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_cstate_aware = 1;
59 #ifdef CONFIG_SCHED_WALT
60 unsigned int sysctl_sched_use_walt_cpu_util = 1;
61 unsigned int sysctl_sched_use_walt_task_util = 1;
62 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
66 * The initial- and re-scaling of tunables is configurable
67 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
70 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
71 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
72 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
74 enum sched_tunable_scaling sysctl_sched_tunable_scaling
75 = SCHED_TUNABLESCALING_LOG;
78 * Minimal preemption granularity for CPU-bound tasks:
79 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
81 unsigned int sysctl_sched_min_granularity = 750000ULL;
82 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
85 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
87 static unsigned int sched_nr_latency = 8;
90 * After fork, child runs first. If set to 0 (default) then
91 * parent will (try to) run first.
93 unsigned int sysctl_sched_child_runs_first __read_mostly;
96 * SCHED_OTHER wake-up granularity.
97 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
99 * This option delays the preemption effects of decoupled workloads
100 * and reduces their over-scheduling. Synchronous workloads will still
101 * have immediate wakeup/sleep latencies.
103 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
104 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
106 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
109 * The exponential sliding window over which load is averaged for shares
113 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
115 #ifdef CONFIG_CFS_BANDWIDTH
117 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
118 * each time a cfs_rq requests quota.
120 * Note: in the case that the slice exceeds the runtime remaining (either due
121 * to consumption or the quota being specified to be smaller than the slice)
122 * we will always only issue the remaining available time.
124 * default: 5 msec, units: microseconds
126 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
130 * The margin used when comparing utilization with CPU capacity:
131 * util * margin < capacity * 1024
133 unsigned int capacity_margin = 1280; /* ~20% */
135 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
141 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
147 static inline void update_load_set(struct load_weight *lw, unsigned long w)
154 * Increase the granularity value when there are more CPUs,
155 * because with more CPUs the 'effective latency' as visible
156 * to users decreases. But the relationship is not linear,
157 * so pick a second-best guess by going with the log2 of the
160 * This idea comes from the SD scheduler of Con Kolivas:
162 static unsigned int get_update_sysctl_factor(void)
164 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
167 switch (sysctl_sched_tunable_scaling) {
168 case SCHED_TUNABLESCALING_NONE:
171 case SCHED_TUNABLESCALING_LINEAR:
174 case SCHED_TUNABLESCALING_LOG:
176 factor = 1 + ilog2(cpus);
183 static void update_sysctl(void)
185 unsigned int factor = get_update_sysctl_factor();
187 #define SET_SYSCTL(name) \
188 (sysctl_##name = (factor) * normalized_sysctl_##name)
189 SET_SYSCTL(sched_min_granularity);
190 SET_SYSCTL(sched_latency);
191 SET_SYSCTL(sched_wakeup_granularity);
195 void sched_init_granularity(void)
200 #define WMULT_CONST (~0U)
201 #define WMULT_SHIFT 32
203 static void __update_inv_weight(struct load_weight *lw)
207 if (likely(lw->inv_weight))
210 w = scale_load_down(lw->weight);
212 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
214 else if (unlikely(!w))
215 lw->inv_weight = WMULT_CONST;
217 lw->inv_weight = WMULT_CONST / w;
221 * delta_exec * weight / lw.weight
223 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
225 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
226 * we're guaranteed shift stays positive because inv_weight is guaranteed to
227 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
229 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
230 * weight/lw.weight <= 1, and therefore our shift will also be positive.
232 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
234 u64 fact = scale_load_down(weight);
235 int shift = WMULT_SHIFT;
237 __update_inv_weight(lw);
239 if (unlikely(fact >> 32)) {
246 /* hint to use a 32x32->64 mul */
247 fact = (u64)(u32)fact * lw->inv_weight;
254 return mul_u64_u32_shr(delta_exec, fact, shift);
258 const struct sched_class fair_sched_class;
260 /**************************************************************
261 * CFS operations on generic schedulable entities:
264 #ifdef CONFIG_FAIR_GROUP_SCHED
266 /* cpu runqueue to which this cfs_rq is attached */
267 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
272 /* An entity is a task if it doesn't "own" a runqueue */
273 #define entity_is_task(se) (!se->my_q)
275 static inline struct task_struct *task_of(struct sched_entity *se)
277 #ifdef CONFIG_SCHED_DEBUG
278 WARN_ON_ONCE(!entity_is_task(se));
280 return container_of(se, struct task_struct, se);
283 /* Walk up scheduling entities hierarchy */
284 #define for_each_sched_entity(se) \
285 for (; se; se = se->parent)
287 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
292 /* runqueue on which this entity is (to be) queued */
293 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
298 /* runqueue "owned" by this group */
299 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
304 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
306 if (!cfs_rq->on_list) {
307 struct rq *rq = rq_of(cfs_rq);
308 int cpu = cpu_of(rq);
310 * Ensure we either appear before our parent (if already
311 * enqueued) or force our parent to appear after us when it is
312 * enqueued. The fact that we always enqueue bottom-up
313 * reduces this to two cases and a special case for the root
314 * cfs_rq. Furthermore, it also means that we will always reset
315 * tmp_alone_branch either when the branch is connected
316 * to a tree or when we reach the beg of the tree
318 if (cfs_rq->tg->parent &&
319 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
321 * If parent is already on the list, we add the child
322 * just before. Thanks to circular linked property of
323 * the list, this means to put the child at the tail
324 * of the list that starts by parent.
326 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
327 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
329 * The branch is now connected to its tree so we can
330 * reset tmp_alone_branch to the beginning of the
333 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
334 } else if (!cfs_rq->tg->parent) {
336 * cfs rq without parent should be put
337 * at the tail of the list.
339 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
340 &rq->leaf_cfs_rq_list);
342 * We have reach the beg of a tree so we can reset
343 * tmp_alone_branch to the beginning of the list.
345 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
348 * The parent has not already been added so we want to
349 * make sure that it will be put after us.
350 * tmp_alone_branch points to the beg of the branch
351 * where we will add parent.
353 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
354 rq->tmp_alone_branch);
356 * update tmp_alone_branch to points to the new beg
359 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
366 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
368 if (cfs_rq->on_list) {
369 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
374 /* Iterate thr' all leaf cfs_rq's on a runqueue */
375 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
376 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
378 /* Do the two (enqueued) entities belong to the same group ? */
379 static inline struct cfs_rq *
380 is_same_group(struct sched_entity *se, struct sched_entity *pse)
382 if (se->cfs_rq == pse->cfs_rq)
388 static inline struct sched_entity *parent_entity(struct sched_entity *se)
394 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
396 int se_depth, pse_depth;
399 * preemption test can be made between sibling entities who are in the
400 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
401 * both tasks until we find their ancestors who are siblings of common
405 /* First walk up until both entities are at same depth */
406 se_depth = (*se)->depth;
407 pse_depth = (*pse)->depth;
409 while (se_depth > pse_depth) {
411 *se = parent_entity(*se);
414 while (pse_depth > se_depth) {
416 *pse = parent_entity(*pse);
419 while (!is_same_group(*se, *pse)) {
420 *se = parent_entity(*se);
421 *pse = parent_entity(*pse);
425 #else /* !CONFIG_FAIR_GROUP_SCHED */
427 static inline struct task_struct *task_of(struct sched_entity *se)
429 return container_of(se, struct task_struct, se);
432 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
434 return container_of(cfs_rq, struct rq, cfs);
437 #define entity_is_task(se) 1
439 #define for_each_sched_entity(se) \
440 for (; se; se = NULL)
442 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
444 return &task_rq(p)->cfs;
447 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
449 struct task_struct *p = task_of(se);
450 struct rq *rq = task_rq(p);
455 /* runqueue "owned" by this group */
456 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
461 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
465 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
469 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
470 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
472 static inline struct sched_entity *parent_entity(struct sched_entity *se)
478 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
482 #endif /* CONFIG_FAIR_GROUP_SCHED */
484 static __always_inline
485 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
487 /**************************************************************
488 * Scheduling class tree data structure manipulation methods:
491 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
493 s64 delta = (s64)(vruntime - max_vruntime);
495 max_vruntime = vruntime;
500 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
502 s64 delta = (s64)(vruntime - min_vruntime);
504 min_vruntime = vruntime;
509 static inline int entity_before(struct sched_entity *a,
510 struct sched_entity *b)
512 return (s64)(a->vruntime - b->vruntime) < 0;
515 static void update_min_vruntime(struct cfs_rq *cfs_rq)
517 u64 vruntime = cfs_rq->min_vruntime;
520 vruntime = cfs_rq->curr->vruntime;
522 if (cfs_rq->rb_leftmost) {
523 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
528 vruntime = se->vruntime;
530 vruntime = min_vruntime(vruntime, se->vruntime);
533 /* ensure we never gain time by being placed backwards. */
534 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
537 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
542 * Enqueue an entity into the rb-tree:
544 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
546 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
547 struct rb_node *parent = NULL;
548 struct sched_entity *entry;
552 * Find the right place in the rbtree:
556 entry = rb_entry(parent, struct sched_entity, run_node);
558 * We dont care about collisions. Nodes with
559 * the same key stay together.
561 if (entity_before(se, entry)) {
562 link = &parent->rb_left;
564 link = &parent->rb_right;
570 * Maintain a cache of leftmost tree entries (it is frequently
574 cfs_rq->rb_leftmost = &se->run_node;
576 rb_link_node(&se->run_node, parent, link);
577 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
580 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
582 if (cfs_rq->rb_leftmost == &se->run_node) {
583 struct rb_node *next_node;
585 next_node = rb_next(&se->run_node);
586 cfs_rq->rb_leftmost = next_node;
589 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
592 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
594 struct rb_node *left = cfs_rq->rb_leftmost;
599 return rb_entry(left, struct sched_entity, run_node);
602 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
604 struct rb_node *next = rb_next(&se->run_node);
609 return rb_entry(next, struct sched_entity, run_node);
612 #ifdef CONFIG_SCHED_DEBUG
613 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
615 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
620 return rb_entry(last, struct sched_entity, run_node);
623 /**************************************************************
624 * Scheduling class statistics methods:
627 int sched_proc_update_handler(struct ctl_table *table, int write,
628 void __user *buffer, size_t *lenp,
631 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
632 unsigned int factor = get_update_sysctl_factor();
637 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
638 sysctl_sched_min_granularity);
640 #define WRT_SYSCTL(name) \
641 (normalized_sysctl_##name = sysctl_##name / (factor))
642 WRT_SYSCTL(sched_min_granularity);
643 WRT_SYSCTL(sched_latency);
644 WRT_SYSCTL(sched_wakeup_granularity);
654 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
656 if (unlikely(se->load.weight != NICE_0_LOAD))
657 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
663 * The idea is to set a period in which each task runs once.
665 * When there are too many tasks (sched_nr_latency) we have to stretch
666 * this period because otherwise the slices get too small.
668 * p = (nr <= nl) ? l : l*nr/nl
670 static u64 __sched_period(unsigned long nr_running)
672 if (unlikely(nr_running > sched_nr_latency))
673 return nr_running * sysctl_sched_min_granularity;
675 return sysctl_sched_latency;
679 * We calculate the wall-time slice from the period by taking a part
680 * proportional to the weight.
684 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
688 for_each_sched_entity(se) {
689 struct load_weight *load;
690 struct load_weight lw;
692 cfs_rq = cfs_rq_of(se);
693 load = &cfs_rq->load;
695 if (unlikely(!se->on_rq)) {
698 update_load_add(&lw, se->load.weight);
701 slice = __calc_delta(slice, se->load.weight, load);
707 * We calculate the vruntime slice of a to-be-inserted task.
711 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
713 return calc_delta_fair(sched_slice(cfs_rq, se), se);
717 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
718 static unsigned long task_h_load(struct task_struct *p);
721 * We choose a half-life close to 1 scheduling period.
722 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
723 * dependent on this value.
725 #define LOAD_AVG_PERIOD 32
726 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
727 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
729 /* Give new sched_entity start runnable values to heavy its load in infant time */
730 void init_entity_runnable_average(struct sched_entity *se)
732 struct sched_avg *sa = &se->avg;
734 sa->last_update_time = 0;
736 * sched_avg's period_contrib should be strictly less then 1024, so
737 * we give it 1023 to make sure it is almost a period (1024us), and
738 * will definitely be update (after enqueue).
740 sa->period_contrib = 1023;
742 * Tasks are intialized with full load to be seen as heavy tasks until
743 * they get a chance to stabilize to their real load level.
744 * Group entities are intialized with zero load to reflect the fact that
745 * nothing has been attached to the task group yet.
747 if (entity_is_task(se))
748 sa->load_avg = scale_load_down(se->load.weight);
749 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
751 * In previous Android versions, we used to have:
752 * sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
753 * sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
754 * However, that functionality has been moved to enqueue.
755 * It is unclear if we should restore this in enqueue.
758 * At this point, util_avg won't be used in select_task_rq_fair anyway
762 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
765 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
766 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
767 static void attach_entity_cfs_rq(struct sched_entity *se);
768 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
771 * With new tasks being created, their initial util_avgs are extrapolated
772 * based on the cfs_rq's current util_avg:
774 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
776 * However, in many cases, the above util_avg does not give a desired
777 * value. Moreover, the sum of the util_avgs may be divergent, such
778 * as when the series is a harmonic series.
780 * To solve this problem, we also cap the util_avg of successive tasks to
781 * only 1/2 of the left utilization budget:
783 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
785 * where n denotes the nth task.
787 * For example, a simplest series from the beginning would be like:
789 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
790 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
792 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
793 * if util_avg > util_avg_cap.
795 void post_init_entity_util_avg(struct sched_entity *se)
797 struct cfs_rq *cfs_rq = cfs_rq_of(se);
798 struct sched_avg *sa = &se->avg;
799 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
802 if (cfs_rq->avg.util_avg != 0) {
803 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
804 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
806 if (sa->util_avg > cap)
812 * If we wish to restore tuning via setting initial util,
813 * this is where we should do it.
815 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
818 if (entity_is_task(se)) {
819 struct task_struct *p = task_of(se);
820 if (p->sched_class != &fair_sched_class) {
822 * For !fair tasks do:
824 update_cfs_rq_load_avg(now, cfs_rq, false);
825 attach_entity_load_avg(cfs_rq, se);
826 switched_from_fair(rq, p);
828 * such that the next switched_to_fair() has the
831 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
836 attach_entity_cfs_rq(se);
839 #else /* !CONFIG_SMP */
840 void init_entity_runnable_average(struct sched_entity *se)
843 void post_init_entity_util_avg(struct sched_entity *se)
846 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
849 #endif /* CONFIG_SMP */
852 * Update the current task's runtime statistics.
854 static void update_curr(struct cfs_rq *cfs_rq)
856 struct sched_entity *curr = cfs_rq->curr;
857 u64 now = rq_clock_task(rq_of(cfs_rq));
863 delta_exec = now - curr->exec_start;
864 if (unlikely((s64)delta_exec <= 0))
867 curr->exec_start = now;
869 schedstat_set(curr->statistics.exec_max,
870 max(delta_exec, curr->statistics.exec_max));
872 curr->sum_exec_runtime += delta_exec;
873 schedstat_add(cfs_rq, exec_clock, delta_exec);
875 curr->vruntime += calc_delta_fair(delta_exec, curr);
876 update_min_vruntime(cfs_rq);
878 if (entity_is_task(curr)) {
879 struct task_struct *curtask = task_of(curr);
881 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
882 cpuacct_charge(curtask, delta_exec);
883 account_group_exec_runtime(curtask, delta_exec);
886 account_cfs_rq_runtime(cfs_rq, delta_exec);
889 static void update_curr_fair(struct rq *rq)
891 update_curr(cfs_rq_of(&rq->curr->se));
895 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
897 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
901 * Task is being enqueued - update stats:
903 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
906 * Are we enqueueing a waiting task? (for current tasks
907 * a dequeue/enqueue event is a NOP)
909 if (se != cfs_rq->curr)
910 update_stats_wait_start(cfs_rq, se);
914 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
916 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
917 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
918 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
919 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
920 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
921 #ifdef CONFIG_SCHEDSTATS
922 if (entity_is_task(se)) {
923 trace_sched_stat_wait(task_of(se),
924 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
927 schedstat_set(se->statistics.wait_start, 0);
931 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
934 * Mark the end of the wait period if dequeueing a
937 if (se != cfs_rq->curr)
938 update_stats_wait_end(cfs_rq, se);
942 * We are picking a new current task - update its stats:
945 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
948 * We are starting a new run period:
950 se->exec_start = rq_clock_task(rq_of(cfs_rq));
953 /**************************************************
954 * Scheduling class queueing methods:
957 #ifdef CONFIG_NUMA_BALANCING
959 * Approximate time to scan a full NUMA task in ms. The task scan period is
960 * calculated based on the tasks virtual memory size and
961 * numa_balancing_scan_size.
963 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
964 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
966 /* Portion of address space to scan in MB */
967 unsigned int sysctl_numa_balancing_scan_size = 256;
969 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
970 unsigned int sysctl_numa_balancing_scan_delay = 1000;
972 static unsigned int task_nr_scan_windows(struct task_struct *p)
974 unsigned long rss = 0;
975 unsigned long nr_scan_pages;
978 * Calculations based on RSS as non-present and empty pages are skipped
979 * by the PTE scanner and NUMA hinting faults should be trapped based
982 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
983 rss = get_mm_rss(p->mm);
987 rss = round_up(rss, nr_scan_pages);
988 return rss / nr_scan_pages;
991 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
992 #define MAX_SCAN_WINDOW 2560
994 static unsigned int task_scan_min(struct task_struct *p)
996 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
997 unsigned int scan, floor;
998 unsigned int windows = 1;
1000 if (scan_size < MAX_SCAN_WINDOW)
1001 windows = MAX_SCAN_WINDOW / scan_size;
1002 floor = 1000 / windows;
1004 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1005 return max_t(unsigned int, floor, scan);
1008 static unsigned int task_scan_max(struct task_struct *p)
1010 unsigned int smin = task_scan_min(p);
1013 /* Watch for min being lower than max due to floor calculations */
1014 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1015 return max(smin, smax);
1018 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1020 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1021 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1024 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1026 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1027 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1033 spinlock_t lock; /* nr_tasks, tasks */
1037 struct rcu_head rcu;
1038 nodemask_t active_nodes;
1039 unsigned long total_faults;
1041 * Faults_cpu is used to decide whether memory should move
1042 * towards the CPU. As a consequence, these stats are weighted
1043 * more by CPU use than by memory faults.
1045 unsigned long *faults_cpu;
1046 unsigned long faults[0];
1049 /* Shared or private faults. */
1050 #define NR_NUMA_HINT_FAULT_TYPES 2
1052 /* Memory and CPU locality */
1053 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1055 /* Averaged statistics, and temporary buffers. */
1056 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1058 pid_t task_numa_group_id(struct task_struct *p)
1060 return p->numa_group ? p->numa_group->gid : 0;
1064 * The averaged statistics, shared & private, memory & cpu,
1065 * occupy the first half of the array. The second half of the
1066 * array is for current counters, which are averaged into the
1067 * first set by task_numa_placement.
1069 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1071 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1074 static inline unsigned long task_faults(struct task_struct *p, int nid)
1076 if (!p->numa_faults)
1079 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1080 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1083 static inline unsigned long group_faults(struct task_struct *p, int nid)
1088 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1089 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1092 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1094 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1095 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1098 /* Handle placement on systems where not all nodes are directly connected. */
1099 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1100 int maxdist, bool task)
1102 unsigned long score = 0;
1106 * All nodes are directly connected, and the same distance
1107 * from each other. No need for fancy placement algorithms.
1109 if (sched_numa_topology_type == NUMA_DIRECT)
1113 * This code is called for each node, introducing N^2 complexity,
1114 * which should be ok given the number of nodes rarely exceeds 8.
1116 for_each_online_node(node) {
1117 unsigned long faults;
1118 int dist = node_distance(nid, node);
1121 * The furthest away nodes in the system are not interesting
1122 * for placement; nid was already counted.
1124 if (dist == sched_max_numa_distance || node == nid)
1128 * On systems with a backplane NUMA topology, compare groups
1129 * of nodes, and move tasks towards the group with the most
1130 * memory accesses. When comparing two nodes at distance
1131 * "hoplimit", only nodes closer by than "hoplimit" are part
1132 * of each group. Skip other nodes.
1134 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1138 /* Add up the faults from nearby nodes. */
1140 faults = task_faults(p, node);
1142 faults = group_faults(p, node);
1145 * On systems with a glueless mesh NUMA topology, there are
1146 * no fixed "groups of nodes". Instead, nodes that are not
1147 * directly connected bounce traffic through intermediate
1148 * nodes; a numa_group can occupy any set of nodes.
1149 * The further away a node is, the less the faults count.
1150 * This seems to result in good task placement.
1152 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1153 faults *= (sched_max_numa_distance - dist);
1154 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1164 * These return the fraction of accesses done by a particular task, or
1165 * task group, on a particular numa node. The group weight is given a
1166 * larger multiplier, in order to group tasks together that are almost
1167 * evenly spread out between numa nodes.
1169 static inline unsigned long task_weight(struct task_struct *p, int nid,
1172 unsigned long faults, total_faults;
1174 if (!p->numa_faults)
1177 total_faults = p->total_numa_faults;
1182 faults = task_faults(p, nid);
1183 faults += score_nearby_nodes(p, nid, dist, true);
1185 return 1000 * faults / total_faults;
1188 static inline unsigned long group_weight(struct task_struct *p, int nid,
1191 unsigned long faults, total_faults;
1196 total_faults = p->numa_group->total_faults;
1201 faults = group_faults(p, nid);
1202 faults += score_nearby_nodes(p, nid, dist, false);
1204 return 1000 * faults / total_faults;
1207 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1208 int src_nid, int dst_cpu)
1210 struct numa_group *ng = p->numa_group;
1211 int dst_nid = cpu_to_node(dst_cpu);
1212 int last_cpupid, this_cpupid;
1214 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1217 * Multi-stage node selection is used in conjunction with a periodic
1218 * migration fault to build a temporal task<->page relation. By using
1219 * a two-stage filter we remove short/unlikely relations.
1221 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1222 * a task's usage of a particular page (n_p) per total usage of this
1223 * page (n_t) (in a given time-span) to a probability.
1225 * Our periodic faults will sample this probability and getting the
1226 * same result twice in a row, given these samples are fully
1227 * independent, is then given by P(n)^2, provided our sample period
1228 * is sufficiently short compared to the usage pattern.
1230 * This quadric squishes small probabilities, making it less likely we
1231 * act on an unlikely task<->page relation.
1233 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1234 if (!cpupid_pid_unset(last_cpupid) &&
1235 cpupid_to_nid(last_cpupid) != dst_nid)
1238 /* Always allow migrate on private faults */
1239 if (cpupid_match_pid(p, last_cpupid))
1242 /* A shared fault, but p->numa_group has not been set up yet. */
1247 * Do not migrate if the destination is not a node that
1248 * is actively used by this numa group.
1250 if (!node_isset(dst_nid, ng->active_nodes))
1254 * Source is a node that is not actively used by this
1255 * numa group, while the destination is. Migrate.
1257 if (!node_isset(src_nid, ng->active_nodes))
1261 * Both source and destination are nodes in active
1262 * use by this numa group. Maximize memory bandwidth
1263 * by migrating from more heavily used groups, to less
1264 * heavily used ones, spreading the load around.
1265 * Use a 1/4 hysteresis to avoid spurious page movement.
1267 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1270 static unsigned long weighted_cpuload(const int cpu);
1271 static unsigned long source_load(int cpu, int type);
1272 static unsigned long target_load(int cpu, int type);
1273 static unsigned long capacity_of(int cpu);
1274 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1276 /* Cached statistics for all CPUs within a node */
1278 unsigned long nr_running;
1281 /* Total compute capacity of CPUs on a node */
1282 unsigned long compute_capacity;
1284 /* Approximate capacity in terms of runnable tasks on a node */
1285 unsigned long task_capacity;
1286 int has_free_capacity;
1290 * XXX borrowed from update_sg_lb_stats
1292 static void update_numa_stats(struct numa_stats *ns, int nid)
1294 int smt, cpu, cpus = 0;
1295 unsigned long capacity;
1297 memset(ns, 0, sizeof(*ns));
1298 for_each_cpu(cpu, cpumask_of_node(nid)) {
1299 struct rq *rq = cpu_rq(cpu);
1301 ns->nr_running += rq->nr_running;
1302 ns->load += weighted_cpuload(cpu);
1303 ns->compute_capacity += capacity_of(cpu);
1309 * If we raced with hotplug and there are no CPUs left in our mask
1310 * the @ns structure is NULL'ed and task_numa_compare() will
1311 * not find this node attractive.
1313 * We'll either bail at !has_free_capacity, or we'll detect a huge
1314 * imbalance and bail there.
1319 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1320 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1321 capacity = cpus / smt; /* cores */
1323 ns->task_capacity = min_t(unsigned, capacity,
1324 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1325 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1328 struct task_numa_env {
1329 struct task_struct *p;
1331 int src_cpu, src_nid;
1332 int dst_cpu, dst_nid;
1334 struct numa_stats src_stats, dst_stats;
1339 struct task_struct *best_task;
1344 static void task_numa_assign(struct task_numa_env *env,
1345 struct task_struct *p, long imp)
1348 put_task_struct(env->best_task);
1351 env->best_imp = imp;
1352 env->best_cpu = env->dst_cpu;
1355 static bool load_too_imbalanced(long src_load, long dst_load,
1356 struct task_numa_env *env)
1359 long orig_src_load, orig_dst_load;
1360 long src_capacity, dst_capacity;
1363 * The load is corrected for the CPU capacity available on each node.
1366 * ------------ vs ---------
1367 * src_capacity dst_capacity
1369 src_capacity = env->src_stats.compute_capacity;
1370 dst_capacity = env->dst_stats.compute_capacity;
1372 /* We care about the slope of the imbalance, not the direction. */
1373 if (dst_load < src_load)
1374 swap(dst_load, src_load);
1376 /* Is the difference below the threshold? */
1377 imb = dst_load * src_capacity * 100 -
1378 src_load * dst_capacity * env->imbalance_pct;
1383 * The imbalance is above the allowed threshold.
1384 * Compare it with the old imbalance.
1386 orig_src_load = env->src_stats.load;
1387 orig_dst_load = env->dst_stats.load;
1389 if (orig_dst_load < orig_src_load)
1390 swap(orig_dst_load, orig_src_load);
1392 old_imb = orig_dst_load * src_capacity * 100 -
1393 orig_src_load * dst_capacity * env->imbalance_pct;
1395 /* Would this change make things worse? */
1396 return (imb > old_imb);
1400 * This checks if the overall compute and NUMA accesses of the system would
1401 * be improved if the source tasks was migrated to the target dst_cpu taking
1402 * into account that it might be best if task running on the dst_cpu should
1403 * be exchanged with the source task
1405 static void task_numa_compare(struct task_numa_env *env,
1406 long taskimp, long groupimp)
1408 struct rq *src_rq = cpu_rq(env->src_cpu);
1409 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1410 struct task_struct *cur;
1411 long src_load, dst_load;
1413 long imp = env->p->numa_group ? groupimp : taskimp;
1415 int dist = env->dist;
1416 bool assigned = false;
1420 raw_spin_lock_irq(&dst_rq->lock);
1423 * No need to move the exiting task or idle task.
1425 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1429 * The task_struct must be protected here to protect the
1430 * p->numa_faults access in the task_weight since the
1431 * numa_faults could already be freed in the following path:
1432 * finish_task_switch()
1433 * --> put_task_struct()
1434 * --> __put_task_struct()
1435 * --> task_numa_free()
1437 get_task_struct(cur);
1440 raw_spin_unlock_irq(&dst_rq->lock);
1443 * Because we have preemption enabled we can get migrated around and
1444 * end try selecting ourselves (current == env->p) as a swap candidate.
1450 * "imp" is the fault differential for the source task between the
1451 * source and destination node. Calculate the total differential for
1452 * the source task and potential destination task. The more negative
1453 * the value is, the more rmeote accesses that would be expected to
1454 * be incurred if the tasks were swapped.
1457 /* Skip this swap candidate if cannot move to the source cpu */
1458 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1462 * If dst and source tasks are in the same NUMA group, or not
1463 * in any group then look only at task weights.
1465 if (cur->numa_group == env->p->numa_group) {
1466 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1467 task_weight(cur, env->dst_nid, dist);
1469 * Add some hysteresis to prevent swapping the
1470 * tasks within a group over tiny differences.
1472 if (cur->numa_group)
1476 * Compare the group weights. If a task is all by
1477 * itself (not part of a group), use the task weight
1480 if (cur->numa_group)
1481 imp += group_weight(cur, env->src_nid, dist) -
1482 group_weight(cur, env->dst_nid, dist);
1484 imp += task_weight(cur, env->src_nid, dist) -
1485 task_weight(cur, env->dst_nid, dist);
1489 if (imp <= env->best_imp && moveimp <= env->best_imp)
1493 /* Is there capacity at our destination? */
1494 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1495 !env->dst_stats.has_free_capacity)
1501 /* Balance doesn't matter much if we're running a task per cpu */
1502 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1503 dst_rq->nr_running == 1)
1507 * In the overloaded case, try and keep the load balanced.
1510 load = task_h_load(env->p);
1511 dst_load = env->dst_stats.load + load;
1512 src_load = env->src_stats.load - load;
1514 if (moveimp > imp && moveimp > env->best_imp) {
1516 * If the improvement from just moving env->p direction is
1517 * better than swapping tasks around, check if a move is
1518 * possible. Store a slightly smaller score than moveimp,
1519 * so an actually idle CPU will win.
1521 if (!load_too_imbalanced(src_load, dst_load, env)) {
1523 put_task_struct(cur);
1529 if (imp <= env->best_imp)
1533 load = task_h_load(cur);
1538 if (load_too_imbalanced(src_load, dst_load, env))
1542 * One idle CPU per node is evaluated for a task numa move.
1543 * Call select_idle_sibling to maybe find a better one.
1546 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1551 task_numa_assign(env, cur, imp);
1555 * The dst_rq->curr isn't assigned. The protection for task_struct is
1558 if (cur && !assigned)
1559 put_task_struct(cur);
1562 static void task_numa_find_cpu(struct task_numa_env *env,
1563 long taskimp, long groupimp)
1567 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1568 /* Skip this CPU if the source task cannot migrate */
1569 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1573 task_numa_compare(env, taskimp, groupimp);
1577 /* Only move tasks to a NUMA node less busy than the current node. */
1578 static bool numa_has_capacity(struct task_numa_env *env)
1580 struct numa_stats *src = &env->src_stats;
1581 struct numa_stats *dst = &env->dst_stats;
1583 if (src->has_free_capacity && !dst->has_free_capacity)
1587 * Only consider a task move if the source has a higher load
1588 * than the destination, corrected for CPU capacity on each node.
1590 * src->load dst->load
1591 * --------------------- vs ---------------------
1592 * src->compute_capacity dst->compute_capacity
1594 if (src->load * dst->compute_capacity * env->imbalance_pct >
1596 dst->load * src->compute_capacity * 100)
1602 static int task_numa_migrate(struct task_struct *p)
1604 struct task_numa_env env = {
1607 .src_cpu = task_cpu(p),
1608 .src_nid = task_node(p),
1610 .imbalance_pct = 112,
1616 struct sched_domain *sd;
1617 unsigned long taskweight, groupweight;
1619 long taskimp, groupimp;
1622 * Pick the lowest SD_NUMA domain, as that would have the smallest
1623 * imbalance and would be the first to start moving tasks about.
1625 * And we want to avoid any moving of tasks about, as that would create
1626 * random movement of tasks -- counter the numa conditions we're trying
1630 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1632 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1636 * Cpusets can break the scheduler domain tree into smaller
1637 * balance domains, some of which do not cross NUMA boundaries.
1638 * Tasks that are "trapped" in such domains cannot be migrated
1639 * elsewhere, so there is no point in (re)trying.
1641 if (unlikely(!sd)) {
1642 p->numa_preferred_nid = task_node(p);
1646 env.dst_nid = p->numa_preferred_nid;
1647 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1648 taskweight = task_weight(p, env.src_nid, dist);
1649 groupweight = group_weight(p, env.src_nid, dist);
1650 update_numa_stats(&env.src_stats, env.src_nid);
1651 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1652 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1653 update_numa_stats(&env.dst_stats, env.dst_nid);
1655 /* Try to find a spot on the preferred nid. */
1656 if (numa_has_capacity(&env))
1657 task_numa_find_cpu(&env, taskimp, groupimp);
1660 * Look at other nodes in these cases:
1661 * - there is no space available on the preferred_nid
1662 * - the task is part of a numa_group that is interleaved across
1663 * multiple NUMA nodes; in order to better consolidate the group,
1664 * we need to check other locations.
1666 if (env.best_cpu == -1 || (p->numa_group &&
1667 nodes_weight(p->numa_group->active_nodes) > 1)) {
1668 for_each_online_node(nid) {
1669 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1672 dist = node_distance(env.src_nid, env.dst_nid);
1673 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1675 taskweight = task_weight(p, env.src_nid, dist);
1676 groupweight = group_weight(p, env.src_nid, dist);
1679 /* Only consider nodes where both task and groups benefit */
1680 taskimp = task_weight(p, nid, dist) - taskweight;
1681 groupimp = group_weight(p, nid, dist) - groupweight;
1682 if (taskimp < 0 && groupimp < 0)
1687 update_numa_stats(&env.dst_stats, env.dst_nid);
1688 if (numa_has_capacity(&env))
1689 task_numa_find_cpu(&env, taskimp, groupimp);
1694 * If the task is part of a workload that spans multiple NUMA nodes,
1695 * and is migrating into one of the workload's active nodes, remember
1696 * this node as the task's preferred numa node, so the workload can
1698 * A task that migrated to a second choice node will be better off
1699 * trying for a better one later. Do not set the preferred node here.
1701 if (p->numa_group) {
1702 if (env.best_cpu == -1)
1707 if (node_isset(nid, p->numa_group->active_nodes))
1708 sched_setnuma(p, env.dst_nid);
1711 /* No better CPU than the current one was found. */
1712 if (env.best_cpu == -1)
1716 * Reset the scan period if the task is being rescheduled on an
1717 * alternative node to recheck if the tasks is now properly placed.
1719 p->numa_scan_period = task_scan_min(p);
1721 if (env.best_task == NULL) {
1722 ret = migrate_task_to(p, env.best_cpu);
1724 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1728 ret = migrate_swap(p, env.best_task);
1730 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1731 put_task_struct(env.best_task);
1735 /* Attempt to migrate a task to a CPU on the preferred node. */
1736 static void numa_migrate_preferred(struct task_struct *p)
1738 unsigned long interval = HZ;
1740 /* This task has no NUMA fault statistics yet */
1741 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1744 /* Periodically retry migrating the task to the preferred node */
1745 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1746 p->numa_migrate_retry = jiffies + interval;
1748 /* Success if task is already running on preferred CPU */
1749 if (task_node(p) == p->numa_preferred_nid)
1752 /* Otherwise, try migrate to a CPU on the preferred node */
1753 task_numa_migrate(p);
1757 * Find the nodes on which the workload is actively running. We do this by
1758 * tracking the nodes from which NUMA hinting faults are triggered. This can
1759 * be different from the set of nodes where the workload's memory is currently
1762 * The bitmask is used to make smarter decisions on when to do NUMA page
1763 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1764 * are added when they cause over 6/16 of the maximum number of faults, but
1765 * only removed when they drop below 3/16.
1767 static void update_numa_active_node_mask(struct numa_group *numa_group)
1769 unsigned long faults, max_faults = 0;
1772 for_each_online_node(nid) {
1773 faults = group_faults_cpu(numa_group, nid);
1774 if (faults > max_faults)
1775 max_faults = faults;
1778 for_each_online_node(nid) {
1779 faults = group_faults_cpu(numa_group, nid);
1780 if (!node_isset(nid, numa_group->active_nodes)) {
1781 if (faults > max_faults * 6 / 16)
1782 node_set(nid, numa_group->active_nodes);
1783 } else if (faults < max_faults * 3 / 16)
1784 node_clear(nid, numa_group->active_nodes);
1789 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1790 * increments. The more local the fault statistics are, the higher the scan
1791 * period will be for the next scan window. If local/(local+remote) ratio is
1792 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1793 * the scan period will decrease. Aim for 70% local accesses.
1795 #define NUMA_PERIOD_SLOTS 10
1796 #define NUMA_PERIOD_THRESHOLD 7
1799 * Increase the scan period (slow down scanning) if the majority of
1800 * our memory is already on our local node, or if the majority of
1801 * the page accesses are shared with other processes.
1802 * Otherwise, decrease the scan period.
1804 static void update_task_scan_period(struct task_struct *p,
1805 unsigned long shared, unsigned long private)
1807 unsigned int period_slot;
1811 unsigned long remote = p->numa_faults_locality[0];
1812 unsigned long local = p->numa_faults_locality[1];
1815 * If there were no record hinting faults then either the task is
1816 * completely idle or all activity is areas that are not of interest
1817 * to automatic numa balancing. Related to that, if there were failed
1818 * migration then it implies we are migrating too quickly or the local
1819 * node is overloaded. In either case, scan slower
1821 if (local + shared == 0 || p->numa_faults_locality[2]) {
1822 p->numa_scan_period = min(p->numa_scan_period_max,
1823 p->numa_scan_period << 1);
1825 p->mm->numa_next_scan = jiffies +
1826 msecs_to_jiffies(p->numa_scan_period);
1832 * Prepare to scale scan period relative to the current period.
1833 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1834 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1835 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1837 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1838 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1839 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1840 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1843 diff = slot * period_slot;
1845 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1848 * Scale scan rate increases based on sharing. There is an
1849 * inverse relationship between the degree of sharing and
1850 * the adjustment made to the scanning period. Broadly
1851 * speaking the intent is that there is little point
1852 * scanning faster if shared accesses dominate as it may
1853 * simply bounce migrations uselessly
1855 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1856 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1859 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1860 task_scan_min(p), task_scan_max(p));
1861 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1865 * Get the fraction of time the task has been running since the last
1866 * NUMA placement cycle. The scheduler keeps similar statistics, but
1867 * decays those on a 32ms period, which is orders of magnitude off
1868 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1869 * stats only if the task is so new there are no NUMA statistics yet.
1871 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1873 u64 runtime, delta, now;
1874 /* Use the start of this time slice to avoid calculations. */
1875 now = p->se.exec_start;
1876 runtime = p->se.sum_exec_runtime;
1878 if (p->last_task_numa_placement) {
1879 delta = runtime - p->last_sum_exec_runtime;
1880 *period = now - p->last_task_numa_placement;
1882 delta = p->se.avg.load_sum / p->se.load.weight;
1883 *period = LOAD_AVG_MAX;
1886 p->last_sum_exec_runtime = runtime;
1887 p->last_task_numa_placement = now;
1893 * Determine the preferred nid for a task in a numa_group. This needs to
1894 * be done in a way that produces consistent results with group_weight,
1895 * otherwise workloads might not converge.
1897 static int preferred_group_nid(struct task_struct *p, int nid)
1902 /* Direct connections between all NUMA nodes. */
1903 if (sched_numa_topology_type == NUMA_DIRECT)
1907 * On a system with glueless mesh NUMA topology, group_weight
1908 * scores nodes according to the number of NUMA hinting faults on
1909 * both the node itself, and on nearby nodes.
1911 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1912 unsigned long score, max_score = 0;
1913 int node, max_node = nid;
1915 dist = sched_max_numa_distance;
1917 for_each_online_node(node) {
1918 score = group_weight(p, node, dist);
1919 if (score > max_score) {
1928 * Finding the preferred nid in a system with NUMA backplane
1929 * interconnect topology is more involved. The goal is to locate
1930 * tasks from numa_groups near each other in the system, and
1931 * untangle workloads from different sides of the system. This requires
1932 * searching down the hierarchy of node groups, recursively searching
1933 * inside the highest scoring group of nodes. The nodemask tricks
1934 * keep the complexity of the search down.
1936 nodes = node_online_map;
1937 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1938 unsigned long max_faults = 0;
1939 nodemask_t max_group = NODE_MASK_NONE;
1942 /* Are there nodes at this distance from each other? */
1943 if (!find_numa_distance(dist))
1946 for_each_node_mask(a, nodes) {
1947 unsigned long faults = 0;
1948 nodemask_t this_group;
1949 nodes_clear(this_group);
1951 /* Sum group's NUMA faults; includes a==b case. */
1952 for_each_node_mask(b, nodes) {
1953 if (node_distance(a, b) < dist) {
1954 faults += group_faults(p, b);
1955 node_set(b, this_group);
1956 node_clear(b, nodes);
1960 /* Remember the top group. */
1961 if (faults > max_faults) {
1962 max_faults = faults;
1963 max_group = this_group;
1965 * subtle: at the smallest distance there is
1966 * just one node left in each "group", the
1967 * winner is the preferred nid.
1972 /* Next round, evaluate the nodes within max_group. */
1980 static void task_numa_placement(struct task_struct *p)
1982 int seq, nid, max_nid = -1, max_group_nid = -1;
1983 unsigned long max_faults = 0, max_group_faults = 0;
1984 unsigned long fault_types[2] = { 0, 0 };
1985 unsigned long total_faults;
1986 u64 runtime, period;
1987 spinlock_t *group_lock = NULL;
1990 * The p->mm->numa_scan_seq field gets updated without
1991 * exclusive access. Use READ_ONCE() here to ensure
1992 * that the field is read in a single access:
1994 seq = READ_ONCE(p->mm->numa_scan_seq);
1995 if (p->numa_scan_seq == seq)
1997 p->numa_scan_seq = seq;
1998 p->numa_scan_period_max = task_scan_max(p);
2000 total_faults = p->numa_faults_locality[0] +
2001 p->numa_faults_locality[1];
2002 runtime = numa_get_avg_runtime(p, &period);
2004 /* If the task is part of a group prevent parallel updates to group stats */
2005 if (p->numa_group) {
2006 group_lock = &p->numa_group->lock;
2007 spin_lock_irq(group_lock);
2010 /* Find the node with the highest number of faults */
2011 for_each_online_node(nid) {
2012 /* Keep track of the offsets in numa_faults array */
2013 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2014 unsigned long faults = 0, group_faults = 0;
2017 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2018 long diff, f_diff, f_weight;
2020 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2021 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2022 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2023 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2025 /* Decay existing window, copy faults since last scan */
2026 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2027 fault_types[priv] += p->numa_faults[membuf_idx];
2028 p->numa_faults[membuf_idx] = 0;
2031 * Normalize the faults_from, so all tasks in a group
2032 * count according to CPU use, instead of by the raw
2033 * number of faults. Tasks with little runtime have
2034 * little over-all impact on throughput, and thus their
2035 * faults are less important.
2037 f_weight = div64_u64(runtime << 16, period + 1);
2038 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2040 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2041 p->numa_faults[cpubuf_idx] = 0;
2043 p->numa_faults[mem_idx] += diff;
2044 p->numa_faults[cpu_idx] += f_diff;
2045 faults += p->numa_faults[mem_idx];
2046 p->total_numa_faults += diff;
2047 if (p->numa_group) {
2049 * safe because we can only change our own group
2051 * mem_idx represents the offset for a given
2052 * nid and priv in a specific region because it
2053 * is at the beginning of the numa_faults array.
2055 p->numa_group->faults[mem_idx] += diff;
2056 p->numa_group->faults_cpu[mem_idx] += f_diff;
2057 p->numa_group->total_faults += diff;
2058 group_faults += p->numa_group->faults[mem_idx];
2062 if (faults > max_faults) {
2063 max_faults = faults;
2067 if (group_faults > max_group_faults) {
2068 max_group_faults = group_faults;
2069 max_group_nid = nid;
2073 update_task_scan_period(p, fault_types[0], fault_types[1]);
2075 if (p->numa_group) {
2076 update_numa_active_node_mask(p->numa_group);
2077 spin_unlock_irq(group_lock);
2078 max_nid = preferred_group_nid(p, max_group_nid);
2082 /* Set the new preferred node */
2083 if (max_nid != p->numa_preferred_nid)
2084 sched_setnuma(p, max_nid);
2086 if (task_node(p) != p->numa_preferred_nid)
2087 numa_migrate_preferred(p);
2091 static inline int get_numa_group(struct numa_group *grp)
2093 return atomic_inc_not_zero(&grp->refcount);
2096 static inline void put_numa_group(struct numa_group *grp)
2098 if (atomic_dec_and_test(&grp->refcount))
2099 kfree_rcu(grp, rcu);
2102 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2105 struct numa_group *grp, *my_grp;
2106 struct task_struct *tsk;
2108 int cpu = cpupid_to_cpu(cpupid);
2111 if (unlikely(!p->numa_group)) {
2112 unsigned int size = sizeof(struct numa_group) +
2113 4*nr_node_ids*sizeof(unsigned long);
2115 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2119 atomic_set(&grp->refcount, 1);
2120 spin_lock_init(&grp->lock);
2122 /* Second half of the array tracks nids where faults happen */
2123 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2126 node_set(task_node(current), grp->active_nodes);
2128 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2129 grp->faults[i] = p->numa_faults[i];
2131 grp->total_faults = p->total_numa_faults;
2134 rcu_assign_pointer(p->numa_group, grp);
2138 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2140 if (!cpupid_match_pid(tsk, cpupid))
2143 grp = rcu_dereference(tsk->numa_group);
2147 my_grp = p->numa_group;
2152 * Only join the other group if its bigger; if we're the bigger group,
2153 * the other task will join us.
2155 if (my_grp->nr_tasks > grp->nr_tasks)
2159 * Tie-break on the grp address.
2161 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2164 /* Always join threads in the same process. */
2165 if (tsk->mm == current->mm)
2168 /* Simple filter to avoid false positives due to PID collisions */
2169 if (flags & TNF_SHARED)
2172 /* Update priv based on whether false sharing was detected */
2175 if (join && !get_numa_group(grp))
2183 BUG_ON(irqs_disabled());
2184 double_lock_irq(&my_grp->lock, &grp->lock);
2186 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2187 my_grp->faults[i] -= p->numa_faults[i];
2188 grp->faults[i] += p->numa_faults[i];
2190 my_grp->total_faults -= p->total_numa_faults;
2191 grp->total_faults += p->total_numa_faults;
2196 spin_unlock(&my_grp->lock);
2197 spin_unlock_irq(&grp->lock);
2199 rcu_assign_pointer(p->numa_group, grp);
2201 put_numa_group(my_grp);
2209 void task_numa_free(struct task_struct *p)
2211 struct numa_group *grp = p->numa_group;
2212 void *numa_faults = p->numa_faults;
2213 unsigned long flags;
2217 spin_lock_irqsave(&grp->lock, flags);
2218 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2219 grp->faults[i] -= p->numa_faults[i];
2220 grp->total_faults -= p->total_numa_faults;
2223 spin_unlock_irqrestore(&grp->lock, flags);
2224 RCU_INIT_POINTER(p->numa_group, NULL);
2225 put_numa_group(grp);
2228 p->numa_faults = NULL;
2233 * Got a PROT_NONE fault for a page on @node.
2235 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2237 struct task_struct *p = current;
2238 bool migrated = flags & TNF_MIGRATED;
2239 int cpu_node = task_node(current);
2240 int local = !!(flags & TNF_FAULT_LOCAL);
2243 if (!static_branch_likely(&sched_numa_balancing))
2246 /* for example, ksmd faulting in a user's mm */
2250 /* Allocate buffer to track faults on a per-node basis */
2251 if (unlikely(!p->numa_faults)) {
2252 int size = sizeof(*p->numa_faults) *
2253 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2255 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2256 if (!p->numa_faults)
2259 p->total_numa_faults = 0;
2260 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2264 * First accesses are treated as private, otherwise consider accesses
2265 * to be private if the accessing pid has not changed
2267 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2270 priv = cpupid_match_pid(p, last_cpupid);
2271 if (!priv && !(flags & TNF_NO_GROUP))
2272 task_numa_group(p, last_cpupid, flags, &priv);
2276 * If a workload spans multiple NUMA nodes, a shared fault that
2277 * occurs wholly within the set of nodes that the workload is
2278 * actively using should be counted as local. This allows the
2279 * scan rate to slow down when a workload has settled down.
2281 if (!priv && !local && p->numa_group &&
2282 node_isset(cpu_node, p->numa_group->active_nodes) &&
2283 node_isset(mem_node, p->numa_group->active_nodes))
2286 task_numa_placement(p);
2289 * Retry task to preferred node migration periodically, in case it
2290 * case it previously failed, or the scheduler moved us.
2292 if (time_after(jiffies, p->numa_migrate_retry))
2293 numa_migrate_preferred(p);
2296 p->numa_pages_migrated += pages;
2297 if (flags & TNF_MIGRATE_FAIL)
2298 p->numa_faults_locality[2] += pages;
2300 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2301 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2302 p->numa_faults_locality[local] += pages;
2305 static void reset_ptenuma_scan(struct task_struct *p)
2308 * We only did a read acquisition of the mmap sem, so
2309 * p->mm->numa_scan_seq is written to without exclusive access
2310 * and the update is not guaranteed to be atomic. That's not
2311 * much of an issue though, since this is just used for
2312 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2313 * expensive, to avoid any form of compiler optimizations:
2315 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2316 p->mm->numa_scan_offset = 0;
2320 * The expensive part of numa migration is done from task_work context.
2321 * Triggered from task_tick_numa().
2323 void task_numa_work(struct callback_head *work)
2325 unsigned long migrate, next_scan, now = jiffies;
2326 struct task_struct *p = current;
2327 struct mm_struct *mm = p->mm;
2328 struct vm_area_struct *vma;
2329 unsigned long start, end;
2330 unsigned long nr_pte_updates = 0;
2331 long pages, virtpages;
2333 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2335 work->next = work; /* protect against double add */
2337 * Who cares about NUMA placement when they're dying.
2339 * NOTE: make sure not to dereference p->mm before this check,
2340 * exit_task_work() happens _after_ exit_mm() so we could be called
2341 * without p->mm even though we still had it when we enqueued this
2344 if (p->flags & PF_EXITING)
2347 if (!mm->numa_next_scan) {
2348 mm->numa_next_scan = now +
2349 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2353 * Enforce maximal scan/migration frequency..
2355 migrate = mm->numa_next_scan;
2356 if (time_before(now, migrate))
2359 if (p->numa_scan_period == 0) {
2360 p->numa_scan_period_max = task_scan_max(p);
2361 p->numa_scan_period = task_scan_min(p);
2364 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2365 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2369 * Delay this task enough that another task of this mm will likely win
2370 * the next time around.
2372 p->node_stamp += 2 * TICK_NSEC;
2374 start = mm->numa_scan_offset;
2375 pages = sysctl_numa_balancing_scan_size;
2376 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2377 virtpages = pages * 8; /* Scan up to this much virtual space */
2382 down_read(&mm->mmap_sem);
2383 vma = find_vma(mm, start);
2385 reset_ptenuma_scan(p);
2389 for (; vma; vma = vma->vm_next) {
2390 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2391 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2396 * Shared library pages mapped by multiple processes are not
2397 * migrated as it is expected they are cache replicated. Avoid
2398 * hinting faults in read-only file-backed mappings or the vdso
2399 * as migrating the pages will be of marginal benefit.
2402 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2406 * Skip inaccessible VMAs to avoid any confusion between
2407 * PROT_NONE and NUMA hinting ptes
2409 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2413 start = max(start, vma->vm_start);
2414 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2415 end = min(end, vma->vm_end);
2416 nr_pte_updates = change_prot_numa(vma, start, end);
2419 * Try to scan sysctl_numa_balancing_size worth of
2420 * hpages that have at least one present PTE that
2421 * is not already pte-numa. If the VMA contains
2422 * areas that are unused or already full of prot_numa
2423 * PTEs, scan up to virtpages, to skip through those
2427 pages -= (end - start) >> PAGE_SHIFT;
2428 virtpages -= (end - start) >> PAGE_SHIFT;
2431 if (pages <= 0 || virtpages <= 0)
2435 } while (end != vma->vm_end);
2440 * It is possible to reach the end of the VMA list but the last few
2441 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2442 * would find the !migratable VMA on the next scan but not reset the
2443 * scanner to the start so check it now.
2446 mm->numa_scan_offset = start;
2448 reset_ptenuma_scan(p);
2449 up_read(&mm->mmap_sem);
2453 * Drive the periodic memory faults..
2455 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2457 struct callback_head *work = &curr->numa_work;
2461 * We don't care about NUMA placement if we don't have memory.
2463 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2467 * Using runtime rather than walltime has the dual advantage that
2468 * we (mostly) drive the selection from busy threads and that the
2469 * task needs to have done some actual work before we bother with
2472 now = curr->se.sum_exec_runtime;
2473 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2475 if (now > curr->node_stamp + period) {
2476 if (!curr->node_stamp)
2477 curr->numa_scan_period = task_scan_min(curr);
2478 curr->node_stamp += period;
2480 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2481 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2482 task_work_add(curr, work, true);
2487 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2491 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2495 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2498 #endif /* CONFIG_NUMA_BALANCING */
2501 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2503 update_load_add(&cfs_rq->load, se->load.weight);
2504 if (!parent_entity(se))
2505 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2507 if (entity_is_task(se)) {
2508 struct rq *rq = rq_of(cfs_rq);
2510 account_numa_enqueue(rq, task_of(se));
2511 list_add(&se->group_node, &rq->cfs_tasks);
2514 cfs_rq->nr_running++;
2518 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2520 update_load_sub(&cfs_rq->load, se->load.weight);
2521 if (!parent_entity(se))
2522 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2523 if (entity_is_task(se)) {
2524 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2525 list_del_init(&se->group_node);
2527 cfs_rq->nr_running--;
2530 #ifdef CONFIG_FAIR_GROUP_SCHED
2532 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2534 long tg_weight, load, shares;
2537 * This really should be: cfs_rq->avg.load_avg, but instead we use
2538 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2539 * the shares for small weight interactive tasks.
2541 load = scale_load_down(cfs_rq->load.weight);
2543 tg_weight = atomic_long_read(&tg->load_avg);
2545 /* Ensure tg_weight >= load */
2546 tg_weight -= cfs_rq->tg_load_avg_contrib;
2549 shares = (tg->shares * load);
2551 shares /= tg_weight;
2553 if (shares < MIN_SHARES)
2554 shares = MIN_SHARES;
2555 if (shares > tg->shares)
2556 shares = tg->shares;
2560 # else /* CONFIG_SMP */
2561 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2565 # endif /* CONFIG_SMP */
2567 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2568 unsigned long weight)
2571 /* commit outstanding execution time */
2572 if (cfs_rq->curr == se)
2573 update_curr(cfs_rq);
2574 account_entity_dequeue(cfs_rq, se);
2577 update_load_set(&se->load, weight);
2580 account_entity_enqueue(cfs_rq, se);
2583 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2585 static void update_cfs_shares(struct sched_entity *se)
2587 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2588 struct task_group *tg;
2594 if (throttled_hierarchy(cfs_rq))
2600 if (likely(se->load.weight == tg->shares))
2603 shares = calc_cfs_shares(cfs_rq, tg);
2605 reweight_entity(cfs_rq_of(se), se, shares);
2608 #else /* CONFIG_FAIR_GROUP_SCHED */
2609 static inline void update_cfs_shares(struct sched_entity *se)
2612 #endif /* CONFIG_FAIR_GROUP_SCHED */
2615 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2616 static const u32 runnable_avg_yN_inv[] = {
2617 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2618 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2619 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2620 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2621 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2622 0x85aac367, 0x82cd8698,
2626 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2627 * over-estimates when re-combining.
2629 static const u32 runnable_avg_yN_sum[] = {
2630 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2631 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2632 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2637 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2639 static __always_inline u64 decay_load(u64 val, u64 n)
2641 unsigned int local_n;
2645 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2648 /* after bounds checking we can collapse to 32-bit */
2652 * As y^PERIOD = 1/2, we can combine
2653 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2654 * With a look-up table which covers y^n (n<PERIOD)
2656 * To achieve constant time decay_load.
2658 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2659 val >>= local_n / LOAD_AVG_PERIOD;
2660 local_n %= LOAD_AVG_PERIOD;
2663 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2668 * For updates fully spanning n periods, the contribution to runnable
2669 * average will be: \Sum 1024*y^n
2671 * We can compute this reasonably efficiently by combining:
2672 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2674 static u32 __compute_runnable_contrib(u64 n)
2678 if (likely(n <= LOAD_AVG_PERIOD))
2679 return runnable_avg_yN_sum[n];
2680 else if (unlikely(n >= LOAD_AVG_MAX_N))
2681 return LOAD_AVG_MAX;
2683 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2685 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2686 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2688 n -= LOAD_AVG_PERIOD;
2689 } while (n > LOAD_AVG_PERIOD);
2691 contrib = decay_load(contrib, n);
2692 return contrib + runnable_avg_yN_sum[n];
2695 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2696 #error "load tracking assumes 2^10 as unit"
2699 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2702 * We can represent the historical contribution to runnable average as the
2703 * coefficients of a geometric series. To do this we sub-divide our runnable
2704 * history into segments of approximately 1ms (1024us); label the segment that
2705 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2707 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2709 * (now) (~1ms ago) (~2ms ago)
2711 * Let u_i denote the fraction of p_i that the entity was runnable.
2713 * We then designate the fractions u_i as our co-efficients, yielding the
2714 * following representation of historical load:
2715 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2717 * We choose y based on the with of a reasonably scheduling period, fixing:
2720 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2721 * approximately half as much as the contribution to load within the last ms
2724 * When a period "rolls over" and we have new u_0`, multiplying the previous
2725 * sum again by y is sufficient to update:
2726 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2727 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2729 static __always_inline int
2730 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2731 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2733 u64 delta, scaled_delta, periods;
2735 unsigned int delta_w, scaled_delta_w, decayed = 0;
2736 unsigned long scale_freq, scale_cpu;
2738 delta = now - sa->last_update_time;
2740 * This should only happen when time goes backwards, which it
2741 * unfortunately does during sched clock init when we swap over to TSC.
2743 if ((s64)delta < 0) {
2744 sa->last_update_time = now;
2749 * Use 1024ns as the unit of measurement since it's a reasonable
2750 * approximation of 1us and fast to compute.
2755 sa->last_update_time = now;
2757 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2758 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2759 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2761 /* delta_w is the amount already accumulated against our next period */
2762 delta_w = sa->period_contrib;
2763 if (delta + delta_w >= 1024) {
2766 /* how much left for next period will start over, we don't know yet */
2767 sa->period_contrib = 0;
2770 * Now that we know we're crossing a period boundary, figure
2771 * out how much from delta we need to complete the current
2772 * period and accrue it.
2774 delta_w = 1024 - delta_w;
2775 scaled_delta_w = cap_scale(delta_w, scale_freq);
2777 sa->load_sum += weight * scaled_delta_w;
2779 cfs_rq->runnable_load_sum +=
2780 weight * scaled_delta_w;
2784 sa->util_sum += scaled_delta_w * scale_cpu;
2788 /* Figure out how many additional periods this update spans */
2789 periods = delta / 1024;
2792 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2794 cfs_rq->runnable_load_sum =
2795 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2797 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2799 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2800 contrib = __compute_runnable_contrib(periods);
2801 contrib = cap_scale(contrib, scale_freq);
2803 sa->load_sum += weight * contrib;
2805 cfs_rq->runnable_load_sum += weight * contrib;
2808 sa->util_sum += contrib * scale_cpu;
2811 /* Remainder of delta accrued against u_0` */
2812 scaled_delta = cap_scale(delta, scale_freq);
2814 sa->load_sum += weight * scaled_delta;
2816 cfs_rq->runnable_load_sum += weight * scaled_delta;
2819 sa->util_sum += scaled_delta * scale_cpu;
2821 sa->period_contrib += delta;
2824 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2826 cfs_rq->runnable_load_avg =
2827 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2829 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2836 * Signed add and clamp on underflow.
2838 * Explicitly do a load-store to ensure the intermediate value never hits
2839 * memory. This allows lockless observations without ever seeing the negative
2842 #define add_positive(_ptr, _val) do { \
2843 typeof(_ptr) ptr = (_ptr); \
2844 typeof(_val) val = (_val); \
2845 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2849 if (val < 0 && res > var) \
2852 WRITE_ONCE(*ptr, res); \
2855 #ifdef CONFIG_FAIR_GROUP_SCHED
2857 * update_tg_load_avg - update the tg's load avg
2858 * @cfs_rq: the cfs_rq whose avg changed
2859 * @force: update regardless of how small the difference
2861 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2862 * However, because tg->load_avg is a global value there are performance
2865 * In order to avoid having to look at the other cfs_rq's, we use a
2866 * differential update where we store the last value we propagated. This in
2867 * turn allows skipping updates if the differential is 'small'.
2869 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2870 * done) and effective_load() (which is not done because it is too costly).
2872 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2874 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2877 * No need to update load_avg for root_task_group as it is not used.
2879 if (cfs_rq->tg == &root_task_group)
2882 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2883 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2884 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2889 * Called within set_task_rq() right before setting a task's cpu. The
2890 * caller only guarantees p->pi_lock is held; no other assumptions,
2891 * including the state of rq->lock, should be made.
2893 void set_task_rq_fair(struct sched_entity *se,
2894 struct cfs_rq *prev, struct cfs_rq *next)
2896 if (!sched_feat(ATTACH_AGE_LOAD))
2900 * We are supposed to update the task to "current" time, then its up to
2901 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2902 * getting what current time is, so simply throw away the out-of-date
2903 * time. This will result in the wakee task is less decayed, but giving
2904 * the wakee more load sounds not bad.
2906 if (se->avg.last_update_time && prev) {
2907 u64 p_last_update_time;
2908 u64 n_last_update_time;
2910 #ifndef CONFIG_64BIT
2911 u64 p_last_update_time_copy;
2912 u64 n_last_update_time_copy;
2915 p_last_update_time_copy = prev->load_last_update_time_copy;
2916 n_last_update_time_copy = next->load_last_update_time_copy;
2920 p_last_update_time = prev->avg.last_update_time;
2921 n_last_update_time = next->avg.last_update_time;
2923 } while (p_last_update_time != p_last_update_time_copy ||
2924 n_last_update_time != n_last_update_time_copy);
2926 p_last_update_time = prev->avg.last_update_time;
2927 n_last_update_time = next->avg.last_update_time;
2929 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2930 &se->avg, 0, 0, NULL);
2931 se->avg.last_update_time = n_last_update_time;
2935 /* Take into account change of utilization of a child task group */
2937 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
2939 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2940 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
2942 /* Nothing to update */
2946 /* Set new sched_entity's utilization */
2947 se->avg.util_avg = gcfs_rq->avg.util_avg;
2948 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
2950 /* Update parent cfs_rq utilization */
2951 add_positive(&cfs_rq->avg.util_avg, delta);
2952 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
2955 /* Take into account change of load of a child task group */
2957 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
2959 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2960 long delta, load = gcfs_rq->avg.load_avg;
2963 * If the load of group cfs_rq is null, the load of the
2964 * sched_entity will also be null so we can skip the formula
2969 /* Get tg's load and ensure tg_load > 0 */
2970 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
2972 /* Ensure tg_load >= load and updated with current load*/
2973 tg_load -= gcfs_rq->tg_load_avg_contrib;
2977 * We need to compute a correction term in the case that the
2978 * task group is consuming more CPU than a task of equal
2979 * weight. A task with a weight equals to tg->shares will have
2980 * a load less or equal to scale_load_down(tg->shares).
2981 * Similarly, the sched_entities that represent the task group
2982 * at parent level, can't have a load higher than
2983 * scale_load_down(tg->shares). And the Sum of sched_entities'
2984 * load must be <= scale_load_down(tg->shares).
2986 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
2987 /* scale gcfs_rq's load into tg's shares*/
2988 load *= scale_load_down(gcfs_rq->tg->shares);
2993 delta = load - se->avg.load_avg;
2995 /* Nothing to update */
2999 /* Set new sched_entity's load */
3000 se->avg.load_avg = load;
3001 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3003 /* Update parent cfs_rq load */
3004 add_positive(&cfs_rq->avg.load_avg, delta);
3005 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3008 * If the sched_entity is already enqueued, we also have to update the
3009 * runnable load avg.
3012 /* Update parent cfs_rq runnable_load_avg */
3013 add_positive(&cfs_rq->runnable_load_avg, delta);
3014 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3018 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3020 cfs_rq->propagate_avg = 1;
3023 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3025 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3027 if (!cfs_rq->propagate_avg)
3030 cfs_rq->propagate_avg = 0;
3034 /* Update task and its cfs_rq load average */
3035 static inline int propagate_entity_load_avg(struct sched_entity *se)
3037 struct cfs_rq *cfs_rq;
3039 if (entity_is_task(se))
3042 if (!test_and_clear_tg_cfs_propagate(se))
3045 cfs_rq = cfs_rq_of(se);
3047 set_tg_cfs_propagate(cfs_rq);
3049 update_tg_cfs_util(cfs_rq, se);
3050 update_tg_cfs_load(cfs_rq, se);
3055 #else /* CONFIG_FAIR_GROUP_SCHED */
3057 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3059 static inline int propagate_entity_load_avg(struct sched_entity *se)
3064 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3066 #endif /* CONFIG_FAIR_GROUP_SCHED */
3068 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3070 if (&this_rq()->cfs == cfs_rq) {
3072 * There are a few boundary cases this might miss but it should
3073 * get called often enough that that should (hopefully) not be
3074 * a real problem -- added to that it only calls on the local
3075 * CPU, so if we enqueue remotely we'll miss an update, but
3076 * the next tick/schedule should update.
3078 * It will not get called when we go idle, because the idle
3079 * thread is a different class (!fair), nor will the utilization
3080 * number include things like RT tasks.
3082 * As is, the util number is not freq-invariant (we'd have to
3083 * implement arch_scale_freq_capacity() for that).
3087 cpufreq_update_util(rq_of(cfs_rq), 0);
3091 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
3094 * Unsigned subtract and clamp on underflow.
3096 * Explicitly do a load-store to ensure the intermediate value never hits
3097 * memory. This allows lockless observations without ever seeing the negative
3100 #define sub_positive(_ptr, _val) do { \
3101 typeof(_ptr) ptr = (_ptr); \
3102 typeof(*ptr) val = (_val); \
3103 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3107 WRITE_ONCE(*ptr, res); \
3111 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3112 * @now: current time, as per cfs_rq_clock_task()
3113 * @cfs_rq: cfs_rq to update
3114 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3116 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3117 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3118 * post_init_entity_util_avg().
3120 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3122 * Returns true if the load decayed or we removed load.
3124 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3125 * call update_tg_load_avg() when this function returns true.
3128 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3130 struct sched_avg *sa = &cfs_rq->avg;
3131 int decayed, removed = 0, removed_util = 0;
3133 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3134 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3135 sub_positive(&sa->load_avg, r);
3136 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3138 set_tg_cfs_propagate(cfs_rq);
3141 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3142 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3143 sub_positive(&sa->util_avg, r);
3144 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3146 set_tg_cfs_propagate(cfs_rq);
3149 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3150 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3152 #ifndef CONFIG_64BIT
3154 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3157 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
3158 if (cfs_rq == &rq_of(cfs_rq)->cfs)
3159 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
3161 if (update_freq && (decayed || removed_util))
3162 cfs_rq_util_change(cfs_rq);
3164 return decayed || removed;
3168 * Optional action to be done while updating the load average
3170 #define UPDATE_TG 0x1
3171 #define SKIP_AGE_LOAD 0x2
3173 /* Update task and its cfs_rq load average */
3174 static inline void update_load_avg(struct sched_entity *se, int flags)
3176 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3177 u64 now = cfs_rq_clock_task(cfs_rq);
3178 int cpu = cpu_of(rq_of(cfs_rq));
3183 * Track task load average for carrying it to new CPU after migrated, and
3184 * track group sched_entity load average for task_h_load calc in migration
3186 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
3187 __update_load_avg(now, cpu, &se->avg,
3188 se->on_rq * scale_load_down(se->load.weight),
3189 cfs_rq->curr == se, NULL);
3192 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3193 decayed |= propagate_entity_load_avg(se);
3195 if (decayed && (flags & UPDATE_TG))
3196 update_tg_load_avg(cfs_rq, 0);
3198 if (entity_is_task(se)) {
3199 #ifdef CONFIG_SCHED_WALT
3200 ptr = (void *)&(task_of(se)->ravg);
3202 trace_sched_load_avg_task(task_of(se), &se->avg, ptr);
3207 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3208 * @cfs_rq: cfs_rq to attach to
3209 * @se: sched_entity to attach
3211 * Must call update_cfs_rq_load_avg() before this, since we rely on
3212 * cfs_rq->avg.last_update_time being current.
3214 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3216 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3217 cfs_rq->avg.load_avg += se->avg.load_avg;
3218 cfs_rq->avg.load_sum += se->avg.load_sum;
3219 cfs_rq->avg.util_avg += se->avg.util_avg;
3220 cfs_rq->avg.util_sum += se->avg.util_sum;
3221 set_tg_cfs_propagate(cfs_rq);
3223 cfs_rq_util_change(cfs_rq);
3227 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3228 * @cfs_rq: cfs_rq to detach from
3229 * @se: sched_entity to detach
3231 * Must call update_cfs_rq_load_avg() before this, since we rely on
3232 * cfs_rq->avg.last_update_time being current.
3234 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3237 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3238 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3239 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3240 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3241 set_tg_cfs_propagate(cfs_rq);
3243 cfs_rq_util_change(cfs_rq);
3246 /* Add the load generated by se into cfs_rq's load average */
3248 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3250 struct sched_avg *sa = &se->avg;
3252 cfs_rq->runnable_load_avg += sa->load_avg;
3253 cfs_rq->runnable_load_sum += sa->load_sum;
3255 if (!sa->last_update_time) {
3256 attach_entity_load_avg(cfs_rq, se);
3257 update_tg_load_avg(cfs_rq, 0);
3261 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3263 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3265 cfs_rq->runnable_load_avg =
3266 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3267 cfs_rq->runnable_load_sum =
3268 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3271 #ifndef CONFIG_64BIT
3272 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3274 u64 last_update_time_copy;
3275 u64 last_update_time;
3278 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3280 last_update_time = cfs_rq->avg.last_update_time;
3281 } while (last_update_time != last_update_time_copy);
3283 return last_update_time;
3286 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3288 return cfs_rq->avg.last_update_time;
3293 * Synchronize entity load avg of dequeued entity without locking
3296 void sync_entity_load_avg(struct sched_entity *se)
3298 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3299 u64 last_update_time;
3301 last_update_time = cfs_rq_last_update_time(cfs_rq);
3302 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3306 * Task first catches up with cfs_rq, and then subtract
3307 * itself from the cfs_rq (task must be off the queue now).
3309 void remove_entity_load_avg(struct sched_entity *se)
3311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3314 * tasks cannot exit without having gone through wake_up_new_task() ->
3315 * post_init_entity_util_avg() which will have added things to the
3316 * cfs_rq, so we can remove unconditionally.
3318 * Similarly for groups, they will have passed through
3319 * post_init_entity_util_avg() before unregister_sched_fair_group()
3323 sync_entity_load_avg(se);
3324 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3325 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3329 * Update the rq's load with the elapsed running time before entering
3330 * idle. if the last scheduled task is not a CFS task, idle_enter will
3331 * be the only way to update the runnable statistic.
3333 void idle_enter_fair(struct rq *this_rq)
3338 * Update the rq's load with the elapsed idle time before a task is
3339 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3340 * be the only way to update the runnable statistic.
3342 void idle_exit_fair(struct rq *this_rq)
3346 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3348 return cfs_rq->runnable_load_avg;
3351 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3353 return cfs_rq->avg.load_avg;
3356 static int idle_balance(struct rq *this_rq);
3358 #else /* CONFIG_SMP */
3361 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3366 #define UPDATE_TG 0x0
3367 #define SKIP_AGE_LOAD 0x0
3369 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3371 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3373 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3374 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3377 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3379 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3381 static inline int idle_balance(struct rq *rq)
3386 #endif /* CONFIG_SMP */
3388 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3390 #ifdef CONFIG_SCHEDSTATS
3391 struct task_struct *tsk = NULL;
3393 if (entity_is_task(se))
3396 if (se->statistics.sleep_start) {
3397 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3402 if (unlikely(delta > se->statistics.sleep_max))
3403 se->statistics.sleep_max = delta;
3405 se->statistics.sleep_start = 0;
3406 se->statistics.sum_sleep_runtime += delta;
3409 account_scheduler_latency(tsk, delta >> 10, 1);
3410 trace_sched_stat_sleep(tsk, delta);
3413 if (se->statistics.block_start) {
3414 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3419 if (unlikely(delta > se->statistics.block_max))
3420 se->statistics.block_max = delta;
3422 se->statistics.block_start = 0;
3423 se->statistics.sum_sleep_runtime += delta;
3426 if (tsk->in_iowait) {
3427 se->statistics.iowait_sum += delta;
3428 se->statistics.iowait_count++;
3429 trace_sched_stat_iowait(tsk, delta);
3432 trace_sched_stat_blocked(tsk, delta);
3433 trace_sched_blocked_reason(tsk);
3436 * Blocking time is in units of nanosecs, so shift by
3437 * 20 to get a milliseconds-range estimation of the
3438 * amount of time that the task spent sleeping:
3440 if (unlikely(prof_on == SLEEP_PROFILING)) {
3441 profile_hits(SLEEP_PROFILING,
3442 (void *)get_wchan(tsk),
3445 account_scheduler_latency(tsk, delta >> 10, 0);
3451 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3453 #ifdef CONFIG_SCHED_DEBUG
3454 s64 d = se->vruntime - cfs_rq->min_vruntime;
3459 if (d > 3*sysctl_sched_latency)
3460 schedstat_inc(cfs_rq, nr_spread_over);
3465 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3467 u64 vruntime = cfs_rq->min_vruntime;
3470 * The 'current' period is already promised to the current tasks,
3471 * however the extra weight of the new task will slow them down a
3472 * little, place the new task so that it fits in the slot that
3473 * stays open at the end.
3475 if (initial && sched_feat(START_DEBIT))
3476 vruntime += sched_vslice(cfs_rq, se);
3478 /* sleeps up to a single latency don't count. */
3480 unsigned long thresh = sysctl_sched_latency;
3483 * Halve their sleep time's effect, to allow
3484 * for a gentler effect of sleepers:
3486 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3492 /* ensure we never gain time by being placed backwards. */
3493 se->vruntime = max_vruntime(se->vruntime, vruntime);
3496 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3499 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3502 * Update the normalized vruntime before updating min_vruntime
3503 * through calling update_curr().
3505 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3506 se->vruntime += cfs_rq->min_vruntime;
3509 * Update run-time statistics of the 'current'.
3511 update_curr(cfs_rq);
3512 update_load_avg(se, UPDATE_TG);
3513 enqueue_entity_load_avg(cfs_rq, se);
3514 update_cfs_shares(se);
3515 account_entity_enqueue(cfs_rq, se);
3517 if (flags & ENQUEUE_WAKEUP) {
3518 place_entity(cfs_rq, se, 0);
3519 enqueue_sleeper(cfs_rq, se);
3522 update_stats_enqueue(cfs_rq, se);
3523 check_spread(cfs_rq, se);
3524 if (se != cfs_rq->curr)
3525 __enqueue_entity(cfs_rq, se);
3528 if (cfs_rq->nr_running == 1) {
3529 list_add_leaf_cfs_rq(cfs_rq);
3530 check_enqueue_throttle(cfs_rq);
3534 static void __clear_buddies_last(struct sched_entity *se)
3536 for_each_sched_entity(se) {
3537 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3538 if (cfs_rq->last != se)
3541 cfs_rq->last = NULL;
3545 static void __clear_buddies_next(struct sched_entity *se)
3547 for_each_sched_entity(se) {
3548 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3549 if (cfs_rq->next != se)
3552 cfs_rq->next = NULL;
3556 static void __clear_buddies_skip(struct sched_entity *se)
3558 for_each_sched_entity(se) {
3559 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3560 if (cfs_rq->skip != se)
3563 cfs_rq->skip = NULL;
3567 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3569 if (cfs_rq->last == se)
3570 __clear_buddies_last(se);
3572 if (cfs_rq->next == se)
3573 __clear_buddies_next(se);
3575 if (cfs_rq->skip == se)
3576 __clear_buddies_skip(se);
3579 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3582 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3585 * Update run-time statistics of the 'current'.
3587 update_curr(cfs_rq);
3590 * When dequeuing a sched_entity, we must:
3591 * - Update loads to have both entity and cfs_rq synced with now.
3592 * - Substract its load from the cfs_rq->runnable_avg.
3593 * - Substract its previous weight from cfs_rq->load.weight.
3594 * - For group entity, update its weight to reflect the new share
3595 * of its group cfs_rq.
3597 update_load_avg(se, UPDATE_TG);
3598 dequeue_entity_load_avg(cfs_rq, se);
3600 update_stats_dequeue(cfs_rq, se);
3601 if (flags & DEQUEUE_SLEEP) {
3602 #ifdef CONFIG_SCHEDSTATS
3603 if (entity_is_task(se)) {
3604 struct task_struct *tsk = task_of(se);
3606 if (tsk->state & TASK_INTERRUPTIBLE)
3607 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3608 if (tsk->state & TASK_UNINTERRUPTIBLE)
3609 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3614 clear_buddies(cfs_rq, se);
3616 if (se != cfs_rq->curr)
3617 __dequeue_entity(cfs_rq, se);
3619 account_entity_dequeue(cfs_rq, se);
3622 * Normalize the entity after updating the min_vruntime because the
3623 * update can refer to the ->curr item and we need to reflect this
3624 * movement in our normalized position.
3626 if (!(flags & DEQUEUE_SLEEP))
3627 se->vruntime -= cfs_rq->min_vruntime;
3629 /* return excess runtime on last dequeue */
3630 return_cfs_rq_runtime(cfs_rq);
3632 update_min_vruntime(cfs_rq);
3633 update_cfs_shares(se);
3637 * Preempt the current task with a newly woken task if needed:
3640 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3642 unsigned long ideal_runtime, delta_exec;
3643 struct sched_entity *se;
3646 ideal_runtime = sched_slice(cfs_rq, curr);
3647 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3648 if (delta_exec > ideal_runtime) {
3649 resched_curr(rq_of(cfs_rq));
3651 * The current task ran long enough, ensure it doesn't get
3652 * re-elected due to buddy favours.
3654 clear_buddies(cfs_rq, curr);
3659 * Ensure that a task that missed wakeup preemption by a
3660 * narrow margin doesn't have to wait for a full slice.
3661 * This also mitigates buddy induced latencies under load.
3663 if (delta_exec < sysctl_sched_min_granularity)
3666 se = __pick_first_entity(cfs_rq);
3667 delta = curr->vruntime - se->vruntime;
3672 if (delta > ideal_runtime)
3673 resched_curr(rq_of(cfs_rq));
3677 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3679 /* 'current' is not kept within the tree. */
3682 * Any task has to be enqueued before it get to execute on
3683 * a CPU. So account for the time it spent waiting on the
3686 update_stats_wait_end(cfs_rq, se);
3687 __dequeue_entity(cfs_rq, se);
3688 update_load_avg(se, UPDATE_TG);
3691 update_stats_curr_start(cfs_rq, se);
3693 #ifdef CONFIG_SCHEDSTATS
3695 * Track our maximum slice length, if the CPU's load is at
3696 * least twice that of our own weight (i.e. dont track it
3697 * when there are only lesser-weight tasks around):
3699 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3700 se->statistics.slice_max = max(se->statistics.slice_max,
3701 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3704 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3708 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3711 * Pick the next process, keeping these things in mind, in this order:
3712 * 1) keep things fair between processes/task groups
3713 * 2) pick the "next" process, since someone really wants that to run
3714 * 3) pick the "last" process, for cache locality
3715 * 4) do not run the "skip" process, if something else is available
3717 static struct sched_entity *
3718 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3720 struct sched_entity *left = __pick_first_entity(cfs_rq);
3721 struct sched_entity *se;
3724 * If curr is set we have to see if its left of the leftmost entity
3725 * still in the tree, provided there was anything in the tree at all.
3727 if (!left || (curr && entity_before(curr, left)))
3730 se = left; /* ideally we run the leftmost entity */
3733 * Avoid running the skip buddy, if running something else can
3734 * be done without getting too unfair.
3736 if (cfs_rq->skip == se) {
3737 struct sched_entity *second;
3740 second = __pick_first_entity(cfs_rq);
3742 second = __pick_next_entity(se);
3743 if (!second || (curr && entity_before(curr, second)))
3747 if (second && wakeup_preempt_entity(second, left) < 1)
3752 * Prefer last buddy, try to return the CPU to a preempted task.
3754 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3758 * Someone really wants this to run. If it's not unfair, run it.
3760 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3763 clear_buddies(cfs_rq, se);
3768 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3770 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3773 * If still on the runqueue then deactivate_task()
3774 * was not called and update_curr() has to be done:
3777 update_curr(cfs_rq);
3779 /* throttle cfs_rqs exceeding runtime */
3780 check_cfs_rq_runtime(cfs_rq);
3782 check_spread(cfs_rq, prev);
3784 update_stats_wait_start(cfs_rq, prev);
3785 /* Put 'current' back into the tree. */
3786 __enqueue_entity(cfs_rq, prev);
3787 /* in !on_rq case, update occurred at dequeue */
3788 update_load_avg(prev, 0);
3790 cfs_rq->curr = NULL;
3794 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3797 * Update run-time statistics of the 'current'.
3799 update_curr(cfs_rq);
3802 * Ensure that runnable average is periodically updated.
3804 update_load_avg(curr, UPDATE_TG);
3805 update_cfs_shares(curr);
3807 #ifdef CONFIG_SCHED_HRTICK
3809 * queued ticks are scheduled to match the slice, so don't bother
3810 * validating it and just reschedule.
3813 resched_curr(rq_of(cfs_rq));
3817 * don't let the period tick interfere with the hrtick preemption
3819 if (!sched_feat(DOUBLE_TICK) &&
3820 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3824 if (cfs_rq->nr_running > 1)
3825 check_preempt_tick(cfs_rq, curr);
3829 /**************************************************
3830 * CFS bandwidth control machinery
3833 #ifdef CONFIG_CFS_BANDWIDTH
3835 #ifdef HAVE_JUMP_LABEL
3836 static struct static_key __cfs_bandwidth_used;
3838 static inline bool cfs_bandwidth_used(void)
3840 return static_key_false(&__cfs_bandwidth_used);
3843 void cfs_bandwidth_usage_inc(void)
3845 static_key_slow_inc(&__cfs_bandwidth_used);
3848 void cfs_bandwidth_usage_dec(void)
3850 static_key_slow_dec(&__cfs_bandwidth_used);
3852 #else /* HAVE_JUMP_LABEL */
3853 static bool cfs_bandwidth_used(void)
3858 void cfs_bandwidth_usage_inc(void) {}
3859 void cfs_bandwidth_usage_dec(void) {}
3860 #endif /* HAVE_JUMP_LABEL */
3863 * default period for cfs group bandwidth.
3864 * default: 0.1s, units: nanoseconds
3866 static inline u64 default_cfs_period(void)
3868 return 100000000ULL;
3871 static inline u64 sched_cfs_bandwidth_slice(void)
3873 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3877 * Replenish runtime according to assigned quota and update expiration time.
3878 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3879 * additional synchronization around rq->lock.
3881 * requires cfs_b->lock
3883 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3887 if (cfs_b->quota == RUNTIME_INF)
3890 now = sched_clock_cpu(smp_processor_id());
3891 cfs_b->runtime = cfs_b->quota;
3892 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3895 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3897 return &tg->cfs_bandwidth;
3900 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3901 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3903 if (unlikely(cfs_rq->throttle_count))
3904 return cfs_rq->throttled_clock_task;
3906 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3909 /* returns 0 on failure to allocate runtime */
3910 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3912 struct task_group *tg = cfs_rq->tg;
3913 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3914 u64 amount = 0, min_amount, expires;
3916 /* note: this is a positive sum as runtime_remaining <= 0 */
3917 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3919 raw_spin_lock(&cfs_b->lock);
3920 if (cfs_b->quota == RUNTIME_INF)
3921 amount = min_amount;
3923 start_cfs_bandwidth(cfs_b);
3925 if (cfs_b->runtime > 0) {
3926 amount = min(cfs_b->runtime, min_amount);
3927 cfs_b->runtime -= amount;
3931 expires = cfs_b->runtime_expires;
3932 raw_spin_unlock(&cfs_b->lock);
3934 cfs_rq->runtime_remaining += amount;
3936 * we may have advanced our local expiration to account for allowed
3937 * spread between our sched_clock and the one on which runtime was
3940 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3941 cfs_rq->runtime_expires = expires;
3943 return cfs_rq->runtime_remaining > 0;
3947 * Note: This depends on the synchronization provided by sched_clock and the
3948 * fact that rq->clock snapshots this value.
3950 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3952 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3954 /* if the deadline is ahead of our clock, nothing to do */
3955 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3958 if (cfs_rq->runtime_remaining < 0)
3962 * If the local deadline has passed we have to consider the
3963 * possibility that our sched_clock is 'fast' and the global deadline
3964 * has not truly expired.
3966 * Fortunately we can check determine whether this the case by checking
3967 * whether the global deadline has advanced. It is valid to compare
3968 * cfs_b->runtime_expires without any locks since we only care about
3969 * exact equality, so a partial write will still work.
3972 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3973 /* extend local deadline, drift is bounded above by 2 ticks */
3974 cfs_rq->runtime_expires += TICK_NSEC;
3976 /* global deadline is ahead, expiration has passed */
3977 cfs_rq->runtime_remaining = 0;
3981 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3983 /* dock delta_exec before expiring quota (as it could span periods) */
3984 cfs_rq->runtime_remaining -= delta_exec;
3985 expire_cfs_rq_runtime(cfs_rq);
3987 if (likely(cfs_rq->runtime_remaining > 0))
3991 * if we're unable to extend our runtime we resched so that the active
3992 * hierarchy can be throttled
3994 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3995 resched_curr(rq_of(cfs_rq));
3998 static __always_inline
3999 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4001 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4004 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4007 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4009 return cfs_bandwidth_used() && cfs_rq->throttled;
4012 /* check whether cfs_rq, or any parent, is throttled */
4013 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4015 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4019 * Ensure that neither of the group entities corresponding to src_cpu or
4020 * dest_cpu are members of a throttled hierarchy when performing group
4021 * load-balance operations.
4023 static inline int throttled_lb_pair(struct task_group *tg,
4024 int src_cpu, int dest_cpu)
4026 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4028 src_cfs_rq = tg->cfs_rq[src_cpu];
4029 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4031 return throttled_hierarchy(src_cfs_rq) ||
4032 throttled_hierarchy(dest_cfs_rq);
4035 /* updated child weight may affect parent so we have to do this bottom up */
4036 static int tg_unthrottle_up(struct task_group *tg, void *data)
4038 struct rq *rq = data;
4039 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4041 cfs_rq->throttle_count--;
4043 if (!cfs_rq->throttle_count) {
4044 /* adjust cfs_rq_clock_task() */
4045 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4046 cfs_rq->throttled_clock_task;
4053 static int tg_throttle_down(struct task_group *tg, void *data)
4055 struct rq *rq = data;
4056 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4058 /* group is entering throttled state, stop time */
4059 if (!cfs_rq->throttle_count)
4060 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4061 cfs_rq->throttle_count++;
4066 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4068 struct rq *rq = rq_of(cfs_rq);
4069 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4070 struct sched_entity *se;
4071 long task_delta, dequeue = 1;
4074 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4076 /* freeze hierarchy runnable averages while throttled */
4078 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4081 task_delta = cfs_rq->h_nr_running;
4082 for_each_sched_entity(se) {
4083 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4084 /* throttled entity or throttle-on-deactivate */
4089 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4090 qcfs_rq->h_nr_running -= task_delta;
4092 if (qcfs_rq->load.weight)
4097 sub_nr_running(rq, task_delta);
4099 cfs_rq->throttled = 1;
4100 cfs_rq->throttled_clock = rq_clock(rq);
4101 raw_spin_lock(&cfs_b->lock);
4102 empty = list_empty(&cfs_b->throttled_cfs_rq);
4105 * Add to the _head_ of the list, so that an already-started
4106 * distribute_cfs_runtime will not see us
4108 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4111 * If we're the first throttled task, make sure the bandwidth
4115 start_cfs_bandwidth(cfs_b);
4117 raw_spin_unlock(&cfs_b->lock);
4120 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4122 struct rq *rq = rq_of(cfs_rq);
4123 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4124 struct sched_entity *se;
4128 se = cfs_rq->tg->se[cpu_of(rq)];
4130 cfs_rq->throttled = 0;
4132 update_rq_clock(rq);
4134 raw_spin_lock(&cfs_b->lock);
4135 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4136 list_del_rcu(&cfs_rq->throttled_list);
4137 raw_spin_unlock(&cfs_b->lock);
4139 /* update hierarchical throttle state */
4140 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4142 if (!cfs_rq->load.weight)
4145 task_delta = cfs_rq->h_nr_running;
4146 for_each_sched_entity(se) {
4150 cfs_rq = cfs_rq_of(se);
4152 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4153 cfs_rq->h_nr_running += task_delta;
4155 if (cfs_rq_throttled(cfs_rq))
4160 add_nr_running(rq, task_delta);
4162 /* determine whether we need to wake up potentially idle cpu */
4163 if (rq->curr == rq->idle && rq->cfs.nr_running)
4167 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4168 u64 remaining, u64 expires)
4170 struct cfs_rq *cfs_rq;
4172 u64 starting_runtime = remaining;
4175 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4177 struct rq *rq = rq_of(cfs_rq);
4179 raw_spin_lock(&rq->lock);
4180 if (!cfs_rq_throttled(cfs_rq))
4183 runtime = -cfs_rq->runtime_remaining + 1;
4184 if (runtime > remaining)
4185 runtime = remaining;
4186 remaining -= runtime;
4188 cfs_rq->runtime_remaining += runtime;
4189 cfs_rq->runtime_expires = expires;
4191 /* we check whether we're throttled above */
4192 if (cfs_rq->runtime_remaining > 0)
4193 unthrottle_cfs_rq(cfs_rq);
4196 raw_spin_unlock(&rq->lock);
4203 return starting_runtime - remaining;
4207 * Responsible for refilling a task_group's bandwidth and unthrottling its
4208 * cfs_rqs as appropriate. If there has been no activity within the last
4209 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4210 * used to track this state.
4212 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4214 u64 runtime, runtime_expires;
4217 /* no need to continue the timer with no bandwidth constraint */
4218 if (cfs_b->quota == RUNTIME_INF)
4219 goto out_deactivate;
4221 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4222 cfs_b->nr_periods += overrun;
4225 * idle depends on !throttled (for the case of a large deficit), and if
4226 * we're going inactive then everything else can be deferred
4228 if (cfs_b->idle && !throttled)
4229 goto out_deactivate;
4231 __refill_cfs_bandwidth_runtime(cfs_b);
4234 /* mark as potentially idle for the upcoming period */
4239 /* account preceding periods in which throttling occurred */
4240 cfs_b->nr_throttled += overrun;
4242 runtime_expires = cfs_b->runtime_expires;
4245 * This check is repeated as we are holding onto the new bandwidth while
4246 * we unthrottle. This can potentially race with an unthrottled group
4247 * trying to acquire new bandwidth from the global pool. This can result
4248 * in us over-using our runtime if it is all used during this loop, but
4249 * only by limited amounts in that extreme case.
4251 while (throttled && cfs_b->runtime > 0) {
4252 runtime = cfs_b->runtime;
4253 raw_spin_unlock(&cfs_b->lock);
4254 /* we can't nest cfs_b->lock while distributing bandwidth */
4255 runtime = distribute_cfs_runtime(cfs_b, runtime,
4257 raw_spin_lock(&cfs_b->lock);
4259 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4261 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4265 * While we are ensured activity in the period following an
4266 * unthrottle, this also covers the case in which the new bandwidth is
4267 * insufficient to cover the existing bandwidth deficit. (Forcing the
4268 * timer to remain active while there are any throttled entities.)
4278 /* a cfs_rq won't donate quota below this amount */
4279 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4280 /* minimum remaining period time to redistribute slack quota */
4281 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4282 /* how long we wait to gather additional slack before distributing */
4283 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4286 * Are we near the end of the current quota period?
4288 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4289 * hrtimer base being cleared by hrtimer_start. In the case of
4290 * migrate_hrtimers, base is never cleared, so we are fine.
4292 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4294 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4297 /* if the call-back is running a quota refresh is already occurring */
4298 if (hrtimer_callback_running(refresh_timer))
4301 /* is a quota refresh about to occur? */
4302 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4303 if (remaining < min_expire)
4309 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4311 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4313 /* if there's a quota refresh soon don't bother with slack */
4314 if (runtime_refresh_within(cfs_b, min_left))
4317 hrtimer_start(&cfs_b->slack_timer,
4318 ns_to_ktime(cfs_bandwidth_slack_period),
4322 /* we know any runtime found here is valid as update_curr() precedes return */
4323 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4325 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4326 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4328 if (slack_runtime <= 0)
4331 raw_spin_lock(&cfs_b->lock);
4332 if (cfs_b->quota != RUNTIME_INF &&
4333 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4334 cfs_b->runtime += slack_runtime;
4336 /* we are under rq->lock, defer unthrottling using a timer */
4337 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4338 !list_empty(&cfs_b->throttled_cfs_rq))
4339 start_cfs_slack_bandwidth(cfs_b);
4341 raw_spin_unlock(&cfs_b->lock);
4343 /* even if it's not valid for return we don't want to try again */
4344 cfs_rq->runtime_remaining -= slack_runtime;
4347 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4349 if (!cfs_bandwidth_used())
4352 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4355 __return_cfs_rq_runtime(cfs_rq);
4359 * This is done with a timer (instead of inline with bandwidth return) since
4360 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4362 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4364 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4367 /* confirm we're still not at a refresh boundary */
4368 raw_spin_lock(&cfs_b->lock);
4369 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4370 raw_spin_unlock(&cfs_b->lock);
4374 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4375 runtime = cfs_b->runtime;
4377 expires = cfs_b->runtime_expires;
4378 raw_spin_unlock(&cfs_b->lock);
4383 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4385 raw_spin_lock(&cfs_b->lock);
4386 if (expires == cfs_b->runtime_expires)
4387 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4388 raw_spin_unlock(&cfs_b->lock);
4392 * When a group wakes up we want to make sure that its quota is not already
4393 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4394 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4396 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4398 if (!cfs_bandwidth_used())
4401 /* Synchronize hierarchical throttle counter: */
4402 if (unlikely(!cfs_rq->throttle_uptodate)) {
4403 struct rq *rq = rq_of(cfs_rq);
4404 struct cfs_rq *pcfs_rq;
4405 struct task_group *tg;
4407 cfs_rq->throttle_uptodate = 1;
4409 /* Get closest up-to-date node, because leaves go first: */
4410 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4411 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4412 if (pcfs_rq->throttle_uptodate)
4416 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4417 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4421 /* an active group must be handled by the update_curr()->put() path */
4422 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4425 /* ensure the group is not already throttled */
4426 if (cfs_rq_throttled(cfs_rq))
4429 /* update runtime allocation */
4430 account_cfs_rq_runtime(cfs_rq, 0);
4431 if (cfs_rq->runtime_remaining <= 0)
4432 throttle_cfs_rq(cfs_rq);
4435 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4436 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4438 if (!cfs_bandwidth_used())
4441 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4445 * it's possible for a throttled entity to be forced into a running
4446 * state (e.g. set_curr_task), in this case we're finished.
4448 if (cfs_rq_throttled(cfs_rq))
4451 throttle_cfs_rq(cfs_rq);
4455 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4457 struct cfs_bandwidth *cfs_b =
4458 container_of(timer, struct cfs_bandwidth, slack_timer);
4460 do_sched_cfs_slack_timer(cfs_b);
4462 return HRTIMER_NORESTART;
4465 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4467 struct cfs_bandwidth *cfs_b =
4468 container_of(timer, struct cfs_bandwidth, period_timer);
4472 raw_spin_lock(&cfs_b->lock);
4474 overrun = hrtimer_forward_now(timer, cfs_b->period);
4478 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4481 cfs_b->period_active = 0;
4482 raw_spin_unlock(&cfs_b->lock);
4484 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4487 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4489 raw_spin_lock_init(&cfs_b->lock);
4491 cfs_b->quota = RUNTIME_INF;
4492 cfs_b->period = ns_to_ktime(default_cfs_period());
4494 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4495 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4496 cfs_b->period_timer.function = sched_cfs_period_timer;
4497 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4498 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4501 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4503 cfs_rq->runtime_enabled = 0;
4504 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4507 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4509 lockdep_assert_held(&cfs_b->lock);
4511 if (!cfs_b->period_active) {
4512 cfs_b->period_active = 1;
4513 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4514 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4518 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4520 /* init_cfs_bandwidth() was not called */
4521 if (!cfs_b->throttled_cfs_rq.next)
4524 hrtimer_cancel(&cfs_b->period_timer);
4525 hrtimer_cancel(&cfs_b->slack_timer);
4528 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4530 struct cfs_rq *cfs_rq;
4532 for_each_leaf_cfs_rq(rq, cfs_rq) {
4533 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4535 raw_spin_lock(&cfs_b->lock);
4536 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4537 raw_spin_unlock(&cfs_b->lock);
4541 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4543 struct cfs_rq *cfs_rq;
4545 for_each_leaf_cfs_rq(rq, cfs_rq) {
4546 if (!cfs_rq->runtime_enabled)
4550 * clock_task is not advancing so we just need to make sure
4551 * there's some valid quota amount
4553 cfs_rq->runtime_remaining = 1;
4555 * Offline rq is schedulable till cpu is completely disabled
4556 * in take_cpu_down(), so we prevent new cfs throttling here.
4558 cfs_rq->runtime_enabled = 0;
4560 if (cfs_rq_throttled(cfs_rq))
4561 unthrottle_cfs_rq(cfs_rq);
4565 #else /* CONFIG_CFS_BANDWIDTH */
4566 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4568 return rq_clock_task(rq_of(cfs_rq));
4571 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4572 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4573 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4574 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4576 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4581 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4586 static inline int throttled_lb_pair(struct task_group *tg,
4587 int src_cpu, int dest_cpu)
4592 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4594 #ifdef CONFIG_FAIR_GROUP_SCHED
4595 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4598 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4602 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4603 static inline void update_runtime_enabled(struct rq *rq) {}
4604 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4606 #endif /* CONFIG_CFS_BANDWIDTH */
4608 /**************************************************
4609 * CFS operations on tasks:
4612 #ifdef CONFIG_SCHED_HRTICK
4613 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4615 struct sched_entity *se = &p->se;
4616 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4618 WARN_ON(task_rq(p) != rq);
4620 if (cfs_rq->nr_running > 1) {
4621 u64 slice = sched_slice(cfs_rq, se);
4622 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4623 s64 delta = slice - ran;
4630 hrtick_start(rq, delta);
4635 * called from enqueue/dequeue and updates the hrtick when the
4636 * current task is from our class and nr_running is low enough
4639 static void hrtick_update(struct rq *rq)
4641 struct task_struct *curr = rq->curr;
4643 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4646 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4647 hrtick_start_fair(rq, curr);
4649 #else /* !CONFIG_SCHED_HRTICK */
4651 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4655 static inline void hrtick_update(struct rq *rq)
4661 static bool __cpu_overutilized(int cpu, int delta);
4662 static bool cpu_overutilized(int cpu);
4663 unsigned long boosted_cpu_util(int cpu);
4665 #define boosted_cpu_util(cpu) cpu_util_freq(cpu)
4669 * The enqueue_task method is called before nr_running is
4670 * increased. Here we update the fair scheduling stats and
4671 * then put the task into the rbtree:
4674 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4676 struct cfs_rq *cfs_rq;
4677 struct sched_entity *se = &p->se;
4679 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4683 * If in_iowait is set, the code below may not trigger any cpufreq
4684 * utilization updates, so do it here explicitly with the IOWAIT flag
4688 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4690 for_each_sched_entity(se) {
4693 cfs_rq = cfs_rq_of(se);
4694 enqueue_entity(cfs_rq, se, flags);
4697 * end evaluation on encountering a throttled cfs_rq
4699 * note: in the case of encountering a throttled cfs_rq we will
4700 * post the final h_nr_running increment below.
4702 if (cfs_rq_throttled(cfs_rq))
4704 cfs_rq->h_nr_running++;
4705 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4707 flags = ENQUEUE_WAKEUP;
4710 for_each_sched_entity(se) {
4711 cfs_rq = cfs_rq_of(se);
4712 cfs_rq->h_nr_running++;
4713 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4715 if (cfs_rq_throttled(cfs_rq))
4718 update_load_avg(se, UPDATE_TG);
4719 update_cfs_shares(se);
4723 add_nr_running(rq, 1);
4728 * Update SchedTune accounting.
4730 * We do it before updating the CPU capacity to ensure the
4731 * boost value of the current task is accounted for in the
4732 * selection of the OPP.
4734 * We do it also in the case where we enqueue a throttled task;
4735 * we could argue that a throttled task should not boost a CPU,
4737 * a) properly implementing CPU boosting considering throttled
4738 * tasks will increase a lot the complexity of the solution
4739 * b) it's not easy to quantify the benefits introduced by
4740 * such a more complex solution.
4741 * Thus, for the time being we go for the simple solution and boost
4742 * also for throttled RQs.
4744 schedtune_enqueue_task(p, cpu_of(rq));
4747 walt_inc_cumulative_runnable_avg(rq, p);
4748 if (!task_new && !rq->rd->overutilized &&
4749 cpu_overutilized(rq->cpu)) {
4750 rq->rd->overutilized = true;
4751 trace_sched_overutilized(true);
4755 #endif /* CONFIG_SMP */
4759 static void set_next_buddy(struct sched_entity *se);
4762 * The dequeue_task method is called before nr_running is
4763 * decreased. We remove the task from the rbtree and
4764 * update the fair scheduling stats:
4766 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4768 struct cfs_rq *cfs_rq;
4769 struct sched_entity *se = &p->se;
4770 int task_sleep = flags & DEQUEUE_SLEEP;
4772 for_each_sched_entity(se) {
4773 cfs_rq = cfs_rq_of(se);
4774 dequeue_entity(cfs_rq, se, flags);
4777 * end evaluation on encountering a throttled cfs_rq
4779 * note: in the case of encountering a throttled cfs_rq we will
4780 * post the final h_nr_running decrement below.
4782 if (cfs_rq_throttled(cfs_rq))
4784 cfs_rq->h_nr_running--;
4785 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4787 /* Don't dequeue parent if it has other entities besides us */
4788 if (cfs_rq->load.weight) {
4789 /* Avoid re-evaluating load for this entity: */
4790 se = parent_entity(se);
4792 * Bias pick_next to pick a task from this cfs_rq, as
4793 * p is sleeping when it is within its sched_slice.
4795 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4799 flags |= DEQUEUE_SLEEP;
4802 for_each_sched_entity(se) {
4803 cfs_rq = cfs_rq_of(se);
4804 cfs_rq->h_nr_running--;
4805 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4807 if (cfs_rq_throttled(cfs_rq))
4810 update_load_avg(se, UPDATE_TG);
4811 update_cfs_shares(se);
4815 sub_nr_running(rq, 1);
4820 * Update SchedTune accounting
4822 * We do it before updating the CPU capacity to ensure the
4823 * boost value of the current task is accounted for in the
4824 * selection of the OPP.
4826 schedtune_dequeue_task(p, cpu_of(rq));
4829 walt_dec_cumulative_runnable_avg(rq, p);
4830 #endif /* CONFIG_SMP */
4838 * per rq 'load' arrray crap; XXX kill this.
4842 * The exact cpuload at various idx values, calculated at every tick would be
4843 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4845 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4846 * on nth tick when cpu may be busy, then we have:
4847 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4848 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4850 * decay_load_missed() below does efficient calculation of
4851 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4852 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4854 * The calculation is approximated on a 128 point scale.
4855 * degrade_zero_ticks is the number of ticks after which load at any
4856 * particular idx is approximated to be zero.
4857 * degrade_factor is a precomputed table, a row for each load idx.
4858 * Each column corresponds to degradation factor for a power of two ticks,
4859 * based on 128 point scale.
4861 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4862 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4864 * With this power of 2 load factors, we can degrade the load n times
4865 * by looking at 1 bits in n and doing as many mult/shift instead of
4866 * n mult/shifts needed by the exact degradation.
4868 #define DEGRADE_SHIFT 7
4869 static const unsigned char
4870 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4871 static const unsigned char
4872 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4873 {0, 0, 0, 0, 0, 0, 0, 0},
4874 {64, 32, 8, 0, 0, 0, 0, 0},
4875 {96, 72, 40, 12, 1, 0, 0},
4876 {112, 98, 75, 43, 15, 1, 0},
4877 {120, 112, 98, 76, 45, 16, 2} };
4880 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4881 * would be when CPU is idle and so we just decay the old load without
4882 * adding any new load.
4884 static unsigned long
4885 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4889 if (!missed_updates)
4892 if (missed_updates >= degrade_zero_ticks[idx])
4896 return load >> missed_updates;
4898 while (missed_updates) {
4899 if (missed_updates % 2)
4900 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4902 missed_updates >>= 1;
4909 * Update rq->cpu_load[] statistics. This function is usually called every
4910 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4911 * every tick. We fix it up based on jiffies.
4913 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4914 unsigned long pending_updates)
4918 this_rq->nr_load_updates++;
4920 /* Update our load: */
4921 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4922 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4923 unsigned long old_load, new_load;
4925 /* scale is effectively 1 << i now, and >> i divides by scale */
4927 old_load = this_rq->cpu_load[i];
4928 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4929 new_load = this_load;
4931 * Round up the averaging division if load is increasing. This
4932 * prevents us from getting stuck on 9 if the load is 10, for
4935 if (new_load > old_load)
4936 new_load += scale - 1;
4938 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4941 sched_avg_update(this_rq);
4944 /* Used instead of source_load when we know the type == 0 */
4945 static unsigned long weighted_cpuload(const int cpu)
4947 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4950 #ifdef CONFIG_NO_HZ_COMMON
4952 * There is no sane way to deal with nohz on smp when using jiffies because the
4953 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4954 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4956 * Therefore we cannot use the delta approach from the regular tick since that
4957 * would seriously skew the load calculation. However we'll make do for those
4958 * updates happening while idle (nohz_idle_balance) or coming out of idle
4959 * (tick_nohz_idle_exit).
4961 * This means we might still be one tick off for nohz periods.
4965 * Called from nohz_idle_balance() to update the load ratings before doing the
4968 static void update_idle_cpu_load(struct rq *this_rq)
4970 unsigned long curr_jiffies = READ_ONCE(jiffies);
4971 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4972 unsigned long pending_updates;
4975 * bail if there's load or we're actually up-to-date.
4977 if (load || curr_jiffies == this_rq->last_load_update_tick)
4980 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4981 this_rq->last_load_update_tick = curr_jiffies;
4983 __update_cpu_load(this_rq, load, pending_updates);
4987 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4989 void update_cpu_load_nohz(void)
4991 struct rq *this_rq = this_rq();
4992 unsigned long curr_jiffies = READ_ONCE(jiffies);
4993 unsigned long pending_updates;
4995 if (curr_jiffies == this_rq->last_load_update_tick)
4998 raw_spin_lock(&this_rq->lock);
4999 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5000 if (pending_updates) {
5001 this_rq->last_load_update_tick = curr_jiffies;
5003 * We were idle, this means load 0, the current load might be
5004 * !0 due to remote wakeups and the sort.
5006 __update_cpu_load(this_rq, 0, pending_updates);
5008 raw_spin_unlock(&this_rq->lock);
5010 #endif /* CONFIG_NO_HZ */
5013 * Called from scheduler_tick()
5015 void update_cpu_load_active(struct rq *this_rq)
5017 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5019 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
5021 this_rq->last_load_update_tick = jiffies;
5022 __update_cpu_load(this_rq, load, 1);
5026 * Return a low guess at the load of a migration-source cpu weighted
5027 * according to the scheduling class and "nice" value.
5029 * We want to under-estimate the load of migration sources, to
5030 * balance conservatively.
5032 static unsigned long source_load(int cpu, int type)
5034 struct rq *rq = cpu_rq(cpu);
5035 unsigned long total = weighted_cpuload(cpu);
5037 if (type == 0 || !sched_feat(LB_BIAS))
5040 return min(rq->cpu_load[type-1], total);
5044 * Return a high guess at the load of a migration-target cpu weighted
5045 * according to the scheduling class and "nice" value.
5047 static unsigned long target_load(int cpu, int type)
5049 struct rq *rq = cpu_rq(cpu);
5050 unsigned long total = weighted_cpuload(cpu);
5052 if (type == 0 || !sched_feat(LB_BIAS))
5055 return max(rq->cpu_load[type-1], total);
5059 static unsigned long cpu_avg_load_per_task(int cpu)
5061 struct rq *rq = cpu_rq(cpu);
5062 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5063 unsigned long load_avg = weighted_cpuload(cpu);
5066 return load_avg / nr_running;
5071 static void record_wakee(struct task_struct *p)
5074 * Rough decay (wiping) for cost saving, don't worry
5075 * about the boundary, really active task won't care
5078 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5079 current->wakee_flips >>= 1;
5080 current->wakee_flip_decay_ts = jiffies;
5083 if (current->last_wakee != p) {
5084 current->last_wakee = p;
5085 current->wakee_flips++;
5089 static void task_waking_fair(struct task_struct *p)
5091 struct sched_entity *se = &p->se;
5092 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5095 #ifndef CONFIG_64BIT
5096 u64 min_vruntime_copy;
5099 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5101 min_vruntime = cfs_rq->min_vruntime;
5102 } while (min_vruntime != min_vruntime_copy);
5104 min_vruntime = cfs_rq->min_vruntime;
5107 se->vruntime -= min_vruntime;
5111 #ifdef CONFIG_FAIR_GROUP_SCHED
5113 * effective_load() calculates the load change as seen from the root_task_group
5115 * Adding load to a group doesn't make a group heavier, but can cause movement
5116 * of group shares between cpus. Assuming the shares were perfectly aligned one
5117 * can calculate the shift in shares.
5119 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5120 * on this @cpu and results in a total addition (subtraction) of @wg to the
5121 * total group weight.
5123 * Given a runqueue weight distribution (rw_i) we can compute a shares
5124 * distribution (s_i) using:
5126 * s_i = rw_i / \Sum rw_j (1)
5128 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5129 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5130 * shares distribution (s_i):
5132 * rw_i = { 2, 4, 1, 0 }
5133 * s_i = { 2/7, 4/7, 1/7, 0 }
5135 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5136 * task used to run on and the CPU the waker is running on), we need to
5137 * compute the effect of waking a task on either CPU and, in case of a sync
5138 * wakeup, compute the effect of the current task going to sleep.
5140 * So for a change of @wl to the local @cpu with an overall group weight change
5141 * of @wl we can compute the new shares distribution (s'_i) using:
5143 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5145 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5146 * differences in waking a task to CPU 0. The additional task changes the
5147 * weight and shares distributions like:
5149 * rw'_i = { 3, 4, 1, 0 }
5150 * s'_i = { 3/8, 4/8, 1/8, 0 }
5152 * We can then compute the difference in effective weight by using:
5154 * dw_i = S * (s'_i - s_i) (3)
5156 * Where 'S' is the group weight as seen by its parent.
5158 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5159 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5160 * 4/7) times the weight of the group.
5162 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5164 struct sched_entity *se = tg->se[cpu];
5166 if (!tg->parent) /* the trivial, non-cgroup case */
5169 for_each_sched_entity(se) {
5170 struct cfs_rq *cfs_rq = se->my_q;
5171 long W, w = cfs_rq_load_avg(cfs_rq);
5176 * W = @wg + \Sum rw_j
5178 W = wg + atomic_long_read(&tg->load_avg);
5180 /* Ensure \Sum rw_j >= rw_i */
5181 W -= cfs_rq->tg_load_avg_contrib;
5190 * wl = S * s'_i; see (2)
5193 wl = (w * (long)tg->shares) / W;
5198 * Per the above, wl is the new se->load.weight value; since
5199 * those are clipped to [MIN_SHARES, ...) do so now. See
5200 * calc_cfs_shares().
5202 if (wl < MIN_SHARES)
5206 * wl = dw_i = S * (s'_i - s_i); see (3)
5208 wl -= se->avg.load_avg;
5211 * Recursively apply this logic to all parent groups to compute
5212 * the final effective load change on the root group. Since
5213 * only the @tg group gets extra weight, all parent groups can
5214 * only redistribute existing shares. @wl is the shift in shares
5215 * resulting from this level per the above.
5224 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5232 * Returns the current capacity of cpu after applying both
5233 * cpu and freq scaling.
5235 unsigned long capacity_curr_of(int cpu)
5237 return cpu_rq(cpu)->cpu_capacity_orig *
5238 arch_scale_freq_capacity(NULL, cpu)
5239 >> SCHED_CAPACITY_SHIFT;
5242 static inline bool energy_aware(void)
5244 return sched_feat(ENERGY_AWARE);
5248 struct sched_group *sg_top;
5249 struct sched_group *sg_cap;
5257 struct task_struct *task;
5271 static int cpu_util_wake(int cpu, struct task_struct *p);
5274 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5275 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE], which is useful for
5276 * energy calculations.
5278 * Since util is a scale-invariant utilization defined as:
5280 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5282 * the normalized util can be found using the specific capacity.
5284 * capacity = capacity_orig * curr_freq/max_freq
5286 * norm_util = running_time/time ~ util/capacity
5288 static unsigned long __cpu_norm_util(unsigned long util, unsigned long capacity)
5290 if (util >= capacity)
5291 return SCHED_CAPACITY_SCALE;
5293 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5296 static unsigned long group_max_util(struct energy_env *eenv)
5298 unsigned long max_util = 0;
5302 for_each_cpu(cpu, sched_group_cpus(eenv->sg_cap)) {
5303 util = cpu_util_wake(cpu, eenv->task);
5306 * If we are looking at the target CPU specified by the eenv,
5307 * then we should add the (estimated) utilization of the task
5308 * assuming we will wake it up on that CPU.
5310 if (unlikely(cpu == eenv->trg_cpu))
5311 util += eenv->util_delta;
5313 max_util = max(max_util, util);
5320 * group_norm_util() returns the approximated group util relative to it's
5321 * current capacity (busy ratio), in the range [0..SCHED_LOAD_SCALE], for use
5322 * in energy calculations.
5324 * Since task executions may or may not overlap in time in the group the true
5325 * normalized util is between MAX(cpu_norm_util(i)) and SUM(cpu_norm_util(i))
5326 * when iterating over all CPUs in the group.
5327 * The latter estimate is used as it leads to a more pessimistic energy
5328 * estimate (more busy).
5331 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5333 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5334 unsigned long util, util_sum = 0;
5337 for_each_cpu(cpu, sched_group_cpus(sg)) {
5338 util = cpu_util_wake(cpu, eenv->task);
5341 * If we are looking at the target CPU specified by the eenv,
5342 * then we should add the (estimated) utilization of the task
5343 * assuming we will wake it up on that CPU.
5345 if (unlikely(cpu == eenv->trg_cpu))
5346 util += eenv->util_delta;
5348 util_sum += __cpu_norm_util(util, capacity);
5351 return min_t(unsigned long, util_sum, SCHED_CAPACITY_SCALE);
5354 static int find_new_capacity(struct energy_env *eenv,
5355 const struct sched_group_energy * const sge)
5357 int idx, max_idx = sge->nr_cap_states - 1;
5358 unsigned long util = group_max_util(eenv);
5360 /* default is max_cap if we don't find a match */
5361 eenv->cap_idx = max_idx;
5363 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5364 if (sge->cap_states[idx].cap >= util) {
5365 eenv->cap_idx = idx;
5370 return eenv->cap_idx;
5373 static int group_idle_state(struct energy_env *eenv, struct sched_group *sg)
5375 int i, state = INT_MAX;
5376 int src_in_grp, dst_in_grp;
5379 /* Find the shallowest idle state in the sched group. */
5380 for_each_cpu(i, sched_group_cpus(sg))
5381 state = min(state, idle_get_state_idx(cpu_rq(i)));
5383 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5386 src_in_grp = cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg));
5387 dst_in_grp = cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg));
5388 if (src_in_grp == dst_in_grp) {
5389 /* both CPUs under consideration are in the same group or not in
5390 * either group, migration should leave idle state the same.
5396 * Try to estimate if a deeper idle state is
5397 * achievable when we move the task.
5399 for_each_cpu(i, sched_group_cpus(sg)) {
5400 grp_util += cpu_util_wake(i, eenv->task);
5401 if (unlikely(i == eenv->trg_cpu))
5402 grp_util += eenv->util_delta;
5406 ((long)sg->sgc->max_capacity * (int)sg->group_weight)) {
5407 /* after moving, this group is at most partly
5408 * occupied, so it should have some idle time.
5410 int max_idle_state_idx = sg->sge->nr_idle_states - 2;
5411 int new_state = grp_util * max_idle_state_idx;
5413 /* group will have no util, use lowest state */
5414 new_state = max_idle_state_idx + 1;
5416 /* for partially idle, linearly map util to idle
5417 * states, excluding the lowest one. This does not
5418 * correspond to the state we expect to enter in
5419 * reality, but an indication of what might happen.
5421 new_state = min(max_idle_state_idx, (int)
5422 (new_state / sg->sgc->max_capacity));
5423 new_state = max_idle_state_idx - new_state;
5427 /* After moving, the group will be fully occupied
5428 * so assume it will not be idle at all.
5437 * sched_group_energy(): Computes the absolute energy consumption of cpus
5438 * belonging to the sched_group including shared resources shared only by
5439 * members of the group. Iterates over all cpus in the hierarchy below the
5440 * sched_group starting from the bottom working it's way up before going to
5441 * the next cpu until all cpus are covered at all levels. The current
5442 * implementation is likely to gather the same util statistics multiple times.
5443 * This can probably be done in a faster but more complex way.
5444 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5446 static int sched_group_energy(struct energy_env *eenv)
5448 struct cpumask visit_cpus;
5449 u64 total_energy = 0;
5451 WARN_ON(!eenv->sg_top->sge);
5453 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5455 while (!cpumask_empty(&visit_cpus)) {
5456 struct sched_group *sg_shared_cap = NULL;
5457 int cpu = cpumask_first(&visit_cpus);
5458 struct sched_domain *sd;
5461 * Is the group utilization affected by cpus outside this
5464 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5466 if (sd && sd->parent)
5467 sg_shared_cap = sd->parent->groups;
5469 for_each_domain(cpu, sd) {
5470 struct sched_group *sg = sd->groups;
5472 /* Has this sched_domain already been visited? */
5473 if (sd->child && group_first_cpu(sg) != cpu)
5477 unsigned long group_util;
5478 int sg_busy_energy, sg_idle_energy;
5479 int cap_idx, idle_idx;
5481 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5482 eenv->sg_cap = sg_shared_cap;
5486 cap_idx = find_new_capacity(eenv, sg->sge);
5488 if (sg->group_weight == 1) {
5489 /* Remove capacity of src CPU (before task move) */
5490 if (eenv->trg_cpu == eenv->src_cpu &&
5491 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5492 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5493 eenv->cap.delta -= eenv->cap.before;
5495 /* Add capacity of dst CPU (after task move) */
5496 if (eenv->trg_cpu == eenv->dst_cpu &&
5497 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5498 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5499 eenv->cap.delta += eenv->cap.after;
5503 idle_idx = group_idle_state(eenv, sg);
5504 group_util = group_norm_util(eenv, sg);
5506 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power);
5507 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5508 * sg->sge->idle_states[idle_idx].power);
5510 total_energy += sg_busy_energy + sg_idle_energy;
5513 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5515 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5518 } while (sg = sg->next, sg != sd->groups);
5522 * If we raced with hotplug and got an sd NULL-pointer;
5523 * returning a wrong energy estimation is better than
5524 * entering an infinite loop.
5526 if (cpumask_test_cpu(cpu, &visit_cpus))
5529 cpumask_clear_cpu(cpu, &visit_cpus);
5533 eenv->energy = total_energy >> SCHED_CAPACITY_SHIFT;
5537 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5539 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5542 static inline unsigned long task_util(struct task_struct *p);
5545 * energy_diff(): Estimate the energy impact of changing the utilization
5546 * distribution. eenv specifies the change: utilisation amount, source, and
5547 * destination cpu. Source or destination cpu may be -1 in which case the
5548 * utilization is removed from or added to the system (e.g. task wake-up). If
5549 * both are specified, the utilization is migrated.
5551 static inline int __energy_diff(struct energy_env *eenv)
5553 struct sched_domain *sd;
5554 struct sched_group *sg;
5555 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5558 struct energy_env eenv_before = {
5559 .util_delta = task_util(eenv->task),
5560 .src_cpu = eenv->src_cpu,
5561 .dst_cpu = eenv->dst_cpu,
5562 .trg_cpu = eenv->src_cpu,
5563 .nrg = { 0, 0, 0, 0},
5568 if (eenv->src_cpu == eenv->dst_cpu)
5571 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5572 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5575 return 0; /* Error */
5580 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5581 eenv_before.sg_top = eenv->sg_top = sg;
5583 if (sched_group_energy(&eenv_before))
5584 return 0; /* Invalid result abort */
5585 energy_before += eenv_before.energy;
5587 /* Keep track of SRC cpu (before) capacity */
5588 eenv->cap.before = eenv_before.cap.before;
5589 eenv->cap.delta = eenv_before.cap.delta;
5591 if (sched_group_energy(eenv))
5592 return 0; /* Invalid result abort */
5593 energy_after += eenv->energy;
5595 } while (sg = sg->next, sg != sd->groups);
5597 eenv->nrg.before = energy_before;
5598 eenv->nrg.after = energy_after;
5599 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5601 #ifndef CONFIG_SCHED_TUNE
5602 trace_sched_energy_diff(eenv->task,
5603 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5604 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5605 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5606 eenv->nrg.delta, eenv->payoff);
5609 * Dead-zone margin preventing too many migrations.
5612 margin = eenv->nrg.before >> 6; /* ~1.56% */
5614 diff = eenv->nrg.after - eenv->nrg.before;
5616 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5618 return eenv->nrg.diff;
5621 #ifdef CONFIG_SCHED_TUNE
5623 struct target_nrg schedtune_target_nrg;
5625 #ifdef CONFIG_CGROUP_SCHEDTUNE
5626 extern bool schedtune_initialized;
5627 #endif /* CONFIG_CGROUP_SCHEDTUNE */
5630 * System energy normalization
5631 * Returns the normalized value, in the range [0..SCHED_CAPACITY_SCALE],
5632 * corresponding to the specified energy variation.
5635 normalize_energy(int energy_diff)
5639 #ifdef CONFIG_CGROUP_SCHEDTUNE
5640 /* during early setup, we don't know the extents */
5641 if (unlikely(!schedtune_initialized))
5642 return energy_diff < 0 ? -1 : 1 ;
5643 #endif /* CONFIG_CGROUP_SCHEDTUNE */
5645 #ifdef CONFIG_SCHED_DEBUG
5649 /* Check for boundaries */
5650 max_delta = schedtune_target_nrg.max_power;
5651 max_delta -= schedtune_target_nrg.min_power;
5652 WARN_ON(abs(energy_diff) >= max_delta);
5656 /* Do scaling using positive numbers to increase the range */
5657 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5659 /* Scale by energy magnitude */
5660 normalized_nrg <<= SCHED_CAPACITY_SHIFT;
5662 /* Normalize on max energy for target platform */
5663 normalized_nrg = reciprocal_divide(
5664 normalized_nrg, schedtune_target_nrg.rdiv);
5666 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5670 energy_diff(struct energy_env *eenv)
5672 int boost = schedtune_task_boost(eenv->task);
5675 /* Conpute "absolute" energy diff */
5676 __energy_diff(eenv);
5678 /* Return energy diff when boost margin is 0 */
5680 trace_sched_energy_diff(eenv->task,
5681 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5682 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5683 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5684 0, -eenv->nrg.diff);
5685 return eenv->nrg.diff;
5688 /* Compute normalized energy diff */
5689 nrg_delta = normalize_energy(eenv->nrg.diff);
5690 eenv->nrg.delta = nrg_delta;
5692 eenv->payoff = schedtune_accept_deltas(
5697 trace_sched_energy_diff(eenv->task,
5698 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5699 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5700 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5701 eenv->nrg.delta, eenv->payoff);
5704 * When SchedTune is enabled, the energy_diff() function will return
5705 * the computed energy payoff value. Since the energy_diff() return
5706 * value is expected to be negative by its callers, this evaluation
5707 * function return a negative value each time the evaluation return a
5708 * positive payoff, which is the condition for the acceptance of
5709 * a scheduling decision
5711 return -eenv->payoff;
5713 #else /* CONFIG_SCHED_TUNE */
5714 #define energy_diff(eenv) __energy_diff(eenv)
5718 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5719 * A waker of many should wake a different task than the one last awakened
5720 * at a frequency roughly N times higher than one of its wakees. In order
5721 * to determine whether we should let the load spread vs consolodating to
5722 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5723 * partner, and a factor of lls_size higher frequency in the other. With
5724 * both conditions met, we can be relatively sure that the relationship is
5725 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5726 * being client/server, worker/dispatcher, interrupt source or whatever is
5727 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5729 static int wake_wide(struct task_struct *p, int sibling_count_hint)
5731 unsigned int master = current->wakee_flips;
5732 unsigned int slave = p->wakee_flips;
5733 int llc_size = this_cpu_read(sd_llc_size);
5735 if (sibling_count_hint >= llc_size)
5739 swap(master, slave);
5740 if (slave < llc_size || master < slave * llc_size)
5745 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5746 int prev_cpu, int sync)
5748 s64 this_load, load;
5749 s64 this_eff_load, prev_eff_load;
5751 struct task_group *tg;
5752 unsigned long weight;
5756 this_cpu = smp_processor_id();
5757 load = source_load(prev_cpu, idx);
5758 this_load = target_load(this_cpu, idx);
5761 * If sync wakeup then subtract the (maximum possible)
5762 * effect of the currently running task from the load
5763 * of the current CPU:
5766 tg = task_group(current);
5767 weight = current->se.avg.load_avg;
5769 this_load += effective_load(tg, this_cpu, -weight, -weight);
5770 load += effective_load(tg, prev_cpu, 0, -weight);
5774 weight = p->se.avg.load_avg;
5777 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5778 * due to the sync cause above having dropped this_load to 0, we'll
5779 * always have an imbalance, but there's really nothing you can do
5780 * about that, so that's good too.
5782 * Otherwise check if either cpus are near enough in load to allow this
5783 * task to be woken on this_cpu.
5785 this_eff_load = 100;
5786 this_eff_load *= capacity_of(prev_cpu);
5788 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5789 prev_eff_load *= capacity_of(this_cpu);
5791 if (this_load > 0) {
5792 this_eff_load *= this_load +
5793 effective_load(tg, this_cpu, weight, weight);
5795 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5798 balanced = this_eff_load <= prev_eff_load;
5800 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5805 schedstat_inc(sd, ttwu_move_affine);
5806 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5811 static inline unsigned long task_util(struct task_struct *p)
5813 #ifdef CONFIG_SCHED_WALT
5814 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5815 unsigned long demand = p->ravg.demand;
5816 return (demand << 10) / walt_ravg_window;
5819 return p->se.avg.util_avg;
5822 static inline unsigned long boosted_task_util(struct task_struct *task);
5824 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5826 unsigned long capacity = capacity_of(cpu);
5828 util += boosted_task_util(p);
5830 return (capacity * 1024) > (util * capacity_margin);
5833 static inline bool task_fits_max(struct task_struct *p, int cpu)
5835 unsigned long capacity = capacity_of(cpu);
5836 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5838 if (capacity == max_capacity)
5841 if (capacity * capacity_margin > max_capacity * 1024)
5844 return __task_fits(p, cpu, 0);
5847 static bool __cpu_overutilized(int cpu, int delta)
5849 return (capacity_of(cpu) * 1024) < ((cpu_util(cpu) + delta) * capacity_margin);
5852 static bool cpu_overutilized(int cpu)
5854 return __cpu_overutilized(cpu, 0);
5857 #ifdef CONFIG_SCHED_TUNE
5859 struct reciprocal_value schedtune_spc_rdiv;
5862 schedtune_margin(unsigned long signal, long boost)
5864 long long margin = 0;
5867 * Signal proportional compensation (SPC)
5869 * The Boost (B) value is used to compute a Margin (M) which is
5870 * proportional to the complement of the original Signal (S):
5871 * M = B * (SCHED_CAPACITY_SCALE - S)
5872 * The obtained M could be used by the caller to "boost" S.
5875 margin = SCHED_CAPACITY_SCALE - signal;
5878 margin = -signal * boost;
5880 margin = reciprocal_divide(margin, schedtune_spc_rdiv);
5888 schedtune_cpu_margin(unsigned long util, int cpu)
5890 int boost = schedtune_cpu_boost(cpu);
5895 return schedtune_margin(util, boost);
5899 schedtune_task_margin(struct task_struct *task)
5901 int boost = schedtune_task_boost(task);
5908 util = task_util(task);
5909 margin = schedtune_margin(util, boost);
5914 #else /* CONFIG_SCHED_TUNE */
5917 schedtune_cpu_margin(unsigned long util, int cpu)
5923 schedtune_task_margin(struct task_struct *task)
5928 #endif /* CONFIG_SCHED_TUNE */
5931 boosted_cpu_util(int cpu)
5933 unsigned long util = cpu_util_freq(cpu);
5934 long margin = schedtune_cpu_margin(util, cpu);
5936 trace_sched_boost_cpu(cpu, util, margin);
5938 return util + margin;
5941 static inline unsigned long
5942 boosted_task_util(struct task_struct *task)
5944 unsigned long util = task_util(task);
5945 long margin = schedtune_task_margin(task);
5947 trace_sched_boost_task(task, util, margin);
5949 return util + margin;
5952 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5954 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5958 * find_idlest_group finds and returns the least busy CPU group within the
5961 * Assumes p is allowed on at least one CPU in sd.
5963 static struct sched_group *
5964 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5965 int this_cpu, int sd_flag)
5967 struct sched_group *idlest = NULL, *group = sd->groups;
5968 struct sched_group *most_spare_sg = NULL;
5969 unsigned long min_load = ULONG_MAX, this_load = ULONG_MAX;
5970 unsigned long most_spare = 0, this_spare = 0;
5971 int load_idx = sd->forkexec_idx;
5972 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5974 if (sd_flag & SD_BALANCE_WAKE)
5975 load_idx = sd->wake_idx;
5978 unsigned long load, avg_load, spare_cap, max_spare_cap;
5982 /* Skip over this group if it has no CPUs allowed */
5983 if (!cpumask_intersects(sched_group_cpus(group),
5984 tsk_cpus_allowed(p)))
5987 local_group = cpumask_test_cpu(this_cpu,
5988 sched_group_cpus(group));
5991 * Tally up the load of all CPUs in the group and find
5992 * the group containing the CPU with most spare capacity.
5997 for_each_cpu(i, sched_group_cpus(group)) {
5998 /* Bias balancing toward cpus of our domain */
6000 load = source_load(i, load_idx);
6002 load = target_load(i, load_idx);
6006 spare_cap = capacity_spare_wake(i, p);
6008 if (spare_cap > max_spare_cap)
6009 max_spare_cap = spare_cap;
6012 /* Adjust by relative CPU capacity of the group */
6013 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
6016 this_load = avg_load;
6017 this_spare = max_spare_cap;
6019 if (avg_load < min_load) {
6020 min_load = avg_load;
6024 if (most_spare < max_spare_cap) {
6025 most_spare = max_spare_cap;
6026 most_spare_sg = group;
6029 } while (group = group->next, group != sd->groups);
6032 * The cross-over point between using spare capacity or least load
6033 * is too conservative for high utilization tasks on partially
6034 * utilized systems if we require spare_capacity > task_util(p),
6035 * so we allow for some task stuffing by using
6036 * spare_capacity > task_util(p)/2.
6038 * Spare capacity can't be used for fork because the utilization has
6039 * not been set yet, we must first select a rq to compute the initial
6042 if (sd_flag & SD_BALANCE_FORK)
6045 if (this_spare > task_util(p) / 2 &&
6046 imbalance*this_spare > 100*most_spare)
6048 else if (most_spare > task_util(p) / 2)
6049 return most_spare_sg;
6052 if (!idlest || 100*this_load < imbalance*min_load)
6058 * find_idlest_group_cpu - find the idlest cpu among the cpus in group.
6061 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6063 unsigned long load, min_load = ULONG_MAX;
6064 unsigned int min_exit_latency = UINT_MAX;
6065 u64 latest_idle_timestamp = 0;
6066 int least_loaded_cpu = this_cpu;
6067 int shallowest_idle_cpu = -1;
6070 /* Check if we have any choice: */
6071 if (group->group_weight == 1)
6072 return cpumask_first(sched_group_cpus(group));
6074 /* Traverse only the allowed CPUs */
6075 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
6077 struct rq *rq = cpu_rq(i);
6078 struct cpuidle_state *idle = idle_get_state(rq);
6079 if (idle && idle->exit_latency < min_exit_latency) {
6081 * We give priority to a CPU whose idle state
6082 * has the smallest exit latency irrespective
6083 * of any idle timestamp.
6085 min_exit_latency = idle->exit_latency;
6086 latest_idle_timestamp = rq->idle_stamp;
6087 shallowest_idle_cpu = i;
6088 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6089 rq->idle_stamp > latest_idle_timestamp) {
6091 * If equal or no active idle state, then
6092 * the most recently idled CPU might have
6095 latest_idle_timestamp = rq->idle_stamp;
6096 shallowest_idle_cpu = i;
6098 } else if (shallowest_idle_cpu == -1) {
6099 load = weighted_cpuload(i);
6100 if (load < min_load || (load == min_load && i == this_cpu)) {
6102 least_loaded_cpu = i;
6107 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6110 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6111 int cpu, int prev_cpu, int sd_flag)
6114 int wu = sd_flag & SD_BALANCE_WAKE;
6118 schedstat_inc(p, se.statistics.nr_wakeups_cas_attempts);
6119 schedstat_inc(this_rq(), eas_stats.cas_attempts);
6122 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6126 struct sched_group *group;
6127 struct sched_domain *tmp;
6131 schedstat_inc(sd, eas_stats.cas_attempts);
6133 if (!(sd->flags & sd_flag)) {
6138 group = find_idlest_group(sd, p, cpu, sd_flag);
6144 new_cpu = find_idlest_group_cpu(group, p, cpu);
6145 if (new_cpu == cpu) {
6146 /* Now try balancing at a lower domain level of cpu */
6151 /* Now try balancing at a lower domain level of new_cpu */
6152 cpu = cas_cpu = new_cpu;
6153 weight = sd->span_weight;
6155 for_each_domain(cpu, tmp) {
6156 if (weight <= tmp->span_weight)
6158 if (tmp->flags & sd_flag)
6161 /* while loop will break here if sd == NULL */
6164 if (wu && (cas_cpu >= 0)) {
6165 schedstat_inc(p, se.statistics.nr_wakeups_cas_count);
6166 schedstat_inc(this_rq(), eas_stats.cas_count);
6173 * Try and locate an idle CPU in the sched_domain.
6175 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6177 struct sched_domain *sd;
6178 struct sched_group *sg;
6179 int best_idle_cpu = -1;
6180 int best_idle_cstate = INT_MAX;
6181 unsigned long best_idle_capacity = ULONG_MAX;
6183 schedstat_inc(p, se.statistics.nr_wakeups_sis_attempts);
6184 schedstat_inc(this_rq(), eas_stats.sis_attempts);
6186 if (!sysctl_sched_cstate_aware) {
6187 if (idle_cpu(target)) {
6188 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle);
6189 schedstat_inc(this_rq(), eas_stats.sis_idle);
6194 * If the prevous cpu is cache affine and idle, don't be stupid.
6196 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) {
6197 schedstat_inc(p, se.statistics.nr_wakeups_sis_cache_affine);
6198 schedstat_inc(this_rq(), eas_stats.sis_cache_affine);
6204 * Otherwise, iterate the domains and find an elegible idle cpu.
6206 sd = rcu_dereference(per_cpu(sd_llc, target));
6207 for_each_lower_domain(sd) {
6211 if (!cpumask_intersects(sched_group_cpus(sg),
6212 tsk_cpus_allowed(p)))
6215 if (sysctl_sched_cstate_aware) {
6216 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6217 int idle_idx = idle_get_state_idx(cpu_rq(i));
6218 unsigned long new_usage = boosted_task_util(p);
6219 unsigned long capacity_orig = capacity_orig_of(i);
6221 if (new_usage > capacity_orig || !idle_cpu(i))
6224 if (i == target && new_usage <= capacity_curr_of(target)) {
6225 schedstat_inc(p, se.statistics.nr_wakeups_sis_suff_cap);
6226 schedstat_inc(this_rq(), eas_stats.sis_suff_cap);
6227 schedstat_inc(sd, eas_stats.sis_suff_cap);
6231 if (idle_idx < best_idle_cstate &&
6232 capacity_orig <= best_idle_capacity) {
6234 best_idle_cstate = idle_idx;
6235 best_idle_capacity = capacity_orig;
6239 for_each_cpu(i, sched_group_cpus(sg)) {
6240 if (i == target || !idle_cpu(i))
6244 target = cpumask_first_and(sched_group_cpus(sg),
6245 tsk_cpus_allowed(p));
6246 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle_cpu);
6247 schedstat_inc(this_rq(), eas_stats.sis_idle_cpu);
6248 schedstat_inc(sd, eas_stats.sis_idle_cpu);
6253 } while (sg != sd->groups);
6256 if (best_idle_cpu >= 0)
6257 target = best_idle_cpu;
6260 schedstat_inc(p, se.statistics.nr_wakeups_sis_count);
6261 schedstat_inc(this_rq(), eas_stats.sis_count);
6267 * cpu_util_wake: Compute cpu utilization with any contributions from
6268 * the waking task p removed. check_for_migration() looks for a better CPU of
6269 * rq->curr. For that case we should return cpu util with contributions from
6270 * currently running task p removed.
6272 static int cpu_util_wake(int cpu, struct task_struct *p)
6274 unsigned long util, capacity;
6276 #ifdef CONFIG_SCHED_WALT
6278 * WALT does not decay idle tasks in the same manner
6279 * as PELT, so it makes little sense to subtract task
6280 * utilization from cpu utilization. Instead just use
6281 * cpu_util for this case.
6283 if (!walt_disabled && sysctl_sched_use_walt_cpu_util &&
6284 p->state == TASK_WAKING)
6285 return cpu_util(cpu);
6287 /* Task has no contribution or is new */
6288 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6289 return cpu_util(cpu);
6291 capacity = capacity_orig_of(cpu);
6292 util = max_t(long, cpu_util(cpu) - task_util(p), 0);
6294 return (util >= capacity) ? capacity : util;
6297 static int start_cpu(bool boosted)
6299 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6301 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
6304 static inline int find_best_target(struct task_struct *p, int *backup_cpu,
6305 bool boosted, bool prefer_idle)
6307 unsigned long best_idle_min_cap_orig = ULONG_MAX;
6308 unsigned long min_util = boosted_task_util(p);
6309 unsigned long target_capacity = ULONG_MAX;
6310 unsigned long min_wake_util = ULONG_MAX;
6311 unsigned long target_max_spare_cap = 0;
6312 unsigned long best_active_util = ULONG_MAX;
6313 int best_idle_cstate = INT_MAX;
6314 struct sched_domain *sd;
6315 struct sched_group *sg;
6316 int best_active_cpu = -1;
6317 int best_idle_cpu = -1;
6318 int target_cpu = -1;
6323 schedstat_inc(p, se.statistics.nr_wakeups_fbt_attempts);
6324 schedstat_inc(this_rq(), eas_stats.fbt_attempts);
6326 /* Find start CPU based on boost value */
6327 cpu = start_cpu(boosted);
6329 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_cpu);
6330 schedstat_inc(this_rq(), eas_stats.fbt_no_cpu);
6334 /* Find SD for the start CPU */
6335 sd = rcu_dereference(per_cpu(sd_ea, cpu));
6337 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_sd);
6338 schedstat_inc(this_rq(), eas_stats.fbt_no_sd);
6342 /* Scan CPUs in all SDs */
6345 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6346 unsigned long capacity_curr = capacity_curr_of(i);
6347 unsigned long capacity_orig = capacity_orig_of(i);
6348 unsigned long wake_util, new_util;
6353 if (walt_cpu_high_irqload(i))
6357 * p's blocked utilization is still accounted for on prev_cpu
6358 * so prev_cpu will receive a negative bias due to the double
6359 * accounting. However, the blocked utilization may be zero.
6361 wake_util = cpu_util_wake(i, p);
6362 new_util = wake_util + task_util(p);
6365 * Ensure minimum capacity to grant the required boost.
6366 * The target CPU can be already at a capacity level higher
6367 * than the one required to boost the task.
6369 new_util = max(min_util, new_util);
6370 if (new_util > capacity_orig)
6374 * Case A) Latency sensitive tasks
6376 * Unconditionally favoring tasks that prefer idle CPU to
6380 * - an idle CPU, whatever its idle_state is, since
6381 * the first CPUs we explore are more likely to be
6382 * reserved for latency sensitive tasks.
6383 * - a non idle CPU where the task fits in its current
6384 * capacity and has the maximum spare capacity.
6385 * - a non idle CPU with lower contention from other
6386 * tasks and running at the lowest possible OPP.
6388 * The last two goals tries to favor a non idle CPU
6389 * where the task can run as if it is "almost alone".
6390 * A maximum spare capacity CPU is favoured since
6391 * the task already fits into that CPU's capacity
6392 * without waiting for an OPP chance.
6394 * The following code path is the only one in the CPUs
6395 * exploration loop which is always used by
6396 * prefer_idle tasks. It exits the loop with wither a
6397 * best_active_cpu or a target_cpu which should
6398 * represent an optimal choice for latency sensitive
6404 * Case A.1: IDLE CPU
6405 * Return the first IDLE CPU we find.
6408 schedstat_inc(p, se.statistics.nr_wakeups_fbt_pref_idle);
6409 schedstat_inc(this_rq(), eas_stats.fbt_pref_idle);
6411 trace_sched_find_best_target(p,
6412 prefer_idle, min_util,
6414 best_active_cpu, i);
6420 * Case A.2: Target ACTIVE CPU
6421 * Favor CPUs with max spare capacity.
6423 if ((capacity_curr > new_util) &&
6424 (capacity_orig - new_util > target_max_spare_cap)) {
6425 target_max_spare_cap = capacity_orig - new_util;
6429 if (target_cpu != -1)
6434 * Case A.3: Backup ACTIVE CPU
6436 * - lower utilization due to other tasks
6437 * - lower utilization with the task in
6439 if (wake_util > min_wake_util)
6441 if (new_util > best_active_util)
6443 min_wake_util = wake_util;
6444 best_active_util = new_util;
6445 best_active_cpu = i;
6452 * For non latency sensitive tasks, skip CPUs that
6453 * will be overutilized by moving the task there.
6455 * The goal here is to remain in EAS mode as long as
6456 * possible at least for !prefer_idle tasks.
6458 if ((new_util * capacity_margin) >
6459 (capacity_orig * SCHED_CAPACITY_SCALE))
6463 * Case B) Non latency sensitive tasks on IDLE CPUs.
6465 * Find an optimal backup IDLE CPU for non latency
6469 * - minimizing the capacity_orig,
6470 * i.e. preferring LITTLE CPUs
6471 * - favoring shallowest idle states
6472 * i.e. avoid to wakeup deep-idle CPUs
6474 * The following code path is used by non latency
6475 * sensitive tasks if IDLE CPUs are available. If at
6476 * least one of such CPUs are available it sets the
6477 * best_idle_cpu to the most suitable idle CPU to be
6480 * If idle CPUs are available, favour these CPUs to
6481 * improve performances by spreading tasks.
6482 * Indeed, the energy_diff() computed by the caller
6483 * will take care to ensure the minimization of energy
6484 * consumptions without affecting performance.
6487 int idle_idx = idle_get_state_idx(cpu_rq(i));
6489 /* Select idle CPU with lower cap_orig */
6490 if (capacity_orig > best_idle_min_cap_orig)
6494 * Skip CPUs in deeper idle state, but only
6495 * if they are also less energy efficient.
6496 * IOW, prefer a deep IDLE LITTLE CPU vs a
6497 * shallow idle big CPU.
6499 if (sysctl_sched_cstate_aware &&
6500 best_idle_cstate <= idle_idx)
6503 /* Keep track of best idle CPU */
6504 best_idle_min_cap_orig = capacity_orig;
6505 best_idle_cstate = idle_idx;
6511 * Case C) Non latency sensitive tasks on ACTIVE CPUs.
6513 * Pack tasks in the most energy efficient capacities.
6515 * This task packing strategy prefers more energy
6516 * efficient CPUs (i.e. pack on smaller maximum
6517 * capacity CPUs) while also trying to spread tasks to
6518 * run them all at the lower OPP.
6520 * This assumes for example that it's more energy
6521 * efficient to run two tasks on two CPUs at a lower
6522 * OPP than packing both on a single CPU but running
6523 * that CPU at an higher OPP.
6525 * Thus, this case keep track of the CPU with the
6526 * smallest maximum capacity and highest spare maximum
6530 /* Favor CPUs with smaller capacity */
6531 if (capacity_orig > target_capacity)
6534 /* Favor CPUs with maximum spare capacity */
6535 if ((capacity_orig - new_util) < target_max_spare_cap)
6538 target_max_spare_cap = capacity_orig - new_util;
6539 target_capacity = capacity_orig;
6543 } while (sg = sg->next, sg != sd->groups);
6546 * For non latency sensitive tasks, cases B and C in the previous loop,
6547 * we pick the best IDLE CPU only if we was not able to find a target
6550 * Policies priorities:
6552 * - prefer_idle tasks:
6554 * a) IDLE CPU available, we return immediately
6555 * b) ACTIVE CPU where task fits and has the bigger maximum spare
6556 * capacity (i.e. target_cpu)
6557 * c) ACTIVE CPU with less contention due to other tasks
6558 * (i.e. best_active_cpu)
6560 * - NON prefer_idle tasks:
6562 * a) ACTIVE CPU: target_cpu
6563 * b) IDLE CPU: best_idle_cpu
6565 if (target_cpu == -1)
6566 target_cpu = prefer_idle
6570 *backup_cpu = prefer_idle
6574 trace_sched_find_best_target(p, prefer_idle, min_util, cpu,
6575 best_idle_cpu, best_active_cpu,
6578 schedstat_inc(p, se.statistics.nr_wakeups_fbt_count);
6579 schedstat_inc(this_rq(), eas_stats.fbt_count);
6585 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6586 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6588 * In that case WAKE_AFFINE doesn't make sense and we'll let
6589 * BALANCE_WAKE sort things out.
6591 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6593 long min_cap, max_cap;
6595 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6596 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6598 /* Minimum capacity is close to max, no need to abort wake_affine */
6599 if (max_cap - min_cap < max_cap >> 3)
6602 /* Bring task utilization in sync with prev_cpu */
6603 sync_entity_load_avg(&p->se);
6605 return min_cap * 1024 < task_util(p) * capacity_margin;
6608 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6610 struct sched_domain *sd;
6611 int target_cpu = prev_cpu, tmp_target, tmp_backup;
6612 bool boosted, prefer_idle;
6614 schedstat_inc(p, se.statistics.nr_wakeups_secb_attempts);
6615 schedstat_inc(this_rq(), eas_stats.secb_attempts);
6617 if (sysctl_sched_sync_hint_enable && sync) {
6618 int cpu = smp_processor_id();
6620 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6621 schedstat_inc(p, se.statistics.nr_wakeups_secb_sync);
6622 schedstat_inc(this_rq(), eas_stats.secb_sync);
6628 #ifdef CONFIG_CGROUP_SCHEDTUNE
6629 boosted = schedtune_task_boost(p) > 0;
6630 prefer_idle = schedtune_prefer_idle(p) > 0;
6632 boosted = get_sysctl_sched_cfs_boost() > 0;
6636 sync_entity_load_avg(&p->se);
6638 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6639 /* Find a cpu with sufficient capacity */
6640 tmp_target = find_best_target(p, &tmp_backup, boosted, prefer_idle);
6644 if (tmp_target >= 0) {
6645 target_cpu = tmp_target;
6646 if ((boosted || prefer_idle) && idle_cpu(target_cpu)) {
6647 schedstat_inc(p, se.statistics.nr_wakeups_secb_idle_bt);
6648 schedstat_inc(this_rq(), eas_stats.secb_idle_bt);
6653 if (target_cpu != prev_cpu) {
6655 struct energy_env eenv = {
6656 .util_delta = task_util(p),
6657 .src_cpu = prev_cpu,
6658 .dst_cpu = target_cpu,
6660 .trg_cpu = target_cpu,
6664 #ifdef CONFIG_SCHED_WALT
6665 if (!walt_disabled && sysctl_sched_use_walt_cpu_util &&
6666 p->state == TASK_WAKING)
6667 delta = task_util(p);
6669 /* Not enough spare capacity on previous cpu */
6670 if (__cpu_overutilized(prev_cpu, delta)) {
6671 schedstat_inc(p, se.statistics.nr_wakeups_secb_insuff_cap);
6672 schedstat_inc(this_rq(), eas_stats.secb_insuff_cap);
6676 if (energy_diff(&eenv) >= 0) {
6677 /* No energy saving for target_cpu, try backup */
6678 target_cpu = tmp_backup;
6679 eenv.dst_cpu = target_cpu;
6680 eenv.trg_cpu = target_cpu;
6681 if (tmp_backup < 0 ||
6682 tmp_backup == prev_cpu ||
6683 energy_diff(&eenv) >= 0) {
6684 schedstat_inc(p, se.statistics.nr_wakeups_secb_no_nrg_sav);
6685 schedstat_inc(this_rq(), eas_stats.secb_no_nrg_sav);
6686 target_cpu = prev_cpu;
6691 schedstat_inc(p, se.statistics.nr_wakeups_secb_nrg_sav);
6692 schedstat_inc(this_rq(), eas_stats.secb_nrg_sav);
6696 schedstat_inc(p, se.statistics.nr_wakeups_secb_count);
6697 schedstat_inc(this_rq(), eas_stats.secb_count);
6706 * select_task_rq_fair: Select target runqueue for the waking task in domains
6707 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6708 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6710 * Balances load by selecting the idlest cpu in the idlest group, or under
6711 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6713 * Returns the target cpu number.
6715 * preempt must be disabled.
6718 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags,
6719 int sibling_count_hint)
6721 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6722 int cpu = smp_processor_id();
6723 int new_cpu = prev_cpu;
6724 int want_affine = 0;
6725 int sync = wake_flags & WF_SYNC;
6727 if (sd_flag & SD_BALANCE_WAKE) {
6729 want_affine = !wake_wide(p, sibling_count_hint) &&
6730 !wake_cap(p, cpu, prev_cpu) &&
6731 cpumask_test_cpu(cpu, &p->cpus_allowed);
6734 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6735 return select_energy_cpu_brute(p, prev_cpu, sync);
6738 for_each_domain(cpu, tmp) {
6739 if (!(tmp->flags & SD_LOAD_BALANCE))
6743 * If both cpu and prev_cpu are part of this domain,
6744 * cpu is a valid SD_WAKE_AFFINE target.
6746 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6747 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6752 if (tmp->flags & sd_flag)
6754 else if (!want_affine)
6759 sd = NULL; /* Prefer wake_affine over balance flags */
6760 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6764 if (sd && !(sd_flag & SD_BALANCE_FORK)) {
6766 * We're going to need the task's util for capacity_spare_wake
6767 * in find_idlest_group. Sync it up to prev_cpu's
6770 sync_entity_load_avg(&p->se);
6774 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6775 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6778 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6786 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6787 * cfs_rq_of(p) references at time of call are still valid and identify the
6788 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6789 * other assumptions, including the state of rq->lock, should be made.
6791 static void migrate_task_rq_fair(struct task_struct *p)
6794 * We are supposed to update the task to "current" time, then its up to date
6795 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6796 * what current time is, so simply throw away the out-of-date time. This
6797 * will result in the wakee task is less decayed, but giving the wakee more
6798 * load sounds not bad.
6800 remove_entity_load_avg(&p->se);
6802 /* Tell new CPU we are migrated */
6803 p->se.avg.last_update_time = 0;
6805 /* We have migrated, no longer consider this task hot */
6806 p->se.exec_start = 0;
6809 static void task_dead_fair(struct task_struct *p)
6811 remove_entity_load_avg(&p->se);
6814 #define task_fits_max(p, cpu) true
6815 #endif /* CONFIG_SMP */
6817 static unsigned long
6818 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6820 unsigned long gran = sysctl_sched_wakeup_granularity;
6823 * Since its curr running now, convert the gran from real-time
6824 * to virtual-time in his units.
6826 * By using 'se' instead of 'curr' we penalize light tasks, so
6827 * they get preempted easier. That is, if 'se' < 'curr' then
6828 * the resulting gran will be larger, therefore penalizing the
6829 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6830 * be smaller, again penalizing the lighter task.
6832 * This is especially important for buddies when the leftmost
6833 * task is higher priority than the buddy.
6835 return calc_delta_fair(gran, se);
6839 * Should 'se' preempt 'curr'.
6853 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6855 s64 gran, vdiff = curr->vruntime - se->vruntime;
6860 gran = wakeup_gran(curr, se);
6867 static void set_last_buddy(struct sched_entity *se)
6869 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6872 for_each_sched_entity(se)
6873 cfs_rq_of(se)->last = se;
6876 static void set_next_buddy(struct sched_entity *se)
6878 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6881 for_each_sched_entity(se)
6882 cfs_rq_of(se)->next = se;
6885 static void set_skip_buddy(struct sched_entity *se)
6887 for_each_sched_entity(se)
6888 cfs_rq_of(se)->skip = se;
6892 * Preempt the current task with a newly woken task if needed:
6894 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6896 struct task_struct *curr = rq->curr;
6897 struct sched_entity *se = &curr->se, *pse = &p->se;
6898 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6899 int scale = cfs_rq->nr_running >= sched_nr_latency;
6900 int next_buddy_marked = 0;
6902 if (unlikely(se == pse))
6906 * This is possible from callers such as attach_tasks(), in which we
6907 * unconditionally check_prempt_curr() after an enqueue (which may have
6908 * lead to a throttle). This both saves work and prevents false
6909 * next-buddy nomination below.
6911 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6914 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6915 set_next_buddy(pse);
6916 next_buddy_marked = 1;
6920 * We can come here with TIF_NEED_RESCHED already set from new task
6923 * Note: this also catches the edge-case of curr being in a throttled
6924 * group (e.g. via set_curr_task), since update_curr() (in the
6925 * enqueue of curr) will have resulted in resched being set. This
6926 * prevents us from potentially nominating it as a false LAST_BUDDY
6929 if (test_tsk_need_resched(curr))
6932 /* Idle tasks are by definition preempted by non-idle tasks. */
6933 if (unlikely(curr->policy == SCHED_IDLE) &&
6934 likely(p->policy != SCHED_IDLE))
6938 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6939 * is driven by the tick):
6941 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6944 find_matching_se(&se, &pse);
6945 update_curr(cfs_rq_of(se));
6947 if (wakeup_preempt_entity(se, pse) == 1) {
6949 * Bias pick_next to pick the sched entity that is
6950 * triggering this preemption.
6952 if (!next_buddy_marked)
6953 set_next_buddy(pse);
6962 * Only set the backward buddy when the current task is still
6963 * on the rq. This can happen when a wakeup gets interleaved
6964 * with schedule on the ->pre_schedule() or idle_balance()
6965 * point, either of which can * drop the rq lock.
6967 * Also, during early boot the idle thread is in the fair class,
6968 * for obvious reasons its a bad idea to schedule back to it.
6970 if (unlikely(!se->on_rq || curr == rq->idle))
6973 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6977 static struct task_struct *
6978 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6980 struct cfs_rq *cfs_rq = &rq->cfs;
6981 struct sched_entity *se;
6982 struct task_struct *p;
6986 #ifdef CONFIG_FAIR_GROUP_SCHED
6987 if (!cfs_rq->nr_running)
6990 if (prev->sched_class != &fair_sched_class)
6994 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6995 * likely that a next task is from the same cgroup as the current.
6997 * Therefore attempt to avoid putting and setting the entire cgroup
6998 * hierarchy, only change the part that actually changes.
7002 struct sched_entity *curr = cfs_rq->curr;
7005 * Since we got here without doing put_prev_entity() we also
7006 * have to consider cfs_rq->curr. If it is still a runnable
7007 * entity, update_curr() will update its vruntime, otherwise
7008 * forget we've ever seen it.
7012 update_curr(cfs_rq);
7017 * This call to check_cfs_rq_runtime() will do the
7018 * throttle and dequeue its entity in the parent(s).
7019 * Therefore the 'simple' nr_running test will indeed
7022 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7026 se = pick_next_entity(cfs_rq, curr);
7027 cfs_rq = group_cfs_rq(se);
7033 * Since we haven't yet done put_prev_entity and if the selected task
7034 * is a different task than we started out with, try and touch the
7035 * least amount of cfs_rqs.
7038 struct sched_entity *pse = &prev->se;
7040 while (!(cfs_rq = is_same_group(se, pse))) {
7041 int se_depth = se->depth;
7042 int pse_depth = pse->depth;
7044 if (se_depth <= pse_depth) {
7045 put_prev_entity(cfs_rq_of(pse), pse);
7046 pse = parent_entity(pse);
7048 if (se_depth >= pse_depth) {
7049 set_next_entity(cfs_rq_of(se), se);
7050 se = parent_entity(se);
7054 put_prev_entity(cfs_rq, pse);
7055 set_next_entity(cfs_rq, se);
7058 if (hrtick_enabled(rq))
7059 hrtick_start_fair(rq, p);
7061 rq->misfit_task = !task_fits_max(p, rq->cpu);
7068 if (!cfs_rq->nr_running)
7071 put_prev_task(rq, prev);
7074 se = pick_next_entity(cfs_rq, NULL);
7075 set_next_entity(cfs_rq, se);
7076 cfs_rq = group_cfs_rq(se);
7081 if (hrtick_enabled(rq))
7082 hrtick_start_fair(rq, p);
7084 rq->misfit_task = !task_fits_max(p, rq->cpu);
7089 rq->misfit_task = 0;
7091 * This is OK, because current is on_cpu, which avoids it being picked
7092 * for load-balance and preemption/IRQs are still disabled avoiding
7093 * further scheduler activity on it and we're being very careful to
7094 * re-start the picking loop.
7096 lockdep_unpin_lock(&rq->lock);
7097 new_tasks = idle_balance(rq);
7098 lockdep_pin_lock(&rq->lock);
7100 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7101 * possible for any higher priority task to appear. In that case we
7102 * must re-start the pick_next_entity() loop.
7114 * Account for a descheduled task:
7116 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7118 struct sched_entity *se = &prev->se;
7119 struct cfs_rq *cfs_rq;
7121 for_each_sched_entity(se) {
7122 cfs_rq = cfs_rq_of(se);
7123 put_prev_entity(cfs_rq, se);
7128 * sched_yield() is very simple
7130 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7132 static void yield_task_fair(struct rq *rq)
7134 struct task_struct *curr = rq->curr;
7135 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7136 struct sched_entity *se = &curr->se;
7139 * Are we the only task in the tree?
7141 if (unlikely(rq->nr_running == 1))
7144 clear_buddies(cfs_rq, se);
7146 if (curr->policy != SCHED_BATCH) {
7147 update_rq_clock(rq);
7149 * Update run-time statistics of the 'current'.
7151 update_curr(cfs_rq);
7153 * Tell update_rq_clock() that we've just updated,
7154 * so we don't do microscopic update in schedule()
7155 * and double the fastpath cost.
7157 rq_clock_skip_update(rq, true);
7163 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7165 struct sched_entity *se = &p->se;
7167 /* throttled hierarchies are not runnable */
7168 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7171 /* Tell the scheduler that we'd really like pse to run next. */
7174 yield_task_fair(rq);
7180 /**************************************************
7181 * Fair scheduling class load-balancing methods.
7185 * The purpose of load-balancing is to achieve the same basic fairness the
7186 * per-cpu scheduler provides, namely provide a proportional amount of compute
7187 * time to each task. This is expressed in the following equation:
7189 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7191 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
7192 * W_i,0 is defined as:
7194 * W_i,0 = \Sum_j w_i,j (2)
7196 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
7197 * is derived from the nice value as per prio_to_weight[].
7199 * The weight average is an exponential decay average of the instantaneous
7202 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7204 * C_i is the compute capacity of cpu i, typically it is the
7205 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7206 * can also include other factors [XXX].
7208 * To achieve this balance we define a measure of imbalance which follows
7209 * directly from (1):
7211 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7213 * We them move tasks around to minimize the imbalance. In the continuous
7214 * function space it is obvious this converges, in the discrete case we get
7215 * a few fun cases generally called infeasible weight scenarios.
7218 * - infeasible weights;
7219 * - local vs global optima in the discrete case. ]
7224 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7225 * for all i,j solution, we create a tree of cpus that follows the hardware
7226 * topology where each level pairs two lower groups (or better). This results
7227 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
7228 * tree to only the first of the previous level and we decrease the frequency
7229 * of load-balance at each level inv. proportional to the number of cpus in
7235 * \Sum { --- * --- * 2^i } = O(n) (5)
7237 * `- size of each group
7238 * | | `- number of cpus doing load-balance
7240 * `- sum over all levels
7242 * Coupled with a limit on how many tasks we can migrate every balance pass,
7243 * this makes (5) the runtime complexity of the balancer.
7245 * An important property here is that each CPU is still (indirectly) connected
7246 * to every other cpu in at most O(log n) steps:
7248 * The adjacency matrix of the resulting graph is given by:
7251 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7254 * And you'll find that:
7256 * A^(log_2 n)_i,j != 0 for all i,j (7)
7258 * Showing there's indeed a path between every cpu in at most O(log n) steps.
7259 * The task movement gives a factor of O(m), giving a convergence complexity
7262 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7267 * In order to avoid CPUs going idle while there's still work to do, new idle
7268 * balancing is more aggressive and has the newly idle cpu iterate up the domain
7269 * tree itself instead of relying on other CPUs to bring it work.
7271 * This adds some complexity to both (5) and (8) but it reduces the total idle
7279 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7282 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7287 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7289 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
7291 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7294 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7295 * rewrite all of this once again.]
7298 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7300 enum fbq_type { regular, remote, all };
7309 #define LBF_ALL_PINNED 0x01
7310 #define LBF_NEED_BREAK 0x02
7311 #define LBF_DST_PINNED 0x04
7312 #define LBF_SOME_PINNED 0x08
7315 struct sched_domain *sd;
7323 struct cpumask *dst_grpmask;
7325 enum cpu_idle_type idle;
7327 unsigned int src_grp_nr_running;
7328 /* The set of CPUs under consideration for load-balancing */
7329 struct cpumask *cpus;
7334 unsigned int loop_break;
7335 unsigned int loop_max;
7337 enum fbq_type fbq_type;
7338 enum group_type busiest_group_type;
7339 struct list_head tasks;
7343 * Is this task likely cache-hot:
7345 static int task_hot(struct task_struct *p, struct lb_env *env)
7349 lockdep_assert_held(&env->src_rq->lock);
7351 if (p->sched_class != &fair_sched_class)
7354 if (unlikely(p->policy == SCHED_IDLE))
7358 * Buddy candidates are cache hot:
7360 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7361 (&p->se == cfs_rq_of(&p->se)->next ||
7362 &p->se == cfs_rq_of(&p->se)->last))
7365 if (sysctl_sched_migration_cost == -1)
7367 if (sysctl_sched_migration_cost == 0)
7370 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7372 return delta < (s64)sysctl_sched_migration_cost;
7375 #ifdef CONFIG_NUMA_BALANCING
7377 * Returns 1, if task migration degrades locality
7378 * Returns 0, if task migration improves locality i.e migration preferred.
7379 * Returns -1, if task migration is not affected by locality.
7381 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7383 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7384 unsigned long src_faults, dst_faults;
7385 int src_nid, dst_nid;
7387 if (!static_branch_likely(&sched_numa_balancing))
7390 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7393 src_nid = cpu_to_node(env->src_cpu);
7394 dst_nid = cpu_to_node(env->dst_cpu);
7396 if (src_nid == dst_nid)
7399 /* Migrating away from the preferred node is always bad. */
7400 if (src_nid == p->numa_preferred_nid) {
7401 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7407 /* Encourage migration to the preferred node. */
7408 if (dst_nid == p->numa_preferred_nid)
7412 src_faults = group_faults(p, src_nid);
7413 dst_faults = group_faults(p, dst_nid);
7415 src_faults = task_faults(p, src_nid);
7416 dst_faults = task_faults(p, dst_nid);
7419 return dst_faults < src_faults;
7423 static inline int migrate_degrades_locality(struct task_struct *p,
7431 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7434 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7438 lockdep_assert_held(&env->src_rq->lock);
7441 * We do not migrate tasks that are:
7442 * 1) throttled_lb_pair, or
7443 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7444 * 3) running (obviously), or
7445 * 4) are cache-hot on their current CPU.
7447 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7450 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
7453 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
7455 env->flags |= LBF_SOME_PINNED;
7458 * Remember if this task can be migrated to any other cpu in
7459 * our sched_group. We may want to revisit it if we couldn't
7460 * meet load balance goals by pulling other tasks on src_cpu.
7462 * Also avoid computing new_dst_cpu if we have already computed
7463 * one in current iteration.
7465 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
7468 /* Prevent to re-select dst_cpu via env's cpus */
7469 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7470 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
7471 env->flags |= LBF_DST_PINNED;
7472 env->new_dst_cpu = cpu;
7480 /* Record that we found atleast one task that could run on dst_cpu */
7481 env->flags &= ~LBF_ALL_PINNED;
7483 if (task_running(env->src_rq, p)) {
7484 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
7489 * Aggressive migration if:
7490 * 1) destination numa is preferred
7491 * 2) task is cache cold, or
7492 * 3) too many balance attempts have failed.
7494 tsk_cache_hot = migrate_degrades_locality(p, env);
7495 if (tsk_cache_hot == -1)
7496 tsk_cache_hot = task_hot(p, env);
7498 if (tsk_cache_hot <= 0 ||
7499 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7500 if (tsk_cache_hot == 1) {
7501 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
7502 schedstat_inc(p, se.statistics.nr_forced_migrations);
7507 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
7512 * detach_task() -- detach the task for the migration specified in env
7514 static void detach_task(struct task_struct *p, struct lb_env *env)
7516 lockdep_assert_held(&env->src_rq->lock);
7518 deactivate_task(env->src_rq, p, 0);
7519 p->on_rq = TASK_ON_RQ_MIGRATING;
7520 double_lock_balance(env->src_rq, env->dst_rq);
7521 set_task_cpu(p, env->dst_cpu);
7522 double_unlock_balance(env->src_rq, env->dst_rq);
7526 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7527 * part of active balancing operations within "domain".
7529 * Returns a task if successful and NULL otherwise.
7531 static struct task_struct *detach_one_task(struct lb_env *env)
7533 struct task_struct *p, *n;
7535 lockdep_assert_held(&env->src_rq->lock);
7537 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7538 if (!can_migrate_task(p, env))
7541 detach_task(p, env);
7544 * Right now, this is only the second place where
7545 * lb_gained[env->idle] is updated (other is detach_tasks)
7546 * so we can safely collect stats here rather than
7547 * inside detach_tasks().
7549 schedstat_inc(env->sd, lb_gained[env->idle]);
7555 static const unsigned int sched_nr_migrate_break = 32;
7558 * detach_tasks() -- tries to detach up to imbalance weighted load from
7559 * busiest_rq, as part of a balancing operation within domain "sd".
7561 * Returns number of detached tasks if successful and 0 otherwise.
7563 static int detach_tasks(struct lb_env *env)
7565 struct list_head *tasks = &env->src_rq->cfs_tasks;
7566 struct task_struct *p;
7570 lockdep_assert_held(&env->src_rq->lock);
7572 if (env->imbalance <= 0)
7575 while (!list_empty(tasks)) {
7577 * We don't want to steal all, otherwise we may be treated likewise,
7578 * which could at worst lead to a livelock crash.
7580 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7583 p = list_first_entry(tasks, struct task_struct, se.group_node);
7586 /* We've more or less seen every task there is, call it quits */
7587 if (env->loop > env->loop_max)
7590 /* take a breather every nr_migrate tasks */
7591 if (env->loop > env->loop_break) {
7592 env->loop_break += sched_nr_migrate_break;
7593 env->flags |= LBF_NEED_BREAK;
7597 if (!can_migrate_task(p, env))
7600 load = task_h_load(p);
7602 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7605 if ((load / 2) > env->imbalance)
7608 detach_task(p, env);
7609 list_add(&p->se.group_node, &env->tasks);
7612 env->imbalance -= load;
7614 #ifdef CONFIG_PREEMPT
7616 * NEWIDLE balancing is a source of latency, so preemptible
7617 * kernels will stop after the first task is detached to minimize
7618 * the critical section.
7620 if (env->idle == CPU_NEWLY_IDLE)
7625 * We only want to steal up to the prescribed amount of
7628 if (env->imbalance <= 0)
7633 list_move_tail(&p->se.group_node, tasks);
7637 * Right now, this is one of only two places we collect this stat
7638 * so we can safely collect detach_one_task() stats here rather
7639 * than inside detach_one_task().
7641 schedstat_add(env->sd, lb_gained[env->idle], detached);
7647 * attach_task() -- attach the task detached by detach_task() to its new rq.
7649 static void attach_task(struct rq *rq, struct task_struct *p)
7651 lockdep_assert_held(&rq->lock);
7653 BUG_ON(task_rq(p) != rq);
7654 p->on_rq = TASK_ON_RQ_QUEUED;
7655 activate_task(rq, p, 0);
7656 check_preempt_curr(rq, p, 0);
7660 * attach_one_task() -- attaches the task returned from detach_one_task() to
7663 static void attach_one_task(struct rq *rq, struct task_struct *p)
7665 raw_spin_lock(&rq->lock);
7667 raw_spin_unlock(&rq->lock);
7671 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7674 static void attach_tasks(struct lb_env *env)
7676 struct list_head *tasks = &env->tasks;
7677 struct task_struct *p;
7679 raw_spin_lock(&env->dst_rq->lock);
7681 while (!list_empty(tasks)) {
7682 p = list_first_entry(tasks, struct task_struct, se.group_node);
7683 list_del_init(&p->se.group_node);
7685 attach_task(env->dst_rq, p);
7688 raw_spin_unlock(&env->dst_rq->lock);
7691 #ifdef CONFIG_FAIR_GROUP_SCHED
7692 static void update_blocked_averages(int cpu)
7694 struct rq *rq = cpu_rq(cpu);
7695 struct cfs_rq *cfs_rq;
7696 unsigned long flags;
7698 raw_spin_lock_irqsave(&rq->lock, flags);
7699 update_rq_clock(rq);
7702 * Iterates the task_group tree in a bottom up fashion, see
7703 * list_add_leaf_cfs_rq() for details.
7705 for_each_leaf_cfs_rq(rq, cfs_rq) {
7706 /* throttled entities do not contribute to load */
7707 if (throttled_hierarchy(cfs_rq))
7710 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7712 update_tg_load_avg(cfs_rq, 0);
7714 /* Propagate pending load changes to the parent */
7715 if (cfs_rq->tg->se[cpu])
7716 update_load_avg(cfs_rq->tg->se[cpu], 0);
7718 raw_spin_unlock_irqrestore(&rq->lock, flags);
7722 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7723 * This needs to be done in a top-down fashion because the load of a child
7724 * group is a fraction of its parents load.
7726 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7728 struct rq *rq = rq_of(cfs_rq);
7729 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7730 unsigned long now = jiffies;
7733 if (cfs_rq->last_h_load_update == now)
7736 cfs_rq->h_load_next = NULL;
7737 for_each_sched_entity(se) {
7738 cfs_rq = cfs_rq_of(se);
7739 cfs_rq->h_load_next = se;
7740 if (cfs_rq->last_h_load_update == now)
7745 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7746 cfs_rq->last_h_load_update = now;
7749 while ((se = cfs_rq->h_load_next) != NULL) {
7750 load = cfs_rq->h_load;
7751 load = div64_ul(load * se->avg.load_avg,
7752 cfs_rq_load_avg(cfs_rq) + 1);
7753 cfs_rq = group_cfs_rq(se);
7754 cfs_rq->h_load = load;
7755 cfs_rq->last_h_load_update = now;
7759 static unsigned long task_h_load(struct task_struct *p)
7761 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7763 update_cfs_rq_h_load(cfs_rq);
7764 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7765 cfs_rq_load_avg(cfs_rq) + 1);
7768 static inline void update_blocked_averages(int cpu)
7770 struct rq *rq = cpu_rq(cpu);
7771 struct cfs_rq *cfs_rq = &rq->cfs;
7772 unsigned long flags;
7774 raw_spin_lock_irqsave(&rq->lock, flags);
7775 update_rq_clock(rq);
7776 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7777 raw_spin_unlock_irqrestore(&rq->lock, flags);
7780 static unsigned long task_h_load(struct task_struct *p)
7782 return p->se.avg.load_avg;
7786 /********** Helpers for find_busiest_group ************************/
7789 * sg_lb_stats - stats of a sched_group required for load_balancing
7791 struct sg_lb_stats {
7792 unsigned long avg_load; /*Avg load across the CPUs of the group */
7793 unsigned long group_load; /* Total load over the CPUs of the group */
7794 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7795 unsigned long load_per_task;
7796 unsigned long group_capacity;
7797 unsigned long group_util; /* Total utilization of the group */
7798 unsigned int sum_nr_running; /* Nr tasks running in the group */
7799 unsigned int idle_cpus;
7800 unsigned int group_weight;
7801 enum group_type group_type;
7802 int group_no_capacity;
7803 int group_misfit_task; /* A cpu has a task too big for its capacity */
7804 #ifdef CONFIG_NUMA_BALANCING
7805 unsigned int nr_numa_running;
7806 unsigned int nr_preferred_running;
7811 * sd_lb_stats - Structure to store the statistics of a sched_domain
7812 * during load balancing.
7814 struct sd_lb_stats {
7815 struct sched_group *busiest; /* Busiest group in this sd */
7816 struct sched_group *local; /* Local group in this sd */
7817 unsigned long total_load; /* Total load of all groups in sd */
7818 unsigned long total_capacity; /* Total capacity of all groups in sd */
7819 unsigned long avg_load; /* Average load across all groups in sd */
7821 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7822 struct sg_lb_stats local_stat; /* Statistics of the local group */
7825 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7828 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7829 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7830 * We must however clear busiest_stat::avg_load because
7831 * update_sd_pick_busiest() reads this before assignment.
7833 *sds = (struct sd_lb_stats){
7837 .total_capacity = 0UL,
7840 .sum_nr_running = 0,
7841 .group_type = group_other,
7847 * get_sd_load_idx - Obtain the load index for a given sched domain.
7848 * @sd: The sched_domain whose load_idx is to be obtained.
7849 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7851 * Return: The load index.
7853 static inline int get_sd_load_idx(struct sched_domain *sd,
7854 enum cpu_idle_type idle)
7860 load_idx = sd->busy_idx;
7863 case CPU_NEWLY_IDLE:
7864 load_idx = sd->newidle_idx;
7867 load_idx = sd->idle_idx;
7874 static unsigned long scale_rt_capacity(int cpu)
7876 struct rq *rq = cpu_rq(cpu);
7877 u64 total, used, age_stamp, avg;
7881 * Since we're reading these variables without serialization make sure
7882 * we read them once before doing sanity checks on them.
7884 age_stamp = READ_ONCE(rq->age_stamp);
7885 avg = READ_ONCE(rq->rt_avg);
7886 delta = __rq_clock_broken(rq) - age_stamp;
7888 if (unlikely(delta < 0))
7891 total = sched_avg_period() + delta;
7893 used = div_u64(avg, total);
7896 * deadline bandwidth is defined at system level so we must
7897 * weight this bandwidth with the max capacity of the system.
7898 * As a reminder, avg_bw is 20bits width and
7899 * scale_cpu_capacity is 10 bits width
7901 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7903 if (likely(used < SCHED_CAPACITY_SCALE))
7904 return SCHED_CAPACITY_SCALE - used;
7909 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7911 raw_spin_lock_init(&mcc->lock);
7916 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7918 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7919 struct sched_group *sdg = sd->groups;
7920 struct max_cpu_capacity *mcc;
7921 unsigned long max_capacity;
7923 unsigned long flags;
7925 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7927 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7929 raw_spin_lock_irqsave(&mcc->lock, flags);
7930 max_capacity = mcc->val;
7931 max_cap_cpu = mcc->cpu;
7933 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7934 (max_capacity < capacity)) {
7935 mcc->val = capacity;
7937 #ifdef CONFIG_SCHED_DEBUG
7938 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7939 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7944 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7946 skip_unlock: __attribute__ ((unused));
7947 capacity *= scale_rt_capacity(cpu);
7948 capacity >>= SCHED_CAPACITY_SHIFT;
7953 cpu_rq(cpu)->cpu_capacity = capacity;
7954 sdg->sgc->capacity = capacity;
7955 sdg->sgc->max_capacity = capacity;
7956 sdg->sgc->min_capacity = capacity;
7959 void update_group_capacity(struct sched_domain *sd, int cpu)
7961 struct sched_domain *child = sd->child;
7962 struct sched_group *group, *sdg = sd->groups;
7963 unsigned long capacity, max_capacity, min_capacity;
7964 unsigned long interval;
7966 interval = msecs_to_jiffies(sd->balance_interval);
7967 interval = clamp(interval, 1UL, max_load_balance_interval);
7968 sdg->sgc->next_update = jiffies + interval;
7971 update_cpu_capacity(sd, cpu);
7977 min_capacity = ULONG_MAX;
7979 if (child->flags & SD_OVERLAP) {
7981 * SD_OVERLAP domains cannot assume that child groups
7982 * span the current group.
7985 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7986 struct sched_group_capacity *sgc;
7987 struct rq *rq = cpu_rq(cpu);
7990 * build_sched_domains() -> init_sched_groups_capacity()
7991 * gets here before we've attached the domains to the
7994 * Use capacity_of(), which is set irrespective of domains
7995 * in update_cpu_capacity().
7997 * This avoids capacity from being 0 and
7998 * causing divide-by-zero issues on boot.
8000 if (unlikely(!rq->sd)) {
8001 capacity += capacity_of(cpu);
8003 sgc = rq->sd->groups->sgc;
8004 capacity += sgc->capacity;
8007 max_capacity = max(capacity, max_capacity);
8008 min_capacity = min(capacity, min_capacity);
8012 * !SD_OVERLAP domains can assume that child groups
8013 * span the current group.
8016 group = child->groups;
8018 struct sched_group_capacity *sgc = group->sgc;
8020 capacity += sgc->capacity;
8021 max_capacity = max(sgc->max_capacity, max_capacity);
8022 min_capacity = min(sgc->min_capacity, min_capacity);
8023 group = group->next;
8024 } while (group != child->groups);
8027 sdg->sgc->capacity = capacity;
8028 sdg->sgc->max_capacity = max_capacity;
8029 sdg->sgc->min_capacity = min_capacity;
8033 * Check whether the capacity of the rq has been noticeably reduced by side
8034 * activity. The imbalance_pct is used for the threshold.
8035 * Return true is the capacity is reduced
8038 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8040 return ((rq->cpu_capacity * sd->imbalance_pct) <
8041 (rq->cpu_capacity_orig * 100));
8045 * Group imbalance indicates (and tries to solve) the problem where balancing
8046 * groups is inadequate due to tsk_cpus_allowed() constraints.
8048 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
8049 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
8052 * { 0 1 2 3 } { 4 5 6 7 }
8055 * If we were to balance group-wise we'd place two tasks in the first group and
8056 * two tasks in the second group. Clearly this is undesired as it will overload
8057 * cpu 3 and leave one of the cpus in the second group unused.
8059 * The current solution to this issue is detecting the skew in the first group
8060 * by noticing the lower domain failed to reach balance and had difficulty
8061 * moving tasks due to affinity constraints.
8063 * When this is so detected; this group becomes a candidate for busiest; see
8064 * update_sd_pick_busiest(). And calculate_imbalance() and
8065 * find_busiest_group() avoid some of the usual balance conditions to allow it
8066 * to create an effective group imbalance.
8068 * This is a somewhat tricky proposition since the next run might not find the
8069 * group imbalance and decide the groups need to be balanced again. A most
8070 * subtle and fragile situation.
8073 static inline int sg_imbalanced(struct sched_group *group)
8075 return group->sgc->imbalance;
8079 * group_has_capacity returns true if the group has spare capacity that could
8080 * be used by some tasks.
8081 * We consider that a group has spare capacity if the * number of task is
8082 * smaller than the number of CPUs or if the utilization is lower than the
8083 * available capacity for CFS tasks.
8084 * For the latter, we use a threshold to stabilize the state, to take into
8085 * account the variance of the tasks' load and to return true if the available
8086 * capacity in meaningful for the load balancer.
8087 * As an example, an available capacity of 1% can appear but it doesn't make
8088 * any benefit for the load balance.
8091 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8093 if (sgs->sum_nr_running < sgs->group_weight)
8096 if ((sgs->group_capacity * 100) >
8097 (sgs->group_util * env->sd->imbalance_pct))
8104 * group_is_overloaded returns true if the group has more tasks than it can
8106 * group_is_overloaded is not equals to !group_has_capacity because a group
8107 * with the exact right number of tasks, has no more spare capacity but is not
8108 * overloaded so both group_has_capacity and group_is_overloaded return
8112 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8114 if (sgs->sum_nr_running <= sgs->group_weight)
8117 if ((sgs->group_capacity * 100) <
8118 (sgs->group_util * env->sd->imbalance_pct))
8126 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
8127 * per-cpu capacity than sched_group ref.
8130 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8132 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
8133 ref->sgc->max_capacity;
8137 group_type group_classify(struct sched_group *group,
8138 struct sg_lb_stats *sgs)
8140 if (sgs->group_no_capacity)
8141 return group_overloaded;
8143 if (sg_imbalanced(group))
8144 return group_imbalanced;
8146 if (sgs->group_misfit_task)
8147 return group_misfit_task;
8152 #ifdef CONFIG_NO_HZ_COMMON
8154 * idle load balancing data
8155 * - used by the nohz balance, but we want it available here
8156 * so that we can see which CPUs have no tick.
8159 cpumask_var_t idle_cpus_mask;
8161 unsigned long next_balance; /* in jiffy units */
8162 } nohz ____cacheline_aligned;
8164 static inline void update_cpu_stats_if_tickless(struct rq *rq)
8166 /* only called from update_sg_lb_stats when irqs are disabled */
8167 if (cpumask_test_cpu(rq->cpu, nohz.idle_cpus_mask)) {
8168 /* rate limit updates to once-per-jiffie at most */
8169 if (READ_ONCE(jiffies) <= rq->last_load_update_tick)
8172 raw_spin_lock(&rq->lock);
8173 update_rq_clock(rq);
8174 update_idle_cpu_load(rq);
8175 update_cfs_rq_load_avg(rq->clock_task, &rq->cfs, false);
8176 raw_spin_unlock(&rq->lock);
8181 static inline void update_cpu_stats_if_tickless(struct rq *rq) { }
8185 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8186 * @env: The load balancing environment.
8187 * @group: sched_group whose statistics are to be updated.
8188 * @load_idx: Load index of sched_domain of this_cpu for load calc.
8189 * @local_group: Does group contain this_cpu.
8190 * @sgs: variable to hold the statistics for this group.
8191 * @overload: Indicate more than one runnable task for any CPU.
8192 * @overutilized: Indicate overutilization for any CPU.
8194 static inline void update_sg_lb_stats(struct lb_env *env,
8195 struct sched_group *group, int load_idx,
8196 int local_group, struct sg_lb_stats *sgs,
8197 bool *overload, bool *overutilized)
8202 memset(sgs, 0, sizeof(*sgs));
8204 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8205 struct rq *rq = cpu_rq(i);
8207 /* if we are entering idle and there are CPUs with
8208 * their tick stopped, do an update for them
8210 if (env->idle == CPU_NEWLY_IDLE)
8211 update_cpu_stats_if_tickless(rq);
8213 /* Bias balancing toward cpus of our domain */
8215 load = target_load(i, load_idx);
8217 load = source_load(i, load_idx);
8219 sgs->group_load += load;
8220 sgs->group_util += cpu_util(i);
8221 sgs->sum_nr_running += rq->cfs.h_nr_running;
8223 nr_running = rq->nr_running;
8227 #ifdef CONFIG_NUMA_BALANCING
8228 sgs->nr_numa_running += rq->nr_numa_running;
8229 sgs->nr_preferred_running += rq->nr_preferred_running;
8231 sgs->sum_weighted_load += weighted_cpuload(i);
8233 * No need to call idle_cpu() if nr_running is not 0
8235 if (!nr_running && idle_cpu(i))
8238 if (cpu_overutilized(i)) {
8239 *overutilized = true;
8240 if (!sgs->group_misfit_task && rq->misfit_task)
8241 sgs->group_misfit_task = capacity_of(i);
8245 /* Adjust by relative CPU capacity of the group */
8246 sgs->group_capacity = group->sgc->capacity;
8247 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8249 if (sgs->sum_nr_running)
8250 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8252 sgs->group_weight = group->group_weight;
8254 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8255 sgs->group_type = group_classify(group, sgs);
8259 * update_sd_pick_busiest - return 1 on busiest group
8260 * @env: The load balancing environment.
8261 * @sds: sched_domain statistics
8262 * @sg: sched_group candidate to be checked for being the busiest
8263 * @sgs: sched_group statistics
8265 * Determine if @sg is a busier group than the previously selected
8268 * Return: %true if @sg is a busier group than the previously selected
8269 * busiest group. %false otherwise.
8271 static bool update_sd_pick_busiest(struct lb_env *env,
8272 struct sd_lb_stats *sds,
8273 struct sched_group *sg,
8274 struct sg_lb_stats *sgs)
8276 struct sg_lb_stats *busiest = &sds->busiest_stat;
8278 if (sgs->group_type > busiest->group_type)
8281 if (sgs->group_type < busiest->group_type)
8285 * Candidate sg doesn't face any serious load-balance problems
8286 * so don't pick it if the local sg is already filled up.
8288 if (sgs->group_type == group_other &&
8289 !group_has_capacity(env, &sds->local_stat))
8292 if (sgs->avg_load <= busiest->avg_load)
8295 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8299 * Candidate sg has no more than one task per CPU and
8300 * has higher per-CPU capacity. Migrating tasks to less
8301 * capable CPUs may harm throughput. Maximize throughput,
8302 * power/energy consequences are not considered.
8304 if (sgs->sum_nr_running <= sgs->group_weight &&
8305 group_smaller_cpu_capacity(sds->local, sg))
8309 /* This is the busiest node in its class. */
8310 if (!(env->sd->flags & SD_ASYM_PACKING))
8314 * ASYM_PACKING needs to move all the work to the lowest
8315 * numbered CPUs in the group, therefore mark all groups
8316 * higher than ourself as busy.
8318 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
8322 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
8329 #ifdef CONFIG_NUMA_BALANCING
8330 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8332 if (sgs->sum_nr_running > sgs->nr_numa_running)
8334 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8339 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8341 if (rq->nr_running > rq->nr_numa_running)
8343 if (rq->nr_running > rq->nr_preferred_running)
8348 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8353 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8357 #endif /* CONFIG_NUMA_BALANCING */
8359 #define lb_sd_parent(sd) \
8360 (sd->parent && sd->parent->groups != sd->parent->groups->next)
8363 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8364 * @env: The load balancing environment.
8365 * @sds: variable to hold the statistics for this sched_domain.
8367 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8369 struct sched_domain *child = env->sd->child;
8370 struct sched_group *sg = env->sd->groups;
8371 struct sg_lb_stats tmp_sgs;
8372 int load_idx, prefer_sibling = 0;
8373 bool overload = false, overutilized = false;
8375 if (child && child->flags & SD_PREFER_SIBLING)
8378 load_idx = get_sd_load_idx(env->sd, env->idle);
8381 struct sg_lb_stats *sgs = &tmp_sgs;
8384 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
8387 sgs = &sds->local_stat;
8389 if (env->idle != CPU_NEWLY_IDLE ||
8390 time_after_eq(jiffies, sg->sgc->next_update))
8391 update_group_capacity(env->sd, env->dst_cpu);
8394 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8395 &overload, &overutilized);
8401 * In case the child domain prefers tasks go to siblings
8402 * first, lower the sg capacity so that we'll try
8403 * and move all the excess tasks away. We lower the capacity
8404 * of a group only if the local group has the capacity to fit
8405 * these excess tasks. The extra check prevents the case where
8406 * you always pull from the heaviest group when it is already
8407 * under-utilized (possible with a large weight task outweighs
8408 * the tasks on the system).
8410 if (prefer_sibling && sds->local &&
8411 group_has_capacity(env, &sds->local_stat) &&
8412 (sgs->sum_nr_running > 1)) {
8413 sgs->group_no_capacity = 1;
8414 sgs->group_type = group_classify(sg, sgs);
8418 * Ignore task groups with misfit tasks if local group has no
8419 * capacity or if per-cpu capacity isn't higher.
8421 if (sgs->group_type == group_misfit_task &&
8422 (!group_has_capacity(env, &sds->local_stat) ||
8423 !group_smaller_cpu_capacity(sg, sds->local)))
8424 sgs->group_type = group_other;
8426 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8428 sds->busiest_stat = *sgs;
8432 /* Now, start updating sd_lb_stats */
8433 sds->total_load += sgs->group_load;
8434 sds->total_capacity += sgs->group_capacity;
8437 } while (sg != env->sd->groups);
8439 if (env->sd->flags & SD_NUMA)
8440 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8442 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
8444 if (!lb_sd_parent(env->sd)) {
8445 /* update overload indicator if we are at root domain */
8446 if (env->dst_rq->rd->overload != overload)
8447 env->dst_rq->rd->overload = overload;
8449 /* Update over-utilization (tipping point, U >= 0) indicator */
8450 if (env->dst_rq->rd->overutilized != overutilized) {
8451 env->dst_rq->rd->overutilized = overutilized;
8452 trace_sched_overutilized(overutilized);
8455 if (!env->dst_rq->rd->overutilized && overutilized) {
8456 env->dst_rq->rd->overutilized = true;
8457 trace_sched_overutilized(true);
8464 * check_asym_packing - Check to see if the group is packed into the
8467 * This is primarily intended to used at the sibling level. Some
8468 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8469 * case of POWER7, it can move to lower SMT modes only when higher
8470 * threads are idle. When in lower SMT modes, the threads will
8471 * perform better since they share less core resources. Hence when we
8472 * have idle threads, we want them to be the higher ones.
8474 * This packing function is run on idle threads. It checks to see if
8475 * the busiest CPU in this domain (core in the P7 case) has a higher
8476 * CPU number than the packing function is being run on. Here we are
8477 * assuming lower CPU number will be equivalent to lower a SMT thread
8480 * Return: 1 when packing is required and a task should be moved to
8481 * this CPU. The amount of the imbalance is returned in *imbalance.
8483 * @env: The load balancing environment.
8484 * @sds: Statistics of the sched_domain which is to be packed
8486 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8490 if (!(env->sd->flags & SD_ASYM_PACKING))
8496 busiest_cpu = group_first_cpu(sds->busiest);
8497 if (env->dst_cpu > busiest_cpu)
8500 env->imbalance = DIV_ROUND_CLOSEST(
8501 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8502 SCHED_CAPACITY_SCALE);
8508 * fix_small_imbalance - Calculate the minor imbalance that exists
8509 * amongst the groups of a sched_domain, during
8511 * @env: The load balancing environment.
8512 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8515 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8517 unsigned long tmp, capa_now = 0, capa_move = 0;
8518 unsigned int imbn = 2;
8519 unsigned long scaled_busy_load_per_task;
8520 struct sg_lb_stats *local, *busiest;
8522 local = &sds->local_stat;
8523 busiest = &sds->busiest_stat;
8525 if (!local->sum_nr_running)
8526 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8527 else if (busiest->load_per_task > local->load_per_task)
8530 scaled_busy_load_per_task =
8531 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8532 busiest->group_capacity;
8534 if (busiest->avg_load + scaled_busy_load_per_task >=
8535 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8536 env->imbalance = busiest->load_per_task;
8541 * OK, we don't have enough imbalance to justify moving tasks,
8542 * however we may be able to increase total CPU capacity used by
8546 capa_now += busiest->group_capacity *
8547 min(busiest->load_per_task, busiest->avg_load);
8548 capa_now += local->group_capacity *
8549 min(local->load_per_task, local->avg_load);
8550 capa_now /= SCHED_CAPACITY_SCALE;
8552 /* Amount of load we'd subtract */
8553 if (busiest->avg_load > scaled_busy_load_per_task) {
8554 capa_move += busiest->group_capacity *
8555 min(busiest->load_per_task,
8556 busiest->avg_load - scaled_busy_load_per_task);
8559 /* Amount of load we'd add */
8560 if (busiest->avg_load * busiest->group_capacity <
8561 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8562 tmp = (busiest->avg_load * busiest->group_capacity) /
8563 local->group_capacity;
8565 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8566 local->group_capacity;
8568 capa_move += local->group_capacity *
8569 min(local->load_per_task, local->avg_load + tmp);
8570 capa_move /= SCHED_CAPACITY_SCALE;
8572 /* Move if we gain throughput */
8573 if (capa_move > capa_now)
8574 env->imbalance = busiest->load_per_task;
8578 * calculate_imbalance - Calculate the amount of imbalance present within the
8579 * groups of a given sched_domain during load balance.
8580 * @env: load balance environment
8581 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8583 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8585 unsigned long max_pull, load_above_capacity = ~0UL;
8586 struct sg_lb_stats *local, *busiest;
8588 local = &sds->local_stat;
8589 busiest = &sds->busiest_stat;
8591 if (busiest->group_type == group_imbalanced) {
8593 * In the group_imb case we cannot rely on group-wide averages
8594 * to ensure cpu-load equilibrium, look at wider averages. XXX
8596 busiest->load_per_task =
8597 min(busiest->load_per_task, sds->avg_load);
8601 * In the presence of smp nice balancing, certain scenarios can have
8602 * max load less than avg load(as we skip the groups at or below
8603 * its cpu_capacity, while calculating max_load..)
8605 if (busiest->avg_load <= sds->avg_load ||
8606 local->avg_load >= sds->avg_load) {
8607 /* Misfitting tasks should be migrated in any case */
8608 if (busiest->group_type == group_misfit_task) {
8609 env->imbalance = busiest->group_misfit_task;
8614 * Busiest group is overloaded, local is not, use the spare
8615 * cycles to maximize throughput
8617 if (busiest->group_type == group_overloaded &&
8618 local->group_type <= group_misfit_task) {
8619 env->imbalance = busiest->load_per_task;
8624 return fix_small_imbalance(env, sds);
8628 * If there aren't any idle cpus, avoid creating some.
8630 if (busiest->group_type == group_overloaded &&
8631 local->group_type == group_overloaded) {
8632 load_above_capacity = busiest->sum_nr_running *
8634 if (load_above_capacity > busiest->group_capacity)
8635 load_above_capacity -= busiest->group_capacity;
8637 load_above_capacity = ~0UL;
8641 * We're trying to get all the cpus to the average_load, so we don't
8642 * want to push ourselves above the average load, nor do we wish to
8643 * reduce the max loaded cpu below the average load. At the same time,
8644 * we also don't want to reduce the group load below the group capacity
8645 * (so that we can implement power-savings policies etc). Thus we look
8646 * for the minimum possible imbalance.
8648 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8650 /* How much load to actually move to equalise the imbalance */
8651 env->imbalance = min(
8652 max_pull * busiest->group_capacity,
8653 (sds->avg_load - local->avg_load) * local->group_capacity
8654 ) / SCHED_CAPACITY_SCALE;
8656 /* Boost imbalance to allow misfit task to be balanced. */
8657 if (busiest->group_type == group_misfit_task)
8658 env->imbalance = max_t(long, env->imbalance,
8659 busiest->group_misfit_task);
8662 * if *imbalance is less than the average load per runnable task
8663 * there is no guarantee that any tasks will be moved so we'll have
8664 * a think about bumping its value to force at least one task to be
8667 if (env->imbalance < busiest->load_per_task)
8668 return fix_small_imbalance(env, sds);
8671 /******* find_busiest_group() helpers end here *********************/
8674 * find_busiest_group - Returns the busiest group within the sched_domain
8675 * if there is an imbalance. If there isn't an imbalance, and
8676 * the user has opted for power-savings, it returns a group whose
8677 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8678 * such a group exists.
8680 * Also calculates the amount of weighted load which should be moved
8681 * to restore balance.
8683 * @env: The load balancing environment.
8685 * Return: - The busiest group if imbalance exists.
8686 * - If no imbalance and user has opted for power-savings balance,
8687 * return the least loaded group whose CPUs can be
8688 * put to idle by rebalancing its tasks onto our group.
8690 static struct sched_group *find_busiest_group(struct lb_env *env)
8692 struct sg_lb_stats *local, *busiest;
8693 struct sd_lb_stats sds;
8695 init_sd_lb_stats(&sds);
8698 * Compute the various statistics relavent for load balancing at
8701 update_sd_lb_stats(env, &sds);
8703 if (energy_aware() && !env->dst_rq->rd->overutilized)
8706 local = &sds.local_stat;
8707 busiest = &sds.busiest_stat;
8709 /* ASYM feature bypasses nice load balance check */
8710 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8711 check_asym_packing(env, &sds))
8714 /* There is no busy sibling group to pull tasks from */
8715 if (!sds.busiest || busiest->sum_nr_running == 0)
8718 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8719 / sds.total_capacity;
8722 * If the busiest group is imbalanced the below checks don't
8723 * work because they assume all things are equal, which typically
8724 * isn't true due to cpus_allowed constraints and the like.
8726 if (busiest->group_type == group_imbalanced)
8730 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8731 * capacities from resulting in underutilization due to avg_load.
8733 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8734 busiest->group_no_capacity)
8737 /* Misfitting tasks should be dealt with regardless of the avg load */
8738 if (busiest->group_type == group_misfit_task) {
8743 * If the local group is busier than the selected busiest group
8744 * don't try and pull any tasks.
8746 if (local->avg_load >= busiest->avg_load)
8750 * Don't pull any tasks if this group is already above the domain
8753 if (local->avg_load >= sds.avg_load)
8756 if (env->idle == CPU_IDLE) {
8758 * This cpu is idle. If the busiest group is not overloaded
8759 * and there is no imbalance between this and busiest group
8760 * wrt idle cpus, it is balanced. The imbalance becomes
8761 * significant if the diff is greater than 1 otherwise we
8762 * might end up to just move the imbalance on another group
8764 if ((busiest->group_type != group_overloaded) &&
8765 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8766 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8770 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8771 * imbalance_pct to be conservative.
8773 if (100 * busiest->avg_load <=
8774 env->sd->imbalance_pct * local->avg_load)
8779 env->busiest_group_type = busiest->group_type;
8780 /* Looks like there is an imbalance. Compute it */
8781 calculate_imbalance(env, &sds);
8790 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8792 static struct rq *find_busiest_queue(struct lb_env *env,
8793 struct sched_group *group)
8795 struct rq *busiest = NULL, *rq;
8796 unsigned long busiest_load = 0, busiest_capacity = 1;
8799 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8800 unsigned long capacity, wl;
8804 rt = fbq_classify_rq(rq);
8807 * We classify groups/runqueues into three groups:
8808 * - regular: there are !numa tasks
8809 * - remote: there are numa tasks that run on the 'wrong' node
8810 * - all: there is no distinction
8812 * In order to avoid migrating ideally placed numa tasks,
8813 * ignore those when there's better options.
8815 * If we ignore the actual busiest queue to migrate another
8816 * task, the next balance pass can still reduce the busiest
8817 * queue by moving tasks around inside the node.
8819 * If we cannot move enough load due to this classification
8820 * the next pass will adjust the group classification and
8821 * allow migration of more tasks.
8823 * Both cases only affect the total convergence complexity.
8825 if (rt > env->fbq_type)
8828 capacity = capacity_of(i);
8830 wl = weighted_cpuload(i);
8833 * When comparing with imbalance, use weighted_cpuload()
8834 * which is not scaled with the cpu capacity.
8837 if (rq->nr_running == 1 && wl > env->imbalance &&
8838 !check_cpu_capacity(rq, env->sd) &&
8839 env->busiest_group_type != group_misfit_task)
8843 * For the load comparisons with the other cpu's, consider
8844 * the weighted_cpuload() scaled with the cpu capacity, so
8845 * that the load can be moved away from the cpu that is
8846 * potentially running at a lower capacity.
8848 * Thus we're looking for max(wl_i / capacity_i), crosswise
8849 * multiplication to rid ourselves of the division works out
8850 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8851 * our previous maximum.
8853 if (wl * busiest_capacity > busiest_load * capacity) {
8855 busiest_capacity = capacity;
8864 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8865 * so long as it is large enough.
8867 #define MAX_PINNED_INTERVAL 512
8869 /* Working cpumask for load_balance and load_balance_newidle. */
8870 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8872 static int need_active_balance(struct lb_env *env)
8874 struct sched_domain *sd = env->sd;
8876 if (env->idle == CPU_NEWLY_IDLE) {
8879 * ASYM_PACKING needs to force migrate tasks from busy but
8880 * higher numbered CPUs in order to pack all tasks in the
8881 * lowest numbered CPUs.
8883 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8888 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8889 * It's worth migrating the task if the src_cpu's capacity is reduced
8890 * because of other sched_class or IRQs if more capacity stays
8891 * available on dst_cpu.
8893 if ((env->idle != CPU_NOT_IDLE) &&
8894 (env->src_rq->cfs.h_nr_running == 1)) {
8895 if ((check_cpu_capacity(env->src_rq, sd)) &&
8896 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8900 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8901 ((capacity_orig_of(env->src_cpu) < capacity_orig_of(env->dst_cpu))) &&
8902 env->src_rq->cfs.h_nr_running == 1 &&
8903 cpu_overutilized(env->src_cpu) &&
8904 !cpu_overutilized(env->dst_cpu)) {
8908 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8911 static int active_load_balance_cpu_stop(void *data);
8913 static int should_we_balance(struct lb_env *env)
8915 struct sched_group *sg = env->sd->groups;
8916 struct cpumask *sg_cpus, *sg_mask;
8917 int cpu, balance_cpu = -1;
8920 * In the newly idle case, we will allow all the cpu's
8921 * to do the newly idle load balance.
8923 if (env->idle == CPU_NEWLY_IDLE)
8926 sg_cpus = sched_group_cpus(sg);
8927 sg_mask = sched_group_mask(sg);
8928 /* Try to find first idle cpu */
8929 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8930 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8937 if (balance_cpu == -1)
8938 balance_cpu = group_balance_cpu(sg);
8941 * First idle cpu or the first cpu(busiest) in this sched group
8942 * is eligible for doing load balancing at this and above domains.
8944 return balance_cpu == env->dst_cpu;
8948 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8949 * tasks if there is an imbalance.
8951 static int load_balance(int this_cpu, struct rq *this_rq,
8952 struct sched_domain *sd, enum cpu_idle_type idle,
8953 int *continue_balancing)
8955 int ld_moved, cur_ld_moved, active_balance = 0;
8956 struct sched_domain *sd_parent = lb_sd_parent(sd) ? sd->parent : NULL;
8957 struct sched_group *group;
8959 unsigned long flags;
8960 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8962 struct lb_env env = {
8964 .dst_cpu = this_cpu,
8966 .dst_grpmask = sched_group_cpus(sd->groups),
8968 .loop_break = sched_nr_migrate_break,
8971 .tasks = LIST_HEAD_INIT(env.tasks),
8975 * For NEWLY_IDLE load_balancing, we don't need to consider
8976 * other cpus in our group
8978 if (idle == CPU_NEWLY_IDLE)
8979 env.dst_grpmask = NULL;
8981 cpumask_copy(cpus, cpu_active_mask);
8983 schedstat_inc(sd, lb_count[idle]);
8986 if (!should_we_balance(&env)) {
8987 *continue_balancing = 0;
8991 group = find_busiest_group(&env);
8993 schedstat_inc(sd, lb_nobusyg[idle]);
8997 busiest = find_busiest_queue(&env, group);
8999 schedstat_inc(sd, lb_nobusyq[idle]);
9003 BUG_ON(busiest == env.dst_rq);
9005 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
9007 env.src_cpu = busiest->cpu;
9008 env.src_rq = busiest;
9011 if (busiest->nr_running > 1) {
9013 * Attempt to move tasks. If find_busiest_group has found
9014 * an imbalance but busiest->nr_running <= 1, the group is
9015 * still unbalanced. ld_moved simply stays zero, so it is
9016 * correctly treated as an imbalance.
9018 env.flags |= LBF_ALL_PINNED;
9019 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9022 raw_spin_lock_irqsave(&busiest->lock, flags);
9023 update_rq_clock(busiest);
9026 * cur_ld_moved - load moved in current iteration
9027 * ld_moved - cumulative load moved across iterations
9029 cur_ld_moved = detach_tasks(&env);
9032 * We've detached some tasks from busiest_rq. Every
9033 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9034 * unlock busiest->lock, and we are able to be sure
9035 * that nobody can manipulate the tasks in parallel.
9036 * See task_rq_lock() family for the details.
9039 raw_spin_unlock(&busiest->lock);
9043 ld_moved += cur_ld_moved;
9046 local_irq_restore(flags);
9048 if (env.flags & LBF_NEED_BREAK) {
9049 env.flags &= ~LBF_NEED_BREAK;
9054 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9055 * us and move them to an alternate dst_cpu in our sched_group
9056 * where they can run. The upper limit on how many times we
9057 * iterate on same src_cpu is dependent on number of cpus in our
9060 * This changes load balance semantics a bit on who can move
9061 * load to a given_cpu. In addition to the given_cpu itself
9062 * (or a ilb_cpu acting on its behalf where given_cpu is
9063 * nohz-idle), we now have balance_cpu in a position to move
9064 * load to given_cpu. In rare situations, this may cause
9065 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9066 * _independently_ and at _same_ time to move some load to
9067 * given_cpu) causing exceess load to be moved to given_cpu.
9068 * This however should not happen so much in practice and
9069 * moreover subsequent load balance cycles should correct the
9070 * excess load moved.
9072 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9074 /* Prevent to re-select dst_cpu via env's cpus */
9075 cpumask_clear_cpu(env.dst_cpu, env.cpus);
9077 env.dst_rq = cpu_rq(env.new_dst_cpu);
9078 env.dst_cpu = env.new_dst_cpu;
9079 env.flags &= ~LBF_DST_PINNED;
9081 env.loop_break = sched_nr_migrate_break;
9084 * Go back to "more_balance" rather than "redo" since we
9085 * need to continue with same src_cpu.
9091 * We failed to reach balance because of affinity.
9094 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9096 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9097 *group_imbalance = 1;
9100 /* All tasks on this runqueue were pinned by CPU affinity */
9101 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9102 cpumask_clear_cpu(cpu_of(busiest), cpus);
9103 if (!cpumask_empty(cpus)) {
9105 env.loop_break = sched_nr_migrate_break;
9108 goto out_all_pinned;
9113 schedstat_inc(sd, lb_failed[idle]);
9115 * Increment the failure counter only on periodic balance.
9116 * We do not want newidle balance, which can be very
9117 * frequent, pollute the failure counter causing
9118 * excessive cache_hot migrations and active balances.
9120 if (idle != CPU_NEWLY_IDLE)
9121 if (env.src_grp_nr_running > 1)
9122 sd->nr_balance_failed++;
9124 if (need_active_balance(&env)) {
9125 raw_spin_lock_irqsave(&busiest->lock, flags);
9127 /* don't kick the active_load_balance_cpu_stop,
9128 * if the curr task on busiest cpu can't be
9131 if (!cpumask_test_cpu(this_cpu,
9132 tsk_cpus_allowed(busiest->curr))) {
9133 raw_spin_unlock_irqrestore(&busiest->lock,
9135 env.flags |= LBF_ALL_PINNED;
9136 goto out_one_pinned;
9140 * ->active_balance synchronizes accesses to
9141 * ->active_balance_work. Once set, it's cleared
9142 * only after active load balance is finished.
9144 if (!busiest->active_balance) {
9145 busiest->active_balance = 1;
9146 busiest->push_cpu = this_cpu;
9149 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9151 if (active_balance) {
9152 stop_one_cpu_nowait(cpu_of(busiest),
9153 active_load_balance_cpu_stop, busiest,
9154 &busiest->active_balance_work);
9158 * We've kicked active balancing, reset the failure
9161 sd->nr_balance_failed = sd->cache_nice_tries+1;
9164 sd->nr_balance_failed = 0;
9166 if (likely(!active_balance)) {
9167 /* We were unbalanced, so reset the balancing interval */
9168 sd->balance_interval = sd->min_interval;
9171 * If we've begun active balancing, start to back off. This
9172 * case may not be covered by the all_pinned logic if there
9173 * is only 1 task on the busy runqueue (because we don't call
9176 if (sd->balance_interval < sd->max_interval)
9177 sd->balance_interval *= 2;
9184 * We reach balance although we may have faced some affinity
9185 * constraints. Clear the imbalance flag if it was set.
9188 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9190 if (*group_imbalance)
9191 *group_imbalance = 0;
9196 * We reach balance because all tasks are pinned at this level so
9197 * we can't migrate them. Let the imbalance flag set so parent level
9198 * can try to migrate them.
9200 schedstat_inc(sd, lb_balanced[idle]);
9202 sd->nr_balance_failed = 0;
9205 /* tune up the balancing interval */
9206 if (((env.flags & LBF_ALL_PINNED) &&
9207 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9208 (sd->balance_interval < sd->max_interval))
9209 sd->balance_interval *= 2;
9216 static inline unsigned long
9217 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9219 unsigned long interval = sd->balance_interval;
9222 interval *= sd->busy_factor;
9224 /* scale ms to jiffies */
9225 interval = msecs_to_jiffies(interval);
9226 interval = clamp(interval, 1UL, max_load_balance_interval);
9232 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
9234 unsigned long interval, next;
9236 interval = get_sd_balance_interval(sd, cpu_busy);
9237 next = sd->last_balance + interval;
9239 if (time_after(*next_balance, next))
9240 *next_balance = next;
9244 * idle_balance is called by schedule() if this_cpu is about to become
9245 * idle. Attempts to pull tasks from other CPUs.
9247 static int idle_balance(struct rq *this_rq)
9249 unsigned long next_balance = jiffies + HZ;
9250 int this_cpu = this_rq->cpu;
9251 struct sched_domain *sd;
9252 int pulled_task = 0;
9255 idle_enter_fair(this_rq);
9258 * We must set idle_stamp _before_ calling idle_balance(), such that we
9259 * measure the duration of idle_balance() as idle time.
9261 this_rq->idle_stamp = rq_clock(this_rq);
9263 if (!energy_aware() &&
9264 (this_rq->avg_idle < sysctl_sched_migration_cost ||
9265 !this_rq->rd->overload)) {
9267 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9269 update_next_balance(sd, 0, &next_balance);
9275 raw_spin_unlock(&this_rq->lock);
9277 update_blocked_averages(this_cpu);
9279 for_each_domain(this_cpu, sd) {
9280 int continue_balancing = 1;
9281 u64 t0, domain_cost;
9283 if (!(sd->flags & SD_LOAD_BALANCE))
9286 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9287 update_next_balance(sd, 0, &next_balance);
9291 if (sd->flags & SD_BALANCE_NEWIDLE) {
9292 t0 = sched_clock_cpu(this_cpu);
9294 pulled_task = load_balance(this_cpu, this_rq,
9296 &continue_balancing);
9298 domain_cost = sched_clock_cpu(this_cpu) - t0;
9299 if (domain_cost > sd->max_newidle_lb_cost)
9300 sd->max_newidle_lb_cost = domain_cost;
9302 curr_cost += domain_cost;
9305 update_next_balance(sd, 0, &next_balance);
9308 * Stop searching for tasks to pull if there are
9309 * now runnable tasks on this rq.
9311 if (pulled_task || this_rq->nr_running > 0)
9316 raw_spin_lock(&this_rq->lock);
9318 if (curr_cost > this_rq->max_idle_balance_cost)
9319 this_rq->max_idle_balance_cost = curr_cost;
9322 * While browsing the domains, we released the rq lock, a task could
9323 * have been enqueued in the meantime. Since we're not going idle,
9324 * pretend we pulled a task.
9326 if (this_rq->cfs.h_nr_running && !pulled_task)
9330 /* Move the next balance forward */
9331 if (time_after(this_rq->next_balance, next_balance))
9332 this_rq->next_balance = next_balance;
9334 /* Is there a task of a high priority class? */
9335 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9339 idle_exit_fair(this_rq);
9340 this_rq->idle_stamp = 0;
9347 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
9348 * running tasks off the busiest CPU onto idle CPUs. It requires at
9349 * least 1 task to be running on each physical CPU where possible, and
9350 * avoids physical / logical imbalances.
9352 static int active_load_balance_cpu_stop(void *data)
9354 struct rq *busiest_rq = data;
9355 int busiest_cpu = cpu_of(busiest_rq);
9356 int target_cpu = busiest_rq->push_cpu;
9357 struct rq *target_rq = cpu_rq(target_cpu);
9358 struct sched_domain *sd = NULL;
9359 struct task_struct *p = NULL;
9360 struct task_struct *push_task;
9361 int push_task_detached = 0;
9362 struct lb_env env = {
9364 .dst_cpu = target_cpu,
9365 .dst_rq = target_rq,
9366 .src_cpu = busiest_rq->cpu,
9367 .src_rq = busiest_rq,
9371 raw_spin_lock_irq(&busiest_rq->lock);
9373 /* make sure the requested cpu hasn't gone down in the meantime */
9374 if (unlikely(busiest_cpu != smp_processor_id() ||
9375 !busiest_rq->active_balance))
9378 /* Is there any task to move? */
9379 if (busiest_rq->nr_running <= 1)
9383 * This condition is "impossible", if it occurs
9384 * we need to fix it. Originally reported by
9385 * Bjorn Helgaas on a 128-cpu setup.
9387 BUG_ON(busiest_rq == target_rq);
9389 push_task = busiest_rq->push_task;
9391 if (task_on_rq_queued(push_task) &&
9392 task_cpu(push_task) == busiest_cpu &&
9393 cpu_online(target_cpu)) {
9394 detach_task(push_task, &env);
9395 push_task_detached = 1;
9400 /* Search for an sd spanning us and the target CPU. */
9402 for_each_domain(target_cpu, sd) {
9403 if ((sd->flags & SD_LOAD_BALANCE) &&
9404 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9410 schedstat_inc(sd, alb_count);
9411 update_rq_clock(busiest_rq);
9413 p = detach_one_task(&env);
9415 schedstat_inc(sd, alb_pushed);
9417 schedstat_inc(sd, alb_failed);
9421 busiest_rq->active_balance = 0;
9424 busiest_rq->push_task = NULL;
9426 raw_spin_unlock(&busiest_rq->lock);
9429 if (push_task_detached)
9430 attach_one_task(target_rq, push_task);
9431 put_task_struct(push_task);
9435 attach_one_task(target_rq, p);
9442 static inline int on_null_domain(struct rq *rq)
9444 return unlikely(!rcu_dereference_sched(rq->sd));
9447 #ifdef CONFIG_NO_HZ_COMMON
9449 * idle load balancing details
9450 * - When one of the busy CPUs notice that there may be an idle rebalancing
9451 * needed, they will kick the idle load balancer, which then does idle
9452 * load balancing for all the idle CPUs.
9454 static inline int find_new_ilb(void)
9456 int ilb = cpumask_first(nohz.idle_cpus_mask);
9458 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9465 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9466 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9467 * CPU (if there is one).
9469 static void nohz_balancer_kick(void)
9473 nohz.next_balance++;
9475 ilb_cpu = find_new_ilb();
9477 if (ilb_cpu >= nr_cpu_ids)
9480 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9483 * Use smp_send_reschedule() instead of resched_cpu().
9484 * This way we generate a sched IPI on the target cpu which
9485 * is idle. And the softirq performing nohz idle load balance
9486 * will be run before returning from the IPI.
9488 smp_send_reschedule(ilb_cpu);
9492 static inline void nohz_balance_exit_idle(int cpu)
9494 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9496 * Completely isolated CPUs don't ever set, so we must test.
9498 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9499 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9500 atomic_dec(&nohz.nr_cpus);
9502 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9506 static inline void set_cpu_sd_state_busy(void)
9508 struct sched_domain *sd;
9509 int cpu = smp_processor_id();
9512 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9514 if (!sd || !sd->nohz_idle)
9518 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
9523 void set_cpu_sd_state_idle(void)
9525 struct sched_domain *sd;
9526 int cpu = smp_processor_id();
9529 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9531 if (!sd || sd->nohz_idle)
9535 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
9541 * This routine will record that the cpu is going idle with tick stopped.
9542 * This info will be used in performing idle load balancing in the future.
9544 void nohz_balance_enter_idle(int cpu)
9547 * If this cpu is going down, then nothing needs to be done.
9549 if (!cpu_active(cpu))
9552 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9556 * If we're a completely isolated CPU, we don't play.
9558 if (on_null_domain(cpu_rq(cpu)))
9561 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9562 atomic_inc(&nohz.nr_cpus);
9563 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9566 static int sched_ilb_notifier(struct notifier_block *nfb,
9567 unsigned long action, void *hcpu)
9569 switch (action & ~CPU_TASKS_FROZEN) {
9571 nohz_balance_exit_idle(smp_processor_id());
9579 static DEFINE_SPINLOCK(balancing);
9582 * Scale the max load_balance interval with the number of CPUs in the system.
9583 * This trades load-balance latency on larger machines for less cross talk.
9585 void update_max_interval(void)
9587 max_load_balance_interval = HZ*num_online_cpus()/10;
9591 * It checks each scheduling domain to see if it is due to be balanced,
9592 * and initiates a balancing operation if so.
9594 * Balancing parameters are set up in init_sched_domains.
9596 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9598 int continue_balancing = 1;
9600 unsigned long interval;
9601 struct sched_domain *sd;
9602 /* Earliest time when we have to do rebalance again */
9603 unsigned long next_balance = jiffies + 60*HZ;
9604 int update_next_balance = 0;
9605 int need_serialize, need_decay = 0;
9608 update_blocked_averages(cpu);
9611 for_each_domain(cpu, sd) {
9613 * Decay the newidle max times here because this is a regular
9614 * visit to all the domains. Decay ~1% per second.
9616 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9617 sd->max_newidle_lb_cost =
9618 (sd->max_newidle_lb_cost * 253) / 256;
9619 sd->next_decay_max_lb_cost = jiffies + HZ;
9622 max_cost += sd->max_newidle_lb_cost;
9624 if (!(sd->flags & SD_LOAD_BALANCE))
9628 * Stop the load balance at this level. There is another
9629 * CPU in our sched group which is doing load balancing more
9632 if (!continue_balancing) {
9638 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9640 need_serialize = sd->flags & SD_SERIALIZE;
9641 if (need_serialize) {
9642 if (!spin_trylock(&balancing))
9646 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9647 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9649 * The LBF_DST_PINNED logic could have changed
9650 * env->dst_cpu, so we can't know our idle
9651 * state even if we migrated tasks. Update it.
9653 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9655 sd->last_balance = jiffies;
9656 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9659 spin_unlock(&balancing);
9661 if (time_after(next_balance, sd->last_balance + interval)) {
9662 next_balance = sd->last_balance + interval;
9663 update_next_balance = 1;
9668 * Ensure the rq-wide value also decays but keep it at a
9669 * reasonable floor to avoid funnies with rq->avg_idle.
9671 rq->max_idle_balance_cost =
9672 max((u64)sysctl_sched_migration_cost, max_cost);
9677 * next_balance will be updated only when there is a need.
9678 * When the cpu is attached to null domain for ex, it will not be
9681 if (likely(update_next_balance)) {
9682 rq->next_balance = next_balance;
9684 #ifdef CONFIG_NO_HZ_COMMON
9686 * If this CPU has been elected to perform the nohz idle
9687 * balance. Other idle CPUs have already rebalanced with
9688 * nohz_idle_balance() and nohz.next_balance has been
9689 * updated accordingly. This CPU is now running the idle load
9690 * balance for itself and we need to update the
9691 * nohz.next_balance accordingly.
9693 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9694 nohz.next_balance = rq->next_balance;
9699 #ifdef CONFIG_NO_HZ_COMMON
9701 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9702 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9704 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9706 int this_cpu = this_rq->cpu;
9709 /* Earliest time when we have to do rebalance again */
9710 unsigned long next_balance = jiffies + 60*HZ;
9711 int update_next_balance = 0;
9713 if (idle != CPU_IDLE ||
9714 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9717 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9718 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9722 * If this cpu gets work to do, stop the load balancing
9723 * work being done for other cpus. Next load
9724 * balancing owner will pick it up.
9729 rq = cpu_rq(balance_cpu);
9732 * If time for next balance is due,
9735 if (time_after_eq(jiffies, rq->next_balance)) {
9736 raw_spin_lock_irq(&rq->lock);
9737 update_rq_clock(rq);
9738 update_idle_cpu_load(rq);
9739 raw_spin_unlock_irq(&rq->lock);
9740 rebalance_domains(rq, CPU_IDLE);
9743 if (time_after(next_balance, rq->next_balance)) {
9744 next_balance = rq->next_balance;
9745 update_next_balance = 1;
9750 * next_balance will be updated only when there is a need.
9751 * When the CPU is attached to null domain for ex, it will not be
9754 if (likely(update_next_balance))
9755 nohz.next_balance = next_balance;
9757 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9761 * Current heuristic for kicking the idle load balancer in the presence
9762 * of an idle cpu in the system.
9763 * - This rq has more than one task.
9764 * - This rq has at least one CFS task and the capacity of the CPU is
9765 * significantly reduced because of RT tasks or IRQs.
9766 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9767 * multiple busy cpu.
9768 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9769 * domain span are idle.
9771 static inline bool nohz_kick_needed(struct rq *rq)
9773 unsigned long now = jiffies;
9774 struct sched_domain *sd;
9775 struct sched_group_capacity *sgc;
9776 int nr_busy, cpu = rq->cpu;
9779 if (unlikely(rq->idle_balance))
9783 * We may be recently in ticked or tickless idle mode. At the first
9784 * busy tick after returning from idle, we will update the busy stats.
9786 set_cpu_sd_state_busy();
9787 nohz_balance_exit_idle(cpu);
9790 * None are in tickless mode and hence no need for NOHZ idle load
9793 if (likely(!atomic_read(&nohz.nr_cpus)))
9796 if (time_before(now, nohz.next_balance))
9799 if (rq->nr_running >= 2 &&
9800 (!energy_aware() || cpu_overutilized(cpu)))
9803 /* Do idle load balance if there have misfit task */
9805 return rq->misfit_task;
9808 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9810 sgc = sd->groups->sgc;
9811 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9820 sd = rcu_dereference(rq->sd);
9822 if ((rq->cfs.h_nr_running >= 1) &&
9823 check_cpu_capacity(rq, sd)) {
9829 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9830 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9831 sched_domain_span(sd)) < cpu)) {
9841 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9845 * run_rebalance_domains is triggered when needed from the scheduler tick.
9846 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9848 static void run_rebalance_domains(struct softirq_action *h)
9850 struct rq *this_rq = this_rq();
9851 enum cpu_idle_type idle = this_rq->idle_balance ?
9852 CPU_IDLE : CPU_NOT_IDLE;
9855 * If this cpu has a pending nohz_balance_kick, then do the
9856 * balancing on behalf of the other idle cpus whose ticks are
9857 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9858 * give the idle cpus a chance to load balance. Else we may
9859 * load balance only within the local sched_domain hierarchy
9860 * and abort nohz_idle_balance altogether if we pull some load.
9862 nohz_idle_balance(this_rq, idle);
9863 rebalance_domains(this_rq, idle);
9867 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9869 void trigger_load_balance(struct rq *rq)
9871 /* Don't need to rebalance while attached to NULL domain */
9872 if (unlikely(on_null_domain(rq)))
9875 if (time_after_eq(jiffies, rq->next_balance))
9876 raise_softirq(SCHED_SOFTIRQ);
9877 #ifdef CONFIG_NO_HZ_COMMON
9878 if (nohz_kick_needed(rq))
9879 nohz_balancer_kick();
9883 static void rq_online_fair(struct rq *rq)
9887 update_runtime_enabled(rq);
9890 static void rq_offline_fair(struct rq *rq)
9894 /* Ensure any throttled groups are reachable by pick_next_task */
9895 unthrottle_offline_cfs_rqs(rq);
9899 kick_active_balance(struct rq *rq, struct task_struct *p, int new_cpu)
9903 /* Invoke active balance to force migrate currently running task */
9904 raw_spin_lock(&rq->lock);
9905 if (!rq->active_balance) {
9906 rq->active_balance = 1;
9907 rq->push_cpu = new_cpu;
9912 raw_spin_unlock(&rq->lock);
9917 void check_for_migration(struct rq *rq, struct task_struct *p)
9921 int cpu = task_cpu(p);
9923 if (rq->misfit_task) {
9924 if (rq->curr->state != TASK_RUNNING ||
9925 rq->curr->nr_cpus_allowed == 1)
9928 new_cpu = select_energy_cpu_brute(p, cpu, 0);
9929 if (capacity_orig_of(new_cpu) > capacity_orig_of(cpu)) {
9930 active_balance = kick_active_balance(rq, p, new_cpu);
9932 stop_one_cpu_nowait(cpu,
9933 active_load_balance_cpu_stop,
9934 rq, &rq->active_balance_work);
9939 #endif /* CONFIG_SMP */
9942 * scheduler tick hitting a task of our scheduling class:
9944 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9946 struct cfs_rq *cfs_rq;
9947 struct sched_entity *se = &curr->se;
9949 for_each_sched_entity(se) {
9950 cfs_rq = cfs_rq_of(se);
9951 entity_tick(cfs_rq, se, queued);
9954 if (static_branch_unlikely(&sched_numa_balancing))
9955 task_tick_numa(rq, curr);
9958 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9959 rq->rd->overutilized = true;
9960 trace_sched_overutilized(true);
9963 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9969 * called on fork with the child task as argument from the parent's context
9970 * - child not yet on the tasklist
9971 * - preemption disabled
9973 static void task_fork_fair(struct task_struct *p)
9975 struct cfs_rq *cfs_rq;
9976 struct sched_entity *se = &p->se, *curr;
9977 struct rq *rq = this_rq();
9979 raw_spin_lock(&rq->lock);
9980 update_rq_clock(rq);
9982 cfs_rq = task_cfs_rq(current);
9983 curr = cfs_rq->curr;
9985 update_curr(cfs_rq);
9986 se->vruntime = curr->vruntime;
9988 place_entity(cfs_rq, se, 1);
9990 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9992 * Upon rescheduling, sched_class::put_prev_task() will place
9993 * 'current' within the tree based on its new key value.
9995 swap(curr->vruntime, se->vruntime);
9999 se->vruntime -= cfs_rq->min_vruntime;
10000 raw_spin_unlock(&rq->lock);
10004 * Priority of the task has changed. Check to see if we preempt
10005 * the current task.
10008 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10010 if (!task_on_rq_queued(p))
10014 * Reschedule if we are currently running on this runqueue and
10015 * our priority decreased, or if we are not currently running on
10016 * this runqueue and our priority is higher than the current's
10018 if (rq->curr == p) {
10019 if (p->prio > oldprio)
10022 check_preempt_curr(rq, p, 0);
10025 static inline bool vruntime_normalized(struct task_struct *p)
10027 struct sched_entity *se = &p->se;
10030 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10031 * the dequeue_entity(.flags=0) will already have normalized the
10038 * When !on_rq, vruntime of the task has usually NOT been normalized.
10039 * But there are some cases where it has already been normalized:
10041 * - A forked child which is waiting for being woken up by
10042 * wake_up_new_task().
10043 * - A task which has been woken up by try_to_wake_up() and
10044 * waiting for actually being woken up by sched_ttwu_pending().
10046 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
10052 #ifdef CONFIG_FAIR_GROUP_SCHED
10054 * Propagate the changes of the sched_entity across the tg tree to make it
10055 * visible to the root
10057 static void propagate_entity_cfs_rq(struct sched_entity *se)
10059 struct cfs_rq *cfs_rq;
10061 /* Start to propagate at parent */
10064 for_each_sched_entity(se) {
10065 cfs_rq = cfs_rq_of(se);
10067 if (cfs_rq_throttled(cfs_rq))
10070 update_load_avg(se, UPDATE_TG);
10074 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10077 static void detach_entity_cfs_rq(struct sched_entity *se)
10079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10081 /* Catch up with the cfs_rq and remove our load when we leave */
10082 update_load_avg(se, 0);
10083 detach_entity_load_avg(cfs_rq, se);
10084 update_tg_load_avg(cfs_rq, false);
10085 propagate_entity_cfs_rq(se);
10088 static void attach_entity_cfs_rq(struct sched_entity *se)
10090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10092 #ifdef CONFIG_FAIR_GROUP_SCHED
10094 * Since the real-depth could have been changed (only FAIR
10095 * class maintain depth value), reset depth properly.
10097 se->depth = se->parent ? se->parent->depth + 1 : 0;
10100 /* Synchronize entity with its cfs_rq */
10101 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10102 attach_entity_load_avg(cfs_rq, se);
10103 update_tg_load_avg(cfs_rq, false);
10104 propagate_entity_cfs_rq(se);
10107 static void detach_task_cfs_rq(struct task_struct *p)
10109 struct sched_entity *se = &p->se;
10110 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10112 if (!vruntime_normalized(p)) {
10114 * Fix up our vruntime so that the current sleep doesn't
10115 * cause 'unlimited' sleep bonus.
10117 place_entity(cfs_rq, se, 0);
10118 se->vruntime -= cfs_rq->min_vruntime;
10121 detach_entity_cfs_rq(se);
10124 static void attach_task_cfs_rq(struct task_struct *p)
10126 struct sched_entity *se = &p->se;
10127 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10129 attach_entity_cfs_rq(se);
10131 if (!vruntime_normalized(p))
10132 se->vruntime += cfs_rq->min_vruntime;
10135 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10137 detach_task_cfs_rq(p);
10140 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10142 attach_task_cfs_rq(p);
10144 if (task_on_rq_queued(p)) {
10146 * We were most likely switched from sched_rt, so
10147 * kick off the schedule if running, otherwise just see
10148 * if we can still preempt the current task.
10153 check_preempt_curr(rq, p, 0);
10157 /* Account for a task changing its policy or group.
10159 * This routine is mostly called to set cfs_rq->curr field when a task
10160 * migrates between groups/classes.
10162 static void set_curr_task_fair(struct rq *rq)
10164 struct sched_entity *se = &rq->curr->se;
10166 for_each_sched_entity(se) {
10167 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10169 set_next_entity(cfs_rq, se);
10170 /* ensure bandwidth has been allocated on our new cfs_rq */
10171 account_cfs_rq_runtime(cfs_rq, 0);
10175 void init_cfs_rq(struct cfs_rq *cfs_rq)
10177 cfs_rq->tasks_timeline = RB_ROOT;
10178 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10179 #ifndef CONFIG_64BIT
10180 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10183 #ifdef CONFIG_FAIR_GROUP_SCHED
10184 cfs_rq->propagate_avg = 0;
10186 atomic_long_set(&cfs_rq->removed_load_avg, 0);
10187 atomic_long_set(&cfs_rq->removed_util_avg, 0);
10191 #ifdef CONFIG_FAIR_GROUP_SCHED
10192 static void task_set_group_fair(struct task_struct *p)
10194 struct sched_entity *se = &p->se;
10196 set_task_rq(p, task_cpu(p));
10197 se->depth = se->parent ? se->parent->depth + 1 : 0;
10200 static void task_move_group_fair(struct task_struct *p)
10202 detach_task_cfs_rq(p);
10203 set_task_rq(p, task_cpu(p));
10206 /* Tell se's cfs_rq has been changed -- migrated */
10207 p->se.avg.last_update_time = 0;
10209 attach_task_cfs_rq(p);
10212 static void task_change_group_fair(struct task_struct *p, int type)
10215 case TASK_SET_GROUP:
10216 task_set_group_fair(p);
10219 case TASK_MOVE_GROUP:
10220 task_move_group_fair(p);
10225 void free_fair_sched_group(struct task_group *tg)
10229 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10231 for_each_possible_cpu(i) {
10233 kfree(tg->cfs_rq[i]);
10236 remove_entity_load_avg(tg->se[i]);
10245 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10247 struct sched_entity *se;
10248 struct cfs_rq *cfs_rq;
10252 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
10255 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
10259 tg->shares = NICE_0_LOAD;
10261 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10263 for_each_possible_cpu(i) {
10266 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10267 GFP_KERNEL, cpu_to_node(i));
10271 se = kzalloc_node(sizeof(struct sched_entity),
10272 GFP_KERNEL, cpu_to_node(i));
10276 init_cfs_rq(cfs_rq);
10277 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10278 init_entity_runnable_average(se);
10280 raw_spin_lock_irq(&rq->lock);
10281 post_init_entity_util_avg(se);
10282 raw_spin_unlock_irq(&rq->lock);
10293 void unregister_fair_sched_group(struct task_group *tg, int cpu)
10295 struct rq *rq = cpu_rq(cpu);
10296 unsigned long flags;
10299 * Only empty task groups can be destroyed; so we can speculatively
10300 * check on_list without danger of it being re-added.
10302 if (!tg->cfs_rq[cpu]->on_list)
10305 raw_spin_lock_irqsave(&rq->lock, flags);
10306 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10307 raw_spin_unlock_irqrestore(&rq->lock, flags);
10310 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10311 struct sched_entity *se, int cpu,
10312 struct sched_entity *parent)
10314 struct rq *rq = cpu_rq(cpu);
10318 init_cfs_rq_runtime(cfs_rq);
10320 tg->cfs_rq[cpu] = cfs_rq;
10323 /* se could be NULL for root_task_group */
10328 se->cfs_rq = &rq->cfs;
10331 se->cfs_rq = parent->my_q;
10332 se->depth = parent->depth + 1;
10336 /* guarantee group entities always have weight */
10337 update_load_set(&se->load, NICE_0_LOAD);
10338 se->parent = parent;
10341 static DEFINE_MUTEX(shares_mutex);
10343 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10346 unsigned long flags;
10349 * We can't change the weight of the root cgroup.
10354 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10356 mutex_lock(&shares_mutex);
10357 if (tg->shares == shares)
10360 tg->shares = shares;
10361 for_each_possible_cpu(i) {
10362 struct rq *rq = cpu_rq(i);
10363 struct sched_entity *se;
10366 /* Propagate contribution to hierarchy */
10367 raw_spin_lock_irqsave(&rq->lock, flags);
10369 /* Possible calls to update_curr() need rq clock */
10370 update_rq_clock(rq);
10371 for_each_sched_entity(se) {
10372 update_load_avg(se, UPDATE_TG);
10373 update_cfs_shares(se);
10375 raw_spin_unlock_irqrestore(&rq->lock, flags);
10379 mutex_unlock(&shares_mutex);
10382 #else /* CONFIG_FAIR_GROUP_SCHED */
10384 void free_fair_sched_group(struct task_group *tg) { }
10386 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10391 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
10393 #endif /* CONFIG_FAIR_GROUP_SCHED */
10396 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10398 struct sched_entity *se = &task->se;
10399 unsigned int rr_interval = 0;
10402 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10405 if (rq->cfs.load.weight)
10406 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10408 return rr_interval;
10412 * All the scheduling class methods:
10414 const struct sched_class fair_sched_class = {
10415 .next = &idle_sched_class,
10416 .enqueue_task = enqueue_task_fair,
10417 .dequeue_task = dequeue_task_fair,
10418 .yield_task = yield_task_fair,
10419 .yield_to_task = yield_to_task_fair,
10421 .check_preempt_curr = check_preempt_wakeup,
10423 .pick_next_task = pick_next_task_fair,
10424 .put_prev_task = put_prev_task_fair,
10427 .select_task_rq = select_task_rq_fair,
10428 .migrate_task_rq = migrate_task_rq_fair,
10430 .rq_online = rq_online_fair,
10431 .rq_offline = rq_offline_fair,
10433 .task_waking = task_waking_fair,
10434 .task_dead = task_dead_fair,
10435 .set_cpus_allowed = set_cpus_allowed_common,
10438 .set_curr_task = set_curr_task_fair,
10439 .task_tick = task_tick_fair,
10440 .task_fork = task_fork_fair,
10442 .prio_changed = prio_changed_fair,
10443 .switched_from = switched_from_fair,
10444 .switched_to = switched_to_fair,
10446 .get_rr_interval = get_rr_interval_fair,
10448 .update_curr = update_curr_fair,
10450 #ifdef CONFIG_FAIR_GROUP_SCHED
10451 .task_change_group = task_change_group_fair,
10455 #ifdef CONFIG_SCHED_DEBUG
10456 void print_cfs_stats(struct seq_file *m, int cpu)
10458 struct cfs_rq *cfs_rq;
10461 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10462 print_cfs_rq(m, cpu, cfs_rq);
10466 #ifdef CONFIG_NUMA_BALANCING
10467 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10470 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10472 for_each_online_node(node) {
10473 if (p->numa_faults) {
10474 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10475 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10477 if (p->numa_group) {
10478 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10479 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10481 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10484 #endif /* CONFIG_NUMA_BALANCING */
10485 #endif /* CONFIG_SCHED_DEBUG */
10487 __init void init_sched_fair_class(void)
10490 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10492 #ifdef CONFIG_NO_HZ_COMMON
10493 nohz.next_balance = jiffies;
10494 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10495 cpu_notifier(sched_ilb_notifier, 0);