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;
743 * Tasks are intialized with full load to be seen as heavy tasks until
744 * they get a chance to stabilize to their real load level.
745 * Group entities are intialized with zero load to reflect the fact that
746 * nothing has been attached to the task group yet.
748 if (entity_is_task(se))
749 sa->load_avg = scale_load_down(se->load.weight);
750 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
752 * In previous Android versions, we used to have:
753 * sa->util_avg = sched_freq() ?
754 * sysctl_sched_initial_task_util :
755 * scale_load_down(SCHED_LOAD_SCALE);
756 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
757 * However, that functionality has been moved to enqueue.
758 * It is unclear if we should restore this in enqueue.
761 * At this point, util_avg won't be used in select_task_rq_fair anyway
765 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
768 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
769 static void attach_entity_cfs_rq(struct sched_entity *se);
772 * With new tasks being created, their initial util_avgs are extrapolated
773 * based on the cfs_rq's current util_avg:
775 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
777 * However, in many cases, the above util_avg does not give a desired
778 * value. Moreover, the sum of the util_avgs may be divergent, such
779 * as when the series is a harmonic series.
781 * To solve this problem, we also cap the util_avg of successive tasks to
782 * only 1/2 of the left utilization budget:
784 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
786 * where n denotes the nth task.
788 * For example, a simplest series from the beginning would be like:
790 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
791 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
793 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
794 * if util_avg > util_avg_cap.
796 void post_init_entity_util_avg(struct sched_entity *se)
798 struct cfs_rq *cfs_rq = cfs_rq_of(se);
799 struct sched_avg *sa = &se->avg;
800 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
803 if (cfs_rq->avg.util_avg != 0) {
804 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
805 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
807 if (sa->util_avg > cap)
813 * If we wish to restore tuning via setting initial util,
814 * this is where we should do it.
816 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
819 if (entity_is_task(se)) {
820 struct task_struct *p = task_of(se);
821 if (p->sched_class != &fair_sched_class) {
823 * For !fair tasks do:
825 update_cfs_rq_load_avg(now, cfs_rq, false);
826 attach_entity_load_avg(cfs_rq, se);
827 switched_from_fair(rq, p);
829 * such that the next switched_to_fair() has the
832 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
837 attach_entity_cfs_rq(se);
841 void init_entity_runnable_average(struct sched_entity *se)
844 void post_init_entity_util_avg(struct sched_entity *se)
847 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
850 #endif /* CONFIG_SMP */
853 * Update the current task's runtime statistics.
855 static void update_curr(struct cfs_rq *cfs_rq)
857 struct sched_entity *curr = cfs_rq->curr;
858 u64 now = rq_clock_task(rq_of(cfs_rq));
864 delta_exec = now - curr->exec_start;
865 if (unlikely((s64)delta_exec <= 0))
868 curr->exec_start = now;
870 schedstat_set(curr->statistics.exec_max,
871 max(delta_exec, curr->statistics.exec_max));
873 curr->sum_exec_runtime += delta_exec;
874 schedstat_add(cfs_rq, exec_clock, delta_exec);
876 curr->vruntime += calc_delta_fair(delta_exec, curr);
877 update_min_vruntime(cfs_rq);
879 if (entity_is_task(curr)) {
880 struct task_struct *curtask = task_of(curr);
882 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
883 cpuacct_charge(curtask, delta_exec);
884 account_group_exec_runtime(curtask, delta_exec);
887 account_cfs_rq_runtime(cfs_rq, delta_exec);
890 static void update_curr_fair(struct rq *rq)
892 update_curr(cfs_rq_of(&rq->curr->se));
896 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
898 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
902 * Task is being enqueued - update stats:
904 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
907 * Are we enqueueing a waiting task? (for current tasks
908 * a dequeue/enqueue event is a NOP)
910 if (se != cfs_rq->curr)
911 update_stats_wait_start(cfs_rq, se);
915 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
917 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
918 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
919 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
920 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
921 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
922 #ifdef CONFIG_SCHEDSTATS
923 if (entity_is_task(se)) {
924 trace_sched_stat_wait(task_of(se),
925 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
928 schedstat_set(se->statistics.wait_start, 0);
932 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
935 * Mark the end of the wait period if dequeueing a
938 if (se != cfs_rq->curr)
939 update_stats_wait_end(cfs_rq, se);
943 * We are picking a new current task - update its stats:
946 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
949 * We are starting a new run period:
951 se->exec_start = rq_clock_task(rq_of(cfs_rq));
954 /**************************************************
955 * Scheduling class queueing methods:
958 #ifdef CONFIG_NUMA_BALANCING
960 * Approximate time to scan a full NUMA task in ms. The task scan period is
961 * calculated based on the tasks virtual memory size and
962 * numa_balancing_scan_size.
964 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
965 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
967 /* Portion of address space to scan in MB */
968 unsigned int sysctl_numa_balancing_scan_size = 256;
970 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
971 unsigned int sysctl_numa_balancing_scan_delay = 1000;
973 static unsigned int task_nr_scan_windows(struct task_struct *p)
975 unsigned long rss = 0;
976 unsigned long nr_scan_pages;
979 * Calculations based on RSS as non-present and empty pages are skipped
980 * by the PTE scanner and NUMA hinting faults should be trapped based
983 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
984 rss = get_mm_rss(p->mm);
988 rss = round_up(rss, nr_scan_pages);
989 return rss / nr_scan_pages;
992 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
993 #define MAX_SCAN_WINDOW 2560
995 static unsigned int task_scan_min(struct task_struct *p)
997 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
998 unsigned int scan, floor;
999 unsigned int windows = 1;
1001 if (scan_size < MAX_SCAN_WINDOW)
1002 windows = MAX_SCAN_WINDOW / scan_size;
1003 floor = 1000 / windows;
1005 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1006 return max_t(unsigned int, floor, scan);
1009 static unsigned int task_scan_max(struct task_struct *p)
1011 unsigned int smin = task_scan_min(p);
1014 /* Watch for min being lower than max due to floor calculations */
1015 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1016 return max(smin, smax);
1019 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1021 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1022 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1025 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1027 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1028 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1034 spinlock_t lock; /* nr_tasks, tasks */
1038 struct rcu_head rcu;
1039 nodemask_t active_nodes;
1040 unsigned long total_faults;
1042 * Faults_cpu is used to decide whether memory should move
1043 * towards the CPU. As a consequence, these stats are weighted
1044 * more by CPU use than by memory faults.
1046 unsigned long *faults_cpu;
1047 unsigned long faults[0];
1050 /* Shared or private faults. */
1051 #define NR_NUMA_HINT_FAULT_TYPES 2
1053 /* Memory and CPU locality */
1054 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1056 /* Averaged statistics, and temporary buffers. */
1057 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1059 pid_t task_numa_group_id(struct task_struct *p)
1061 return p->numa_group ? p->numa_group->gid : 0;
1065 * The averaged statistics, shared & private, memory & cpu,
1066 * occupy the first half of the array. The second half of the
1067 * array is for current counters, which are averaged into the
1068 * first set by task_numa_placement.
1070 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1072 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1075 static inline unsigned long task_faults(struct task_struct *p, int nid)
1077 if (!p->numa_faults)
1080 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1081 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1084 static inline unsigned long group_faults(struct task_struct *p, int nid)
1089 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1090 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1093 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1095 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1096 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1099 /* Handle placement on systems where not all nodes are directly connected. */
1100 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1101 int maxdist, bool task)
1103 unsigned long score = 0;
1107 * All nodes are directly connected, and the same distance
1108 * from each other. No need for fancy placement algorithms.
1110 if (sched_numa_topology_type == NUMA_DIRECT)
1114 * This code is called for each node, introducing N^2 complexity,
1115 * which should be ok given the number of nodes rarely exceeds 8.
1117 for_each_online_node(node) {
1118 unsigned long faults;
1119 int dist = node_distance(nid, node);
1122 * The furthest away nodes in the system are not interesting
1123 * for placement; nid was already counted.
1125 if (dist == sched_max_numa_distance || node == nid)
1129 * On systems with a backplane NUMA topology, compare groups
1130 * of nodes, and move tasks towards the group with the most
1131 * memory accesses. When comparing two nodes at distance
1132 * "hoplimit", only nodes closer by than "hoplimit" are part
1133 * of each group. Skip other nodes.
1135 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1139 /* Add up the faults from nearby nodes. */
1141 faults = task_faults(p, node);
1143 faults = group_faults(p, node);
1146 * On systems with a glueless mesh NUMA topology, there are
1147 * no fixed "groups of nodes". Instead, nodes that are not
1148 * directly connected bounce traffic through intermediate
1149 * nodes; a numa_group can occupy any set of nodes.
1150 * The further away a node is, the less the faults count.
1151 * This seems to result in good task placement.
1153 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1154 faults *= (sched_max_numa_distance - dist);
1155 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1165 * These return the fraction of accesses done by a particular task, or
1166 * task group, on a particular numa node. The group weight is given a
1167 * larger multiplier, in order to group tasks together that are almost
1168 * evenly spread out between numa nodes.
1170 static inline unsigned long task_weight(struct task_struct *p, int nid,
1173 unsigned long faults, total_faults;
1175 if (!p->numa_faults)
1178 total_faults = p->total_numa_faults;
1183 faults = task_faults(p, nid);
1184 faults += score_nearby_nodes(p, nid, dist, true);
1186 return 1000 * faults / total_faults;
1189 static inline unsigned long group_weight(struct task_struct *p, int nid,
1192 unsigned long faults, total_faults;
1197 total_faults = p->numa_group->total_faults;
1202 faults = group_faults(p, nid);
1203 faults += score_nearby_nodes(p, nid, dist, false);
1205 return 1000 * faults / total_faults;
1208 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1209 int src_nid, int dst_cpu)
1211 struct numa_group *ng = p->numa_group;
1212 int dst_nid = cpu_to_node(dst_cpu);
1213 int last_cpupid, this_cpupid;
1215 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1218 * Multi-stage node selection is used in conjunction with a periodic
1219 * migration fault to build a temporal task<->page relation. By using
1220 * a two-stage filter we remove short/unlikely relations.
1222 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1223 * a task's usage of a particular page (n_p) per total usage of this
1224 * page (n_t) (in a given time-span) to a probability.
1226 * Our periodic faults will sample this probability and getting the
1227 * same result twice in a row, given these samples are fully
1228 * independent, is then given by P(n)^2, provided our sample period
1229 * is sufficiently short compared to the usage pattern.
1231 * This quadric squishes small probabilities, making it less likely we
1232 * act on an unlikely task<->page relation.
1234 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1235 if (!cpupid_pid_unset(last_cpupid) &&
1236 cpupid_to_nid(last_cpupid) != dst_nid)
1239 /* Always allow migrate on private faults */
1240 if (cpupid_match_pid(p, last_cpupid))
1243 /* A shared fault, but p->numa_group has not been set up yet. */
1248 * Do not migrate if the destination is not a node that
1249 * is actively used by this numa group.
1251 if (!node_isset(dst_nid, ng->active_nodes))
1255 * Source is a node that is not actively used by this
1256 * numa group, while the destination is. Migrate.
1258 if (!node_isset(src_nid, ng->active_nodes))
1262 * Both source and destination are nodes in active
1263 * use by this numa group. Maximize memory bandwidth
1264 * by migrating from more heavily used groups, to less
1265 * heavily used ones, spreading the load around.
1266 * Use a 1/4 hysteresis to avoid spurious page movement.
1268 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1271 static unsigned long weighted_cpuload(const int cpu);
1272 static unsigned long source_load(int cpu, int type);
1273 static unsigned long target_load(int cpu, int type);
1274 static unsigned long capacity_of(int cpu);
1275 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1277 /* Cached statistics for all CPUs within a node */
1279 unsigned long nr_running;
1282 /* Total compute capacity of CPUs on a node */
1283 unsigned long compute_capacity;
1285 /* Approximate capacity in terms of runnable tasks on a node */
1286 unsigned long task_capacity;
1287 int has_free_capacity;
1291 * XXX borrowed from update_sg_lb_stats
1293 static void update_numa_stats(struct numa_stats *ns, int nid)
1295 int smt, cpu, cpus = 0;
1296 unsigned long capacity;
1298 memset(ns, 0, sizeof(*ns));
1299 for_each_cpu(cpu, cpumask_of_node(nid)) {
1300 struct rq *rq = cpu_rq(cpu);
1302 ns->nr_running += rq->nr_running;
1303 ns->load += weighted_cpuload(cpu);
1304 ns->compute_capacity += capacity_of(cpu);
1310 * If we raced with hotplug and there are no CPUs left in our mask
1311 * the @ns structure is NULL'ed and task_numa_compare() will
1312 * not find this node attractive.
1314 * We'll either bail at !has_free_capacity, or we'll detect a huge
1315 * imbalance and bail there.
1320 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1321 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1322 capacity = cpus / smt; /* cores */
1324 ns->task_capacity = min_t(unsigned, capacity,
1325 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1326 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1329 struct task_numa_env {
1330 struct task_struct *p;
1332 int src_cpu, src_nid;
1333 int dst_cpu, dst_nid;
1335 struct numa_stats src_stats, dst_stats;
1340 struct task_struct *best_task;
1345 static void task_numa_assign(struct task_numa_env *env,
1346 struct task_struct *p, long imp)
1349 put_task_struct(env->best_task);
1352 env->best_imp = imp;
1353 env->best_cpu = env->dst_cpu;
1356 static bool load_too_imbalanced(long src_load, long dst_load,
1357 struct task_numa_env *env)
1360 long orig_src_load, orig_dst_load;
1361 long src_capacity, dst_capacity;
1364 * The load is corrected for the CPU capacity available on each node.
1367 * ------------ vs ---------
1368 * src_capacity dst_capacity
1370 src_capacity = env->src_stats.compute_capacity;
1371 dst_capacity = env->dst_stats.compute_capacity;
1373 /* We care about the slope of the imbalance, not the direction. */
1374 if (dst_load < src_load)
1375 swap(dst_load, src_load);
1377 /* Is the difference below the threshold? */
1378 imb = dst_load * src_capacity * 100 -
1379 src_load * dst_capacity * env->imbalance_pct;
1384 * The imbalance is above the allowed threshold.
1385 * Compare it with the old imbalance.
1387 orig_src_load = env->src_stats.load;
1388 orig_dst_load = env->dst_stats.load;
1390 if (orig_dst_load < orig_src_load)
1391 swap(orig_dst_load, orig_src_load);
1393 old_imb = orig_dst_load * src_capacity * 100 -
1394 orig_src_load * dst_capacity * env->imbalance_pct;
1396 /* Would this change make things worse? */
1397 return (imb > old_imb);
1401 * This checks if the overall compute and NUMA accesses of the system would
1402 * be improved if the source tasks was migrated to the target dst_cpu taking
1403 * into account that it might be best if task running on the dst_cpu should
1404 * be exchanged with the source task
1406 static void task_numa_compare(struct task_numa_env *env,
1407 long taskimp, long groupimp)
1409 struct rq *src_rq = cpu_rq(env->src_cpu);
1410 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1411 struct task_struct *cur;
1412 long src_load, dst_load;
1414 long imp = env->p->numa_group ? groupimp : taskimp;
1416 int dist = env->dist;
1417 bool assigned = false;
1421 raw_spin_lock_irq(&dst_rq->lock);
1424 * No need to move the exiting task or idle task.
1426 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1430 * The task_struct must be protected here to protect the
1431 * p->numa_faults access in the task_weight since the
1432 * numa_faults could already be freed in the following path:
1433 * finish_task_switch()
1434 * --> put_task_struct()
1435 * --> __put_task_struct()
1436 * --> task_numa_free()
1438 get_task_struct(cur);
1441 raw_spin_unlock_irq(&dst_rq->lock);
1444 * Because we have preemption enabled we can get migrated around and
1445 * end try selecting ourselves (current == env->p) as a swap candidate.
1451 * "imp" is the fault differential for the source task between the
1452 * source and destination node. Calculate the total differential for
1453 * the source task and potential destination task. The more negative
1454 * the value is, the more rmeote accesses that would be expected to
1455 * be incurred if the tasks were swapped.
1458 /* Skip this swap candidate if cannot move to the source cpu */
1459 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1463 * If dst and source tasks are in the same NUMA group, or not
1464 * in any group then look only at task weights.
1466 if (cur->numa_group == env->p->numa_group) {
1467 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1468 task_weight(cur, env->dst_nid, dist);
1470 * Add some hysteresis to prevent swapping the
1471 * tasks within a group over tiny differences.
1473 if (cur->numa_group)
1477 * Compare the group weights. If a task is all by
1478 * itself (not part of a group), use the task weight
1481 if (cur->numa_group)
1482 imp += group_weight(cur, env->src_nid, dist) -
1483 group_weight(cur, env->dst_nid, dist);
1485 imp += task_weight(cur, env->src_nid, dist) -
1486 task_weight(cur, env->dst_nid, dist);
1490 if (imp <= env->best_imp && moveimp <= env->best_imp)
1494 /* Is there capacity at our destination? */
1495 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1496 !env->dst_stats.has_free_capacity)
1502 /* Balance doesn't matter much if we're running a task per cpu */
1503 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1504 dst_rq->nr_running == 1)
1508 * In the overloaded case, try and keep the load balanced.
1511 load = task_h_load(env->p);
1512 dst_load = env->dst_stats.load + load;
1513 src_load = env->src_stats.load - load;
1515 if (moveimp > imp && moveimp > env->best_imp) {
1517 * If the improvement from just moving env->p direction is
1518 * better than swapping tasks around, check if a move is
1519 * possible. Store a slightly smaller score than moveimp,
1520 * so an actually idle CPU will win.
1522 if (!load_too_imbalanced(src_load, dst_load, env)) {
1524 put_task_struct(cur);
1530 if (imp <= env->best_imp)
1534 load = task_h_load(cur);
1539 if (load_too_imbalanced(src_load, dst_load, env))
1543 * One idle CPU per node is evaluated for a task numa move.
1544 * Call select_idle_sibling to maybe find a better one.
1547 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1552 task_numa_assign(env, cur, imp);
1556 * The dst_rq->curr isn't assigned. The protection for task_struct is
1559 if (cur && !assigned)
1560 put_task_struct(cur);
1563 static void task_numa_find_cpu(struct task_numa_env *env,
1564 long taskimp, long groupimp)
1568 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1569 /* Skip this CPU if the source task cannot migrate */
1570 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1574 task_numa_compare(env, taskimp, groupimp);
1578 /* Only move tasks to a NUMA node less busy than the current node. */
1579 static bool numa_has_capacity(struct task_numa_env *env)
1581 struct numa_stats *src = &env->src_stats;
1582 struct numa_stats *dst = &env->dst_stats;
1584 if (src->has_free_capacity && !dst->has_free_capacity)
1588 * Only consider a task move if the source has a higher load
1589 * than the destination, corrected for CPU capacity on each node.
1591 * src->load dst->load
1592 * --------------------- vs ---------------------
1593 * src->compute_capacity dst->compute_capacity
1595 if (src->load * dst->compute_capacity * env->imbalance_pct >
1597 dst->load * src->compute_capacity * 100)
1603 static int task_numa_migrate(struct task_struct *p)
1605 struct task_numa_env env = {
1608 .src_cpu = task_cpu(p),
1609 .src_nid = task_node(p),
1611 .imbalance_pct = 112,
1617 struct sched_domain *sd;
1618 unsigned long taskweight, groupweight;
1620 long taskimp, groupimp;
1623 * Pick the lowest SD_NUMA domain, as that would have the smallest
1624 * imbalance and would be the first to start moving tasks about.
1626 * And we want to avoid any moving of tasks about, as that would create
1627 * random movement of tasks -- counter the numa conditions we're trying
1631 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1633 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1637 * Cpusets can break the scheduler domain tree into smaller
1638 * balance domains, some of which do not cross NUMA boundaries.
1639 * Tasks that are "trapped" in such domains cannot be migrated
1640 * elsewhere, so there is no point in (re)trying.
1642 if (unlikely(!sd)) {
1643 p->numa_preferred_nid = task_node(p);
1647 env.dst_nid = p->numa_preferred_nid;
1648 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1649 taskweight = task_weight(p, env.src_nid, dist);
1650 groupweight = group_weight(p, env.src_nid, dist);
1651 update_numa_stats(&env.src_stats, env.src_nid);
1652 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1653 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1654 update_numa_stats(&env.dst_stats, env.dst_nid);
1656 /* Try to find a spot on the preferred nid. */
1657 if (numa_has_capacity(&env))
1658 task_numa_find_cpu(&env, taskimp, groupimp);
1661 * Look at other nodes in these cases:
1662 * - there is no space available on the preferred_nid
1663 * - the task is part of a numa_group that is interleaved across
1664 * multiple NUMA nodes; in order to better consolidate the group,
1665 * we need to check other locations.
1667 if (env.best_cpu == -1 || (p->numa_group &&
1668 nodes_weight(p->numa_group->active_nodes) > 1)) {
1669 for_each_online_node(nid) {
1670 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1673 dist = node_distance(env.src_nid, env.dst_nid);
1674 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1676 taskweight = task_weight(p, env.src_nid, dist);
1677 groupweight = group_weight(p, env.src_nid, dist);
1680 /* Only consider nodes where both task and groups benefit */
1681 taskimp = task_weight(p, nid, dist) - taskweight;
1682 groupimp = group_weight(p, nid, dist) - groupweight;
1683 if (taskimp < 0 && groupimp < 0)
1688 update_numa_stats(&env.dst_stats, env.dst_nid);
1689 if (numa_has_capacity(&env))
1690 task_numa_find_cpu(&env, taskimp, groupimp);
1695 * If the task is part of a workload that spans multiple NUMA nodes,
1696 * and is migrating into one of the workload's active nodes, remember
1697 * this node as the task's preferred numa node, so the workload can
1699 * A task that migrated to a second choice node will be better off
1700 * trying for a better one later. Do not set the preferred node here.
1702 if (p->numa_group) {
1703 if (env.best_cpu == -1)
1708 if (node_isset(nid, p->numa_group->active_nodes))
1709 sched_setnuma(p, env.dst_nid);
1712 /* No better CPU than the current one was found. */
1713 if (env.best_cpu == -1)
1717 * Reset the scan period if the task is being rescheduled on an
1718 * alternative node to recheck if the tasks is now properly placed.
1720 p->numa_scan_period = task_scan_min(p);
1722 if (env.best_task == NULL) {
1723 ret = migrate_task_to(p, env.best_cpu);
1725 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1729 ret = migrate_swap(p, env.best_task);
1731 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1732 put_task_struct(env.best_task);
1736 /* Attempt to migrate a task to a CPU on the preferred node. */
1737 static void numa_migrate_preferred(struct task_struct *p)
1739 unsigned long interval = HZ;
1741 /* This task has no NUMA fault statistics yet */
1742 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1745 /* Periodically retry migrating the task to the preferred node */
1746 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1747 p->numa_migrate_retry = jiffies + interval;
1749 /* Success if task is already running on preferred CPU */
1750 if (task_node(p) == p->numa_preferred_nid)
1753 /* Otherwise, try migrate to a CPU on the preferred node */
1754 task_numa_migrate(p);
1758 * Find the nodes on which the workload is actively running. We do this by
1759 * tracking the nodes from which NUMA hinting faults are triggered. This can
1760 * be different from the set of nodes where the workload's memory is currently
1763 * The bitmask is used to make smarter decisions on when to do NUMA page
1764 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1765 * are added when they cause over 6/16 of the maximum number of faults, but
1766 * only removed when they drop below 3/16.
1768 static void update_numa_active_node_mask(struct numa_group *numa_group)
1770 unsigned long faults, max_faults = 0;
1773 for_each_online_node(nid) {
1774 faults = group_faults_cpu(numa_group, nid);
1775 if (faults > max_faults)
1776 max_faults = faults;
1779 for_each_online_node(nid) {
1780 faults = group_faults_cpu(numa_group, nid);
1781 if (!node_isset(nid, numa_group->active_nodes)) {
1782 if (faults > max_faults * 6 / 16)
1783 node_set(nid, numa_group->active_nodes);
1784 } else if (faults < max_faults * 3 / 16)
1785 node_clear(nid, numa_group->active_nodes);
1790 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1791 * increments. The more local the fault statistics are, the higher the scan
1792 * period will be for the next scan window. If local/(local+remote) ratio is
1793 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1794 * the scan period will decrease. Aim for 70% local accesses.
1796 #define NUMA_PERIOD_SLOTS 10
1797 #define NUMA_PERIOD_THRESHOLD 7
1800 * Increase the scan period (slow down scanning) if the majority of
1801 * our memory is already on our local node, or if the majority of
1802 * the page accesses are shared with other processes.
1803 * Otherwise, decrease the scan period.
1805 static void update_task_scan_period(struct task_struct *p,
1806 unsigned long shared, unsigned long private)
1808 unsigned int period_slot;
1812 unsigned long remote = p->numa_faults_locality[0];
1813 unsigned long local = p->numa_faults_locality[1];
1816 * If there were no record hinting faults then either the task is
1817 * completely idle or all activity is areas that are not of interest
1818 * to automatic numa balancing. Related to that, if there were failed
1819 * migration then it implies we are migrating too quickly or the local
1820 * node is overloaded. In either case, scan slower
1822 if (local + shared == 0 || p->numa_faults_locality[2]) {
1823 p->numa_scan_period = min(p->numa_scan_period_max,
1824 p->numa_scan_period << 1);
1826 p->mm->numa_next_scan = jiffies +
1827 msecs_to_jiffies(p->numa_scan_period);
1833 * Prepare to scale scan period relative to the current period.
1834 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1835 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1836 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1838 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1839 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1840 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1841 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1844 diff = slot * period_slot;
1846 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1849 * Scale scan rate increases based on sharing. There is an
1850 * inverse relationship between the degree of sharing and
1851 * the adjustment made to the scanning period. Broadly
1852 * speaking the intent is that there is little point
1853 * scanning faster if shared accesses dominate as it may
1854 * simply bounce migrations uselessly
1856 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1857 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1860 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1861 task_scan_min(p), task_scan_max(p));
1862 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1866 * Get the fraction of time the task has been running since the last
1867 * NUMA placement cycle. The scheduler keeps similar statistics, but
1868 * decays those on a 32ms period, which is orders of magnitude off
1869 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1870 * stats only if the task is so new there are no NUMA statistics yet.
1872 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1874 u64 runtime, delta, now;
1875 /* Use the start of this time slice to avoid calculations. */
1876 now = p->se.exec_start;
1877 runtime = p->se.sum_exec_runtime;
1879 if (p->last_task_numa_placement) {
1880 delta = runtime - p->last_sum_exec_runtime;
1881 *period = now - p->last_task_numa_placement;
1883 delta = p->se.avg.load_sum / p->se.load.weight;
1884 *period = LOAD_AVG_MAX;
1887 p->last_sum_exec_runtime = runtime;
1888 p->last_task_numa_placement = now;
1894 * Determine the preferred nid for a task in a numa_group. This needs to
1895 * be done in a way that produces consistent results with group_weight,
1896 * otherwise workloads might not converge.
1898 static int preferred_group_nid(struct task_struct *p, int nid)
1903 /* Direct connections between all NUMA nodes. */
1904 if (sched_numa_topology_type == NUMA_DIRECT)
1908 * On a system with glueless mesh NUMA topology, group_weight
1909 * scores nodes according to the number of NUMA hinting faults on
1910 * both the node itself, and on nearby nodes.
1912 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1913 unsigned long score, max_score = 0;
1914 int node, max_node = nid;
1916 dist = sched_max_numa_distance;
1918 for_each_online_node(node) {
1919 score = group_weight(p, node, dist);
1920 if (score > max_score) {
1929 * Finding the preferred nid in a system with NUMA backplane
1930 * interconnect topology is more involved. The goal is to locate
1931 * tasks from numa_groups near each other in the system, and
1932 * untangle workloads from different sides of the system. This requires
1933 * searching down the hierarchy of node groups, recursively searching
1934 * inside the highest scoring group of nodes. The nodemask tricks
1935 * keep the complexity of the search down.
1937 nodes = node_online_map;
1938 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1939 unsigned long max_faults = 0;
1940 nodemask_t max_group = NODE_MASK_NONE;
1943 /* Are there nodes at this distance from each other? */
1944 if (!find_numa_distance(dist))
1947 for_each_node_mask(a, nodes) {
1948 unsigned long faults = 0;
1949 nodemask_t this_group;
1950 nodes_clear(this_group);
1952 /* Sum group's NUMA faults; includes a==b case. */
1953 for_each_node_mask(b, nodes) {
1954 if (node_distance(a, b) < dist) {
1955 faults += group_faults(p, b);
1956 node_set(b, this_group);
1957 node_clear(b, nodes);
1961 /* Remember the top group. */
1962 if (faults > max_faults) {
1963 max_faults = faults;
1964 max_group = this_group;
1966 * subtle: at the smallest distance there is
1967 * just one node left in each "group", the
1968 * winner is the preferred nid.
1973 /* Next round, evaluate the nodes within max_group. */
1981 static void task_numa_placement(struct task_struct *p)
1983 int seq, nid, max_nid = -1, max_group_nid = -1;
1984 unsigned long max_faults = 0, max_group_faults = 0;
1985 unsigned long fault_types[2] = { 0, 0 };
1986 unsigned long total_faults;
1987 u64 runtime, period;
1988 spinlock_t *group_lock = NULL;
1991 * The p->mm->numa_scan_seq field gets updated without
1992 * exclusive access. Use READ_ONCE() here to ensure
1993 * that the field is read in a single access:
1995 seq = READ_ONCE(p->mm->numa_scan_seq);
1996 if (p->numa_scan_seq == seq)
1998 p->numa_scan_seq = seq;
1999 p->numa_scan_period_max = task_scan_max(p);
2001 total_faults = p->numa_faults_locality[0] +
2002 p->numa_faults_locality[1];
2003 runtime = numa_get_avg_runtime(p, &period);
2005 /* If the task is part of a group prevent parallel updates to group stats */
2006 if (p->numa_group) {
2007 group_lock = &p->numa_group->lock;
2008 spin_lock_irq(group_lock);
2011 /* Find the node with the highest number of faults */
2012 for_each_online_node(nid) {
2013 /* Keep track of the offsets in numa_faults array */
2014 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2015 unsigned long faults = 0, group_faults = 0;
2018 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2019 long diff, f_diff, f_weight;
2021 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2022 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2023 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2024 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2026 /* Decay existing window, copy faults since last scan */
2027 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2028 fault_types[priv] += p->numa_faults[membuf_idx];
2029 p->numa_faults[membuf_idx] = 0;
2032 * Normalize the faults_from, so all tasks in a group
2033 * count according to CPU use, instead of by the raw
2034 * number of faults. Tasks with little runtime have
2035 * little over-all impact on throughput, and thus their
2036 * faults are less important.
2038 f_weight = div64_u64(runtime << 16, period + 1);
2039 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2041 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2042 p->numa_faults[cpubuf_idx] = 0;
2044 p->numa_faults[mem_idx] += diff;
2045 p->numa_faults[cpu_idx] += f_diff;
2046 faults += p->numa_faults[mem_idx];
2047 p->total_numa_faults += diff;
2048 if (p->numa_group) {
2050 * safe because we can only change our own group
2052 * mem_idx represents the offset for a given
2053 * nid and priv in a specific region because it
2054 * is at the beginning of the numa_faults array.
2056 p->numa_group->faults[mem_idx] += diff;
2057 p->numa_group->faults_cpu[mem_idx] += f_diff;
2058 p->numa_group->total_faults += diff;
2059 group_faults += p->numa_group->faults[mem_idx];
2063 if (faults > max_faults) {
2064 max_faults = faults;
2068 if (group_faults > max_group_faults) {
2069 max_group_faults = group_faults;
2070 max_group_nid = nid;
2074 update_task_scan_period(p, fault_types[0], fault_types[1]);
2076 if (p->numa_group) {
2077 update_numa_active_node_mask(p->numa_group);
2078 spin_unlock_irq(group_lock);
2079 max_nid = preferred_group_nid(p, max_group_nid);
2083 /* Set the new preferred node */
2084 if (max_nid != p->numa_preferred_nid)
2085 sched_setnuma(p, max_nid);
2087 if (task_node(p) != p->numa_preferred_nid)
2088 numa_migrate_preferred(p);
2092 static inline int get_numa_group(struct numa_group *grp)
2094 return atomic_inc_not_zero(&grp->refcount);
2097 static inline void put_numa_group(struct numa_group *grp)
2099 if (atomic_dec_and_test(&grp->refcount))
2100 kfree_rcu(grp, rcu);
2103 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2106 struct numa_group *grp, *my_grp;
2107 struct task_struct *tsk;
2109 int cpu = cpupid_to_cpu(cpupid);
2112 if (unlikely(!p->numa_group)) {
2113 unsigned int size = sizeof(struct numa_group) +
2114 4*nr_node_ids*sizeof(unsigned long);
2116 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2120 atomic_set(&grp->refcount, 1);
2121 spin_lock_init(&grp->lock);
2123 /* Second half of the array tracks nids where faults happen */
2124 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2127 node_set(task_node(current), grp->active_nodes);
2129 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2130 grp->faults[i] = p->numa_faults[i];
2132 grp->total_faults = p->total_numa_faults;
2135 rcu_assign_pointer(p->numa_group, grp);
2139 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2141 if (!cpupid_match_pid(tsk, cpupid))
2144 grp = rcu_dereference(tsk->numa_group);
2148 my_grp = p->numa_group;
2153 * Only join the other group if its bigger; if we're the bigger group,
2154 * the other task will join us.
2156 if (my_grp->nr_tasks > grp->nr_tasks)
2160 * Tie-break on the grp address.
2162 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2165 /* Always join threads in the same process. */
2166 if (tsk->mm == current->mm)
2169 /* Simple filter to avoid false positives due to PID collisions */
2170 if (flags & TNF_SHARED)
2173 /* Update priv based on whether false sharing was detected */
2176 if (join && !get_numa_group(grp))
2184 BUG_ON(irqs_disabled());
2185 double_lock_irq(&my_grp->lock, &grp->lock);
2187 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2188 my_grp->faults[i] -= p->numa_faults[i];
2189 grp->faults[i] += p->numa_faults[i];
2191 my_grp->total_faults -= p->total_numa_faults;
2192 grp->total_faults += p->total_numa_faults;
2197 spin_unlock(&my_grp->lock);
2198 spin_unlock_irq(&grp->lock);
2200 rcu_assign_pointer(p->numa_group, grp);
2202 put_numa_group(my_grp);
2210 void task_numa_free(struct task_struct *p)
2212 struct numa_group *grp = p->numa_group;
2213 void *numa_faults = p->numa_faults;
2214 unsigned long flags;
2218 spin_lock_irqsave(&grp->lock, flags);
2219 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2220 grp->faults[i] -= p->numa_faults[i];
2221 grp->total_faults -= p->total_numa_faults;
2224 spin_unlock_irqrestore(&grp->lock, flags);
2225 RCU_INIT_POINTER(p->numa_group, NULL);
2226 put_numa_group(grp);
2229 p->numa_faults = NULL;
2234 * Got a PROT_NONE fault for a page on @node.
2236 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2238 struct task_struct *p = current;
2239 bool migrated = flags & TNF_MIGRATED;
2240 int cpu_node = task_node(current);
2241 int local = !!(flags & TNF_FAULT_LOCAL);
2244 if (!static_branch_likely(&sched_numa_balancing))
2247 /* for example, ksmd faulting in a user's mm */
2251 /* Allocate buffer to track faults on a per-node basis */
2252 if (unlikely(!p->numa_faults)) {
2253 int size = sizeof(*p->numa_faults) *
2254 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2256 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2257 if (!p->numa_faults)
2260 p->total_numa_faults = 0;
2261 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2265 * First accesses are treated as private, otherwise consider accesses
2266 * to be private if the accessing pid has not changed
2268 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2271 priv = cpupid_match_pid(p, last_cpupid);
2272 if (!priv && !(flags & TNF_NO_GROUP))
2273 task_numa_group(p, last_cpupid, flags, &priv);
2277 * If a workload spans multiple NUMA nodes, a shared fault that
2278 * occurs wholly within the set of nodes that the workload is
2279 * actively using should be counted as local. This allows the
2280 * scan rate to slow down when a workload has settled down.
2282 if (!priv && !local && p->numa_group &&
2283 node_isset(cpu_node, p->numa_group->active_nodes) &&
2284 node_isset(mem_node, p->numa_group->active_nodes))
2287 task_numa_placement(p);
2290 * Retry task to preferred node migration periodically, in case it
2291 * case it previously failed, or the scheduler moved us.
2293 if (time_after(jiffies, p->numa_migrate_retry))
2294 numa_migrate_preferred(p);
2297 p->numa_pages_migrated += pages;
2298 if (flags & TNF_MIGRATE_FAIL)
2299 p->numa_faults_locality[2] += pages;
2301 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2302 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2303 p->numa_faults_locality[local] += pages;
2306 static void reset_ptenuma_scan(struct task_struct *p)
2309 * We only did a read acquisition of the mmap sem, so
2310 * p->mm->numa_scan_seq is written to without exclusive access
2311 * and the update is not guaranteed to be atomic. That's not
2312 * much of an issue though, since this is just used for
2313 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2314 * expensive, to avoid any form of compiler optimizations:
2316 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2317 p->mm->numa_scan_offset = 0;
2321 * The expensive part of numa migration is done from task_work context.
2322 * Triggered from task_tick_numa().
2324 void task_numa_work(struct callback_head *work)
2326 unsigned long migrate, next_scan, now = jiffies;
2327 struct task_struct *p = current;
2328 struct mm_struct *mm = p->mm;
2329 struct vm_area_struct *vma;
2330 unsigned long start, end;
2331 unsigned long nr_pte_updates = 0;
2332 long pages, virtpages;
2334 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2336 work->next = work; /* protect against double add */
2338 * Who cares about NUMA placement when they're dying.
2340 * NOTE: make sure not to dereference p->mm before this check,
2341 * exit_task_work() happens _after_ exit_mm() so we could be called
2342 * without p->mm even though we still had it when we enqueued this
2345 if (p->flags & PF_EXITING)
2348 if (!mm->numa_next_scan) {
2349 mm->numa_next_scan = now +
2350 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2354 * Enforce maximal scan/migration frequency..
2356 migrate = mm->numa_next_scan;
2357 if (time_before(now, migrate))
2360 if (p->numa_scan_period == 0) {
2361 p->numa_scan_period_max = task_scan_max(p);
2362 p->numa_scan_period = task_scan_min(p);
2365 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2366 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2370 * Delay this task enough that another task of this mm will likely win
2371 * the next time around.
2373 p->node_stamp += 2 * TICK_NSEC;
2375 start = mm->numa_scan_offset;
2376 pages = sysctl_numa_balancing_scan_size;
2377 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2378 virtpages = pages * 8; /* Scan up to this much virtual space */
2383 down_read(&mm->mmap_sem);
2384 vma = find_vma(mm, start);
2386 reset_ptenuma_scan(p);
2390 for (; vma; vma = vma->vm_next) {
2391 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2392 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2397 * Shared library pages mapped by multiple processes are not
2398 * migrated as it is expected they are cache replicated. Avoid
2399 * hinting faults in read-only file-backed mappings or the vdso
2400 * as migrating the pages will be of marginal benefit.
2403 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2407 * Skip inaccessible VMAs to avoid any confusion between
2408 * PROT_NONE and NUMA hinting ptes
2410 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2414 start = max(start, vma->vm_start);
2415 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2416 end = min(end, vma->vm_end);
2417 nr_pte_updates = change_prot_numa(vma, start, end);
2420 * Try to scan sysctl_numa_balancing_size worth of
2421 * hpages that have at least one present PTE that
2422 * is not already pte-numa. If the VMA contains
2423 * areas that are unused or already full of prot_numa
2424 * PTEs, scan up to virtpages, to skip through those
2428 pages -= (end - start) >> PAGE_SHIFT;
2429 virtpages -= (end - start) >> PAGE_SHIFT;
2432 if (pages <= 0 || virtpages <= 0)
2436 } while (end != vma->vm_end);
2441 * It is possible to reach the end of the VMA list but the last few
2442 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2443 * would find the !migratable VMA on the next scan but not reset the
2444 * scanner to the start so check it now.
2447 mm->numa_scan_offset = start;
2449 reset_ptenuma_scan(p);
2450 up_read(&mm->mmap_sem);
2454 * Drive the periodic memory faults..
2456 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2458 struct callback_head *work = &curr->numa_work;
2462 * We don't care about NUMA placement if we don't have memory.
2464 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2468 * Using runtime rather than walltime has the dual advantage that
2469 * we (mostly) drive the selection from busy threads and that the
2470 * task needs to have done some actual work before we bother with
2473 now = curr->se.sum_exec_runtime;
2474 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2476 if (now > curr->node_stamp + period) {
2477 if (!curr->node_stamp)
2478 curr->numa_scan_period = task_scan_min(curr);
2479 curr->node_stamp += period;
2481 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2482 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2483 task_work_add(curr, work, true);
2488 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2492 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2496 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2499 #endif /* CONFIG_NUMA_BALANCING */
2502 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2504 update_load_add(&cfs_rq->load, se->load.weight);
2505 if (!parent_entity(se))
2506 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2508 if (entity_is_task(se)) {
2509 struct rq *rq = rq_of(cfs_rq);
2511 account_numa_enqueue(rq, task_of(se));
2512 list_add(&se->group_node, &rq->cfs_tasks);
2515 cfs_rq->nr_running++;
2519 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2521 update_load_sub(&cfs_rq->load, se->load.weight);
2522 if (!parent_entity(se))
2523 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2524 if (entity_is_task(se)) {
2525 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2526 list_del_init(&se->group_node);
2528 cfs_rq->nr_running--;
2531 #ifdef CONFIG_FAIR_GROUP_SCHED
2533 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2535 long tg_weight, load, shares;
2538 * This really should be: cfs_rq->avg.load_avg, but instead we use
2539 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2540 * the shares for small weight interactive tasks.
2542 load = scale_load_down(cfs_rq->load.weight);
2544 tg_weight = atomic_long_read(&tg->load_avg);
2546 /* Ensure tg_weight >= load */
2547 tg_weight -= cfs_rq->tg_load_avg_contrib;
2550 shares = (tg->shares * load);
2552 shares /= tg_weight;
2554 if (shares < MIN_SHARES)
2555 shares = MIN_SHARES;
2556 if (shares > tg->shares)
2557 shares = tg->shares;
2561 # else /* CONFIG_SMP */
2562 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2566 # 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 sched_entity *se)
2588 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2589 struct task_group *tg;
2595 if (throttled_hierarchy(cfs_rq))
2601 if (likely(se->load.weight == tg->shares))
2604 shares = calc_cfs_shares(cfs_rq, tg);
2606 reweight_entity(cfs_rq_of(se), se, shares);
2609 #else /* CONFIG_FAIR_GROUP_SCHED */
2610 static inline void update_cfs_shares(struct sched_entity *se)
2613 #endif /* CONFIG_FAIR_GROUP_SCHED */
2616 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2617 static const u32 runnable_avg_yN_inv[] = {
2618 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2619 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2620 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2621 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2622 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2623 0x85aac367, 0x82cd8698,
2627 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2628 * over-estimates when re-combining.
2630 static const u32 runnable_avg_yN_sum[] = {
2631 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2632 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2633 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2638 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2640 static __always_inline u64 decay_load(u64 val, u64 n)
2642 unsigned int local_n;
2646 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2649 /* after bounds checking we can collapse to 32-bit */
2653 * As y^PERIOD = 1/2, we can combine
2654 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2655 * With a look-up table which covers y^n (n<PERIOD)
2657 * To achieve constant time decay_load.
2659 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2660 val >>= local_n / LOAD_AVG_PERIOD;
2661 local_n %= LOAD_AVG_PERIOD;
2664 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2669 * For updates fully spanning n periods, the contribution to runnable
2670 * average will be: \Sum 1024*y^n
2672 * We can compute this reasonably efficiently by combining:
2673 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2675 static u32 __compute_runnable_contrib(u64 n)
2679 if (likely(n <= LOAD_AVG_PERIOD))
2680 return runnable_avg_yN_sum[n];
2681 else if (unlikely(n >= LOAD_AVG_MAX_N))
2682 return LOAD_AVG_MAX;
2684 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2686 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2687 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2689 n -= LOAD_AVG_PERIOD;
2690 } while (n > LOAD_AVG_PERIOD);
2692 contrib = decay_load(contrib, n);
2693 return contrib + runnable_avg_yN_sum[n];
2696 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2697 #error "load tracking assumes 2^10 as unit"
2700 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2703 * We can represent the historical contribution to runnable average as the
2704 * coefficients of a geometric series. To do this we sub-divide our runnable
2705 * history into segments of approximately 1ms (1024us); label the segment that
2706 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2708 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2710 * (now) (~1ms ago) (~2ms ago)
2712 * Let u_i denote the fraction of p_i that the entity was runnable.
2714 * We then designate the fractions u_i as our co-efficients, yielding the
2715 * following representation of historical load:
2716 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2718 * We choose y based on the with of a reasonably scheduling period, fixing:
2721 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2722 * approximately half as much as the contribution to load within the last ms
2725 * When a period "rolls over" and we have new u_0`, multiplying the previous
2726 * sum again by y is sufficient to update:
2727 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2728 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2730 static __always_inline int
2731 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2732 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2734 u64 delta, scaled_delta, periods;
2736 unsigned int delta_w, scaled_delta_w, decayed = 0;
2737 unsigned long scale_freq, scale_cpu;
2739 delta = now - sa->last_update_time;
2741 * This should only happen when time goes backwards, which it
2742 * unfortunately does during sched clock init when we swap over to TSC.
2744 if ((s64)delta < 0) {
2745 sa->last_update_time = now;
2750 * Use 1024ns as the unit of measurement since it's a reasonable
2751 * approximation of 1us and fast to compute.
2756 sa->last_update_time = now;
2758 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2759 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2760 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2762 /* delta_w is the amount already accumulated against our next period */
2763 delta_w = sa->period_contrib;
2764 if (delta + delta_w >= 1024) {
2767 /* how much left for next period will start over, we don't know yet */
2768 sa->period_contrib = 0;
2771 * Now that we know we're crossing a period boundary, figure
2772 * out how much from delta we need to complete the current
2773 * period and accrue it.
2775 delta_w = 1024 - delta_w;
2776 scaled_delta_w = cap_scale(delta_w, scale_freq);
2778 sa->load_sum += weight * scaled_delta_w;
2780 cfs_rq->runnable_load_sum +=
2781 weight * scaled_delta_w;
2785 sa->util_sum += scaled_delta_w * scale_cpu;
2789 /* Figure out how many additional periods this update spans */
2790 periods = delta / 1024;
2793 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2795 cfs_rq->runnable_load_sum =
2796 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2798 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2800 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2801 contrib = __compute_runnable_contrib(periods);
2802 contrib = cap_scale(contrib, scale_freq);
2804 sa->load_sum += weight * contrib;
2806 cfs_rq->runnable_load_sum += weight * contrib;
2809 sa->util_sum += contrib * scale_cpu;
2812 /* Remainder of delta accrued against u_0` */
2813 scaled_delta = cap_scale(delta, scale_freq);
2815 sa->load_sum += weight * scaled_delta;
2817 cfs_rq->runnable_load_sum += weight * scaled_delta;
2820 sa->util_sum += scaled_delta * scale_cpu;
2822 sa->period_contrib += delta;
2825 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2827 cfs_rq->runnable_load_avg =
2828 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2830 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2837 * Signed add and clamp on underflow.
2839 * Explicitly do a load-store to ensure the intermediate value never hits
2840 * memory. This allows lockless observations without ever seeing the negative
2843 #define add_positive(_ptr, _val) do { \
2844 typeof(_ptr) ptr = (_ptr); \
2845 typeof(_val) val = (_val); \
2846 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2850 if (val < 0 && res > var) \
2853 WRITE_ONCE(*ptr, res); \
2856 #ifdef CONFIG_FAIR_GROUP_SCHED
2858 * update_tg_load_avg - update the tg's load avg
2859 * @cfs_rq: the cfs_rq whose avg changed
2860 * @force: update regardless of how small the difference
2862 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2863 * However, because tg->load_avg is a global value there are performance
2866 * In order to avoid having to look at the other cfs_rq's, we use a
2867 * differential update where we store the last value we propagated. This in
2868 * turn allows skipping updates if the differential is 'small'.
2870 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2871 * done) and effective_load() (which is not done because it is too costly).
2873 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2875 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2878 * No need to update load_avg for root_task_group as it is not used.
2880 if (cfs_rq->tg == &root_task_group)
2883 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2884 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2885 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2890 * Called within set_task_rq() right before setting a task's cpu. The
2891 * caller only guarantees p->pi_lock is held; no other assumptions,
2892 * including the state of rq->lock, should be made.
2894 void set_task_rq_fair(struct sched_entity *se,
2895 struct cfs_rq *prev, struct cfs_rq *next)
2897 if (!sched_feat(ATTACH_AGE_LOAD))
2901 * We are supposed to update the task to "current" time, then its up to
2902 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2903 * getting what current time is, so simply throw away the out-of-date
2904 * time. This will result in the wakee task is less decayed, but giving
2905 * the wakee more load sounds not bad.
2907 if (se->avg.last_update_time && prev) {
2908 u64 p_last_update_time;
2909 u64 n_last_update_time;
2911 #ifndef CONFIG_64BIT
2912 u64 p_last_update_time_copy;
2913 u64 n_last_update_time_copy;
2916 p_last_update_time_copy = prev->load_last_update_time_copy;
2917 n_last_update_time_copy = next->load_last_update_time_copy;
2921 p_last_update_time = prev->avg.last_update_time;
2922 n_last_update_time = next->avg.last_update_time;
2924 } while (p_last_update_time != p_last_update_time_copy ||
2925 n_last_update_time != n_last_update_time_copy);
2927 p_last_update_time = prev->avg.last_update_time;
2928 n_last_update_time = next->avg.last_update_time;
2930 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2931 &se->avg, 0, 0, NULL);
2932 se->avg.last_update_time = n_last_update_time;
2936 /* Take into account change of utilization of a child task group */
2938 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
2940 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2941 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
2943 /* Nothing to update */
2947 /* Set new sched_entity's utilization */
2948 se->avg.util_avg = gcfs_rq->avg.util_avg;
2949 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
2951 /* Update parent cfs_rq utilization */
2952 add_positive(&cfs_rq->avg.util_avg, delta);
2953 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
2956 /* Take into account change of load of a child task group */
2958 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
2960 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2961 long delta, load = gcfs_rq->avg.load_avg;
2964 * If the load of group cfs_rq is null, the load of the
2965 * sched_entity will also be null so we can skip the formula
2970 /* Get tg's load and ensure tg_load > 0 */
2971 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
2973 /* Ensure tg_load >= load and updated with current load*/
2974 tg_load -= gcfs_rq->tg_load_avg_contrib;
2978 * We need to compute a correction term in the case that the
2979 * task group is consuming more CPU than a task of equal
2980 * weight. A task with a weight equals to tg->shares will have
2981 * a load less or equal to scale_load_down(tg->shares).
2982 * Similarly, the sched_entities that represent the task group
2983 * at parent level, can't have a load higher than
2984 * scale_load_down(tg->shares). And the Sum of sched_entities'
2985 * load must be <= scale_load_down(tg->shares).
2987 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
2988 /* scale gcfs_rq's load into tg's shares*/
2989 load *= scale_load_down(gcfs_rq->tg->shares);
2994 delta = load - se->avg.load_avg;
2996 /* Nothing to update */
3000 /* Set new sched_entity's load */
3001 se->avg.load_avg = load;
3002 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3004 /* Update parent cfs_rq load */
3005 add_positive(&cfs_rq->avg.load_avg, delta);
3006 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3009 * If the sched_entity is already enqueued, we also have to update the
3010 * runnable load avg.
3013 /* Update parent cfs_rq runnable_load_avg */
3014 add_positive(&cfs_rq->runnable_load_avg, delta);
3015 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3019 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3021 cfs_rq->propagate_avg = 1;
3024 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3026 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3028 if (!cfs_rq->propagate_avg)
3031 cfs_rq->propagate_avg = 0;
3035 /* Update task and its cfs_rq load average */
3036 static inline int propagate_entity_load_avg(struct sched_entity *se)
3038 struct cfs_rq *cfs_rq;
3040 if (entity_is_task(se))
3043 if (!test_and_clear_tg_cfs_propagate(se))
3046 cfs_rq = cfs_rq_of(se);
3048 set_tg_cfs_propagate(cfs_rq);
3050 update_tg_cfs_util(cfs_rq, se);
3051 update_tg_cfs_load(cfs_rq, se);
3056 #else /* CONFIG_FAIR_GROUP_SCHED */
3058 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3060 static inline int propagate_entity_load_avg(struct sched_entity *se)
3065 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3067 #endif /* CONFIG_FAIR_GROUP_SCHED */
3069 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3071 if (&this_rq()->cfs == cfs_rq) {
3073 * There are a few boundary cases this might miss but it should
3074 * get called often enough that that should (hopefully) not be
3075 * a real problem -- added to that it only calls on the local
3076 * CPU, so if we enqueue remotely we'll miss an update, but
3077 * the next tick/schedule should update.
3079 * It will not get called when we go idle, because the idle
3080 * thread is a different class (!fair), nor will the utilization
3081 * number include things like RT tasks.
3083 * As is, the util number is not freq-invariant (we'd have to
3084 * implement arch_scale_freq_capacity() for that).
3088 cpufreq_update_util(rq_of(cfs_rq), 0);
3092 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
3095 * Unsigned subtract and clamp on underflow.
3097 * Explicitly do a load-store to ensure the intermediate value never hits
3098 * memory. This allows lockless observations without ever seeing the negative
3101 #define sub_positive(_ptr, _val) do { \
3102 typeof(_ptr) ptr = (_ptr); \
3103 typeof(*ptr) val = (_val); \
3104 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3108 WRITE_ONCE(*ptr, res); \
3112 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3113 * @now: current time, as per cfs_rq_clock_task()
3114 * @cfs_rq: cfs_rq to update
3115 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3117 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3118 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3119 * post_init_entity_util_avg().
3121 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3123 * Returns true if the load decayed or we removed load.
3125 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3126 * call update_tg_load_avg() when this function returns true.
3129 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3131 struct sched_avg *sa = &cfs_rq->avg;
3132 int decayed, removed = 0, removed_util = 0;
3134 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3135 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3136 sub_positive(&sa->load_avg, r);
3137 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3139 set_tg_cfs_propagate(cfs_rq);
3142 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3143 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3144 sub_positive(&sa->util_avg, r);
3145 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3147 set_tg_cfs_propagate(cfs_rq);
3150 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3151 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3153 #ifndef CONFIG_64BIT
3155 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3158 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
3159 if (cfs_rq == &rq_of(cfs_rq)->cfs)
3160 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
3162 if (update_freq && (decayed || removed_util))
3163 cfs_rq_util_change(cfs_rq);
3165 return decayed || removed;
3169 * Optional action to be done while updating the load average
3171 #define UPDATE_TG 0x1
3172 #define SKIP_AGE_LOAD 0x2
3174 /* Update task and its cfs_rq load average */
3175 static inline void update_load_avg(struct sched_entity *se, int flags)
3177 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3178 u64 now = cfs_rq_clock_task(cfs_rq);
3179 int cpu = cpu_of(rq_of(cfs_rq));
3184 * Track task load average for carrying it to new CPU after migrated, and
3185 * track group sched_entity load average for task_h_load calc in migration
3187 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
3188 __update_load_avg(now, cpu, &se->avg,
3189 se->on_rq * scale_load_down(se->load.weight),
3190 cfs_rq->curr == se, NULL);
3193 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3194 decayed |= propagate_entity_load_avg(se);
3196 if (decayed && (flags & UPDATE_TG))
3197 update_tg_load_avg(cfs_rq, 0);
3199 if (entity_is_task(se)) {
3200 #ifdef CONFIG_SCHED_WALT
3201 ptr = (void *)&(task_of(se)->ravg);
3203 trace_sched_load_avg_task(task_of(se), &se->avg, ptr);
3208 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3209 * @cfs_rq: cfs_rq to attach to
3210 * @se: sched_entity to attach
3212 * Must call update_cfs_rq_load_avg() before this, since we rely on
3213 * cfs_rq->avg.last_update_time being current.
3215 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3217 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3218 cfs_rq->avg.load_avg += se->avg.load_avg;
3219 cfs_rq->avg.load_sum += se->avg.load_sum;
3220 cfs_rq->avg.util_avg += se->avg.util_avg;
3221 cfs_rq->avg.util_sum += se->avg.util_sum;
3222 set_tg_cfs_propagate(cfs_rq);
3224 cfs_rq_util_change(cfs_rq);
3228 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3229 * @cfs_rq: cfs_rq to detach from
3230 * @se: sched_entity to detach
3232 * Must call update_cfs_rq_load_avg() before this, since we rely on
3233 * cfs_rq->avg.last_update_time being current.
3235 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3238 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3239 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3240 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3241 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3242 set_tg_cfs_propagate(cfs_rq);
3244 cfs_rq_util_change(cfs_rq);
3247 /* Add the load generated by se into cfs_rq's load average */
3249 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3251 struct sched_avg *sa = &se->avg;
3253 cfs_rq->runnable_load_avg += sa->load_avg;
3254 cfs_rq->runnable_load_sum += sa->load_sum;
3256 if (!sa->last_update_time) {
3257 attach_entity_load_avg(cfs_rq, se);
3258 update_tg_load_avg(cfs_rq, 0);
3262 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3264 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3266 cfs_rq->runnable_load_avg =
3267 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3268 cfs_rq->runnable_load_sum =
3269 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3272 #ifndef CONFIG_64BIT
3273 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3275 u64 last_update_time_copy;
3276 u64 last_update_time;
3279 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3281 last_update_time = cfs_rq->avg.last_update_time;
3282 } while (last_update_time != last_update_time_copy);
3284 return last_update_time;
3287 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3289 return cfs_rq->avg.last_update_time;
3294 * Synchronize entity load avg of dequeued entity without locking
3297 void sync_entity_load_avg(struct sched_entity *se)
3299 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3300 u64 last_update_time;
3302 last_update_time = cfs_rq_last_update_time(cfs_rq);
3303 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3307 * Task first catches up with cfs_rq, and then subtract
3308 * itself from the cfs_rq (task must be off the queue now).
3310 void remove_entity_load_avg(struct sched_entity *se)
3312 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3315 * Newly created task or never used group entity should not be removed
3316 * from its (source) cfs_rq
3318 if (se->avg.last_update_time == 0)
3321 sync_entity_load_avg(se);
3322 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3323 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3327 * Update the rq's load with the elapsed running time before entering
3328 * idle. if the last scheduled task is not a CFS task, idle_enter will
3329 * be the only way to update the runnable statistic.
3331 void idle_enter_fair(struct rq *this_rq)
3336 * Update the rq's load with the elapsed idle time before a task is
3337 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3338 * be the only way to update the runnable statistic.
3340 void idle_exit_fair(struct rq *this_rq)
3344 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3346 return cfs_rq->runnable_load_avg;
3349 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3351 return cfs_rq->avg.load_avg;
3354 static int idle_balance(struct rq *this_rq);
3356 #else /* CONFIG_SMP */
3359 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3364 #define UPDATE_TG 0x0
3365 #define SKIP_AGE_LOAD 0x0
3367 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3369 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3371 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3372 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3375 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3377 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3379 static inline int idle_balance(struct rq *rq)
3384 #endif /* CONFIG_SMP */
3386 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3388 #ifdef CONFIG_SCHEDSTATS
3389 struct task_struct *tsk = NULL;
3391 if (entity_is_task(se))
3394 if (se->statistics.sleep_start) {
3395 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3400 if (unlikely(delta > se->statistics.sleep_max))
3401 se->statistics.sleep_max = delta;
3403 se->statistics.sleep_start = 0;
3404 se->statistics.sum_sleep_runtime += delta;
3407 account_scheduler_latency(tsk, delta >> 10, 1);
3408 trace_sched_stat_sleep(tsk, delta);
3411 if (se->statistics.block_start) {
3412 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3417 if (unlikely(delta > se->statistics.block_max))
3418 se->statistics.block_max = delta;
3420 se->statistics.block_start = 0;
3421 se->statistics.sum_sleep_runtime += delta;
3424 if (tsk->in_iowait) {
3425 se->statistics.iowait_sum += delta;
3426 se->statistics.iowait_count++;
3427 trace_sched_stat_iowait(tsk, delta);
3430 trace_sched_stat_blocked(tsk, delta);
3431 trace_sched_blocked_reason(tsk);
3434 * Blocking time is in units of nanosecs, so shift by
3435 * 20 to get a milliseconds-range estimation of the
3436 * amount of time that the task spent sleeping:
3438 if (unlikely(prof_on == SLEEP_PROFILING)) {
3439 profile_hits(SLEEP_PROFILING,
3440 (void *)get_wchan(tsk),
3443 account_scheduler_latency(tsk, delta >> 10, 0);
3449 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3451 #ifdef CONFIG_SCHED_DEBUG
3452 s64 d = se->vruntime - cfs_rq->min_vruntime;
3457 if (d > 3*sysctl_sched_latency)
3458 schedstat_inc(cfs_rq, nr_spread_over);
3463 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3465 u64 vruntime = cfs_rq->min_vruntime;
3468 * The 'current' period is already promised to the current tasks,
3469 * however the extra weight of the new task will slow them down a
3470 * little, place the new task so that it fits in the slot that
3471 * stays open at the end.
3473 if (initial && sched_feat(START_DEBIT))
3474 vruntime += sched_vslice(cfs_rq, se);
3476 /* sleeps up to a single latency don't count. */
3478 unsigned long thresh = sysctl_sched_latency;
3481 * Halve their sleep time's effect, to allow
3482 * for a gentler effect of sleepers:
3484 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3490 /* ensure we never gain time by being placed backwards. */
3491 se->vruntime = max_vruntime(se->vruntime, vruntime);
3494 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3497 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3500 * Update the normalized vruntime before updating min_vruntime
3501 * through calling update_curr().
3503 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3504 se->vruntime += cfs_rq->min_vruntime;
3507 * Update run-time statistics of the 'current'.
3509 update_curr(cfs_rq);
3510 update_load_avg(se, UPDATE_TG);
3511 enqueue_entity_load_avg(cfs_rq, se);
3512 update_cfs_shares(se);
3513 account_entity_enqueue(cfs_rq, se);
3515 if (flags & ENQUEUE_WAKEUP) {
3516 place_entity(cfs_rq, se, 0);
3517 enqueue_sleeper(cfs_rq, se);
3520 update_stats_enqueue(cfs_rq, se);
3521 check_spread(cfs_rq, se);
3522 if (se != cfs_rq->curr)
3523 __enqueue_entity(cfs_rq, se);
3526 if (cfs_rq->nr_running == 1) {
3527 list_add_leaf_cfs_rq(cfs_rq);
3528 check_enqueue_throttle(cfs_rq);
3532 static void __clear_buddies_last(struct sched_entity *se)
3534 for_each_sched_entity(se) {
3535 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3536 if (cfs_rq->last != se)
3539 cfs_rq->last = NULL;
3543 static void __clear_buddies_next(struct sched_entity *se)
3545 for_each_sched_entity(se) {
3546 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3547 if (cfs_rq->next != se)
3550 cfs_rq->next = NULL;
3554 static void __clear_buddies_skip(struct sched_entity *se)
3556 for_each_sched_entity(se) {
3557 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3558 if (cfs_rq->skip != se)
3561 cfs_rq->skip = NULL;
3565 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3567 if (cfs_rq->last == se)
3568 __clear_buddies_last(se);
3570 if (cfs_rq->next == se)
3571 __clear_buddies_next(se);
3573 if (cfs_rq->skip == se)
3574 __clear_buddies_skip(se);
3577 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3580 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3583 * Update run-time statistics of the 'current'.
3585 update_curr(cfs_rq);
3588 * When dequeuing a sched_entity, we must:
3589 * - Update loads to have both entity and cfs_rq synced with now.
3590 * - Substract its load from the cfs_rq->runnable_avg.
3591 * - Substract its previous weight from cfs_rq->load.weight.
3592 * - For group entity, update its weight to reflect the new share
3593 * of its group cfs_rq.
3595 update_load_avg(se, UPDATE_TG);
3596 dequeue_entity_load_avg(cfs_rq, se);
3598 update_stats_dequeue(cfs_rq, se);
3599 if (flags & DEQUEUE_SLEEP) {
3600 #ifdef CONFIG_SCHEDSTATS
3601 if (entity_is_task(se)) {
3602 struct task_struct *tsk = task_of(se);
3604 if (tsk->state & TASK_INTERRUPTIBLE)
3605 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3606 if (tsk->state & TASK_UNINTERRUPTIBLE)
3607 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3612 clear_buddies(cfs_rq, se);
3614 if (se != cfs_rq->curr)
3615 __dequeue_entity(cfs_rq, se);
3617 account_entity_dequeue(cfs_rq, se);
3620 * Normalize the entity after updating the min_vruntime because the
3621 * update can refer to the ->curr item and we need to reflect this
3622 * movement in our normalized position.
3624 if (!(flags & DEQUEUE_SLEEP))
3625 se->vruntime -= cfs_rq->min_vruntime;
3627 /* return excess runtime on last dequeue */
3628 return_cfs_rq_runtime(cfs_rq);
3630 update_min_vruntime(cfs_rq);
3631 update_cfs_shares(se);
3635 * Preempt the current task with a newly woken task if needed:
3638 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3640 unsigned long ideal_runtime, delta_exec;
3641 struct sched_entity *se;
3644 ideal_runtime = sched_slice(cfs_rq, curr);
3645 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3646 if (delta_exec > ideal_runtime) {
3647 resched_curr(rq_of(cfs_rq));
3649 * The current task ran long enough, ensure it doesn't get
3650 * re-elected due to buddy favours.
3652 clear_buddies(cfs_rq, curr);
3657 * Ensure that a task that missed wakeup preemption by a
3658 * narrow margin doesn't have to wait for a full slice.
3659 * This also mitigates buddy induced latencies under load.
3661 if (delta_exec < sysctl_sched_min_granularity)
3664 se = __pick_first_entity(cfs_rq);
3665 delta = curr->vruntime - se->vruntime;
3670 if (delta > ideal_runtime)
3671 resched_curr(rq_of(cfs_rq));
3675 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3677 /* 'current' is not kept within the tree. */
3680 * Any task has to be enqueued before it get to execute on
3681 * a CPU. So account for the time it spent waiting on the
3684 update_stats_wait_end(cfs_rq, se);
3685 __dequeue_entity(cfs_rq, se);
3686 update_load_avg(se, UPDATE_TG);
3689 update_stats_curr_start(cfs_rq, se);
3691 #ifdef CONFIG_SCHEDSTATS
3693 * Track our maximum slice length, if the CPU's load is at
3694 * least twice that of our own weight (i.e. dont track it
3695 * when there are only lesser-weight tasks around):
3697 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3698 se->statistics.slice_max = max(se->statistics.slice_max,
3699 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3702 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3706 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3709 * Pick the next process, keeping these things in mind, in this order:
3710 * 1) keep things fair between processes/task groups
3711 * 2) pick the "next" process, since someone really wants that to run
3712 * 3) pick the "last" process, for cache locality
3713 * 4) do not run the "skip" process, if something else is available
3715 static struct sched_entity *
3716 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3718 struct sched_entity *left = __pick_first_entity(cfs_rq);
3719 struct sched_entity *se;
3722 * If curr is set we have to see if its left of the leftmost entity
3723 * still in the tree, provided there was anything in the tree at all.
3725 if (!left || (curr && entity_before(curr, left)))
3728 se = left; /* ideally we run the leftmost entity */
3731 * Avoid running the skip buddy, if running something else can
3732 * be done without getting too unfair.
3734 if (cfs_rq->skip == se) {
3735 struct sched_entity *second;
3738 second = __pick_first_entity(cfs_rq);
3740 second = __pick_next_entity(se);
3741 if (!second || (curr && entity_before(curr, second)))
3745 if (second && wakeup_preempt_entity(second, left) < 1)
3750 * Prefer last buddy, try to return the CPU to a preempted task.
3752 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3756 * Someone really wants this to run. If it's not unfair, run it.
3758 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3761 clear_buddies(cfs_rq, se);
3766 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3768 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3771 * If still on the runqueue then deactivate_task()
3772 * was not called and update_curr() has to be done:
3775 update_curr(cfs_rq);
3777 /* throttle cfs_rqs exceeding runtime */
3778 check_cfs_rq_runtime(cfs_rq);
3780 check_spread(cfs_rq, prev);
3782 update_stats_wait_start(cfs_rq, prev);
3783 /* Put 'current' back into the tree. */
3784 __enqueue_entity(cfs_rq, prev);
3785 /* in !on_rq case, update occurred at dequeue */
3786 update_load_avg(prev, 0);
3788 cfs_rq->curr = NULL;
3792 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3795 * Update run-time statistics of the 'current'.
3797 update_curr(cfs_rq);
3800 * Ensure that runnable average is periodically updated.
3802 update_load_avg(curr, UPDATE_TG);
3803 update_cfs_shares(curr);
3805 #ifdef CONFIG_SCHED_HRTICK
3807 * queued ticks are scheduled to match the slice, so don't bother
3808 * validating it and just reschedule.
3811 resched_curr(rq_of(cfs_rq));
3815 * don't let the period tick interfere with the hrtick preemption
3817 if (!sched_feat(DOUBLE_TICK) &&
3818 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3822 if (cfs_rq->nr_running > 1)
3823 check_preempt_tick(cfs_rq, curr);
3827 /**************************************************
3828 * CFS bandwidth control machinery
3831 #ifdef CONFIG_CFS_BANDWIDTH
3833 #ifdef HAVE_JUMP_LABEL
3834 static struct static_key __cfs_bandwidth_used;
3836 static inline bool cfs_bandwidth_used(void)
3838 return static_key_false(&__cfs_bandwidth_used);
3841 void cfs_bandwidth_usage_inc(void)
3843 static_key_slow_inc(&__cfs_bandwidth_used);
3846 void cfs_bandwidth_usage_dec(void)
3848 static_key_slow_dec(&__cfs_bandwidth_used);
3850 #else /* HAVE_JUMP_LABEL */
3851 static bool cfs_bandwidth_used(void)
3856 void cfs_bandwidth_usage_inc(void) {}
3857 void cfs_bandwidth_usage_dec(void) {}
3858 #endif /* HAVE_JUMP_LABEL */
3861 * default period for cfs group bandwidth.
3862 * default: 0.1s, units: nanoseconds
3864 static inline u64 default_cfs_period(void)
3866 return 100000000ULL;
3869 static inline u64 sched_cfs_bandwidth_slice(void)
3871 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3875 * Replenish runtime according to assigned quota and update expiration time.
3876 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3877 * additional synchronization around rq->lock.
3879 * requires cfs_b->lock
3881 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3885 if (cfs_b->quota == RUNTIME_INF)
3888 now = sched_clock_cpu(smp_processor_id());
3889 cfs_b->runtime = cfs_b->quota;
3890 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3893 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3895 return &tg->cfs_bandwidth;
3898 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3899 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3901 if (unlikely(cfs_rq->throttle_count))
3902 return cfs_rq->throttled_clock_task;
3904 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3907 /* returns 0 on failure to allocate runtime */
3908 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3910 struct task_group *tg = cfs_rq->tg;
3911 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3912 u64 amount = 0, min_amount, expires;
3914 /* note: this is a positive sum as runtime_remaining <= 0 */
3915 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3917 raw_spin_lock(&cfs_b->lock);
3918 if (cfs_b->quota == RUNTIME_INF)
3919 amount = min_amount;
3921 start_cfs_bandwidth(cfs_b);
3923 if (cfs_b->runtime > 0) {
3924 amount = min(cfs_b->runtime, min_amount);
3925 cfs_b->runtime -= amount;
3929 expires = cfs_b->runtime_expires;
3930 raw_spin_unlock(&cfs_b->lock);
3932 cfs_rq->runtime_remaining += amount;
3934 * we may have advanced our local expiration to account for allowed
3935 * spread between our sched_clock and the one on which runtime was
3938 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3939 cfs_rq->runtime_expires = expires;
3941 return cfs_rq->runtime_remaining > 0;
3945 * Note: This depends on the synchronization provided by sched_clock and the
3946 * fact that rq->clock snapshots this value.
3948 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3950 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3952 /* if the deadline is ahead of our clock, nothing to do */
3953 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3956 if (cfs_rq->runtime_remaining < 0)
3960 * If the local deadline has passed we have to consider the
3961 * possibility that our sched_clock is 'fast' and the global deadline
3962 * has not truly expired.
3964 * Fortunately we can check determine whether this the case by checking
3965 * whether the global deadline has advanced. It is valid to compare
3966 * cfs_b->runtime_expires without any locks since we only care about
3967 * exact equality, so a partial write will still work.
3970 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3971 /* extend local deadline, drift is bounded above by 2 ticks */
3972 cfs_rq->runtime_expires += TICK_NSEC;
3974 /* global deadline is ahead, expiration has passed */
3975 cfs_rq->runtime_remaining = 0;
3979 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3981 /* dock delta_exec before expiring quota (as it could span periods) */
3982 cfs_rq->runtime_remaining -= delta_exec;
3983 expire_cfs_rq_runtime(cfs_rq);
3985 if (likely(cfs_rq->runtime_remaining > 0))
3989 * if we're unable to extend our runtime we resched so that the active
3990 * hierarchy can be throttled
3992 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3993 resched_curr(rq_of(cfs_rq));
3996 static __always_inline
3997 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3999 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4002 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4005 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4007 return cfs_bandwidth_used() && cfs_rq->throttled;
4010 /* check whether cfs_rq, or any parent, is throttled */
4011 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4013 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4017 * Ensure that neither of the group entities corresponding to src_cpu or
4018 * dest_cpu are members of a throttled hierarchy when performing group
4019 * load-balance operations.
4021 static inline int throttled_lb_pair(struct task_group *tg,
4022 int src_cpu, int dest_cpu)
4024 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4026 src_cfs_rq = tg->cfs_rq[src_cpu];
4027 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4029 return throttled_hierarchy(src_cfs_rq) ||
4030 throttled_hierarchy(dest_cfs_rq);
4033 /* updated child weight may affect parent so we have to do this bottom up */
4034 static int tg_unthrottle_up(struct task_group *tg, void *data)
4036 struct rq *rq = data;
4037 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4039 cfs_rq->throttle_count--;
4041 if (!cfs_rq->throttle_count) {
4042 /* adjust cfs_rq_clock_task() */
4043 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4044 cfs_rq->throttled_clock_task;
4051 static int tg_throttle_down(struct task_group *tg, void *data)
4053 struct rq *rq = data;
4054 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4056 /* group is entering throttled state, stop time */
4057 if (!cfs_rq->throttle_count)
4058 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4059 cfs_rq->throttle_count++;
4064 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4066 struct rq *rq = rq_of(cfs_rq);
4067 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4068 struct sched_entity *se;
4069 long task_delta, dequeue = 1;
4072 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4074 /* freeze hierarchy runnable averages while throttled */
4076 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4079 task_delta = cfs_rq->h_nr_running;
4080 for_each_sched_entity(se) {
4081 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4082 /* throttled entity or throttle-on-deactivate */
4087 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4088 qcfs_rq->h_nr_running -= task_delta;
4090 if (qcfs_rq->load.weight)
4095 sub_nr_running(rq, task_delta);
4097 cfs_rq->throttled = 1;
4098 cfs_rq->throttled_clock = rq_clock(rq);
4099 raw_spin_lock(&cfs_b->lock);
4100 empty = list_empty(&cfs_b->throttled_cfs_rq);
4103 * Add to the _head_ of the list, so that an already-started
4104 * distribute_cfs_runtime will not see us
4106 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4109 * If we're the first throttled task, make sure the bandwidth
4113 start_cfs_bandwidth(cfs_b);
4115 raw_spin_unlock(&cfs_b->lock);
4118 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4120 struct rq *rq = rq_of(cfs_rq);
4121 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4122 struct sched_entity *se;
4126 se = cfs_rq->tg->se[cpu_of(rq)];
4128 cfs_rq->throttled = 0;
4130 update_rq_clock(rq);
4132 raw_spin_lock(&cfs_b->lock);
4133 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4134 list_del_rcu(&cfs_rq->throttled_list);
4135 raw_spin_unlock(&cfs_b->lock);
4137 /* update hierarchical throttle state */
4138 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4140 if (!cfs_rq->load.weight)
4143 task_delta = cfs_rq->h_nr_running;
4144 for_each_sched_entity(se) {
4148 cfs_rq = cfs_rq_of(se);
4150 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4151 cfs_rq->h_nr_running += task_delta;
4153 if (cfs_rq_throttled(cfs_rq))
4158 add_nr_running(rq, task_delta);
4160 /* determine whether we need to wake up potentially idle cpu */
4161 if (rq->curr == rq->idle && rq->cfs.nr_running)
4165 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4166 u64 remaining, u64 expires)
4168 struct cfs_rq *cfs_rq;
4170 u64 starting_runtime = remaining;
4173 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4175 struct rq *rq = rq_of(cfs_rq);
4177 raw_spin_lock(&rq->lock);
4178 if (!cfs_rq_throttled(cfs_rq))
4181 runtime = -cfs_rq->runtime_remaining + 1;
4182 if (runtime > remaining)
4183 runtime = remaining;
4184 remaining -= runtime;
4186 cfs_rq->runtime_remaining += runtime;
4187 cfs_rq->runtime_expires = expires;
4189 /* we check whether we're throttled above */
4190 if (cfs_rq->runtime_remaining > 0)
4191 unthrottle_cfs_rq(cfs_rq);
4194 raw_spin_unlock(&rq->lock);
4201 return starting_runtime - remaining;
4205 * Responsible for refilling a task_group's bandwidth and unthrottling its
4206 * cfs_rqs as appropriate. If there has been no activity within the last
4207 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4208 * used to track this state.
4210 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4212 u64 runtime, runtime_expires;
4215 /* no need to continue the timer with no bandwidth constraint */
4216 if (cfs_b->quota == RUNTIME_INF)
4217 goto out_deactivate;
4219 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4220 cfs_b->nr_periods += overrun;
4223 * idle depends on !throttled (for the case of a large deficit), and if
4224 * we're going inactive then everything else can be deferred
4226 if (cfs_b->idle && !throttled)
4227 goto out_deactivate;
4229 __refill_cfs_bandwidth_runtime(cfs_b);
4232 /* mark as potentially idle for the upcoming period */
4237 /* account preceding periods in which throttling occurred */
4238 cfs_b->nr_throttled += overrun;
4240 runtime_expires = cfs_b->runtime_expires;
4243 * This check is repeated as we are holding onto the new bandwidth while
4244 * we unthrottle. This can potentially race with an unthrottled group
4245 * trying to acquire new bandwidth from the global pool. This can result
4246 * in us over-using our runtime if it is all used during this loop, but
4247 * only by limited amounts in that extreme case.
4249 while (throttled && cfs_b->runtime > 0) {
4250 runtime = cfs_b->runtime;
4251 raw_spin_unlock(&cfs_b->lock);
4252 /* we can't nest cfs_b->lock while distributing bandwidth */
4253 runtime = distribute_cfs_runtime(cfs_b, runtime,
4255 raw_spin_lock(&cfs_b->lock);
4257 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4259 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4263 * While we are ensured activity in the period following an
4264 * unthrottle, this also covers the case in which the new bandwidth is
4265 * insufficient to cover the existing bandwidth deficit. (Forcing the
4266 * timer to remain active while there are any throttled entities.)
4276 /* a cfs_rq won't donate quota below this amount */
4277 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4278 /* minimum remaining period time to redistribute slack quota */
4279 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4280 /* how long we wait to gather additional slack before distributing */
4281 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4284 * Are we near the end of the current quota period?
4286 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4287 * hrtimer base being cleared by hrtimer_start. In the case of
4288 * migrate_hrtimers, base is never cleared, so we are fine.
4290 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4292 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4295 /* if the call-back is running a quota refresh is already occurring */
4296 if (hrtimer_callback_running(refresh_timer))
4299 /* is a quota refresh about to occur? */
4300 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4301 if (remaining < min_expire)
4307 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4309 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4311 /* if there's a quota refresh soon don't bother with slack */
4312 if (runtime_refresh_within(cfs_b, min_left))
4315 hrtimer_start(&cfs_b->slack_timer,
4316 ns_to_ktime(cfs_bandwidth_slack_period),
4320 /* we know any runtime found here is valid as update_curr() precedes return */
4321 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4323 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4324 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4326 if (slack_runtime <= 0)
4329 raw_spin_lock(&cfs_b->lock);
4330 if (cfs_b->quota != RUNTIME_INF &&
4331 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4332 cfs_b->runtime += slack_runtime;
4334 /* we are under rq->lock, defer unthrottling using a timer */
4335 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4336 !list_empty(&cfs_b->throttled_cfs_rq))
4337 start_cfs_slack_bandwidth(cfs_b);
4339 raw_spin_unlock(&cfs_b->lock);
4341 /* even if it's not valid for return we don't want to try again */
4342 cfs_rq->runtime_remaining -= slack_runtime;
4345 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4347 if (!cfs_bandwidth_used())
4350 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4353 __return_cfs_rq_runtime(cfs_rq);
4357 * This is done with a timer (instead of inline with bandwidth return) since
4358 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4360 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4362 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4365 /* confirm we're still not at a refresh boundary */
4366 raw_spin_lock(&cfs_b->lock);
4367 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4368 raw_spin_unlock(&cfs_b->lock);
4372 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4373 runtime = cfs_b->runtime;
4375 expires = cfs_b->runtime_expires;
4376 raw_spin_unlock(&cfs_b->lock);
4381 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4383 raw_spin_lock(&cfs_b->lock);
4384 if (expires == cfs_b->runtime_expires)
4385 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4386 raw_spin_unlock(&cfs_b->lock);
4390 * When a group wakes up we want to make sure that its quota is not already
4391 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4392 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4394 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4396 if (!cfs_bandwidth_used())
4399 /* Synchronize hierarchical throttle counter: */
4400 if (unlikely(!cfs_rq->throttle_uptodate)) {
4401 struct rq *rq = rq_of(cfs_rq);
4402 struct cfs_rq *pcfs_rq;
4403 struct task_group *tg;
4405 cfs_rq->throttle_uptodate = 1;
4407 /* Get closest up-to-date node, because leaves go first: */
4408 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4409 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4410 if (pcfs_rq->throttle_uptodate)
4414 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4415 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4419 /* an active group must be handled by the update_curr()->put() path */
4420 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4423 /* ensure the group is not already throttled */
4424 if (cfs_rq_throttled(cfs_rq))
4427 /* update runtime allocation */
4428 account_cfs_rq_runtime(cfs_rq, 0);
4429 if (cfs_rq->runtime_remaining <= 0)
4430 throttle_cfs_rq(cfs_rq);
4433 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4434 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4436 if (!cfs_bandwidth_used())
4439 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4443 * it's possible for a throttled entity to be forced into a running
4444 * state (e.g. set_curr_task), in this case we're finished.
4446 if (cfs_rq_throttled(cfs_rq))
4449 throttle_cfs_rq(cfs_rq);
4453 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4455 struct cfs_bandwidth *cfs_b =
4456 container_of(timer, struct cfs_bandwidth, slack_timer);
4458 do_sched_cfs_slack_timer(cfs_b);
4460 return HRTIMER_NORESTART;
4463 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4465 struct cfs_bandwidth *cfs_b =
4466 container_of(timer, struct cfs_bandwidth, period_timer);
4470 raw_spin_lock(&cfs_b->lock);
4472 overrun = hrtimer_forward_now(timer, cfs_b->period);
4476 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4479 cfs_b->period_active = 0;
4480 raw_spin_unlock(&cfs_b->lock);
4482 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4485 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4487 raw_spin_lock_init(&cfs_b->lock);
4489 cfs_b->quota = RUNTIME_INF;
4490 cfs_b->period = ns_to_ktime(default_cfs_period());
4492 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4493 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4494 cfs_b->period_timer.function = sched_cfs_period_timer;
4495 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4496 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4499 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4501 cfs_rq->runtime_enabled = 0;
4502 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4505 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4507 lockdep_assert_held(&cfs_b->lock);
4509 if (!cfs_b->period_active) {
4510 cfs_b->period_active = 1;
4511 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4512 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4516 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4518 /* init_cfs_bandwidth() was not called */
4519 if (!cfs_b->throttled_cfs_rq.next)
4522 hrtimer_cancel(&cfs_b->period_timer);
4523 hrtimer_cancel(&cfs_b->slack_timer);
4526 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4528 struct cfs_rq *cfs_rq;
4530 for_each_leaf_cfs_rq(rq, cfs_rq) {
4531 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4533 raw_spin_lock(&cfs_b->lock);
4534 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4535 raw_spin_unlock(&cfs_b->lock);
4539 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4541 struct cfs_rq *cfs_rq;
4543 for_each_leaf_cfs_rq(rq, cfs_rq) {
4544 if (!cfs_rq->runtime_enabled)
4548 * clock_task is not advancing so we just need to make sure
4549 * there's some valid quota amount
4551 cfs_rq->runtime_remaining = 1;
4553 * Offline rq is schedulable till cpu is completely disabled
4554 * in take_cpu_down(), so we prevent new cfs throttling here.
4556 cfs_rq->runtime_enabled = 0;
4558 if (cfs_rq_throttled(cfs_rq))
4559 unthrottle_cfs_rq(cfs_rq);
4563 #else /* CONFIG_CFS_BANDWIDTH */
4564 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4566 return rq_clock_task(rq_of(cfs_rq));
4569 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4570 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4571 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4572 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4574 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4579 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4584 static inline int throttled_lb_pair(struct task_group *tg,
4585 int src_cpu, int dest_cpu)
4590 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4592 #ifdef CONFIG_FAIR_GROUP_SCHED
4593 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4596 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4600 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4601 static inline void update_runtime_enabled(struct rq *rq) {}
4602 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4604 #endif /* CONFIG_CFS_BANDWIDTH */
4606 /**************************************************
4607 * CFS operations on tasks:
4610 #ifdef CONFIG_SCHED_HRTICK
4611 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4613 struct sched_entity *se = &p->se;
4614 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4616 WARN_ON(task_rq(p) != rq);
4618 if (cfs_rq->nr_running > 1) {
4619 u64 slice = sched_slice(cfs_rq, se);
4620 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4621 s64 delta = slice - ran;
4628 hrtick_start(rq, delta);
4633 * called from enqueue/dequeue and updates the hrtick when the
4634 * current task is from our class and nr_running is low enough
4637 static void hrtick_update(struct rq *rq)
4639 struct task_struct *curr = rq->curr;
4641 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4644 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4645 hrtick_start_fair(rq, curr);
4647 #else /* !CONFIG_SCHED_HRTICK */
4649 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4653 static inline void hrtick_update(struct rq *rq)
4659 static bool cpu_overutilized(int cpu);
4660 unsigned long boosted_cpu_util(int cpu);
4662 #define boosted_cpu_util(cpu) cpu_util(cpu)
4666 static void update_capacity_of(int cpu)
4668 unsigned long req_cap;
4673 /* Convert scale-invariant capacity to cpu. */
4674 req_cap = boosted_cpu_util(cpu);
4675 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4676 set_cfs_cpu_capacity(cpu, true, req_cap);
4681 * The enqueue_task method is called before nr_running is
4682 * increased. Here we update the fair scheduling stats and
4683 * then put the task into the rbtree:
4686 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4688 struct cfs_rq *cfs_rq;
4689 struct sched_entity *se = &p->se;
4691 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4692 int task_wakeup = flags & ENQUEUE_WAKEUP;
4696 * If in_iowait is set, the code below may not trigger any cpufreq
4697 * utilization updates, so do it here explicitly with the IOWAIT flag
4701 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4703 for_each_sched_entity(se) {
4706 cfs_rq = cfs_rq_of(se);
4707 enqueue_entity(cfs_rq, se, flags);
4710 * end evaluation on encountering a throttled cfs_rq
4712 * note: in the case of encountering a throttled cfs_rq we will
4713 * post the final h_nr_running increment below.
4715 if (cfs_rq_throttled(cfs_rq))
4717 cfs_rq->h_nr_running++;
4718 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4720 flags = ENQUEUE_WAKEUP;
4723 for_each_sched_entity(se) {
4724 cfs_rq = cfs_rq_of(se);
4725 cfs_rq->h_nr_running++;
4726 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4728 if (cfs_rq_throttled(cfs_rq))
4731 update_load_avg(se, UPDATE_TG);
4732 update_cfs_shares(se);
4736 add_nr_running(rq, 1);
4741 * Update SchedTune accounting.
4743 * We do it before updating the CPU capacity to ensure the
4744 * boost value of the current task is accounted for in the
4745 * selection of the OPP.
4747 * We do it also in the case where we enqueue a throttled task;
4748 * we could argue that a throttled task should not boost a CPU,
4750 * a) properly implementing CPU boosting considering throttled
4751 * tasks will increase a lot the complexity of the solution
4752 * b) it's not easy to quantify the benefits introduced by
4753 * such a more complex solution.
4754 * Thus, for the time being we go for the simple solution and boost
4755 * also for throttled RQs.
4757 schedtune_enqueue_task(p, cpu_of(rq));
4760 walt_inc_cumulative_runnable_avg(rq, p);
4761 if (!task_new && !rq->rd->overutilized &&
4762 cpu_overutilized(rq->cpu)) {
4763 rq->rd->overutilized = true;
4764 trace_sched_overutilized(true);
4768 * We want to potentially trigger a freq switch
4769 * request only for tasks that are waking up; this is
4770 * because we get here also during load balancing, but
4771 * in these cases it seems wise to trigger as single
4772 * request after load balancing is done.
4774 if (task_new || task_wakeup)
4775 update_capacity_of(cpu_of(rq));
4778 #endif /* CONFIG_SMP */
4782 static void set_next_buddy(struct sched_entity *se);
4785 * The dequeue_task method is called before nr_running is
4786 * decreased. We remove the task from the rbtree and
4787 * update the fair scheduling stats:
4789 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4791 struct cfs_rq *cfs_rq;
4792 struct sched_entity *se = &p->se;
4793 int task_sleep = flags & DEQUEUE_SLEEP;
4795 for_each_sched_entity(se) {
4796 cfs_rq = cfs_rq_of(se);
4797 dequeue_entity(cfs_rq, se, flags);
4800 * end evaluation on encountering a throttled cfs_rq
4802 * note: in the case of encountering a throttled cfs_rq we will
4803 * post the final h_nr_running decrement below.
4805 if (cfs_rq_throttled(cfs_rq))
4807 cfs_rq->h_nr_running--;
4808 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4810 /* Don't dequeue parent if it has other entities besides us */
4811 if (cfs_rq->load.weight) {
4812 /* Avoid re-evaluating load for this entity: */
4813 se = parent_entity(se);
4815 * Bias pick_next to pick a task from this cfs_rq, as
4816 * p is sleeping when it is within its sched_slice.
4818 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4822 flags |= DEQUEUE_SLEEP;
4825 for_each_sched_entity(se) {
4826 cfs_rq = cfs_rq_of(se);
4827 cfs_rq->h_nr_running--;
4828 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4830 if (cfs_rq_throttled(cfs_rq))
4833 update_load_avg(se, UPDATE_TG);
4834 update_cfs_shares(se);
4838 sub_nr_running(rq, 1);
4843 * Update SchedTune accounting
4845 * We do it before updating the CPU capacity to ensure the
4846 * boost value of the current task is accounted for in the
4847 * selection of the OPP.
4849 schedtune_dequeue_task(p, cpu_of(rq));
4852 walt_dec_cumulative_runnable_avg(rq, p);
4855 * We want to potentially trigger a freq switch
4856 * request only for tasks that are going to sleep;
4857 * this is because we get here also during load
4858 * balancing, but in these cases it seems wise to
4859 * trigger as single request after load balancing is
4863 if (rq->cfs.nr_running)
4864 update_capacity_of(cpu_of(rq));
4865 else if (sched_freq())
4866 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4870 #endif /* CONFIG_SMP */
4878 * per rq 'load' arrray crap; XXX kill this.
4882 * The exact cpuload at various idx values, calculated at every tick would be
4883 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4885 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4886 * on nth tick when cpu may be busy, then we have:
4887 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4888 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4890 * decay_load_missed() below does efficient calculation of
4891 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4892 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4894 * The calculation is approximated on a 128 point scale.
4895 * degrade_zero_ticks is the number of ticks after which load at any
4896 * particular idx is approximated to be zero.
4897 * degrade_factor is a precomputed table, a row for each load idx.
4898 * Each column corresponds to degradation factor for a power of two ticks,
4899 * based on 128 point scale.
4901 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4902 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4904 * With this power of 2 load factors, we can degrade the load n times
4905 * by looking at 1 bits in n and doing as many mult/shift instead of
4906 * n mult/shifts needed by the exact degradation.
4908 #define DEGRADE_SHIFT 7
4909 static const unsigned char
4910 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4911 static const unsigned char
4912 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4913 {0, 0, 0, 0, 0, 0, 0, 0},
4914 {64, 32, 8, 0, 0, 0, 0, 0},
4915 {96, 72, 40, 12, 1, 0, 0},
4916 {112, 98, 75, 43, 15, 1, 0},
4917 {120, 112, 98, 76, 45, 16, 2} };
4920 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4921 * would be when CPU is idle and so we just decay the old load without
4922 * adding any new load.
4924 static unsigned long
4925 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4929 if (!missed_updates)
4932 if (missed_updates >= degrade_zero_ticks[idx])
4936 return load >> missed_updates;
4938 while (missed_updates) {
4939 if (missed_updates % 2)
4940 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4942 missed_updates >>= 1;
4949 * Update rq->cpu_load[] statistics. This function is usually called every
4950 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4951 * every tick. We fix it up based on jiffies.
4953 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4954 unsigned long pending_updates)
4958 this_rq->nr_load_updates++;
4960 /* Update our load: */
4961 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4962 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4963 unsigned long old_load, new_load;
4965 /* scale is effectively 1 << i now, and >> i divides by scale */
4967 old_load = this_rq->cpu_load[i];
4968 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4969 new_load = this_load;
4971 * Round up the averaging division if load is increasing. This
4972 * prevents us from getting stuck on 9 if the load is 10, for
4975 if (new_load > old_load)
4976 new_load += scale - 1;
4978 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4981 sched_avg_update(this_rq);
4984 /* Used instead of source_load when we know the type == 0 */
4985 static unsigned long weighted_cpuload(const int cpu)
4987 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4990 #ifdef CONFIG_NO_HZ_COMMON
4992 * There is no sane way to deal with nohz on smp when using jiffies because the
4993 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4994 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4996 * Therefore we cannot use the delta approach from the regular tick since that
4997 * would seriously skew the load calculation. However we'll make do for those
4998 * updates happening while idle (nohz_idle_balance) or coming out of idle
4999 * (tick_nohz_idle_exit).
5001 * This means we might still be one tick off for nohz periods.
5005 * Called from nohz_idle_balance() to update the load ratings before doing the
5008 static void update_idle_cpu_load(struct rq *this_rq)
5010 unsigned long curr_jiffies = READ_ONCE(jiffies);
5011 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5012 unsigned long pending_updates;
5015 * bail if there's load or we're actually up-to-date.
5017 if (load || curr_jiffies == this_rq->last_load_update_tick)
5020 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5021 this_rq->last_load_update_tick = curr_jiffies;
5023 __update_cpu_load(this_rq, load, pending_updates);
5027 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
5029 void update_cpu_load_nohz(void)
5031 struct rq *this_rq = this_rq();
5032 unsigned long curr_jiffies = READ_ONCE(jiffies);
5033 unsigned long pending_updates;
5035 if (curr_jiffies == this_rq->last_load_update_tick)
5038 raw_spin_lock(&this_rq->lock);
5039 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5040 if (pending_updates) {
5041 this_rq->last_load_update_tick = curr_jiffies;
5043 * We were idle, this means load 0, the current load might be
5044 * !0 due to remote wakeups and the sort.
5046 __update_cpu_load(this_rq, 0, pending_updates);
5048 raw_spin_unlock(&this_rq->lock);
5050 #endif /* CONFIG_NO_HZ */
5053 * Called from scheduler_tick()
5055 void update_cpu_load_active(struct rq *this_rq)
5057 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5059 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
5061 this_rq->last_load_update_tick = jiffies;
5062 __update_cpu_load(this_rq, load, 1);
5066 * Return a low guess at the load of a migration-source cpu weighted
5067 * according to the scheduling class and "nice" value.
5069 * We want to under-estimate the load of migration sources, to
5070 * balance conservatively.
5072 static unsigned long source_load(int cpu, int type)
5074 struct rq *rq = cpu_rq(cpu);
5075 unsigned long total = weighted_cpuload(cpu);
5077 if (type == 0 || !sched_feat(LB_BIAS))
5080 return min(rq->cpu_load[type-1], total);
5084 * Return a high guess at the load of a migration-target cpu weighted
5085 * according to the scheduling class and "nice" value.
5087 static unsigned long target_load(int cpu, int type)
5089 struct rq *rq = cpu_rq(cpu);
5090 unsigned long total = weighted_cpuload(cpu);
5092 if (type == 0 || !sched_feat(LB_BIAS))
5095 return max(rq->cpu_load[type-1], total);
5099 static unsigned long cpu_avg_load_per_task(int cpu)
5101 struct rq *rq = cpu_rq(cpu);
5102 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5103 unsigned long load_avg = weighted_cpuload(cpu);
5106 return load_avg / nr_running;
5111 static void record_wakee(struct task_struct *p)
5114 * Rough decay (wiping) for cost saving, don't worry
5115 * about the boundary, really active task won't care
5118 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5119 current->wakee_flips >>= 1;
5120 current->wakee_flip_decay_ts = jiffies;
5123 if (current->last_wakee != p) {
5124 current->last_wakee = p;
5125 current->wakee_flips++;
5129 static void task_waking_fair(struct task_struct *p)
5131 struct sched_entity *se = &p->se;
5132 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5135 #ifndef CONFIG_64BIT
5136 u64 min_vruntime_copy;
5139 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5141 min_vruntime = cfs_rq->min_vruntime;
5142 } while (min_vruntime != min_vruntime_copy);
5144 min_vruntime = cfs_rq->min_vruntime;
5147 se->vruntime -= min_vruntime;
5151 #ifdef CONFIG_FAIR_GROUP_SCHED
5153 * effective_load() calculates the load change as seen from the root_task_group
5155 * Adding load to a group doesn't make a group heavier, but can cause movement
5156 * of group shares between cpus. Assuming the shares were perfectly aligned one
5157 * can calculate the shift in shares.
5159 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5160 * on this @cpu and results in a total addition (subtraction) of @wg to the
5161 * total group weight.
5163 * Given a runqueue weight distribution (rw_i) we can compute a shares
5164 * distribution (s_i) using:
5166 * s_i = rw_i / \Sum rw_j (1)
5168 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5169 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5170 * shares distribution (s_i):
5172 * rw_i = { 2, 4, 1, 0 }
5173 * s_i = { 2/7, 4/7, 1/7, 0 }
5175 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5176 * task used to run on and the CPU the waker is running on), we need to
5177 * compute the effect of waking a task on either CPU and, in case of a sync
5178 * wakeup, compute the effect of the current task going to sleep.
5180 * So for a change of @wl to the local @cpu with an overall group weight change
5181 * of @wl we can compute the new shares distribution (s'_i) using:
5183 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5185 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5186 * differences in waking a task to CPU 0. The additional task changes the
5187 * weight and shares distributions like:
5189 * rw'_i = { 3, 4, 1, 0 }
5190 * s'_i = { 3/8, 4/8, 1/8, 0 }
5192 * We can then compute the difference in effective weight by using:
5194 * dw_i = S * (s'_i - s_i) (3)
5196 * Where 'S' is the group weight as seen by its parent.
5198 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5199 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5200 * 4/7) times the weight of the group.
5202 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5204 struct sched_entity *se = tg->se[cpu];
5206 if (!tg->parent) /* the trivial, non-cgroup case */
5209 for_each_sched_entity(se) {
5210 struct cfs_rq *cfs_rq = se->my_q;
5211 long W, w = cfs_rq_load_avg(cfs_rq);
5216 * W = @wg + \Sum rw_j
5218 W = wg + atomic_long_read(&tg->load_avg);
5220 /* Ensure \Sum rw_j >= rw_i */
5221 W -= cfs_rq->tg_load_avg_contrib;
5230 * wl = S * s'_i; see (2)
5233 wl = (w * (long)tg->shares) / W;
5238 * Per the above, wl is the new se->load.weight value; since
5239 * those are clipped to [MIN_SHARES, ...) do so now. See
5240 * calc_cfs_shares().
5242 if (wl < MIN_SHARES)
5246 * wl = dw_i = S * (s'_i - s_i); see (3)
5248 wl -= se->avg.load_avg;
5251 * Recursively apply this logic to all parent groups to compute
5252 * the final effective load change on the root group. Since
5253 * only the @tg group gets extra weight, all parent groups can
5254 * only redistribute existing shares. @wl is the shift in shares
5255 * resulting from this level per the above.
5264 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5272 * Returns the current capacity of cpu after applying both
5273 * cpu and freq scaling.
5275 unsigned long capacity_curr_of(int cpu)
5277 return cpu_rq(cpu)->cpu_capacity_orig *
5278 arch_scale_freq_capacity(NULL, cpu)
5279 >> SCHED_CAPACITY_SHIFT;
5282 static inline bool energy_aware(void)
5284 return sched_feat(ENERGY_AWARE);
5288 struct sched_group *sg_top;
5289 struct sched_group *sg_cap;
5296 struct task_struct *task;
5311 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5312 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
5313 * energy calculations. Using the scale-invariant util returned by
5314 * cpu_util() and approximating scale-invariant util by:
5316 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5318 * the normalized util can be found using the specific capacity.
5320 * capacity = capacity_orig * curr_freq/max_freq
5322 * norm_util = running_time/time ~ util/capacity
5324 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
5326 int util = __cpu_util(cpu, delta);
5328 if (util >= capacity)
5329 return SCHED_CAPACITY_SCALE;
5331 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5334 static int calc_util_delta(struct energy_env *eenv, int cpu)
5336 if (cpu == eenv->src_cpu)
5337 return -eenv->util_delta;
5338 if (cpu == eenv->dst_cpu)
5339 return eenv->util_delta;
5344 unsigned long group_max_util(struct energy_env *eenv)
5347 unsigned long max_util = 0;
5349 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
5350 delta = calc_util_delta(eenv, i);
5351 max_util = max(max_util, __cpu_util(i, delta));
5358 * group_norm_util() returns the approximated group util relative to it's
5359 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
5360 * energy calculations. Since task executions may or may not overlap in time in
5361 * the group the true normalized util is between max(cpu_norm_util(i)) and
5362 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
5363 * latter is used as the estimate as it leads to a more pessimistic energy
5364 * estimate (more busy).
5367 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5370 unsigned long util_sum = 0;
5371 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5373 for_each_cpu(i, sched_group_cpus(sg)) {
5374 delta = calc_util_delta(eenv, i);
5375 util_sum += __cpu_norm_util(i, capacity, delta);
5378 if (util_sum > SCHED_CAPACITY_SCALE)
5379 return SCHED_CAPACITY_SCALE;
5383 static int find_new_capacity(struct energy_env *eenv,
5384 const struct sched_group_energy * const sge)
5387 unsigned long util = group_max_util(eenv);
5389 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5390 if (sge->cap_states[idx].cap >= util)
5394 eenv->cap_idx = idx;
5399 static int group_idle_state(struct sched_group *sg)
5401 int i, state = INT_MAX;
5403 /* Find the shallowest idle state in the sched group. */
5404 for_each_cpu(i, sched_group_cpus(sg))
5405 state = min(state, idle_get_state_idx(cpu_rq(i)));
5407 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5414 * sched_group_energy(): Computes the absolute energy consumption of cpus
5415 * belonging to the sched_group including shared resources shared only by
5416 * members of the group. Iterates over all cpus in the hierarchy below the
5417 * sched_group starting from the bottom working it's way up before going to
5418 * the next cpu until all cpus are covered at all levels. The current
5419 * implementation is likely to gather the same util statistics multiple times.
5420 * This can probably be done in a faster but more complex way.
5421 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5423 static int sched_group_energy(struct energy_env *eenv)
5425 struct sched_domain *sd;
5426 int cpu, total_energy = 0;
5427 struct cpumask visit_cpus;
5428 struct sched_group *sg;
5430 WARN_ON(!eenv->sg_top->sge);
5432 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5434 while (!cpumask_empty(&visit_cpus)) {
5435 struct sched_group *sg_shared_cap = NULL;
5437 cpu = cpumask_first(&visit_cpus);
5440 * Is the group utilization affected by cpus outside this
5443 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5445 if (sd && sd->parent)
5446 sg_shared_cap = sd->parent->groups;
5448 for_each_domain(cpu, sd) {
5451 /* Has this sched_domain already been visited? */
5452 if (sd->child && group_first_cpu(sg) != cpu)
5456 unsigned long group_util;
5457 int sg_busy_energy, sg_idle_energy;
5458 int cap_idx, idle_idx;
5460 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5461 eenv->sg_cap = sg_shared_cap;
5465 cap_idx = find_new_capacity(eenv, sg->sge);
5467 if (sg->group_weight == 1) {
5468 /* Remove capacity of src CPU (before task move) */
5469 if (eenv->util_delta == 0 &&
5470 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5471 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5472 eenv->cap.delta -= eenv->cap.before;
5474 /* Add capacity of dst CPU (after task move) */
5475 if (eenv->util_delta != 0 &&
5476 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5477 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5478 eenv->cap.delta += eenv->cap.after;
5482 idle_idx = group_idle_state(sg);
5483 group_util = group_norm_util(eenv, sg);
5484 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5485 >> SCHED_CAPACITY_SHIFT;
5486 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5487 * sg->sge->idle_states[idle_idx].power)
5488 >> SCHED_CAPACITY_SHIFT;
5490 total_energy += sg_busy_energy + sg_idle_energy;
5493 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5495 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5498 } while (sg = sg->next, sg != sd->groups);
5502 * If we raced with hotplug and got an sd NULL-pointer;
5503 * returning a wrong energy estimation is better than
5504 * entering an infinite loop.
5506 if (cpumask_test_cpu(cpu, &visit_cpus))
5509 cpumask_clear_cpu(cpu, &visit_cpus);
5513 eenv->energy = total_energy;
5517 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5519 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5523 * energy_diff(): Estimate the energy impact of changing the utilization
5524 * distribution. eenv specifies the change: utilisation amount, source, and
5525 * destination cpu. Source or destination cpu may be -1 in which case the
5526 * utilization is removed from or added to the system (e.g. task wake-up). If
5527 * both are specified, the utilization is migrated.
5529 static inline int __energy_diff(struct energy_env *eenv)
5531 struct sched_domain *sd;
5532 struct sched_group *sg;
5533 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5536 struct energy_env eenv_before = {
5538 .src_cpu = eenv->src_cpu,
5539 .dst_cpu = eenv->dst_cpu,
5540 .nrg = { 0, 0, 0, 0},
5544 if (eenv->src_cpu == eenv->dst_cpu)
5547 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5548 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5551 return 0; /* Error */
5556 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5557 eenv_before.sg_top = eenv->sg_top = sg;
5559 if (sched_group_energy(&eenv_before))
5560 return 0; /* Invalid result abort */
5561 energy_before += eenv_before.energy;
5563 /* Keep track of SRC cpu (before) capacity */
5564 eenv->cap.before = eenv_before.cap.before;
5565 eenv->cap.delta = eenv_before.cap.delta;
5567 if (sched_group_energy(eenv))
5568 return 0; /* Invalid result abort */
5569 energy_after += eenv->energy;
5571 } while (sg = sg->next, sg != sd->groups);
5573 eenv->nrg.before = energy_before;
5574 eenv->nrg.after = energy_after;
5575 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5577 #ifndef CONFIG_SCHED_TUNE
5578 trace_sched_energy_diff(eenv->task,
5579 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5580 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5581 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5582 eenv->nrg.delta, eenv->payoff);
5585 * Dead-zone margin preventing too many migrations.
5588 margin = eenv->nrg.before >> 6; /* ~1.56% */
5590 diff = eenv->nrg.after - eenv->nrg.before;
5592 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5594 return eenv->nrg.diff;
5597 #ifdef CONFIG_SCHED_TUNE
5599 struct target_nrg schedtune_target_nrg;
5600 extern bool schedtune_initialized;
5602 * System energy normalization
5603 * Returns the normalized value, in the range [0..SCHED_CAPACITY_SCALE],
5604 * corresponding to the specified energy variation.
5607 normalize_energy(int energy_diff)
5611 /* during early setup, we don't know the extents */
5612 if (unlikely(!schedtune_initialized))
5613 return energy_diff < 0 ? -1 : 1 ;
5615 #ifdef CONFIG_SCHED_DEBUG
5619 /* Check for boundaries */
5620 max_delta = schedtune_target_nrg.max_power;
5621 max_delta -= schedtune_target_nrg.min_power;
5622 WARN_ON(abs(energy_diff) >= max_delta);
5626 /* Do scaling using positive numbers to increase the range */
5627 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5629 /* Scale by energy magnitude */
5630 normalized_nrg <<= SCHED_CAPACITY_SHIFT;
5632 /* Normalize on max energy for target platform */
5633 normalized_nrg = reciprocal_divide(
5634 normalized_nrg, schedtune_target_nrg.rdiv);
5636 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5640 energy_diff(struct energy_env *eenv)
5642 int boost = schedtune_task_boost(eenv->task);
5645 /* Conpute "absolute" energy diff */
5646 __energy_diff(eenv);
5648 /* Return energy diff when boost margin is 0 */
5650 return eenv->nrg.diff;
5652 /* Compute normalized energy diff */
5653 nrg_delta = normalize_energy(eenv->nrg.diff);
5654 eenv->nrg.delta = nrg_delta;
5656 eenv->payoff = schedtune_accept_deltas(
5661 trace_sched_energy_diff(eenv->task,
5662 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5663 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5664 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5665 eenv->nrg.delta, eenv->payoff);
5668 * When SchedTune is enabled, the energy_diff() function will return
5669 * the computed energy payoff value. Since the energy_diff() return
5670 * value is expected to be negative by its callers, this evaluation
5671 * function return a negative value each time the evaluation return a
5672 * positive payoff, which is the condition for the acceptance of
5673 * a scheduling decision
5675 return -eenv->payoff;
5677 #else /* CONFIG_SCHED_TUNE */
5678 #define energy_diff(eenv) __energy_diff(eenv)
5682 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5683 * A waker of many should wake a different task than the one last awakened
5684 * at a frequency roughly N times higher than one of its wakees. In order
5685 * to determine whether we should let the load spread vs consolodating to
5686 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5687 * partner, and a factor of lls_size higher frequency in the other. With
5688 * both conditions met, we can be relatively sure that the relationship is
5689 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5690 * being client/server, worker/dispatcher, interrupt source or whatever is
5691 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5693 static int wake_wide(struct task_struct *p)
5695 unsigned int master = current->wakee_flips;
5696 unsigned int slave = p->wakee_flips;
5697 int factor = this_cpu_read(sd_llc_size);
5700 swap(master, slave);
5701 if (slave < factor || master < slave * factor)
5706 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5707 int prev_cpu, int sync)
5709 s64 this_load, load;
5710 s64 this_eff_load, prev_eff_load;
5712 struct task_group *tg;
5713 unsigned long weight;
5717 this_cpu = smp_processor_id();
5718 load = source_load(prev_cpu, idx);
5719 this_load = target_load(this_cpu, idx);
5722 * If sync wakeup then subtract the (maximum possible)
5723 * effect of the currently running task from the load
5724 * of the current CPU:
5727 tg = task_group(current);
5728 weight = current->se.avg.load_avg;
5730 this_load += effective_load(tg, this_cpu, -weight, -weight);
5731 load += effective_load(tg, prev_cpu, 0, -weight);
5735 weight = p->se.avg.load_avg;
5738 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5739 * due to the sync cause above having dropped this_load to 0, we'll
5740 * always have an imbalance, but there's really nothing you can do
5741 * about that, so that's good too.
5743 * Otherwise check if either cpus are near enough in load to allow this
5744 * task to be woken on this_cpu.
5746 this_eff_load = 100;
5747 this_eff_load *= capacity_of(prev_cpu);
5749 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5750 prev_eff_load *= capacity_of(this_cpu);
5752 if (this_load > 0) {
5753 this_eff_load *= this_load +
5754 effective_load(tg, this_cpu, weight, weight);
5756 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5759 balanced = this_eff_load <= prev_eff_load;
5761 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5766 schedstat_inc(sd, ttwu_move_affine);
5767 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5772 static inline unsigned long task_util(struct task_struct *p)
5774 #ifdef CONFIG_SCHED_WALT
5775 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5776 unsigned long demand = p->ravg.demand;
5777 return (demand << 10) / walt_ravg_window;
5780 return p->se.avg.util_avg;
5783 static inline unsigned long boosted_task_util(struct task_struct *task);
5785 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5787 unsigned long capacity = capacity_of(cpu);
5789 util += boosted_task_util(p);
5791 return (capacity * 1024) > (util * capacity_margin);
5794 static inline bool task_fits_max(struct task_struct *p, int cpu)
5796 unsigned long capacity = capacity_of(cpu);
5797 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5799 if (capacity == max_capacity)
5802 if (capacity * capacity_margin > max_capacity * 1024)
5805 return __task_fits(p, cpu, 0);
5808 static bool cpu_overutilized(int cpu)
5810 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5813 #ifdef CONFIG_SCHED_TUNE
5815 struct reciprocal_value schedtune_spc_rdiv;
5818 schedtune_margin(unsigned long signal, long boost)
5820 long long margin = 0;
5823 * Signal proportional compensation (SPC)
5825 * The Boost (B) value is used to compute a Margin (M) which is
5826 * proportional to the complement of the original Signal (S):
5827 * M = B * (SCHED_CAPACITY_SCALE - S)
5828 * The obtained M could be used by the caller to "boost" S.
5831 margin = SCHED_CAPACITY_SCALE - signal;
5834 margin = -signal * boost;
5836 margin = reciprocal_divide(margin, schedtune_spc_rdiv);
5844 schedtune_cpu_margin(unsigned long util, int cpu)
5846 int boost = schedtune_cpu_boost(cpu);
5851 return schedtune_margin(util, boost);
5855 schedtune_task_margin(struct task_struct *task)
5857 int boost = schedtune_task_boost(task);
5864 util = task_util(task);
5865 margin = schedtune_margin(util, boost);
5870 #else /* CONFIG_SCHED_TUNE */
5873 schedtune_cpu_margin(unsigned long util, int cpu)
5879 schedtune_task_margin(struct task_struct *task)
5884 #endif /* CONFIG_SCHED_TUNE */
5887 boosted_cpu_util(int cpu)
5889 unsigned long util = cpu_util(cpu);
5890 long margin = schedtune_cpu_margin(util, cpu);
5892 trace_sched_boost_cpu(cpu, util, margin);
5894 return util + margin;
5897 static inline unsigned long
5898 boosted_task_util(struct task_struct *task)
5900 unsigned long util = task_util(task);
5901 long margin = schedtune_task_margin(task);
5903 trace_sched_boost_task(task, util, margin);
5905 return util + margin;
5908 static int cpu_util_wake(int cpu, struct task_struct *p);
5910 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5912 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5916 * find_idlest_group finds and returns the least busy CPU group within the
5919 static struct sched_group *
5920 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5921 int this_cpu, int sd_flag)
5923 struct sched_group *idlest = NULL, *group = sd->groups;
5924 struct sched_group *most_spare_sg = NULL;
5925 unsigned long min_load = ULONG_MAX, this_load = 0;
5926 unsigned long most_spare = 0, this_spare = 0;
5927 int load_idx = sd->forkexec_idx;
5928 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5930 if (sd_flag & SD_BALANCE_WAKE)
5931 load_idx = sd->wake_idx;
5934 unsigned long load, avg_load, spare_cap, max_spare_cap;
5938 /* Skip over this group if it has no CPUs allowed */
5939 if (!cpumask_intersects(sched_group_cpus(group),
5940 tsk_cpus_allowed(p)))
5943 local_group = cpumask_test_cpu(this_cpu,
5944 sched_group_cpus(group));
5947 * Tally up the load of all CPUs in the group and find
5948 * the group containing the CPU with most spare capacity.
5953 for_each_cpu(i, sched_group_cpus(group)) {
5954 /* Bias balancing toward cpus of our domain */
5956 load = source_load(i, load_idx);
5958 load = target_load(i, load_idx);
5962 spare_cap = capacity_spare_wake(i, p);
5964 if (spare_cap > max_spare_cap)
5965 max_spare_cap = spare_cap;
5968 /* Adjust by relative CPU capacity of the group */
5969 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5972 this_load = avg_load;
5973 this_spare = max_spare_cap;
5975 if (avg_load < min_load) {
5976 min_load = avg_load;
5980 if (most_spare < max_spare_cap) {
5981 most_spare = max_spare_cap;
5982 most_spare_sg = group;
5985 } while (group = group->next, group != sd->groups);
5988 * The cross-over point between using spare capacity or least load
5989 * is too conservative for high utilization tasks on partially
5990 * utilized systems if we require spare_capacity > task_util(p),
5991 * so we allow for some task stuffing by using
5992 * spare_capacity > task_util(p)/2.
5994 if (this_spare > task_util(p) / 2 &&
5995 imbalance*this_spare > 100*most_spare)
5997 else if (most_spare > task_util(p) / 2)
5998 return most_spare_sg;
6000 if (!idlest || 100*this_load < imbalance*min_load)
6006 * find_idlest_cpu - find the idlest cpu among the cpus in group.
6009 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6011 unsigned long load, min_load = ULONG_MAX;
6012 unsigned int min_exit_latency = UINT_MAX;
6013 u64 latest_idle_timestamp = 0;
6014 int least_loaded_cpu = this_cpu;
6015 int shallowest_idle_cpu = -1;
6018 /* Check if we have any choice: */
6019 if (group->group_weight == 1)
6020 return cpumask_first(sched_group_cpus(group));
6022 /* Traverse only the allowed CPUs */
6023 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
6025 struct rq *rq = cpu_rq(i);
6026 struct cpuidle_state *idle = idle_get_state(rq);
6027 if (idle && idle->exit_latency < min_exit_latency) {
6029 * We give priority to a CPU whose idle state
6030 * has the smallest exit latency irrespective
6031 * of any idle timestamp.
6033 min_exit_latency = idle->exit_latency;
6034 latest_idle_timestamp = rq->idle_stamp;
6035 shallowest_idle_cpu = i;
6036 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6037 rq->idle_stamp > latest_idle_timestamp) {
6039 * If equal or no active idle state, then
6040 * the most recently idled CPU might have
6043 latest_idle_timestamp = rq->idle_stamp;
6044 shallowest_idle_cpu = i;
6046 } else if (shallowest_idle_cpu == -1) {
6047 load = weighted_cpuload(i);
6048 if (load < min_load || (load == min_load && i == this_cpu)) {
6050 least_loaded_cpu = i;
6055 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6059 * Try and locate an idle CPU in the sched_domain.
6061 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6063 struct sched_domain *sd;
6064 struct sched_group *sg;
6065 int best_idle_cpu = -1;
6066 int best_idle_cstate = INT_MAX;
6067 unsigned long best_idle_capacity = ULONG_MAX;
6069 schedstat_inc(p, se.statistics.nr_wakeups_sis_attempts);
6070 schedstat_inc(this_rq(), eas_stats.sis_attempts);
6072 if (!sysctl_sched_cstate_aware) {
6073 if (idle_cpu(target)) {
6074 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle);
6075 schedstat_inc(this_rq(), eas_stats.sis_idle);
6080 * If the prevous cpu is cache affine and idle, don't be stupid.
6082 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) {
6083 schedstat_inc(p, se.statistics.nr_wakeups_sis_cache_affine);
6084 schedstat_inc(this_rq(), eas_stats.sis_cache_affine);
6090 * Otherwise, iterate the domains and find an elegible idle cpu.
6092 sd = rcu_dereference(per_cpu(sd_llc, target));
6093 for_each_lower_domain(sd) {
6097 if (!cpumask_intersects(sched_group_cpus(sg),
6098 tsk_cpus_allowed(p)))
6101 if (sysctl_sched_cstate_aware) {
6102 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6103 int idle_idx = idle_get_state_idx(cpu_rq(i));
6104 unsigned long new_usage = boosted_task_util(p);
6105 unsigned long capacity_orig = capacity_orig_of(i);
6107 if (new_usage > capacity_orig || !idle_cpu(i))
6110 if (i == target && new_usage <= capacity_curr_of(target)) {
6111 schedstat_inc(p, se.statistics.nr_wakeups_sis_suff_cap);
6112 schedstat_inc(this_rq(), eas_stats.sis_suff_cap);
6113 schedstat_inc(sd, eas_stats.sis_suff_cap);
6117 if (idle_idx < best_idle_cstate &&
6118 capacity_orig <= best_idle_capacity) {
6120 best_idle_cstate = idle_idx;
6121 best_idle_capacity = capacity_orig;
6125 for_each_cpu(i, sched_group_cpus(sg)) {
6126 if (i == target || !idle_cpu(i))
6130 target = cpumask_first_and(sched_group_cpus(sg),
6131 tsk_cpus_allowed(p));
6132 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle_cpu);
6133 schedstat_inc(this_rq(), eas_stats.sis_idle_cpu);
6134 schedstat_inc(sd, eas_stats.sis_idle_cpu);
6139 } while (sg != sd->groups);
6142 if (best_idle_cpu >= 0)
6143 target = best_idle_cpu;
6146 schedstat_inc(p, se.statistics.nr_wakeups_sis_count);
6147 schedstat_inc(this_rq(), eas_stats.sis_count);
6153 * cpu_util_wake: Compute cpu utilization with any contributions from
6154 * the waking task p removed.
6156 static int cpu_util_wake(int cpu, struct task_struct *p)
6158 unsigned long util, capacity;
6160 #ifdef CONFIG_SCHED_WALT
6162 * WALT does not decay idle tasks in the same manner
6163 * as PELT, so it makes little sense to subtract task
6164 * utilization from cpu utilization. Instead just use
6165 * cpu_util for this case.
6167 if (!walt_disabled && sysctl_sched_use_walt_cpu_util)
6168 return cpu_util(cpu);
6170 /* Task has no contribution or is new */
6171 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6172 return cpu_util(cpu);
6174 capacity = capacity_orig_of(cpu);
6175 util = max_t(long, cpu_util(cpu) - task_util(p), 0);
6177 return (util >= capacity) ? capacity : util;
6180 static int start_cpu(bool boosted)
6182 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6184 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
6185 "sched RCU must be held");
6187 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
6190 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
6192 int target_cpu = -1;
6193 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
6194 unsigned long backup_capacity = ULONG_MAX;
6195 int best_idle_cpu = -1;
6196 int best_idle_cstate = INT_MAX;
6197 int backup_cpu = -1;
6198 unsigned long min_util = boosted_task_util(p);
6199 struct sched_domain *sd;
6200 struct sched_group *sg;
6201 int cpu = start_cpu(boosted);
6203 schedstat_inc(p, se.statistics.nr_wakeups_fbt_attempts);
6204 schedstat_inc(this_rq(), eas_stats.fbt_attempts);
6207 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_cpu);
6208 schedstat_inc(this_rq(), eas_stats.fbt_no_cpu);
6212 sd = rcu_dereference(per_cpu(sd_ea, cpu));
6215 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_sd);
6216 schedstat_inc(this_rq(), eas_stats.fbt_no_sd);
6225 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6226 unsigned long cur_capacity, new_util, wake_util;
6227 unsigned long min_wake_util = ULONG_MAX;
6233 * p's blocked utilization is still accounted for on prev_cpu
6234 * so prev_cpu will receive a negative bias due to the double
6235 * accounting. However, the blocked utilization may be zero.
6237 wake_util = cpu_util_wake(i, p);
6238 new_util = wake_util + task_util(p);
6241 * Ensure minimum capacity to grant the required boost.
6242 * The target CPU can be already at a capacity level higher
6243 * than the one required to boost the task.
6245 new_util = max(min_util, new_util);
6247 if (new_util > capacity_orig_of(i))
6250 #ifdef CONFIG_SCHED_WALT
6251 if (walt_cpu_high_irqload(i))
6256 * Unconditionally favoring tasks that prefer idle cpus to
6259 if (idle_cpu(i) && prefer_idle) {
6260 schedstat_inc(p, se.statistics.nr_wakeups_fbt_pref_idle);
6261 schedstat_inc(this_rq(), eas_stats.fbt_pref_idle);
6265 cur_capacity = capacity_curr_of(i);
6267 if (new_util < cur_capacity) {
6268 if (cpu_rq(i)->nr_running) {
6270 * Find a target cpu with the lowest/highest
6271 * utilization if prefer_idle/!prefer_idle.
6274 /* Favor the CPU that last ran the task */
6275 if (new_util > target_util ||
6276 wake_util > min_wake_util)
6278 min_wake_util = wake_util;
6279 target_util = new_util;
6281 } else if (target_util < new_util) {
6282 target_util = new_util;
6285 } else if (!prefer_idle) {
6286 int idle_idx = idle_get_state_idx(cpu_rq(i));
6288 if (best_idle_cpu < 0 ||
6289 (sysctl_sched_cstate_aware &&
6290 best_idle_cstate > idle_idx)) {
6291 best_idle_cstate = idle_idx;
6295 } else if (backup_capacity > cur_capacity) {
6296 /* Find a backup cpu with least capacity. */
6297 backup_capacity = cur_capacity;
6301 } while (sg = sg->next, sg != sd->groups);
6304 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
6306 if (target_cpu >= 0) {
6307 schedstat_inc(p, se.statistics.nr_wakeups_fbt_count);
6308 schedstat_inc(this_rq(), eas_stats.fbt_count);
6315 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6316 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6318 * In that case WAKE_AFFINE doesn't make sense and we'll let
6319 * BALANCE_WAKE sort things out.
6321 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6323 long min_cap, max_cap;
6325 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6326 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6328 /* Minimum capacity is close to max, no need to abort wake_affine */
6329 if (max_cap - min_cap < max_cap >> 3)
6332 /* Bring task utilization in sync with prev_cpu */
6333 sync_entity_load_avg(&p->se);
6335 return min_cap * 1024 < task_util(p) * capacity_margin;
6338 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6340 struct sched_domain *sd;
6341 int target_cpu = prev_cpu, tmp_target;
6342 bool boosted, prefer_idle;
6344 schedstat_inc(p, se.statistics.nr_wakeups_secb_attempts);
6345 schedstat_inc(this_rq(), eas_stats.secb_attempts);
6347 if (sysctl_sched_sync_hint_enable && sync) {
6348 int cpu = smp_processor_id();
6350 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6351 schedstat_inc(p, se.statistics.nr_wakeups_secb_sync);
6352 schedstat_inc(this_rq(), eas_stats.secb_sync);
6358 #ifdef CONFIG_CGROUP_SCHEDTUNE
6359 boosted = schedtune_task_boost(p) > 0;
6360 prefer_idle = schedtune_prefer_idle(p) > 0;
6362 boosted = get_sysctl_sched_cfs_boost() > 0;
6366 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6367 /* Find a cpu with sufficient capacity */
6368 tmp_target = find_best_target(p, boosted, prefer_idle);
6372 if (tmp_target >= 0) {
6373 target_cpu = tmp_target;
6374 if ((boosted || prefer_idle) && idle_cpu(target_cpu)) {
6375 schedstat_inc(p, se.statistics.nr_wakeups_secb_idle_bt);
6376 schedstat_inc(this_rq(), eas_stats.secb_idle_bt);
6381 if (target_cpu != prev_cpu) {
6382 struct energy_env eenv = {
6383 .util_delta = task_util(p),
6384 .src_cpu = prev_cpu,
6385 .dst_cpu = target_cpu,
6389 /* Not enough spare capacity on previous cpu */
6390 if (cpu_overutilized(prev_cpu)) {
6391 schedstat_inc(p, se.statistics.nr_wakeups_secb_insuff_cap);
6392 schedstat_inc(this_rq(), eas_stats.secb_insuff_cap);
6396 if (energy_diff(&eenv) >= 0) {
6397 schedstat_inc(p, se.statistics.nr_wakeups_secb_no_nrg_sav);
6398 schedstat_inc(this_rq(), eas_stats.secb_no_nrg_sav);
6399 target_cpu = prev_cpu;
6403 schedstat_inc(p, se.statistics.nr_wakeups_secb_nrg_sav);
6404 schedstat_inc(this_rq(), eas_stats.secb_nrg_sav);
6408 schedstat_inc(p, se.statistics.nr_wakeups_secb_count);
6409 schedstat_inc(this_rq(), eas_stats.secb_count);
6418 * select_task_rq_fair: Select target runqueue for the waking task in domains
6419 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6420 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6422 * Balances load by selecting the idlest cpu in the idlest group, or under
6423 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6425 * Returns the target cpu number.
6427 * preempt must be disabled.
6430 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6432 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6433 int cpu = smp_processor_id();
6434 int new_cpu = prev_cpu;
6435 int want_affine = 0;
6436 int sync = wake_flags & WF_SYNC;
6438 if (sd_flag & SD_BALANCE_WAKE) {
6440 * do wake_cap unconditionally as it causes task and cpu
6441 * utilization to be synced, and we need that for energy
6444 int _wake_cap = wake_cap(p, cpu, prev_cpu);
6445 want_affine = !wake_wide(p) && !_wake_cap
6446 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
6449 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6450 return select_energy_cpu_brute(p, prev_cpu, sync);
6453 for_each_domain(cpu, tmp) {
6454 if (!(tmp->flags & SD_LOAD_BALANCE))
6458 * If both cpu and prev_cpu are part of this domain,
6459 * cpu is a valid SD_WAKE_AFFINE target.
6461 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6462 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6467 if (tmp->flags & sd_flag)
6469 else if (!want_affine)
6474 sd = NULL; /* Prefer wake_affine over balance flags */
6475 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6480 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6481 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6484 int wu = sd_flag & SD_BALANCE_WAKE;
6488 schedstat_inc(p, se.statistics.nr_wakeups_cas_attempts);
6489 schedstat_inc(this_rq(), eas_stats.cas_attempts);
6493 struct sched_group *group;
6497 schedstat_inc(sd, eas_stats.cas_attempts);
6499 if (!(sd->flags & sd_flag)) {
6504 group = find_idlest_group(sd, p, cpu, sd_flag);
6510 new_cpu = find_idlest_cpu(group, p, cpu);
6511 if (new_cpu == -1 || new_cpu == cpu) {
6512 /* Now try balancing at a lower domain level of cpu */
6517 /* Now try balancing at a lower domain level of new_cpu */
6518 cpu = cas_cpu = new_cpu;
6519 weight = sd->span_weight;
6521 for_each_domain(cpu, tmp) {
6522 if (weight <= tmp->span_weight)
6524 if (tmp->flags & sd_flag)
6527 /* while loop will break here if sd == NULL */
6530 if (wu && (cas_cpu >= 0)) {
6531 schedstat_inc(p, se.statistics.nr_wakeups_cas_count);
6532 schedstat_inc(this_rq(), eas_stats.cas_count);
6541 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6542 * cfs_rq_of(p) references at time of call are still valid and identify the
6543 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6544 * other assumptions, including the state of rq->lock, should be made.
6546 static void migrate_task_rq_fair(struct task_struct *p)
6549 * We are supposed to update the task to "current" time, then its up to date
6550 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6551 * what current time is, so simply throw away the out-of-date time. This
6552 * will result in the wakee task is less decayed, but giving the wakee more
6553 * load sounds not bad.
6555 remove_entity_load_avg(&p->se);
6557 /* Tell new CPU we are migrated */
6558 p->se.avg.last_update_time = 0;
6560 /* We have migrated, no longer consider this task hot */
6561 p->se.exec_start = 0;
6564 static void task_dead_fair(struct task_struct *p)
6566 remove_entity_load_avg(&p->se);
6569 #define task_fits_max(p, cpu) true
6570 #endif /* CONFIG_SMP */
6572 static unsigned long
6573 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6575 unsigned long gran = sysctl_sched_wakeup_granularity;
6578 * Since its curr running now, convert the gran from real-time
6579 * to virtual-time in his units.
6581 * By using 'se' instead of 'curr' we penalize light tasks, so
6582 * they get preempted easier. That is, if 'se' < 'curr' then
6583 * the resulting gran will be larger, therefore penalizing the
6584 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6585 * be smaller, again penalizing the lighter task.
6587 * This is especially important for buddies when the leftmost
6588 * task is higher priority than the buddy.
6590 return calc_delta_fair(gran, se);
6594 * Should 'se' preempt 'curr'.
6608 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6610 s64 gran, vdiff = curr->vruntime - se->vruntime;
6615 gran = wakeup_gran(curr, se);
6622 static void set_last_buddy(struct sched_entity *se)
6624 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6627 for_each_sched_entity(se)
6628 cfs_rq_of(se)->last = se;
6631 static void set_next_buddy(struct sched_entity *se)
6633 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6636 for_each_sched_entity(se)
6637 cfs_rq_of(se)->next = se;
6640 static void set_skip_buddy(struct sched_entity *se)
6642 for_each_sched_entity(se)
6643 cfs_rq_of(se)->skip = se;
6647 * Preempt the current task with a newly woken task if needed:
6649 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6651 struct task_struct *curr = rq->curr;
6652 struct sched_entity *se = &curr->se, *pse = &p->se;
6653 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6654 int scale = cfs_rq->nr_running >= sched_nr_latency;
6655 int next_buddy_marked = 0;
6657 if (unlikely(se == pse))
6661 * This is possible from callers such as attach_tasks(), in which we
6662 * unconditionally check_prempt_curr() after an enqueue (which may have
6663 * lead to a throttle). This both saves work and prevents false
6664 * next-buddy nomination below.
6666 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6669 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6670 set_next_buddy(pse);
6671 next_buddy_marked = 1;
6675 * We can come here with TIF_NEED_RESCHED already set from new task
6678 * Note: this also catches the edge-case of curr being in a throttled
6679 * group (e.g. via set_curr_task), since update_curr() (in the
6680 * enqueue of curr) will have resulted in resched being set. This
6681 * prevents us from potentially nominating it as a false LAST_BUDDY
6684 if (test_tsk_need_resched(curr))
6687 /* Idle tasks are by definition preempted by non-idle tasks. */
6688 if (unlikely(curr->policy == SCHED_IDLE) &&
6689 likely(p->policy != SCHED_IDLE))
6693 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6694 * is driven by the tick):
6696 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6699 find_matching_se(&se, &pse);
6700 update_curr(cfs_rq_of(se));
6702 if (wakeup_preempt_entity(se, pse) == 1) {
6704 * Bias pick_next to pick the sched entity that is
6705 * triggering this preemption.
6707 if (!next_buddy_marked)
6708 set_next_buddy(pse);
6717 * Only set the backward buddy when the current task is still
6718 * on the rq. This can happen when a wakeup gets interleaved
6719 * with schedule on the ->pre_schedule() or idle_balance()
6720 * point, either of which can * drop the rq lock.
6722 * Also, during early boot the idle thread is in the fair class,
6723 * for obvious reasons its a bad idea to schedule back to it.
6725 if (unlikely(!se->on_rq || curr == rq->idle))
6728 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6732 static struct task_struct *
6733 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6735 struct cfs_rq *cfs_rq = &rq->cfs;
6736 struct sched_entity *se;
6737 struct task_struct *p;
6741 #ifdef CONFIG_FAIR_GROUP_SCHED
6742 if (!cfs_rq->nr_running)
6745 if (prev->sched_class != &fair_sched_class)
6749 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6750 * likely that a next task is from the same cgroup as the current.
6752 * Therefore attempt to avoid putting and setting the entire cgroup
6753 * hierarchy, only change the part that actually changes.
6757 struct sched_entity *curr = cfs_rq->curr;
6760 * Since we got here without doing put_prev_entity() we also
6761 * have to consider cfs_rq->curr. If it is still a runnable
6762 * entity, update_curr() will update its vruntime, otherwise
6763 * forget we've ever seen it.
6767 update_curr(cfs_rq);
6772 * This call to check_cfs_rq_runtime() will do the
6773 * throttle and dequeue its entity in the parent(s).
6774 * Therefore the 'simple' nr_running test will indeed
6777 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6781 se = pick_next_entity(cfs_rq, curr);
6782 cfs_rq = group_cfs_rq(se);
6788 * Since we haven't yet done put_prev_entity and if the selected task
6789 * is a different task than we started out with, try and touch the
6790 * least amount of cfs_rqs.
6793 struct sched_entity *pse = &prev->se;
6795 while (!(cfs_rq = is_same_group(se, pse))) {
6796 int se_depth = se->depth;
6797 int pse_depth = pse->depth;
6799 if (se_depth <= pse_depth) {
6800 put_prev_entity(cfs_rq_of(pse), pse);
6801 pse = parent_entity(pse);
6803 if (se_depth >= pse_depth) {
6804 set_next_entity(cfs_rq_of(se), se);
6805 se = parent_entity(se);
6809 put_prev_entity(cfs_rq, pse);
6810 set_next_entity(cfs_rq, se);
6813 if (hrtick_enabled(rq))
6814 hrtick_start_fair(rq, p);
6816 rq->misfit_task = !task_fits_max(p, rq->cpu);
6823 if (!cfs_rq->nr_running)
6826 put_prev_task(rq, prev);
6829 se = pick_next_entity(cfs_rq, NULL);
6830 set_next_entity(cfs_rq, se);
6831 cfs_rq = group_cfs_rq(se);
6836 if (hrtick_enabled(rq))
6837 hrtick_start_fair(rq, p);
6839 rq->misfit_task = !task_fits_max(p, rq->cpu);
6844 rq->misfit_task = 0;
6846 * This is OK, because current is on_cpu, which avoids it being picked
6847 * for load-balance and preemption/IRQs are still disabled avoiding
6848 * further scheduler activity on it and we're being very careful to
6849 * re-start the picking loop.
6851 lockdep_unpin_lock(&rq->lock);
6852 new_tasks = idle_balance(rq);
6853 lockdep_pin_lock(&rq->lock);
6855 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6856 * possible for any higher priority task to appear. In that case we
6857 * must re-start the pick_next_entity() loop.
6869 * Account for a descheduled task:
6871 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6873 struct sched_entity *se = &prev->se;
6874 struct cfs_rq *cfs_rq;
6876 for_each_sched_entity(se) {
6877 cfs_rq = cfs_rq_of(se);
6878 put_prev_entity(cfs_rq, se);
6883 * sched_yield() is very simple
6885 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6887 static void yield_task_fair(struct rq *rq)
6889 struct task_struct *curr = rq->curr;
6890 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6891 struct sched_entity *se = &curr->se;
6894 * Are we the only task in the tree?
6896 if (unlikely(rq->nr_running == 1))
6899 clear_buddies(cfs_rq, se);
6901 if (curr->policy != SCHED_BATCH) {
6902 update_rq_clock(rq);
6904 * Update run-time statistics of the 'current'.
6906 update_curr(cfs_rq);
6908 * Tell update_rq_clock() that we've just updated,
6909 * so we don't do microscopic update in schedule()
6910 * and double the fastpath cost.
6912 rq_clock_skip_update(rq, true);
6918 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6920 struct sched_entity *se = &p->se;
6922 /* throttled hierarchies are not runnable */
6923 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6926 /* Tell the scheduler that we'd really like pse to run next. */
6929 yield_task_fair(rq);
6935 /**************************************************
6936 * Fair scheduling class load-balancing methods.
6940 * The purpose of load-balancing is to achieve the same basic fairness the
6941 * per-cpu scheduler provides, namely provide a proportional amount of compute
6942 * time to each task. This is expressed in the following equation:
6944 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6946 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6947 * W_i,0 is defined as:
6949 * W_i,0 = \Sum_j w_i,j (2)
6951 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6952 * is derived from the nice value as per prio_to_weight[].
6954 * The weight average is an exponential decay average of the instantaneous
6957 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6959 * C_i is the compute capacity of cpu i, typically it is the
6960 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6961 * can also include other factors [XXX].
6963 * To achieve this balance we define a measure of imbalance which follows
6964 * directly from (1):
6966 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6968 * We them move tasks around to minimize the imbalance. In the continuous
6969 * function space it is obvious this converges, in the discrete case we get
6970 * a few fun cases generally called infeasible weight scenarios.
6973 * - infeasible weights;
6974 * - local vs global optima in the discrete case. ]
6979 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6980 * for all i,j solution, we create a tree of cpus that follows the hardware
6981 * topology where each level pairs two lower groups (or better). This results
6982 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6983 * tree to only the first of the previous level and we decrease the frequency
6984 * of load-balance at each level inv. proportional to the number of cpus in
6990 * \Sum { --- * --- * 2^i } = O(n) (5)
6992 * `- size of each group
6993 * | | `- number of cpus doing load-balance
6995 * `- sum over all levels
6997 * Coupled with a limit on how many tasks we can migrate every balance pass,
6998 * this makes (5) the runtime complexity of the balancer.
7000 * An important property here is that each CPU is still (indirectly) connected
7001 * to every other cpu in at most O(log n) steps:
7003 * The adjacency matrix of the resulting graph is given by:
7006 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7009 * And you'll find that:
7011 * A^(log_2 n)_i,j != 0 for all i,j (7)
7013 * Showing there's indeed a path between every cpu in at most O(log n) steps.
7014 * The task movement gives a factor of O(m), giving a convergence complexity
7017 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7022 * In order to avoid CPUs going idle while there's still work to do, new idle
7023 * balancing is more aggressive and has the newly idle cpu iterate up the domain
7024 * tree itself instead of relying on other CPUs to bring it work.
7026 * This adds some complexity to both (5) and (8) but it reduces the total idle
7034 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7037 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7042 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7044 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
7046 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7049 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7050 * rewrite all of this once again.]
7053 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7055 enum fbq_type { regular, remote, all };
7064 #define LBF_ALL_PINNED 0x01
7065 #define LBF_NEED_BREAK 0x02
7066 #define LBF_DST_PINNED 0x04
7067 #define LBF_SOME_PINNED 0x08
7070 struct sched_domain *sd;
7078 struct cpumask *dst_grpmask;
7080 enum cpu_idle_type idle;
7082 unsigned int src_grp_nr_running;
7083 /* The set of CPUs under consideration for load-balancing */
7084 struct cpumask *cpus;
7089 unsigned int loop_break;
7090 unsigned int loop_max;
7092 enum fbq_type fbq_type;
7093 enum group_type busiest_group_type;
7094 struct list_head tasks;
7098 * Is this task likely cache-hot:
7100 static int task_hot(struct task_struct *p, struct lb_env *env)
7104 lockdep_assert_held(&env->src_rq->lock);
7106 if (p->sched_class != &fair_sched_class)
7109 if (unlikely(p->policy == SCHED_IDLE))
7113 * Buddy candidates are cache hot:
7115 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7116 (&p->se == cfs_rq_of(&p->se)->next ||
7117 &p->se == cfs_rq_of(&p->se)->last))
7120 if (sysctl_sched_migration_cost == -1)
7122 if (sysctl_sched_migration_cost == 0)
7125 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7127 return delta < (s64)sysctl_sched_migration_cost;
7130 #ifdef CONFIG_NUMA_BALANCING
7132 * Returns 1, if task migration degrades locality
7133 * Returns 0, if task migration improves locality i.e migration preferred.
7134 * Returns -1, if task migration is not affected by locality.
7136 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7138 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7139 unsigned long src_faults, dst_faults;
7140 int src_nid, dst_nid;
7142 if (!static_branch_likely(&sched_numa_balancing))
7145 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7148 src_nid = cpu_to_node(env->src_cpu);
7149 dst_nid = cpu_to_node(env->dst_cpu);
7151 if (src_nid == dst_nid)
7154 /* Migrating away from the preferred node is always bad. */
7155 if (src_nid == p->numa_preferred_nid) {
7156 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7162 /* Encourage migration to the preferred node. */
7163 if (dst_nid == p->numa_preferred_nid)
7167 src_faults = group_faults(p, src_nid);
7168 dst_faults = group_faults(p, dst_nid);
7170 src_faults = task_faults(p, src_nid);
7171 dst_faults = task_faults(p, dst_nid);
7174 return dst_faults < src_faults;
7178 static inline int migrate_degrades_locality(struct task_struct *p,
7186 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7189 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7193 lockdep_assert_held(&env->src_rq->lock);
7196 * We do not migrate tasks that are:
7197 * 1) throttled_lb_pair, or
7198 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7199 * 3) running (obviously), or
7200 * 4) are cache-hot on their current CPU.
7202 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7205 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
7208 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
7210 env->flags |= LBF_SOME_PINNED;
7213 * Remember if this task can be migrated to any other cpu in
7214 * our sched_group. We may want to revisit it if we couldn't
7215 * meet load balance goals by pulling other tasks on src_cpu.
7217 * Also avoid computing new_dst_cpu if we have already computed
7218 * one in current iteration.
7220 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
7223 /* Prevent to re-select dst_cpu via env's cpus */
7224 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7225 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
7226 env->flags |= LBF_DST_PINNED;
7227 env->new_dst_cpu = cpu;
7235 /* Record that we found atleast one task that could run on dst_cpu */
7236 env->flags &= ~LBF_ALL_PINNED;
7238 if (task_running(env->src_rq, p)) {
7239 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
7244 * Aggressive migration if:
7245 * 1) destination numa is preferred
7246 * 2) task is cache cold, or
7247 * 3) too many balance attempts have failed.
7249 tsk_cache_hot = migrate_degrades_locality(p, env);
7250 if (tsk_cache_hot == -1)
7251 tsk_cache_hot = task_hot(p, env);
7253 if (tsk_cache_hot <= 0 ||
7254 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7255 if (tsk_cache_hot == 1) {
7256 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
7257 schedstat_inc(p, se.statistics.nr_forced_migrations);
7262 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
7267 * detach_task() -- detach the task for the migration specified in env
7269 static void detach_task(struct task_struct *p, struct lb_env *env)
7271 lockdep_assert_held(&env->src_rq->lock);
7273 deactivate_task(env->src_rq, p, 0);
7274 p->on_rq = TASK_ON_RQ_MIGRATING;
7275 double_lock_balance(env->src_rq, env->dst_rq);
7276 set_task_cpu(p, env->dst_cpu);
7277 double_unlock_balance(env->src_rq, env->dst_rq);
7281 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7282 * part of active balancing operations within "domain".
7284 * Returns a task if successful and NULL otherwise.
7286 static struct task_struct *detach_one_task(struct lb_env *env)
7288 struct task_struct *p, *n;
7290 lockdep_assert_held(&env->src_rq->lock);
7292 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7293 if (!can_migrate_task(p, env))
7296 detach_task(p, env);
7299 * Right now, this is only the second place where
7300 * lb_gained[env->idle] is updated (other is detach_tasks)
7301 * so we can safely collect stats here rather than
7302 * inside detach_tasks().
7304 schedstat_inc(env->sd, lb_gained[env->idle]);
7310 static const unsigned int sched_nr_migrate_break = 32;
7313 * detach_tasks() -- tries to detach up to imbalance weighted load from
7314 * busiest_rq, as part of a balancing operation within domain "sd".
7316 * Returns number of detached tasks if successful and 0 otherwise.
7318 static int detach_tasks(struct lb_env *env)
7320 struct list_head *tasks = &env->src_rq->cfs_tasks;
7321 struct task_struct *p;
7325 lockdep_assert_held(&env->src_rq->lock);
7327 if (env->imbalance <= 0)
7330 while (!list_empty(tasks)) {
7332 * We don't want to steal all, otherwise we may be treated likewise,
7333 * which could at worst lead to a livelock crash.
7335 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7338 p = list_first_entry(tasks, struct task_struct, se.group_node);
7341 /* We've more or less seen every task there is, call it quits */
7342 if (env->loop > env->loop_max)
7345 /* take a breather every nr_migrate tasks */
7346 if (env->loop > env->loop_break) {
7347 env->loop_break += sched_nr_migrate_break;
7348 env->flags |= LBF_NEED_BREAK;
7352 if (!can_migrate_task(p, env))
7355 load = task_h_load(p);
7357 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7360 if ((load / 2) > env->imbalance)
7363 detach_task(p, env);
7364 list_add(&p->se.group_node, &env->tasks);
7367 env->imbalance -= load;
7369 #ifdef CONFIG_PREEMPT
7371 * NEWIDLE balancing is a source of latency, so preemptible
7372 * kernels will stop after the first task is detached to minimize
7373 * the critical section.
7375 if (env->idle == CPU_NEWLY_IDLE)
7380 * We only want to steal up to the prescribed amount of
7383 if (env->imbalance <= 0)
7388 list_move_tail(&p->se.group_node, tasks);
7392 * Right now, this is one of only two places we collect this stat
7393 * so we can safely collect detach_one_task() stats here rather
7394 * than inside detach_one_task().
7396 schedstat_add(env->sd, lb_gained[env->idle], detached);
7402 * attach_task() -- attach the task detached by detach_task() to its new rq.
7404 static void attach_task(struct rq *rq, struct task_struct *p)
7406 lockdep_assert_held(&rq->lock);
7408 BUG_ON(task_rq(p) != rq);
7409 p->on_rq = TASK_ON_RQ_QUEUED;
7410 activate_task(rq, p, 0);
7411 check_preempt_curr(rq, p, 0);
7415 * attach_one_task() -- attaches the task returned from detach_one_task() to
7418 static void attach_one_task(struct rq *rq, struct task_struct *p)
7420 raw_spin_lock(&rq->lock);
7423 * We want to potentially raise target_cpu's OPP.
7425 update_capacity_of(cpu_of(rq));
7426 raw_spin_unlock(&rq->lock);
7430 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7433 static void attach_tasks(struct lb_env *env)
7435 struct list_head *tasks = &env->tasks;
7436 struct task_struct *p;
7438 raw_spin_lock(&env->dst_rq->lock);
7440 while (!list_empty(tasks)) {
7441 p = list_first_entry(tasks, struct task_struct, se.group_node);
7442 list_del_init(&p->se.group_node);
7444 attach_task(env->dst_rq, p);
7448 * We want to potentially raise env.dst_cpu's OPP.
7450 update_capacity_of(env->dst_cpu);
7452 raw_spin_unlock(&env->dst_rq->lock);
7455 #ifdef CONFIG_FAIR_GROUP_SCHED
7456 static void update_blocked_averages(int cpu)
7458 struct rq *rq = cpu_rq(cpu);
7459 struct cfs_rq *cfs_rq;
7460 unsigned long flags;
7462 raw_spin_lock_irqsave(&rq->lock, flags);
7463 update_rq_clock(rq);
7466 * Iterates the task_group tree in a bottom up fashion, see
7467 * list_add_leaf_cfs_rq() for details.
7469 for_each_leaf_cfs_rq(rq, cfs_rq) {
7470 /* throttled entities do not contribute to load */
7471 if (throttled_hierarchy(cfs_rq))
7474 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7476 update_tg_load_avg(cfs_rq, 0);
7478 /* Propagate pending load changes to the parent */
7479 if (cfs_rq->tg->se[cpu])
7480 update_load_avg(cfs_rq->tg->se[cpu], 0);
7482 raw_spin_unlock_irqrestore(&rq->lock, flags);
7486 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7487 * This needs to be done in a top-down fashion because the load of a child
7488 * group is a fraction of its parents load.
7490 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7492 struct rq *rq = rq_of(cfs_rq);
7493 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7494 unsigned long now = jiffies;
7497 if (cfs_rq->last_h_load_update == now)
7500 cfs_rq->h_load_next = NULL;
7501 for_each_sched_entity(se) {
7502 cfs_rq = cfs_rq_of(se);
7503 cfs_rq->h_load_next = se;
7504 if (cfs_rq->last_h_load_update == now)
7509 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7510 cfs_rq->last_h_load_update = now;
7513 while ((se = cfs_rq->h_load_next) != NULL) {
7514 load = cfs_rq->h_load;
7515 load = div64_ul(load * se->avg.load_avg,
7516 cfs_rq_load_avg(cfs_rq) + 1);
7517 cfs_rq = group_cfs_rq(se);
7518 cfs_rq->h_load = load;
7519 cfs_rq->last_h_load_update = now;
7523 static unsigned long task_h_load(struct task_struct *p)
7525 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7527 update_cfs_rq_h_load(cfs_rq);
7528 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7529 cfs_rq_load_avg(cfs_rq) + 1);
7532 static inline void update_blocked_averages(int cpu)
7534 struct rq *rq = cpu_rq(cpu);
7535 struct cfs_rq *cfs_rq = &rq->cfs;
7536 unsigned long flags;
7538 raw_spin_lock_irqsave(&rq->lock, flags);
7539 update_rq_clock(rq);
7540 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7541 raw_spin_unlock_irqrestore(&rq->lock, flags);
7544 static unsigned long task_h_load(struct task_struct *p)
7546 return p->se.avg.load_avg;
7550 /********** Helpers for find_busiest_group ************************/
7553 * sg_lb_stats - stats of a sched_group required for load_balancing
7555 struct sg_lb_stats {
7556 unsigned long avg_load; /*Avg load across the CPUs of the group */
7557 unsigned long group_load; /* Total load over the CPUs of the group */
7558 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7559 unsigned long load_per_task;
7560 unsigned long group_capacity;
7561 unsigned long group_util; /* Total utilization of the group */
7562 unsigned int sum_nr_running; /* Nr tasks running in the group */
7563 unsigned int idle_cpus;
7564 unsigned int group_weight;
7565 enum group_type group_type;
7566 int group_no_capacity;
7567 int group_misfit_task; /* A cpu has a task too big for its capacity */
7568 #ifdef CONFIG_NUMA_BALANCING
7569 unsigned int nr_numa_running;
7570 unsigned int nr_preferred_running;
7575 * sd_lb_stats - Structure to store the statistics of a sched_domain
7576 * during load balancing.
7578 struct sd_lb_stats {
7579 struct sched_group *busiest; /* Busiest group in this sd */
7580 struct sched_group *local; /* Local group in this sd */
7581 unsigned long total_load; /* Total load of all groups in sd */
7582 unsigned long total_capacity; /* Total capacity of all groups in sd */
7583 unsigned long avg_load; /* Average load across all groups in sd */
7585 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7586 struct sg_lb_stats local_stat; /* Statistics of the local group */
7589 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7592 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7593 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7594 * We must however clear busiest_stat::avg_load because
7595 * update_sd_pick_busiest() reads this before assignment.
7597 *sds = (struct sd_lb_stats){
7601 .total_capacity = 0UL,
7604 .sum_nr_running = 0,
7605 .group_type = group_other,
7611 * get_sd_load_idx - Obtain the load index for a given sched domain.
7612 * @sd: The sched_domain whose load_idx is to be obtained.
7613 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7615 * Return: The load index.
7617 static inline int get_sd_load_idx(struct sched_domain *sd,
7618 enum cpu_idle_type idle)
7624 load_idx = sd->busy_idx;
7627 case CPU_NEWLY_IDLE:
7628 load_idx = sd->newidle_idx;
7631 load_idx = sd->idle_idx;
7638 static unsigned long scale_rt_capacity(int cpu)
7640 struct rq *rq = cpu_rq(cpu);
7641 u64 total, used, age_stamp, avg;
7645 * Since we're reading these variables without serialization make sure
7646 * we read them once before doing sanity checks on them.
7648 age_stamp = READ_ONCE(rq->age_stamp);
7649 avg = READ_ONCE(rq->rt_avg);
7650 delta = __rq_clock_broken(rq) - age_stamp;
7652 if (unlikely(delta < 0))
7655 total = sched_avg_period() + delta;
7657 used = div_u64(avg, total);
7660 * deadline bandwidth is defined at system level so we must
7661 * weight this bandwidth with the max capacity of the system.
7662 * As a reminder, avg_bw is 20bits width and
7663 * scale_cpu_capacity is 10 bits width
7665 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7667 if (likely(used < SCHED_CAPACITY_SCALE))
7668 return SCHED_CAPACITY_SCALE - used;
7673 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7675 raw_spin_lock_init(&mcc->lock);
7680 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7682 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7683 struct sched_group *sdg = sd->groups;
7684 struct max_cpu_capacity *mcc;
7685 unsigned long max_capacity;
7687 unsigned long flags;
7689 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7691 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7693 raw_spin_lock_irqsave(&mcc->lock, flags);
7694 max_capacity = mcc->val;
7695 max_cap_cpu = mcc->cpu;
7697 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7698 (max_capacity < capacity)) {
7699 mcc->val = capacity;
7701 #ifdef CONFIG_SCHED_DEBUG
7702 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7703 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7708 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7710 skip_unlock: __attribute__ ((unused));
7711 capacity *= scale_rt_capacity(cpu);
7712 capacity >>= SCHED_CAPACITY_SHIFT;
7717 cpu_rq(cpu)->cpu_capacity = capacity;
7718 sdg->sgc->capacity = capacity;
7719 sdg->sgc->max_capacity = capacity;
7720 sdg->sgc->min_capacity = capacity;
7723 void update_group_capacity(struct sched_domain *sd, int cpu)
7725 struct sched_domain *child = sd->child;
7726 struct sched_group *group, *sdg = sd->groups;
7727 unsigned long capacity, max_capacity, min_capacity;
7728 unsigned long interval;
7730 interval = msecs_to_jiffies(sd->balance_interval);
7731 interval = clamp(interval, 1UL, max_load_balance_interval);
7732 sdg->sgc->next_update = jiffies + interval;
7735 update_cpu_capacity(sd, cpu);
7741 min_capacity = ULONG_MAX;
7743 if (child->flags & SD_OVERLAP) {
7745 * SD_OVERLAP domains cannot assume that child groups
7746 * span the current group.
7749 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7750 struct sched_group_capacity *sgc;
7751 struct rq *rq = cpu_rq(cpu);
7754 * build_sched_domains() -> init_sched_groups_capacity()
7755 * gets here before we've attached the domains to the
7758 * Use capacity_of(), which is set irrespective of domains
7759 * in update_cpu_capacity().
7761 * This avoids capacity from being 0 and
7762 * causing divide-by-zero issues on boot.
7764 if (unlikely(!rq->sd)) {
7765 capacity += capacity_of(cpu);
7767 sgc = rq->sd->groups->sgc;
7768 capacity += sgc->capacity;
7771 max_capacity = max(capacity, max_capacity);
7772 min_capacity = min(capacity, min_capacity);
7776 * !SD_OVERLAP domains can assume that child groups
7777 * span the current group.
7780 group = child->groups;
7782 struct sched_group_capacity *sgc = group->sgc;
7784 capacity += sgc->capacity;
7785 max_capacity = max(sgc->max_capacity, max_capacity);
7786 min_capacity = min(sgc->min_capacity, min_capacity);
7787 group = group->next;
7788 } while (group != child->groups);
7791 sdg->sgc->capacity = capacity;
7792 sdg->sgc->max_capacity = max_capacity;
7793 sdg->sgc->min_capacity = min_capacity;
7797 * Check whether the capacity of the rq has been noticeably reduced by side
7798 * activity. The imbalance_pct is used for the threshold.
7799 * Return true is the capacity is reduced
7802 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7804 return ((rq->cpu_capacity * sd->imbalance_pct) <
7805 (rq->cpu_capacity_orig * 100));
7809 * Group imbalance indicates (and tries to solve) the problem where balancing
7810 * groups is inadequate due to tsk_cpus_allowed() constraints.
7812 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7813 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7816 * { 0 1 2 3 } { 4 5 6 7 }
7819 * If we were to balance group-wise we'd place two tasks in the first group and
7820 * two tasks in the second group. Clearly this is undesired as it will overload
7821 * cpu 3 and leave one of the cpus in the second group unused.
7823 * The current solution to this issue is detecting the skew in the first group
7824 * by noticing the lower domain failed to reach balance and had difficulty
7825 * moving tasks due to affinity constraints.
7827 * When this is so detected; this group becomes a candidate for busiest; see
7828 * update_sd_pick_busiest(). And calculate_imbalance() and
7829 * find_busiest_group() avoid some of the usual balance conditions to allow it
7830 * to create an effective group imbalance.
7832 * This is a somewhat tricky proposition since the next run might not find the
7833 * group imbalance and decide the groups need to be balanced again. A most
7834 * subtle and fragile situation.
7837 static inline int sg_imbalanced(struct sched_group *group)
7839 return group->sgc->imbalance;
7843 * group_has_capacity returns true if the group has spare capacity that could
7844 * be used by some tasks.
7845 * We consider that a group has spare capacity if the * number of task is
7846 * smaller than the number of CPUs or if the utilization is lower than the
7847 * available capacity for CFS tasks.
7848 * For the latter, we use a threshold to stabilize the state, to take into
7849 * account the variance of the tasks' load and to return true if the available
7850 * capacity in meaningful for the load balancer.
7851 * As an example, an available capacity of 1% can appear but it doesn't make
7852 * any benefit for the load balance.
7855 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7857 if (sgs->sum_nr_running < sgs->group_weight)
7860 if ((sgs->group_capacity * 100) >
7861 (sgs->group_util * env->sd->imbalance_pct))
7868 * group_is_overloaded returns true if the group has more tasks than it can
7870 * group_is_overloaded is not equals to !group_has_capacity because a group
7871 * with the exact right number of tasks, has no more spare capacity but is not
7872 * overloaded so both group_has_capacity and group_is_overloaded return
7876 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7878 if (sgs->sum_nr_running <= sgs->group_weight)
7881 if ((sgs->group_capacity * 100) <
7882 (sgs->group_util * env->sd->imbalance_pct))
7890 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7891 * per-cpu capacity than sched_group ref.
7894 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7896 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7897 ref->sgc->max_capacity;
7901 group_type group_classify(struct sched_group *group,
7902 struct sg_lb_stats *sgs)
7904 if (sgs->group_no_capacity)
7905 return group_overloaded;
7907 if (sg_imbalanced(group))
7908 return group_imbalanced;
7910 if (sgs->group_misfit_task)
7911 return group_misfit_task;
7917 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7918 * @env: The load balancing environment.
7919 * @group: sched_group whose statistics are to be updated.
7920 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7921 * @local_group: Does group contain this_cpu.
7922 * @sgs: variable to hold the statistics for this group.
7923 * @overload: Indicate more than one runnable task for any CPU.
7924 * @overutilized: Indicate overutilization for any CPU.
7926 static inline void update_sg_lb_stats(struct lb_env *env,
7927 struct sched_group *group, int load_idx,
7928 int local_group, struct sg_lb_stats *sgs,
7929 bool *overload, bool *overutilized)
7934 memset(sgs, 0, sizeof(*sgs));
7936 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7937 struct rq *rq = cpu_rq(i);
7939 /* Bias balancing toward cpus of our domain */
7941 load = target_load(i, load_idx);
7943 load = source_load(i, load_idx);
7945 sgs->group_load += load;
7946 sgs->group_util += cpu_util(i);
7947 sgs->sum_nr_running += rq->cfs.h_nr_running;
7949 nr_running = rq->nr_running;
7953 #ifdef CONFIG_NUMA_BALANCING
7954 sgs->nr_numa_running += rq->nr_numa_running;
7955 sgs->nr_preferred_running += rq->nr_preferred_running;
7957 sgs->sum_weighted_load += weighted_cpuload(i);
7959 * No need to call idle_cpu() if nr_running is not 0
7961 if (!nr_running && idle_cpu(i))
7964 if (cpu_overutilized(i)) {
7965 *overutilized = true;
7966 if (!sgs->group_misfit_task && rq->misfit_task)
7967 sgs->group_misfit_task = capacity_of(i);
7971 /* Adjust by relative CPU capacity of the group */
7972 sgs->group_capacity = group->sgc->capacity;
7973 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7975 if (sgs->sum_nr_running)
7976 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7978 sgs->group_weight = group->group_weight;
7980 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7981 sgs->group_type = group_classify(group, sgs);
7985 * update_sd_pick_busiest - return 1 on busiest group
7986 * @env: The load balancing environment.
7987 * @sds: sched_domain statistics
7988 * @sg: sched_group candidate to be checked for being the busiest
7989 * @sgs: sched_group statistics
7991 * Determine if @sg is a busier group than the previously selected
7994 * Return: %true if @sg is a busier group than the previously selected
7995 * busiest group. %false otherwise.
7997 static bool update_sd_pick_busiest(struct lb_env *env,
7998 struct sd_lb_stats *sds,
7999 struct sched_group *sg,
8000 struct sg_lb_stats *sgs)
8002 struct sg_lb_stats *busiest = &sds->busiest_stat;
8004 if (sgs->group_type > busiest->group_type)
8007 if (sgs->group_type < busiest->group_type)
8011 * Candidate sg doesn't face any serious load-balance problems
8012 * so don't pick it if the local sg is already filled up.
8014 if (sgs->group_type == group_other &&
8015 !group_has_capacity(env, &sds->local_stat))
8018 if (sgs->avg_load <= busiest->avg_load)
8021 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8025 * Candidate sg has no more than one task per CPU and
8026 * has higher per-CPU capacity. Migrating tasks to less
8027 * capable CPUs may harm throughput. Maximize throughput,
8028 * power/energy consequences are not considered.
8030 if (sgs->sum_nr_running <= sgs->group_weight &&
8031 group_smaller_cpu_capacity(sds->local, sg))
8035 /* This is the busiest node in its class. */
8036 if (!(env->sd->flags & SD_ASYM_PACKING))
8040 * ASYM_PACKING needs to move all the work to the lowest
8041 * numbered CPUs in the group, therefore mark all groups
8042 * higher than ourself as busy.
8044 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
8048 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
8055 #ifdef CONFIG_NUMA_BALANCING
8056 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8058 if (sgs->sum_nr_running > sgs->nr_numa_running)
8060 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8065 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8067 if (rq->nr_running > rq->nr_numa_running)
8069 if (rq->nr_running > rq->nr_preferred_running)
8074 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8079 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8083 #endif /* CONFIG_NUMA_BALANCING */
8085 #define lb_sd_parent(sd) \
8086 (sd->parent && sd->parent->groups != sd->parent->groups->next)
8089 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8090 * @env: The load balancing environment.
8091 * @sds: variable to hold the statistics for this sched_domain.
8093 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8095 struct sched_domain *child = env->sd->child;
8096 struct sched_group *sg = env->sd->groups;
8097 struct sg_lb_stats tmp_sgs;
8098 int load_idx, prefer_sibling = 0;
8099 bool overload = false, overutilized = false;
8101 if (child && child->flags & SD_PREFER_SIBLING)
8104 load_idx = get_sd_load_idx(env->sd, env->idle);
8107 struct sg_lb_stats *sgs = &tmp_sgs;
8110 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
8113 sgs = &sds->local_stat;
8115 if (env->idle != CPU_NEWLY_IDLE ||
8116 time_after_eq(jiffies, sg->sgc->next_update))
8117 update_group_capacity(env->sd, env->dst_cpu);
8120 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8121 &overload, &overutilized);
8127 * In case the child domain prefers tasks go to siblings
8128 * first, lower the sg capacity so that we'll try
8129 * and move all the excess tasks away. We lower the capacity
8130 * of a group only if the local group has the capacity to fit
8131 * these excess tasks. The extra check prevents the case where
8132 * you always pull from the heaviest group when it is already
8133 * under-utilized (possible with a large weight task outweighs
8134 * the tasks on the system).
8136 if (prefer_sibling && sds->local &&
8137 group_has_capacity(env, &sds->local_stat) &&
8138 (sgs->sum_nr_running > 1)) {
8139 sgs->group_no_capacity = 1;
8140 sgs->group_type = group_classify(sg, sgs);
8144 * Ignore task groups with misfit tasks if local group has no
8145 * capacity or if per-cpu capacity isn't higher.
8147 if (sgs->group_type == group_misfit_task &&
8148 (!group_has_capacity(env, &sds->local_stat) ||
8149 !group_smaller_cpu_capacity(sg, sds->local)))
8150 sgs->group_type = group_other;
8152 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8154 sds->busiest_stat = *sgs;
8158 /* Now, start updating sd_lb_stats */
8159 sds->total_load += sgs->group_load;
8160 sds->total_capacity += sgs->group_capacity;
8163 } while (sg != env->sd->groups);
8165 if (env->sd->flags & SD_NUMA)
8166 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8168 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
8170 if (!lb_sd_parent(env->sd)) {
8171 /* update overload indicator if we are at root domain */
8172 if (env->dst_rq->rd->overload != overload)
8173 env->dst_rq->rd->overload = overload;
8175 /* Update over-utilization (tipping point, U >= 0) indicator */
8176 if (env->dst_rq->rd->overutilized != overutilized) {
8177 env->dst_rq->rd->overutilized = overutilized;
8178 trace_sched_overutilized(overutilized);
8181 if (!env->dst_rq->rd->overutilized && overutilized) {
8182 env->dst_rq->rd->overutilized = true;
8183 trace_sched_overutilized(true);
8190 * check_asym_packing - Check to see if the group is packed into the
8193 * This is primarily intended to used at the sibling level. Some
8194 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8195 * case of POWER7, it can move to lower SMT modes only when higher
8196 * threads are idle. When in lower SMT modes, the threads will
8197 * perform better since they share less core resources. Hence when we
8198 * have idle threads, we want them to be the higher ones.
8200 * This packing function is run on idle threads. It checks to see if
8201 * the busiest CPU in this domain (core in the P7 case) has a higher
8202 * CPU number than the packing function is being run on. Here we are
8203 * assuming lower CPU number will be equivalent to lower a SMT thread
8206 * Return: 1 when packing is required and a task should be moved to
8207 * this CPU. The amount of the imbalance is returned in *imbalance.
8209 * @env: The load balancing environment.
8210 * @sds: Statistics of the sched_domain which is to be packed
8212 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8216 if (!(env->sd->flags & SD_ASYM_PACKING))
8222 busiest_cpu = group_first_cpu(sds->busiest);
8223 if (env->dst_cpu > busiest_cpu)
8226 env->imbalance = DIV_ROUND_CLOSEST(
8227 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8228 SCHED_CAPACITY_SCALE);
8234 * fix_small_imbalance - Calculate the minor imbalance that exists
8235 * amongst the groups of a sched_domain, during
8237 * @env: The load balancing environment.
8238 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8241 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8243 unsigned long tmp, capa_now = 0, capa_move = 0;
8244 unsigned int imbn = 2;
8245 unsigned long scaled_busy_load_per_task;
8246 struct sg_lb_stats *local, *busiest;
8248 local = &sds->local_stat;
8249 busiest = &sds->busiest_stat;
8251 if (!local->sum_nr_running)
8252 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8253 else if (busiest->load_per_task > local->load_per_task)
8256 scaled_busy_load_per_task =
8257 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8258 busiest->group_capacity;
8260 if (busiest->avg_load + scaled_busy_load_per_task >=
8261 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8262 env->imbalance = busiest->load_per_task;
8267 * OK, we don't have enough imbalance to justify moving tasks,
8268 * however we may be able to increase total CPU capacity used by
8272 capa_now += busiest->group_capacity *
8273 min(busiest->load_per_task, busiest->avg_load);
8274 capa_now += local->group_capacity *
8275 min(local->load_per_task, local->avg_load);
8276 capa_now /= SCHED_CAPACITY_SCALE;
8278 /* Amount of load we'd subtract */
8279 if (busiest->avg_load > scaled_busy_load_per_task) {
8280 capa_move += busiest->group_capacity *
8281 min(busiest->load_per_task,
8282 busiest->avg_load - scaled_busy_load_per_task);
8285 /* Amount of load we'd add */
8286 if (busiest->avg_load * busiest->group_capacity <
8287 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8288 tmp = (busiest->avg_load * busiest->group_capacity) /
8289 local->group_capacity;
8291 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8292 local->group_capacity;
8294 capa_move += local->group_capacity *
8295 min(local->load_per_task, local->avg_load + tmp);
8296 capa_move /= SCHED_CAPACITY_SCALE;
8298 /* Move if we gain throughput */
8299 if (capa_move > capa_now)
8300 env->imbalance = busiest->load_per_task;
8304 * calculate_imbalance - Calculate the amount of imbalance present within the
8305 * groups of a given sched_domain during load balance.
8306 * @env: load balance environment
8307 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8309 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8311 unsigned long max_pull, load_above_capacity = ~0UL;
8312 struct sg_lb_stats *local, *busiest;
8314 local = &sds->local_stat;
8315 busiest = &sds->busiest_stat;
8317 if (busiest->group_type == group_imbalanced) {
8319 * In the group_imb case we cannot rely on group-wide averages
8320 * to ensure cpu-load equilibrium, look at wider averages. XXX
8322 busiest->load_per_task =
8323 min(busiest->load_per_task, sds->avg_load);
8327 * In the presence of smp nice balancing, certain scenarios can have
8328 * max load less than avg load(as we skip the groups at or below
8329 * its cpu_capacity, while calculating max_load..)
8331 if (busiest->avg_load <= sds->avg_load ||
8332 local->avg_load >= sds->avg_load) {
8333 /* Misfitting tasks should be migrated in any case */
8334 if (busiest->group_type == group_misfit_task) {
8335 env->imbalance = busiest->group_misfit_task;
8340 * Busiest group is overloaded, local is not, use the spare
8341 * cycles to maximize throughput
8343 if (busiest->group_type == group_overloaded &&
8344 local->group_type <= group_misfit_task) {
8345 env->imbalance = busiest->load_per_task;
8350 return fix_small_imbalance(env, sds);
8354 * If there aren't any idle cpus, avoid creating some.
8356 if (busiest->group_type == group_overloaded &&
8357 local->group_type == group_overloaded) {
8358 load_above_capacity = busiest->sum_nr_running *
8360 if (load_above_capacity > busiest->group_capacity)
8361 load_above_capacity -= busiest->group_capacity;
8363 load_above_capacity = ~0UL;
8367 * We're trying to get all the cpus to the average_load, so we don't
8368 * want to push ourselves above the average load, nor do we wish to
8369 * reduce the max loaded cpu below the average load. At the same time,
8370 * we also don't want to reduce the group load below the group capacity
8371 * (so that we can implement power-savings policies etc). Thus we look
8372 * for the minimum possible imbalance.
8374 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8376 /* How much load to actually move to equalise the imbalance */
8377 env->imbalance = min(
8378 max_pull * busiest->group_capacity,
8379 (sds->avg_load - local->avg_load) * local->group_capacity
8380 ) / SCHED_CAPACITY_SCALE;
8382 /* Boost imbalance to allow misfit task to be balanced. */
8383 if (busiest->group_type == group_misfit_task)
8384 env->imbalance = max_t(long, env->imbalance,
8385 busiest->group_misfit_task);
8388 * if *imbalance is less than the average load per runnable task
8389 * there is no guarantee that any tasks will be moved so we'll have
8390 * a think about bumping its value to force at least one task to be
8393 if (env->imbalance < busiest->load_per_task)
8394 return fix_small_imbalance(env, sds);
8397 /******* find_busiest_group() helpers end here *********************/
8400 * find_busiest_group - Returns the busiest group within the sched_domain
8401 * if there is an imbalance. If there isn't an imbalance, and
8402 * the user has opted for power-savings, it returns a group whose
8403 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8404 * such a group exists.
8406 * Also calculates the amount of weighted load which should be moved
8407 * to restore balance.
8409 * @env: The load balancing environment.
8411 * Return: - The busiest group if imbalance exists.
8412 * - If no imbalance and user has opted for power-savings balance,
8413 * return the least loaded group whose CPUs can be
8414 * put to idle by rebalancing its tasks onto our group.
8416 static struct sched_group *find_busiest_group(struct lb_env *env)
8418 struct sg_lb_stats *local, *busiest;
8419 struct sd_lb_stats sds;
8421 init_sd_lb_stats(&sds);
8424 * Compute the various statistics relavent for load balancing at
8427 update_sd_lb_stats(env, &sds);
8429 if (energy_aware() && !env->dst_rq->rd->overutilized)
8432 local = &sds.local_stat;
8433 busiest = &sds.busiest_stat;
8435 /* ASYM feature bypasses nice load balance check */
8436 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8437 check_asym_packing(env, &sds))
8440 /* There is no busy sibling group to pull tasks from */
8441 if (!sds.busiest || busiest->sum_nr_running == 0)
8444 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8445 / sds.total_capacity;
8448 * If the busiest group is imbalanced the below checks don't
8449 * work because they assume all things are equal, which typically
8450 * isn't true due to cpus_allowed constraints and the like.
8452 if (busiest->group_type == group_imbalanced)
8455 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
8456 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
8457 busiest->group_no_capacity)
8460 /* Misfitting tasks should be dealt with regardless of the avg load */
8461 if (busiest->group_type == group_misfit_task) {
8466 * If the local group is busier than the selected busiest group
8467 * don't try and pull any tasks.
8469 if (local->avg_load >= busiest->avg_load)
8473 * Don't pull any tasks if this group is already above the domain
8476 if (local->avg_load >= sds.avg_load)
8479 if (env->idle == CPU_IDLE) {
8481 * This cpu is idle. If the busiest group is not overloaded
8482 * and there is no imbalance between this and busiest group
8483 * wrt idle cpus, it is balanced. The imbalance becomes
8484 * significant if the diff is greater than 1 otherwise we
8485 * might end up to just move the imbalance on another group
8487 if ((busiest->group_type != group_overloaded) &&
8488 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8489 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8493 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8494 * imbalance_pct to be conservative.
8496 if (100 * busiest->avg_load <=
8497 env->sd->imbalance_pct * local->avg_load)
8502 env->busiest_group_type = busiest->group_type;
8503 /* Looks like there is an imbalance. Compute it */
8504 calculate_imbalance(env, &sds);
8513 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8515 static struct rq *find_busiest_queue(struct lb_env *env,
8516 struct sched_group *group)
8518 struct rq *busiest = NULL, *rq;
8519 unsigned long busiest_load = 0, busiest_capacity = 1;
8522 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8523 unsigned long capacity, wl;
8527 rt = fbq_classify_rq(rq);
8530 * We classify groups/runqueues into three groups:
8531 * - regular: there are !numa tasks
8532 * - remote: there are numa tasks that run on the 'wrong' node
8533 * - all: there is no distinction
8535 * In order to avoid migrating ideally placed numa tasks,
8536 * ignore those when there's better options.
8538 * If we ignore the actual busiest queue to migrate another
8539 * task, the next balance pass can still reduce the busiest
8540 * queue by moving tasks around inside the node.
8542 * If we cannot move enough load due to this classification
8543 * the next pass will adjust the group classification and
8544 * allow migration of more tasks.
8546 * Both cases only affect the total convergence complexity.
8548 if (rt > env->fbq_type)
8551 capacity = capacity_of(i);
8553 wl = weighted_cpuload(i);
8556 * When comparing with imbalance, use weighted_cpuload()
8557 * which is not scaled with the cpu capacity.
8560 if (rq->nr_running == 1 && wl > env->imbalance &&
8561 !check_cpu_capacity(rq, env->sd) &&
8562 env->busiest_group_type != group_misfit_task)
8566 * For the load comparisons with the other cpu's, consider
8567 * the weighted_cpuload() scaled with the cpu capacity, so
8568 * that the load can be moved away from the cpu that is
8569 * potentially running at a lower capacity.
8571 * Thus we're looking for max(wl_i / capacity_i), crosswise
8572 * multiplication to rid ourselves of the division works out
8573 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8574 * our previous maximum.
8576 if (wl * busiest_capacity > busiest_load * capacity) {
8578 busiest_capacity = capacity;
8587 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8588 * so long as it is large enough.
8590 #define MAX_PINNED_INTERVAL 512
8592 /* Working cpumask for load_balance and load_balance_newidle. */
8593 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8595 static int need_active_balance(struct lb_env *env)
8597 struct sched_domain *sd = env->sd;
8599 if (env->idle == CPU_NEWLY_IDLE) {
8602 * ASYM_PACKING needs to force migrate tasks from busy but
8603 * higher numbered CPUs in order to pack all tasks in the
8604 * lowest numbered CPUs.
8606 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8611 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8612 * It's worth migrating the task if the src_cpu's capacity is reduced
8613 * because of other sched_class or IRQs if more capacity stays
8614 * available on dst_cpu.
8616 if ((env->idle != CPU_NOT_IDLE) &&
8617 (env->src_rq->cfs.h_nr_running == 1)) {
8618 if ((check_cpu_capacity(env->src_rq, sd)) &&
8619 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8623 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8624 env->src_rq->cfs.h_nr_running == 1 &&
8625 cpu_overutilized(env->src_cpu) &&
8626 !cpu_overutilized(env->dst_cpu)) {
8630 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8633 static int active_load_balance_cpu_stop(void *data);
8635 static int should_we_balance(struct lb_env *env)
8637 struct sched_group *sg = env->sd->groups;
8638 struct cpumask *sg_cpus, *sg_mask;
8639 int cpu, balance_cpu = -1;
8642 * In the newly idle case, we will allow all the cpu's
8643 * to do the newly idle load balance.
8645 if (env->idle == CPU_NEWLY_IDLE)
8648 sg_cpus = sched_group_cpus(sg);
8649 sg_mask = sched_group_mask(sg);
8650 /* Try to find first idle cpu */
8651 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8652 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8659 if (balance_cpu == -1)
8660 balance_cpu = group_balance_cpu(sg);
8663 * First idle cpu or the first cpu(busiest) in this sched group
8664 * is eligible for doing load balancing at this and above domains.
8666 return balance_cpu == env->dst_cpu;
8670 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8671 * tasks if there is an imbalance.
8673 static int load_balance(int this_cpu, struct rq *this_rq,
8674 struct sched_domain *sd, enum cpu_idle_type idle,
8675 int *continue_balancing)
8677 int ld_moved, cur_ld_moved, active_balance = 0;
8678 struct sched_domain *sd_parent = lb_sd_parent(sd) ? sd->parent : NULL;
8679 struct sched_group *group;
8681 unsigned long flags;
8682 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8684 struct lb_env env = {
8686 .dst_cpu = this_cpu,
8688 .dst_grpmask = sched_group_cpus(sd->groups),
8690 .loop_break = sched_nr_migrate_break,
8693 .tasks = LIST_HEAD_INIT(env.tasks),
8697 * For NEWLY_IDLE load_balancing, we don't need to consider
8698 * other cpus in our group
8700 if (idle == CPU_NEWLY_IDLE)
8701 env.dst_grpmask = NULL;
8703 cpumask_copy(cpus, cpu_active_mask);
8705 schedstat_inc(sd, lb_count[idle]);
8708 if (!should_we_balance(&env)) {
8709 *continue_balancing = 0;
8713 group = find_busiest_group(&env);
8715 schedstat_inc(sd, lb_nobusyg[idle]);
8719 busiest = find_busiest_queue(&env, group);
8721 schedstat_inc(sd, lb_nobusyq[idle]);
8725 BUG_ON(busiest == env.dst_rq);
8727 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8729 env.src_cpu = busiest->cpu;
8730 env.src_rq = busiest;
8733 if (busiest->nr_running > 1) {
8735 * Attempt to move tasks. If find_busiest_group has found
8736 * an imbalance but busiest->nr_running <= 1, the group is
8737 * still unbalanced. ld_moved simply stays zero, so it is
8738 * correctly treated as an imbalance.
8740 env.flags |= LBF_ALL_PINNED;
8741 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8744 raw_spin_lock_irqsave(&busiest->lock, flags);
8747 * cur_ld_moved - load moved in current iteration
8748 * ld_moved - cumulative load moved across iterations
8750 cur_ld_moved = detach_tasks(&env);
8752 * We want to potentially lower env.src_cpu's OPP.
8755 update_capacity_of(env.src_cpu);
8758 * We've detached some tasks from busiest_rq. Every
8759 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8760 * unlock busiest->lock, and we are able to be sure
8761 * that nobody can manipulate the tasks in parallel.
8762 * See task_rq_lock() family for the details.
8765 raw_spin_unlock(&busiest->lock);
8769 ld_moved += cur_ld_moved;
8772 local_irq_restore(flags);
8774 if (env.flags & LBF_NEED_BREAK) {
8775 env.flags &= ~LBF_NEED_BREAK;
8780 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8781 * us and move them to an alternate dst_cpu in our sched_group
8782 * where they can run. The upper limit on how many times we
8783 * iterate on same src_cpu is dependent on number of cpus in our
8786 * This changes load balance semantics a bit on who can move
8787 * load to a given_cpu. In addition to the given_cpu itself
8788 * (or a ilb_cpu acting on its behalf where given_cpu is
8789 * nohz-idle), we now have balance_cpu in a position to move
8790 * load to given_cpu. In rare situations, this may cause
8791 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8792 * _independently_ and at _same_ time to move some load to
8793 * given_cpu) causing exceess load to be moved to given_cpu.
8794 * This however should not happen so much in practice and
8795 * moreover subsequent load balance cycles should correct the
8796 * excess load moved.
8798 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8800 /* Prevent to re-select dst_cpu via env's cpus */
8801 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8803 env.dst_rq = cpu_rq(env.new_dst_cpu);
8804 env.dst_cpu = env.new_dst_cpu;
8805 env.flags &= ~LBF_DST_PINNED;
8807 env.loop_break = sched_nr_migrate_break;
8810 * Go back to "more_balance" rather than "redo" since we
8811 * need to continue with same src_cpu.
8817 * We failed to reach balance because of affinity.
8820 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8822 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8823 *group_imbalance = 1;
8826 /* All tasks on this runqueue were pinned by CPU affinity */
8827 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8828 cpumask_clear_cpu(cpu_of(busiest), cpus);
8829 if (!cpumask_empty(cpus)) {
8831 env.loop_break = sched_nr_migrate_break;
8834 goto out_all_pinned;
8839 schedstat_inc(sd, lb_failed[idle]);
8841 * Increment the failure counter only on periodic balance.
8842 * We do not want newidle balance, which can be very
8843 * frequent, pollute the failure counter causing
8844 * excessive cache_hot migrations and active balances.
8846 if (idle != CPU_NEWLY_IDLE)
8847 if (env.src_grp_nr_running > 1)
8848 sd->nr_balance_failed++;
8850 if (need_active_balance(&env)) {
8851 raw_spin_lock_irqsave(&busiest->lock, flags);
8853 /* don't kick the active_load_balance_cpu_stop,
8854 * if the curr task on busiest cpu can't be
8857 if (!cpumask_test_cpu(this_cpu,
8858 tsk_cpus_allowed(busiest->curr))) {
8859 raw_spin_unlock_irqrestore(&busiest->lock,
8861 env.flags |= LBF_ALL_PINNED;
8862 goto out_one_pinned;
8866 * ->active_balance synchronizes accesses to
8867 * ->active_balance_work. Once set, it's cleared
8868 * only after active load balance is finished.
8870 if (!busiest->active_balance) {
8871 busiest->active_balance = 1;
8872 busiest->push_cpu = this_cpu;
8875 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8877 if (active_balance) {
8878 stop_one_cpu_nowait(cpu_of(busiest),
8879 active_load_balance_cpu_stop, busiest,
8880 &busiest->active_balance_work);
8884 * We've kicked active balancing, reset the failure
8887 sd->nr_balance_failed = sd->cache_nice_tries+1;
8890 sd->nr_balance_failed = 0;
8892 if (likely(!active_balance)) {
8893 /* We were unbalanced, so reset the balancing interval */
8894 sd->balance_interval = sd->min_interval;
8897 * If we've begun active balancing, start to back off. This
8898 * case may not be covered by the all_pinned logic if there
8899 * is only 1 task on the busy runqueue (because we don't call
8902 if (sd->balance_interval < sd->max_interval)
8903 sd->balance_interval *= 2;
8910 * We reach balance although we may have faced some affinity
8911 * constraints. Clear the imbalance flag if it was set.
8914 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8916 if (*group_imbalance)
8917 *group_imbalance = 0;
8922 * We reach balance because all tasks are pinned at this level so
8923 * we can't migrate them. Let the imbalance flag set so parent level
8924 * can try to migrate them.
8926 schedstat_inc(sd, lb_balanced[idle]);
8928 sd->nr_balance_failed = 0;
8931 /* tune up the balancing interval */
8932 if (((env.flags & LBF_ALL_PINNED) &&
8933 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8934 (sd->balance_interval < sd->max_interval))
8935 sd->balance_interval *= 2;
8942 static inline unsigned long
8943 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8945 unsigned long interval = sd->balance_interval;
8948 interval *= sd->busy_factor;
8950 /* scale ms to jiffies */
8951 interval = msecs_to_jiffies(interval);
8952 interval = clamp(interval, 1UL, max_load_balance_interval);
8958 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8960 unsigned long interval, next;
8962 interval = get_sd_balance_interval(sd, cpu_busy);
8963 next = sd->last_balance + interval;
8965 if (time_after(*next_balance, next))
8966 *next_balance = next;
8970 * idle_balance is called by schedule() if this_cpu is about to become
8971 * idle. Attempts to pull tasks from other CPUs.
8973 static int idle_balance(struct rq *this_rq)
8975 unsigned long next_balance = jiffies + HZ;
8976 int this_cpu = this_rq->cpu;
8977 struct sched_domain *sd;
8978 int pulled_task = 0;
8980 long removed_util=0;
8982 idle_enter_fair(this_rq);
8985 * We must set idle_stamp _before_ calling idle_balance(), such that we
8986 * measure the duration of idle_balance() as idle time.
8988 this_rq->idle_stamp = rq_clock(this_rq);
8990 if (!energy_aware() &&
8991 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8992 !this_rq->rd->overload)) {
8994 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8996 update_next_balance(sd, 0, &next_balance);
9002 raw_spin_unlock(&this_rq->lock);
9005 * If removed_util_avg is !0 we most probably migrated some task away
9006 * from this_cpu. In this case we might be willing to trigger an OPP
9007 * update, but we want to do so if we don't find anybody else to pull
9008 * here (we will trigger an OPP update with the pulled task's enqueue
9011 * Record removed_util before calling update_blocked_averages, and use
9012 * it below (before returning) to see if an OPP update is required.
9014 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
9015 update_blocked_averages(this_cpu);
9017 for_each_domain(this_cpu, sd) {
9018 int continue_balancing = 1;
9019 u64 t0, domain_cost;
9021 if (!(sd->flags & SD_LOAD_BALANCE))
9024 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9025 update_next_balance(sd, 0, &next_balance);
9029 if (sd->flags & SD_BALANCE_NEWIDLE) {
9030 t0 = sched_clock_cpu(this_cpu);
9032 pulled_task = load_balance(this_cpu, this_rq,
9034 &continue_balancing);
9036 domain_cost = sched_clock_cpu(this_cpu) - t0;
9037 if (domain_cost > sd->max_newidle_lb_cost)
9038 sd->max_newidle_lb_cost = domain_cost;
9040 curr_cost += domain_cost;
9043 update_next_balance(sd, 0, &next_balance);
9046 * Stop searching for tasks to pull if there are
9047 * now runnable tasks on this rq.
9049 if (pulled_task || this_rq->nr_running > 0)
9054 raw_spin_lock(&this_rq->lock);
9056 if (curr_cost > this_rq->max_idle_balance_cost)
9057 this_rq->max_idle_balance_cost = curr_cost;
9060 * While browsing the domains, we released the rq lock, a task could
9061 * have been enqueued in the meantime. Since we're not going idle,
9062 * pretend we pulled a task.
9064 if (this_rq->cfs.h_nr_running && !pulled_task)
9068 /* Move the next balance forward */
9069 if (time_after(this_rq->next_balance, next_balance))
9070 this_rq->next_balance = next_balance;
9072 /* Is there a task of a high priority class? */
9073 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9077 idle_exit_fair(this_rq);
9078 this_rq->idle_stamp = 0;
9079 } else if (removed_util) {
9081 * No task pulled and someone has been migrated away.
9082 * Good case to trigger an OPP update.
9084 update_capacity_of(this_cpu);
9091 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
9092 * running tasks off the busiest CPU onto idle CPUs. It requires at
9093 * least 1 task to be running on each physical CPU where possible, and
9094 * avoids physical / logical imbalances.
9096 static int active_load_balance_cpu_stop(void *data)
9098 struct rq *busiest_rq = data;
9099 int busiest_cpu = cpu_of(busiest_rq);
9100 int target_cpu = busiest_rq->push_cpu;
9101 struct rq *target_rq = cpu_rq(target_cpu);
9102 struct sched_domain *sd;
9103 struct task_struct *p = NULL;
9105 raw_spin_lock_irq(&busiest_rq->lock);
9107 /* make sure the requested cpu hasn't gone down in the meantime */
9108 if (unlikely(busiest_cpu != smp_processor_id() ||
9109 !busiest_rq->active_balance))
9112 /* Is there any task to move? */
9113 if (busiest_rq->nr_running <= 1)
9117 * This condition is "impossible", if it occurs
9118 * we need to fix it. Originally reported by
9119 * Bjorn Helgaas on a 128-cpu setup.
9121 BUG_ON(busiest_rq == target_rq);
9123 /* Search for an sd spanning us and the target CPU. */
9125 for_each_domain(target_cpu, sd) {
9126 if ((sd->flags & SD_LOAD_BALANCE) &&
9127 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9132 struct lb_env env = {
9134 .dst_cpu = target_cpu,
9135 .dst_rq = target_rq,
9136 .src_cpu = busiest_rq->cpu,
9137 .src_rq = busiest_rq,
9141 schedstat_inc(sd, alb_count);
9143 p = detach_one_task(&env);
9145 schedstat_inc(sd, alb_pushed);
9147 * We want to potentially lower env.src_cpu's OPP.
9149 update_capacity_of(env.src_cpu);
9152 schedstat_inc(sd, alb_failed);
9156 busiest_rq->active_balance = 0;
9157 raw_spin_unlock(&busiest_rq->lock);
9160 attach_one_task(target_rq, p);
9167 static inline int on_null_domain(struct rq *rq)
9169 return unlikely(!rcu_dereference_sched(rq->sd));
9172 #ifdef CONFIG_NO_HZ_COMMON
9174 * idle load balancing details
9175 * - When one of the busy CPUs notice that there may be an idle rebalancing
9176 * needed, they will kick the idle load balancer, which then does idle
9177 * load balancing for all the idle CPUs.
9180 cpumask_var_t idle_cpus_mask;
9182 unsigned long next_balance; /* in jiffy units */
9183 } nohz ____cacheline_aligned;
9185 static inline int find_new_ilb(void)
9187 int ilb = cpumask_first(nohz.idle_cpus_mask);
9189 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9196 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9197 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9198 * CPU (if there is one).
9200 static void nohz_balancer_kick(void)
9204 nohz.next_balance++;
9206 ilb_cpu = find_new_ilb();
9208 if (ilb_cpu >= nr_cpu_ids)
9211 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9214 * Use smp_send_reschedule() instead of resched_cpu().
9215 * This way we generate a sched IPI on the target cpu which
9216 * is idle. And the softirq performing nohz idle load balance
9217 * will be run before returning from the IPI.
9219 smp_send_reschedule(ilb_cpu);
9223 static inline void nohz_balance_exit_idle(int cpu)
9225 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9227 * Completely isolated CPUs don't ever set, so we must test.
9229 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9230 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9231 atomic_dec(&nohz.nr_cpus);
9233 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9237 static inline void set_cpu_sd_state_busy(void)
9239 struct sched_domain *sd;
9240 int cpu = smp_processor_id();
9243 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9245 if (!sd || !sd->nohz_idle)
9249 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
9254 void set_cpu_sd_state_idle(void)
9256 struct sched_domain *sd;
9257 int cpu = smp_processor_id();
9260 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9262 if (!sd || sd->nohz_idle)
9266 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
9272 * This routine will record that the cpu is going idle with tick stopped.
9273 * This info will be used in performing idle load balancing in the future.
9275 void nohz_balance_enter_idle(int cpu)
9278 * If this cpu is going down, then nothing needs to be done.
9280 if (!cpu_active(cpu))
9283 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9287 * If we're a completely isolated CPU, we don't play.
9289 if (on_null_domain(cpu_rq(cpu)))
9292 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9293 atomic_inc(&nohz.nr_cpus);
9294 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9297 static int sched_ilb_notifier(struct notifier_block *nfb,
9298 unsigned long action, void *hcpu)
9300 switch (action & ~CPU_TASKS_FROZEN) {
9302 nohz_balance_exit_idle(smp_processor_id());
9310 static DEFINE_SPINLOCK(balancing);
9313 * Scale the max load_balance interval with the number of CPUs in the system.
9314 * This trades load-balance latency on larger machines for less cross talk.
9316 void update_max_interval(void)
9318 max_load_balance_interval = HZ*num_online_cpus()/10;
9322 * It checks each scheduling domain to see if it is due to be balanced,
9323 * and initiates a balancing operation if so.
9325 * Balancing parameters are set up in init_sched_domains.
9327 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9329 int continue_balancing = 1;
9331 unsigned long interval;
9332 struct sched_domain *sd;
9333 /* Earliest time when we have to do rebalance again */
9334 unsigned long next_balance = jiffies + 60*HZ;
9335 int update_next_balance = 0;
9336 int need_serialize, need_decay = 0;
9339 update_blocked_averages(cpu);
9342 for_each_domain(cpu, sd) {
9344 * Decay the newidle max times here because this is a regular
9345 * visit to all the domains. Decay ~1% per second.
9347 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9348 sd->max_newidle_lb_cost =
9349 (sd->max_newidle_lb_cost * 253) / 256;
9350 sd->next_decay_max_lb_cost = jiffies + HZ;
9353 max_cost += sd->max_newidle_lb_cost;
9355 if (!(sd->flags & SD_LOAD_BALANCE))
9359 * Stop the load balance at this level. There is another
9360 * CPU in our sched group which is doing load balancing more
9363 if (!continue_balancing) {
9369 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9371 need_serialize = sd->flags & SD_SERIALIZE;
9372 if (need_serialize) {
9373 if (!spin_trylock(&balancing))
9377 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9378 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9380 * The LBF_DST_PINNED logic could have changed
9381 * env->dst_cpu, so we can't know our idle
9382 * state even if we migrated tasks. Update it.
9384 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9386 sd->last_balance = jiffies;
9387 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9390 spin_unlock(&balancing);
9392 if (time_after(next_balance, sd->last_balance + interval)) {
9393 next_balance = sd->last_balance + interval;
9394 update_next_balance = 1;
9399 * Ensure the rq-wide value also decays but keep it at a
9400 * reasonable floor to avoid funnies with rq->avg_idle.
9402 rq->max_idle_balance_cost =
9403 max((u64)sysctl_sched_migration_cost, max_cost);
9408 * next_balance will be updated only when there is a need.
9409 * When the cpu is attached to null domain for ex, it will not be
9412 if (likely(update_next_balance)) {
9413 rq->next_balance = next_balance;
9415 #ifdef CONFIG_NO_HZ_COMMON
9417 * If this CPU has been elected to perform the nohz idle
9418 * balance. Other idle CPUs have already rebalanced with
9419 * nohz_idle_balance() and nohz.next_balance has been
9420 * updated accordingly. This CPU is now running the idle load
9421 * balance for itself and we need to update the
9422 * nohz.next_balance accordingly.
9424 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9425 nohz.next_balance = rq->next_balance;
9430 #ifdef CONFIG_NO_HZ_COMMON
9432 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9433 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9435 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9437 int this_cpu = this_rq->cpu;
9440 /* Earliest time when we have to do rebalance again */
9441 unsigned long next_balance = jiffies + 60*HZ;
9442 int update_next_balance = 0;
9444 if (idle != CPU_IDLE ||
9445 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9448 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9449 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9453 * If this cpu gets work to do, stop the load balancing
9454 * work being done for other cpus. Next load
9455 * balancing owner will pick it up.
9460 rq = cpu_rq(balance_cpu);
9463 * If time for next balance is due,
9466 if (time_after_eq(jiffies, rq->next_balance)) {
9467 raw_spin_lock_irq(&rq->lock);
9468 update_rq_clock(rq);
9469 update_idle_cpu_load(rq);
9470 raw_spin_unlock_irq(&rq->lock);
9471 rebalance_domains(rq, CPU_IDLE);
9474 if (time_after(next_balance, rq->next_balance)) {
9475 next_balance = rq->next_balance;
9476 update_next_balance = 1;
9481 * next_balance will be updated only when there is a need.
9482 * When the CPU is attached to null domain for ex, it will not be
9485 if (likely(update_next_balance))
9486 nohz.next_balance = next_balance;
9488 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9492 * Current heuristic for kicking the idle load balancer in the presence
9493 * of an idle cpu in the system.
9494 * - This rq has more than one task.
9495 * - This rq has at least one CFS task and the capacity of the CPU is
9496 * significantly reduced because of RT tasks or IRQs.
9497 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9498 * multiple busy cpu.
9499 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9500 * domain span are idle.
9502 static inline bool nohz_kick_needed(struct rq *rq)
9504 unsigned long now = jiffies;
9505 struct sched_domain *sd;
9506 struct sched_group_capacity *sgc;
9507 int nr_busy, cpu = rq->cpu;
9510 if (unlikely(rq->idle_balance))
9514 * We may be recently in ticked or tickless idle mode. At the first
9515 * busy tick after returning from idle, we will update the busy stats.
9517 set_cpu_sd_state_busy();
9518 nohz_balance_exit_idle(cpu);
9521 * None are in tickless mode and hence no need for NOHZ idle load
9524 if (likely(!atomic_read(&nohz.nr_cpus)))
9527 if (time_before(now, nohz.next_balance))
9530 if (rq->nr_running >= 2 &&
9531 (!energy_aware() || cpu_overutilized(cpu)))
9535 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9536 if (sd && !energy_aware()) {
9537 sgc = sd->groups->sgc;
9538 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9547 sd = rcu_dereference(rq->sd);
9549 if ((rq->cfs.h_nr_running >= 1) &&
9550 check_cpu_capacity(rq, sd)) {
9556 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9557 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9558 sched_domain_span(sd)) < cpu)) {
9568 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9572 * run_rebalance_domains is triggered when needed from the scheduler tick.
9573 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9575 static void run_rebalance_domains(struct softirq_action *h)
9577 struct rq *this_rq = this_rq();
9578 enum cpu_idle_type idle = this_rq->idle_balance ?
9579 CPU_IDLE : CPU_NOT_IDLE;
9582 * If this cpu has a pending nohz_balance_kick, then do the
9583 * balancing on behalf of the other idle cpus whose ticks are
9584 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9585 * give the idle cpus a chance to load balance. Else we may
9586 * load balance only within the local sched_domain hierarchy
9587 * and abort nohz_idle_balance altogether if we pull some load.
9589 nohz_idle_balance(this_rq, idle);
9590 rebalance_domains(this_rq, idle);
9594 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9596 void trigger_load_balance(struct rq *rq)
9598 /* Don't need to rebalance while attached to NULL domain */
9599 if (unlikely(on_null_domain(rq)))
9602 if (time_after_eq(jiffies, rq->next_balance))
9603 raise_softirq(SCHED_SOFTIRQ);
9604 #ifdef CONFIG_NO_HZ_COMMON
9605 if (nohz_kick_needed(rq))
9606 nohz_balancer_kick();
9610 static void rq_online_fair(struct rq *rq)
9614 update_runtime_enabled(rq);
9617 static void rq_offline_fair(struct rq *rq)
9621 /* Ensure any throttled groups are reachable by pick_next_task */
9622 unthrottle_offline_cfs_rqs(rq);
9625 #endif /* CONFIG_SMP */
9628 * scheduler tick hitting a task of our scheduling class:
9630 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9632 struct cfs_rq *cfs_rq;
9633 struct sched_entity *se = &curr->se;
9635 for_each_sched_entity(se) {
9636 cfs_rq = cfs_rq_of(se);
9637 entity_tick(cfs_rq, se, queued);
9640 if (static_branch_unlikely(&sched_numa_balancing))
9641 task_tick_numa(rq, curr);
9644 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9645 rq->rd->overutilized = true;
9646 trace_sched_overutilized(true);
9649 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9655 * called on fork with the child task as argument from the parent's context
9656 * - child not yet on the tasklist
9657 * - preemption disabled
9659 static void task_fork_fair(struct task_struct *p)
9661 struct cfs_rq *cfs_rq;
9662 struct sched_entity *se = &p->se, *curr;
9663 int this_cpu = smp_processor_id();
9664 struct rq *rq = this_rq();
9665 unsigned long flags;
9667 raw_spin_lock_irqsave(&rq->lock, flags);
9669 update_rq_clock(rq);
9671 cfs_rq = task_cfs_rq(current);
9672 curr = cfs_rq->curr;
9675 * Not only the cpu but also the task_group of the parent might have
9676 * been changed after parent->se.parent,cfs_rq were copied to
9677 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9678 * of child point to valid ones.
9681 __set_task_cpu(p, this_cpu);
9684 update_curr(cfs_rq);
9687 se->vruntime = curr->vruntime;
9688 place_entity(cfs_rq, se, 1);
9690 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9692 * Upon rescheduling, sched_class::put_prev_task() will place
9693 * 'current' within the tree based on its new key value.
9695 swap(curr->vruntime, se->vruntime);
9699 se->vruntime -= cfs_rq->min_vruntime;
9701 raw_spin_unlock_irqrestore(&rq->lock, flags);
9705 * Priority of the task has changed. Check to see if we preempt
9709 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9711 if (!task_on_rq_queued(p))
9715 * Reschedule if we are currently running on this runqueue and
9716 * our priority decreased, or if we are not currently running on
9717 * this runqueue and our priority is higher than the current's
9719 if (rq->curr == p) {
9720 if (p->prio > oldprio)
9723 check_preempt_curr(rq, p, 0);
9726 static inline bool vruntime_normalized(struct task_struct *p)
9728 struct sched_entity *se = &p->se;
9731 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9732 * the dequeue_entity(.flags=0) will already have normalized the
9739 * When !on_rq, vruntime of the task has usually NOT been normalized.
9740 * But there are some cases where it has already been normalized:
9742 * - A forked child which is waiting for being woken up by
9743 * wake_up_new_task().
9744 * - A task which has been woken up by try_to_wake_up() and
9745 * waiting for actually being woken up by sched_ttwu_pending().
9747 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9753 #ifdef CONFIG_FAIR_GROUP_SCHED
9755 * Propagate the changes of the sched_entity across the tg tree to make it
9756 * visible to the root
9758 static void propagate_entity_cfs_rq(struct sched_entity *se)
9760 struct cfs_rq *cfs_rq;
9762 /* Start to propagate at parent */
9765 for_each_sched_entity(se) {
9766 cfs_rq = cfs_rq_of(se);
9768 if (cfs_rq_throttled(cfs_rq))
9771 update_load_avg(se, UPDATE_TG);
9775 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9778 static void detach_entity_cfs_rq(struct sched_entity *se)
9780 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9782 /* Catch up with the cfs_rq and remove our load when we leave */
9783 update_load_avg(se, 0);
9784 detach_entity_load_avg(cfs_rq, se);
9785 update_tg_load_avg(cfs_rq, false);
9786 propagate_entity_cfs_rq(se);
9789 static void attach_entity_cfs_rq(struct sched_entity *se)
9791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9793 #ifdef CONFIG_FAIR_GROUP_SCHED
9795 * Since the real-depth could have been changed (only FAIR
9796 * class maintain depth value), reset depth properly.
9798 se->depth = se->parent ? se->parent->depth + 1 : 0;
9801 /* Synchronize entity with its cfs_rq */
9802 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9803 attach_entity_load_avg(cfs_rq, se);
9804 update_tg_load_avg(cfs_rq, false);
9805 propagate_entity_cfs_rq(se);
9808 static void detach_task_cfs_rq(struct task_struct *p)
9810 struct sched_entity *se = &p->se;
9811 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9813 if (!vruntime_normalized(p)) {
9815 * Fix up our vruntime so that the current sleep doesn't
9816 * cause 'unlimited' sleep bonus.
9818 place_entity(cfs_rq, se, 0);
9819 se->vruntime -= cfs_rq->min_vruntime;
9822 detach_entity_cfs_rq(se);
9825 static void attach_task_cfs_rq(struct task_struct *p)
9827 struct sched_entity *se = &p->se;
9828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9830 attach_entity_cfs_rq(se);
9832 if (!vruntime_normalized(p))
9833 se->vruntime += cfs_rq->min_vruntime;
9836 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9838 detach_task_cfs_rq(p);
9841 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9843 attach_task_cfs_rq(p);
9845 if (task_on_rq_queued(p)) {
9847 * We were most likely switched from sched_rt, so
9848 * kick off the schedule if running, otherwise just see
9849 * if we can still preempt the current task.
9854 check_preempt_curr(rq, p, 0);
9858 /* Account for a task changing its policy or group.
9860 * This routine is mostly called to set cfs_rq->curr field when a task
9861 * migrates between groups/classes.
9863 static void set_curr_task_fair(struct rq *rq)
9865 struct sched_entity *se = &rq->curr->se;
9867 for_each_sched_entity(se) {
9868 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9870 set_next_entity(cfs_rq, se);
9871 /* ensure bandwidth has been allocated on our new cfs_rq */
9872 account_cfs_rq_runtime(cfs_rq, 0);
9876 void init_cfs_rq(struct cfs_rq *cfs_rq)
9878 cfs_rq->tasks_timeline = RB_ROOT;
9879 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9880 #ifndef CONFIG_64BIT
9881 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9884 #ifdef CONFIG_FAIR_GROUP_SCHED
9885 cfs_rq->propagate_avg = 0;
9887 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9888 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9892 #ifdef CONFIG_FAIR_GROUP_SCHED
9893 static void task_move_group_fair(struct task_struct *p)
9895 detach_task_cfs_rq(p);
9896 set_task_rq(p, task_cpu(p));
9899 /* Tell se's cfs_rq has been changed -- migrated */
9900 p->se.avg.last_update_time = 0;
9902 attach_task_cfs_rq(p);
9905 void free_fair_sched_group(struct task_group *tg)
9909 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9911 for_each_possible_cpu(i) {
9913 kfree(tg->cfs_rq[i]);
9916 remove_entity_load_avg(tg->se[i]);
9925 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9927 struct sched_entity *se;
9928 struct cfs_rq *cfs_rq;
9932 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9935 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9939 tg->shares = NICE_0_LOAD;
9941 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9943 for_each_possible_cpu(i) {
9946 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9947 GFP_KERNEL, cpu_to_node(i));
9951 se = kzalloc_node(sizeof(struct sched_entity),
9952 GFP_KERNEL, cpu_to_node(i));
9956 init_cfs_rq(cfs_rq);
9957 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9958 init_entity_runnable_average(se);
9960 raw_spin_lock_irq(&rq->lock);
9961 post_init_entity_util_avg(se);
9962 raw_spin_unlock_irq(&rq->lock);
9973 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9975 struct rq *rq = cpu_rq(cpu);
9976 unsigned long flags;
9979 * Only empty task groups can be destroyed; so we can speculatively
9980 * check on_list without danger of it being re-added.
9982 if (!tg->cfs_rq[cpu]->on_list)
9985 raw_spin_lock_irqsave(&rq->lock, flags);
9986 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9987 raw_spin_unlock_irqrestore(&rq->lock, flags);
9990 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9991 struct sched_entity *se, int cpu,
9992 struct sched_entity *parent)
9994 struct rq *rq = cpu_rq(cpu);
9998 init_cfs_rq_runtime(cfs_rq);
10000 tg->cfs_rq[cpu] = cfs_rq;
10003 /* se could be NULL for root_task_group */
10008 se->cfs_rq = &rq->cfs;
10011 se->cfs_rq = parent->my_q;
10012 se->depth = parent->depth + 1;
10016 /* guarantee group entities always have weight */
10017 update_load_set(&se->load, NICE_0_LOAD);
10018 se->parent = parent;
10021 static DEFINE_MUTEX(shares_mutex);
10023 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10026 unsigned long flags;
10029 * We can't change the weight of the root cgroup.
10034 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10036 mutex_lock(&shares_mutex);
10037 if (tg->shares == shares)
10040 tg->shares = shares;
10041 for_each_possible_cpu(i) {
10042 struct rq *rq = cpu_rq(i);
10043 struct sched_entity *se;
10046 /* Propagate contribution to hierarchy */
10047 raw_spin_lock_irqsave(&rq->lock, flags);
10049 /* Possible calls to update_curr() need rq clock */
10050 update_rq_clock(rq);
10051 for_each_sched_entity(se) {
10052 update_load_avg(se, UPDATE_TG);
10053 update_cfs_shares(se);
10055 raw_spin_unlock_irqrestore(&rq->lock, flags);
10059 mutex_unlock(&shares_mutex);
10062 #else /* CONFIG_FAIR_GROUP_SCHED */
10064 void free_fair_sched_group(struct task_group *tg) { }
10066 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10071 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
10073 #endif /* CONFIG_FAIR_GROUP_SCHED */
10076 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10078 struct sched_entity *se = &task->se;
10079 unsigned int rr_interval = 0;
10082 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10085 if (rq->cfs.load.weight)
10086 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10088 return rr_interval;
10092 * All the scheduling class methods:
10094 const struct sched_class fair_sched_class = {
10095 .next = &idle_sched_class,
10096 .enqueue_task = enqueue_task_fair,
10097 .dequeue_task = dequeue_task_fair,
10098 .yield_task = yield_task_fair,
10099 .yield_to_task = yield_to_task_fair,
10101 .check_preempt_curr = check_preempt_wakeup,
10103 .pick_next_task = pick_next_task_fair,
10104 .put_prev_task = put_prev_task_fair,
10107 .select_task_rq = select_task_rq_fair,
10108 .migrate_task_rq = migrate_task_rq_fair,
10110 .rq_online = rq_online_fair,
10111 .rq_offline = rq_offline_fair,
10113 .task_waking = task_waking_fair,
10114 .task_dead = task_dead_fair,
10115 .set_cpus_allowed = set_cpus_allowed_common,
10118 .set_curr_task = set_curr_task_fair,
10119 .task_tick = task_tick_fair,
10120 .task_fork = task_fork_fair,
10122 .prio_changed = prio_changed_fair,
10123 .switched_from = switched_from_fair,
10124 .switched_to = switched_to_fair,
10126 .get_rr_interval = get_rr_interval_fair,
10128 .update_curr = update_curr_fair,
10130 #ifdef CONFIG_FAIR_GROUP_SCHED
10131 .task_move_group = task_move_group_fair,
10135 #ifdef CONFIG_SCHED_DEBUG
10136 void print_cfs_stats(struct seq_file *m, int cpu)
10138 struct cfs_rq *cfs_rq;
10141 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10142 print_cfs_rq(m, cpu, cfs_rq);
10146 #ifdef CONFIG_NUMA_BALANCING
10147 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10150 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10152 for_each_online_node(node) {
10153 if (p->numa_faults) {
10154 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10155 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10157 if (p->numa_group) {
10158 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10159 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10161 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10164 #endif /* CONFIG_NUMA_BALANCING */
10165 #endif /* CONFIG_SCHED_DEBUG */
10167 __init void init_sched_fair_class(void)
10170 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10172 #ifdef CONFIG_NO_HZ_COMMON
10173 nohz.next_balance = jiffies;
10174 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10175 cpu_notifier(sched_ilb_notifier, 0);