance = rq->next_balance;
9088 #ifdef CONFIG_NO_HZ_COMMON
9090 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9091 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9093 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9095 int this_cpu = this_rq->cpu;
9098 /* Earliest time when we have to do rebalance again */
9099 unsigned long next_balance = jiffies + 60*HZ;
9100 int update_next_balance = 0;
9102 if (idle != CPU_IDLE ||
9103 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9106 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9107 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9111 * If this cpu gets work to do, stop the load balancing
9112 * work being done for other cpus. Next load
9113 * balancing owner will pick it up.
9118 rq = cpu_rq(balance_cpu);
9121 * If time for next balance is due,
9124 if (time_after_eq(jiffies, rq->next_balance)) {
9125 raw_spin_lock_irq(&rq->lock);
9126 update_rq_clock(rq);
9127 update_idle_cpu_load(rq);
9128 raw_spin_unlock_irq(&rq->lock);
9129 rebalance_domains(rq, CPU_IDLE);
9132 if (time_after(next_balance, rq->next_balance)) {
9133 next_balance = rq->next_balance;
9134 update_next_balance = 1;
9139 * next_balance will be updated only when there is a need.
9140 * When the CPU is attached to null domain for ex, it will not be
9143 if (likely(update_next_balance))
9144 nohz.next_balance = next_balance;
9146 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9150 * Current heuristic for kicking the idle load balancer in the presence
9151 * of an idle cpu in the system.
9152 * - This rq has more than one task.
9153 * - This rq has at least one CFS task and the capacity of the CPU is
9154 * significantly reduced because of RT tasks or IRQs.
9155 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9156 * multiple busy cpu.
9157 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9158 * domain span are idle.
9160 static inline bool nohz_kick_needed(struct rq *rq)
9162 unsigned long now = jiffies;
9163 struct sched_domain *sd;
9164 struct sched_group_capacity *sgc;
9165 int nr_busy, cpu = rq->cpu;
9168 if (unlikely(rq->idle_balance))
9172 * We may be recently in ticked or tickless idle mode. At the first
9173 * busy tick after returning from idle, we will update the busy stats.
9175 set_cpu_sd_state_busy();
9176 nohz_balance_exit_idle(cpu);
9179 * None are in tickless mode and hence no need for NOHZ idle load
9182 if (likely(!atomic_read(&nohz.nr_cpus)))
9185 if (time_before(now, nohz.next_balance))
9188 if (rq->nr_running >= 2 &&
9189 (!energy_aware() || cpu_overutilized(cpu)))
9193 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9194 if (sd && !energy_aware()) {
9195 sgc = sd->groups->sgc;
9196 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9205 sd = rcu_dereference(rq->sd);
9207 if ((rq->cfs.h_nr_running >= 1) &&
9208 check_cpu_capacity(rq, sd)) {
9214 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9215 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9216 sched_domain_span(sd)) < cpu)) {
9226 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9230 * run_rebalance_domains is triggered when needed from the scheduler tick.
9231 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9233 static void run_rebalance_domains(struct softirq_action *h)
9235 struct rq *this_rq = this_rq();
9236 enum cpu_idle_type idle = this_rq->idle_balance ?
9237 CPU_IDLE : CPU_NOT_IDLE;
9240 * If this cpu has a pending nohz_balance_kick, then do the
9241 * balancing on behalf of the other idle cpus whose ticks are
9242 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9243 * give the idle cpus a chance to load balance. Else we may
9244 * load balance only within the local sched_domain hierarchy
9245 * and abort nohz_idle_balance altogether if we pull some load.
9247 nohz_idle_balance(this_rq, idle);
9248 rebalance_domains(this_rq, idle);
9252 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9254 void trigger_load_balance(struct rq *rq)
9256 /* Don't need to rebalance while attached to NULL domain */
9257 if (unlikely(on_null_domain(rq)))
9260 if (time_after_eq(jiffies, rq->next_balance))
9261 raise_softirq(SCHED_SOFTIRQ);
9262 #ifdef CONFIG_NO_HZ_COMMON
9263 if (nohz_kick_needed(rq))
9264 nohz_balancer_kick();
9268 static void rq_online_fair(struct rq *rq)
9272 update_runtime_enabled(rq);
9275 static void rq_offline_fair(struct rq *rq)
9279 /* Ensure any throttled groups are reachable by pick_next_task */
9280 unthrottle_offline_cfs_rqs(rq);
9283 #endif /* CONFIG_SMP */
9286 * scheduler tick hitting a task of our scheduling class:
9288 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9290 struct cfs_rq *cfs_rq;
9291 struct sched_entity *se = &curr->se;
9293 for_each_sched_entity(se) {
9294 cfs_rq = cfs_rq_of(se);
9295 entity_tick(cfs_rq, se, queued);
9298 if (static_branch_unlikely(&sched_numa_balancing))
9299 task_tick_numa(rq, curr);
9302 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9303 rq->rd->overutilized = true;
9304 trace_sched_overutilized(true);
9307 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9313 * called on fork with the child task as argument from the parent's context
9314 * - child not yet on the tasklist
9315 * - preemption disabled
9317 static void task_fork_fair(struct task_struct *p)
9319 struct cfs_rq *cfs_rq;
9320 struct sched_entity *se = &p->se, *curr;
9321 int this_cpu = smp_processor_id();
9322 struct rq *rq = this_rq();
9323 unsigned long flags;
9325 raw_spin_lock_irqsave(&rq->lock, flags);
9327 update_rq_clock(rq);
9329 cfs_rq = task_cfs_rq(current);
9330 curr = cfs_rq->curr;
9333 * Not only the cpu but also the task_group of the parent might have
9334 * been changed after parent->se.parent,cfs_rq were copied to
9335 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9336 * of child point to valid ones.
9339 __set_task_cpu(p, this_cpu);
9342 update_curr(cfs_rq);
9345 se->vruntime = curr->vruntime;
9346 place_entity(cfs_rq, se, 1);
9348 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9350 * Upon rescheduling, sched_class::put_prev_task() will place
9351 * 'current' within the tree based on its new key value.
9353 swap(curr->vruntime, se->vruntime);
9357 se->vruntime -= cfs_rq->min_vruntime;
9359 raw_spin_unlock_irqrestore(&rq->lock, flags);
9363 * Priority of the task has changed. Check to see if we preempt
9367 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9369 if (!task_on_rq_queued(p))
9373 * Reschedule if we are currently running on this runqueue and
9374 * our priority decreased, or if we are not currently running on
9375 * this runqueue and our priority is higher than the current's
9377 if (rq->curr == p) {
9378 if (p->prio > oldprio)
9381 check_preempt_curr(rq, p, 0);
9384 static inline bool vruntime_normalized(struct task_struct *p)
9386 struct sched_entity *se = &p->se;
9389 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9390 * the dequeue_entity(.flags=0) will already have normalized the
9397 * When !on_rq, vruntime of the task has usually NOT been normalized.
9398 * But there are some cases where it has already been normalized:
9400 * - A forked child which is waiting for being woken up by
9401 * wake_up_new_task().
9402 * - A task which has been woken up by try_to_wake_up() and
9403 * waiting for actually being woken up by sched_ttwu_pending().
9405 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9411 static void detach_entity_cfs_rq(struct sched_entity *se)
9413 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9415 /* Catch up with the cfs_rq and remove our load when we leave */
9416 update_load_avg(se, 0);
9417 detach_entity_load_avg(cfs_rq, se);
9418 update_tg_load_avg(cfs_rq, false);
9421 static void attach_entity_cfs_rq(struct sched_entity *se)
9423 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9425 #ifdef CONFIG_FAIR_GROUP_SCHED
9427 * Since the real-depth could have been changed (only FAIR
9428 * class maintain depth value), reset depth properly.
9430 se->depth = se->parent ? se->parent->depth + 1 : 0;
9433 /* Synchronize entity with its cfs_rq */
9434 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9435 attach_entity_load_avg(cfs_rq, se);
9436 update_tg_load_avg(cfs_rq, false);
9439 static void detach_task_cfs_rq(struct task_struct *p)
9441 struct sched_entity *se = &p->se;
9442 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9444 if (!vruntime_normalized(p)) {
9446 * Fix up our vruntime so that the current sleep doesn't
9447 * cause 'unlimited' sleep bonus.
9449 place_entity(cfs_rq, se, 0);
9450 se->vruntime -= cfs_rq->min_vruntime;
9453 detach_entity_cfs_rq(se);
9456 static void attach_task_cfs_rq(struct task_struct *p)
9458 struct sched_entity *se = &p->se;
9459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9461 attach_entity_cfs_rq(se);
9463 if (!vruntime_normalized(p))
9464 se->vruntime += cfs_rq->min_vruntime;
9467 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9469 detach_task_cfs_rq(p);
9472 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9474 attach_task_cfs_rq(p);
9476 if (task_on_rq_queued(p)) {
9478 * We were most likely switched from sched_rt, so
9479 * kick off the schedule if running, otherwise just see
9480 * if we can still preempt the current task.
9485 check_preempt_curr(rq, p, 0);
9489 /* Account for a task changing its policy or group.
9491 * This routine is mostly called to set cfs_rq->curr field when a task
9492 * migrates between groups/classes.
9494 static void set_curr_task_fair(struct rq *rq)
9496 struct sched_entity *se = &rq->curr->se;
9498 for_each_sched_entity(se) {
9499 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9501 set_next_entity(cfs_rq, se);
9502 /* ensure bandwidth has been allocated on our new cfs_rq */
9503 account_cfs_rq_runtime(cfs_rq, 0);
9507 void init_cfs_rq(struct cfs_rq *cfs_rq)
9509 cfs_rq->tasks_timeline = RB_ROOT;
9510 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9511 #ifndef CONFIG_64BIT
9512 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9515 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9516 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9520 #ifdef CONFIG_FAIR_GROUP_SCHED
9521 static void task_move_group_fair(struct task_struct *p)
9523 detach_task_cfs_rq(p);
9524 set_task_rq(p, task_cpu(p));
9527 /* Tell se's cfs_rq has been changed -- migrated */
9528 p->se.avg.last_update_time = 0;
9530 attach_task_cfs_rq(p);
9533 void free_fair_sched_group(struct task_group *tg)
9537 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9539 for_each_possible_cpu(i) {
9541 kfree(tg->cfs_rq[i]);
9544 remove_entity_load_avg(tg->se[i]);
9553 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9555 struct sched_entity *se;
9556 struct cfs_rq *cfs_rq;
9560 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9563 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9567 tg->shares = NICE_0_LOAD;
9569 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9571 for_each_possible_cpu(i) {
9574 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9575 GFP_KERNEL, cpu_to_node(i));
9579 se = kzalloc_node(sizeof(struct sched_entity),
9580 GFP_KERNEL, cpu_to_node(i));
9584 init_cfs_rq(cfs_rq);
9585 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9586 init_entity_runnable_average(se);
9588 raw_spin_lock_irq(&rq->lock);
9589 post_init_entity_util_avg(se);
9590 raw_spin_unlock_irq(&rq->lock);
9601 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9603 struct rq *rq = cpu_rq(cpu);
9604 unsigned long flags;
9607 * Only empty task groups can be destroyed; so we can speculatively
9608 * check on_list without danger of it being re-added.
9610 if (!tg->cfs_rq[cpu]->on_list)
9613 raw_spin_lock_irqsave(&rq->lock, flags);
9614 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9615 raw_spin_unlock_irqrestore(&rq->lock, flags);
9618 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9619 struct sched_entity *se, int cpu,
9620 struct sched_entity *parent)
9622 struct rq *rq = cpu_rq(cpu);
9626 init_cfs_rq_runtime(cfs_rq);
9628 tg->cfs_rq[cpu] = cfs_rq;
9631 /* se could be NULL for root_task_group */
9636 se->cfs_rq = &rq->cfs;
9639 se->cfs_rq = parent->my_q;
9640 se->depth = parent->depth + 1;
9644 /* guarantee group entities always have weight */
9645 update_load_set(&se->load, NICE_0_LOAD);
9646 se->parent = parent;
9649 static DEFINE_MUTEX(shares_mutex);
9651 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9654 unsigned long flags;
9657 * We can't change the weight of the root cgroup.
9662 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9664 mutex_lock(&shares_mutex);
9665 if (tg->shares == shares)
9668 tg->shares = shares;
9669 for_each_possible_cpu(i) {
9670 struct rq *rq = cpu_rq(i);
9671 struct sched_entity *se;
9674 /* Propagate contribution to hierarchy */
9675 raw_spin_lock_irqsave(&rq->lock, flags);
9677 /* Possible calls to update_curr() need rq clock */
9678 update_rq_clock(rq);
9679 for_each_sched_entity(se)
9680 update_cfs_shares(group_cfs_rq(se));
9681 raw_spin_unlock_irqrestore(&rq->lock, flags);
9685 mutex_unlock(&shares_mutex);
9688 #else /* CONFIG_FAIR_GROUP_SCHED */
9690 void free_fair_sched_group(struct task_group *tg) { }
9692 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9697 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9699 #endif /* CONFIG_FAIR_GROUP_SCHED */
9702 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9704 struct sched_entity *se = &task->se;
9705 unsigned int rr_interval = 0;
9708 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9711 if (rq->cfs.load.weight)
9712 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9718 * All the scheduling class methods:
9720 const struct sched_class fair_sched_class = {
9721 .next = &idle_sched_class,
9722 .enqueue_task = enqueue_task_fair,
9723 .dequeue_task = dequeue_task_fair,
9724 .yield_task = yield_task_fair,
9725 .yield_to_task = yield_to_task_fair,
9727 .check_preempt_curr = check_preempt_wakeup,
9729 .pick_next_task = pick_next_task_fair,
9730 .put_prev_task = put_prev_task_fair,
9733 .select_task_rq = select_task_rq_fair,
9734 .migrate_task_rq = migrate_task_rq_fair,
9736 .rq_online = rq_online_fair,
9737 .rq_offline = rq_offline_fair,
9739 .task_waking = task_waking_fair,
9740 .task_dead = task_dead_fair,
9741 .set_cpus_allowed = set_cpus_allowed_common,
9744 .set_curr_task = set_curr_task_fair,
9745 .task_tick = task_tick_fair,
9746 .task_fork = task_fork_fair,
9748 .prio_changed = prio_changed_fair,
9749 .switched_from = switched_from_fair,
9750 .switched_to = switched_to_fair,
9752 .get_rr_interval = get_rr_interval_fair,
9754 .update_curr = update_curr_fair,
9756 #ifdef CONFIG_FAIR_GROUP_SCHED
9757 .task_move_group = task_move_group_fair,
9761 #ifdef CONFIG_SCHED_DEBUG
9762 void print_cfs_stats(struct seq_file *m, int cpu)
9764 struct cfs_rq *cfs_rq;
9767 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9768 print_cfs_rq(m, cpu, cfs_rq);
9772 #ifdef CONFIG_NUMA_BALANCING
9773 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9776 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9778 for_each_online_node(node) {
9779 if (p->numa_faults) {
9780 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9781 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9783 if (p->numa_group) {
9784 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9785 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9787 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9790 #endif /* CONFIG_NUMA_BALANCING */
9791 #endif /* CONFIG_SCHED_DEBUG */
9793 __init void init_sched_fair_class(void)
9796 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9798 #ifdef CONFIG_NO_HZ_COMMON
9799 nohz.next_balance = jiffies;
9800 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9801 cpu_notifier(sched_ilb_notifier, 0);