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 /* Avoid time going backwards, prevent potential divide error: */
1883 if (unlikely((s64)*period < 0))
1886 delta = p->se.avg.load_sum / p->se.load.weight;
1887 *period = LOAD_AVG_MAX;
1890 p->last_sum_exec_runtime = runtime;
1891 p->last_task_numa_placement = now;
1897 * Determine the preferred nid for a task in a numa_group. This needs to
1898 * be done in a way that produces consistent results with group_weight,
1899 * otherwise workloads might not converge.
1901 static int preferred_group_nid(struct task_struct *p, int nid)
1906 /* Direct connections between all NUMA nodes. */
1907 if (sched_numa_topology_type == NUMA_DIRECT)
1911 * On a system with glueless mesh NUMA topology, group_weight
1912 * scores nodes according to the number of NUMA hinting faults on
1913 * both the node itself, and on nearby nodes.
1915 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1916 unsigned long score, max_score = 0;
1917 int node, max_node = nid;
1919 dist = sched_max_numa_distance;
1921 for_each_online_node(node) {
1922 score = group_weight(p, node, dist);
1923 if (score > max_score) {
1932 * Finding the preferred nid in a system with NUMA backplane
1933 * interconnect topology is more involved. The goal is to locate
1934 * tasks from numa_groups near each other in the system, and
1935 * untangle workloads from different sides of the system. This requires
1936 * searching down the hierarchy of node groups, recursively searching
1937 * inside the highest scoring group of nodes. The nodemask tricks
1938 * keep the complexity of the search down.
1940 nodes = node_online_map;
1941 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1942 unsigned long max_faults = 0;
1943 nodemask_t max_group = NODE_MASK_NONE;
1946 /* Are there nodes at this distance from each other? */
1947 if (!find_numa_distance(dist))
1950 for_each_node_mask(a, nodes) {
1951 unsigned long faults = 0;
1952 nodemask_t this_group;
1953 nodes_clear(this_group);
1955 /* Sum group's NUMA faults; includes a==b case. */
1956 for_each_node_mask(b, nodes) {
1957 if (node_distance(a, b) < dist) {
1958 faults += group_faults(p, b);
1959 node_set(b, this_group);
1960 node_clear(b, nodes);
1964 /* Remember the top group. */
1965 if (faults > max_faults) {
1966 max_faults = faults;
1967 max_group = this_group;
1969 * subtle: at the smallest distance there is
1970 * just one node left in each "group", the
1971 * winner is the preferred nid.
1976 /* Next round, evaluate the nodes within max_group. */
1984 static void task_numa_placement(struct task_struct *p)
1986 int seq, nid, max_nid = -1, max_group_nid = -1;
1987 unsigned long max_faults = 0, max_group_faults = 0;
1988 unsigned long fault_types[2] = { 0, 0 };
1989 unsigned long total_faults;
1990 u64 runtime, period;
1991 spinlock_t *group_lock = NULL;
1994 * The p->mm->numa_scan_seq field gets updated without
1995 * exclusive access. Use READ_ONCE() here to ensure
1996 * that the field is read in a single access:
1998 seq = READ_ONCE(p->mm->numa_scan_seq);
1999 if (p->numa_scan_seq == seq)
2001 p->numa_scan_seq = seq;
2002 p->numa_scan_period_max = task_scan_max(p);
2004 total_faults = p->numa_faults_locality[0] +
2005 p->numa_faults_locality[1];
2006 runtime = numa_get_avg_runtime(p, &period);
2008 /* If the task is part of a group prevent parallel updates to group stats */
2009 if (p->numa_group) {
2010 group_lock = &p->numa_group->lock;
2011 spin_lock_irq(group_lock);
2014 /* Find the node with the highest number of faults */
2015 for_each_online_node(nid) {
2016 /* Keep track of the offsets in numa_faults array */
2017 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2018 unsigned long faults = 0, group_faults = 0;
2021 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2022 long diff, f_diff, f_weight;
2024 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2025 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2026 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2027 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2029 /* Decay existing window, copy faults since last scan */
2030 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2031 fault_types[priv] += p->numa_faults[membuf_idx];
2032 p->numa_faults[membuf_idx] = 0;
2035 * Normalize the faults_from, so all tasks in a group
2036 * count according to CPU use, instead of by the raw
2037 * number of faults. Tasks with little runtime have
2038 * little over-all impact on throughput, and thus their
2039 * faults are less important.
2041 f_weight = div64_u64(runtime << 16, period + 1);
2042 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2044 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2045 p->numa_faults[cpubuf_idx] = 0;
2047 p->numa_faults[mem_idx] += diff;
2048 p->numa_faults[cpu_idx] += f_diff;
2049 faults += p->numa_faults[mem_idx];
2050 p->total_numa_faults += diff;
2051 if (p->numa_group) {
2053 * safe because we can only change our own group
2055 * mem_idx represents the offset for a given
2056 * nid and priv in a specific region because it
2057 * is at the beginning of the numa_faults array.
2059 p->numa_group->faults[mem_idx] += diff;
2060 p->numa_group->faults_cpu[mem_idx] += f_diff;
2061 p->numa_group->total_faults += diff;
2062 group_faults += p->numa_group->faults[mem_idx];
2066 if (faults > max_faults) {
2067 max_faults = faults;
2071 if (group_faults > max_group_faults) {
2072 max_group_faults = group_faults;
2073 max_group_nid = nid;
2077 update_task_scan_period(p, fault_types[0], fault_types[1]);
2079 if (p->numa_group) {
2080 update_numa_active_node_mask(p->numa_group);
2081 spin_unlock_irq(group_lock);
2082 max_nid = preferred_group_nid(p, max_group_nid);
2086 /* Set the new preferred node */
2087 if (max_nid != p->numa_preferred_nid)
2088 sched_setnuma(p, max_nid);
2090 if (task_node(p) != p->numa_preferred_nid)
2091 numa_migrate_preferred(p);
2095 static inline int get_numa_group(struct numa_group *grp)
2097 return atomic_inc_not_zero(&grp->refcount);
2100 static inline void put_numa_group(struct numa_group *grp)
2102 if (atomic_dec_and_test(&grp->refcount))
2103 kfree_rcu(grp, rcu);
2106 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2109 struct numa_group *grp, *my_grp;
2110 struct task_struct *tsk;
2112 int cpu = cpupid_to_cpu(cpupid);
2115 if (unlikely(!p->numa_group)) {
2116 unsigned int size = sizeof(struct numa_group) +
2117 4*nr_node_ids*sizeof(unsigned long);
2119 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2123 atomic_set(&grp->refcount, 1);
2124 spin_lock_init(&grp->lock);
2126 /* Second half of the array tracks nids where faults happen */
2127 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2130 node_set(task_node(current), grp->active_nodes);
2132 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2133 grp->faults[i] = p->numa_faults[i];
2135 grp->total_faults = p->total_numa_faults;
2138 rcu_assign_pointer(p->numa_group, grp);
2142 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2144 if (!cpupid_match_pid(tsk, cpupid))
2147 grp = rcu_dereference(tsk->numa_group);
2151 my_grp = p->numa_group;
2156 * Only join the other group if its bigger; if we're the bigger group,
2157 * the other task will join us.
2159 if (my_grp->nr_tasks > grp->nr_tasks)
2163 * Tie-break on the grp address.
2165 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2168 /* Always join threads in the same process. */
2169 if (tsk->mm == current->mm)
2172 /* Simple filter to avoid false positives due to PID collisions */
2173 if (flags & TNF_SHARED)
2176 /* Update priv based on whether false sharing was detected */
2179 if (join && !get_numa_group(grp))
2187 BUG_ON(irqs_disabled());
2188 double_lock_irq(&my_grp->lock, &grp->lock);
2190 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2191 my_grp->faults[i] -= p->numa_faults[i];
2192 grp->faults[i] += p->numa_faults[i];
2194 my_grp->total_faults -= p->total_numa_faults;
2195 grp->total_faults += p->total_numa_faults;
2200 spin_unlock(&my_grp->lock);
2201 spin_unlock_irq(&grp->lock);
2203 rcu_assign_pointer(p->numa_group, grp);
2205 put_numa_group(my_grp);
2214 * Get rid of NUMA staticstics associated with a task (either current or dead).
2215 * If @final is set, the task is dead and has reached refcount zero, so we can
2216 * safely free all relevant data structures. Otherwise, there might be
2217 * concurrent reads from places like load balancing and procfs, and we should
2218 * reset the data back to default state without freeing ->numa_faults.
2220 void task_numa_free(struct task_struct *p, bool final)
2222 struct numa_group *grp = p->numa_group;
2223 unsigned long *numa_faults = p->numa_faults;
2224 unsigned long flags;
2231 spin_lock_irqsave(&grp->lock, flags);
2232 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2233 grp->faults[i] -= p->numa_faults[i];
2234 grp->total_faults -= p->total_numa_faults;
2237 spin_unlock_irqrestore(&grp->lock, flags);
2238 RCU_INIT_POINTER(p->numa_group, NULL);
2239 put_numa_group(grp);
2243 p->numa_faults = NULL;
2246 p->total_numa_faults = 0;
2247 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2253 * Got a PROT_NONE fault for a page on @node.
2255 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2257 struct task_struct *p = current;
2258 bool migrated = flags & TNF_MIGRATED;
2259 int cpu_node = task_node(current);
2260 int local = !!(flags & TNF_FAULT_LOCAL);
2263 if (!static_branch_likely(&sched_numa_balancing))
2266 /* for example, ksmd faulting in a user's mm */
2270 /* Allocate buffer to track faults on a per-node basis */
2271 if (unlikely(!p->numa_faults)) {
2272 int size = sizeof(*p->numa_faults) *
2273 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2275 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2276 if (!p->numa_faults)
2279 p->total_numa_faults = 0;
2280 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2284 * First accesses are treated as private, otherwise consider accesses
2285 * to be private if the accessing pid has not changed
2287 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2290 priv = cpupid_match_pid(p, last_cpupid);
2291 if (!priv && !(flags & TNF_NO_GROUP))
2292 task_numa_group(p, last_cpupid, flags, &priv);
2296 * If a workload spans multiple NUMA nodes, a shared fault that
2297 * occurs wholly within the set of nodes that the workload is
2298 * actively using should be counted as local. This allows the
2299 * scan rate to slow down when a workload has settled down.
2301 if (!priv && !local && p->numa_group &&
2302 node_isset(cpu_node, p->numa_group->active_nodes) &&
2303 node_isset(mem_node, p->numa_group->active_nodes))
2306 task_numa_placement(p);
2309 * Retry task to preferred node migration periodically, in case it
2310 * case it previously failed, or the scheduler moved us.
2312 if (time_after(jiffies, p->numa_migrate_retry))
2313 numa_migrate_preferred(p);
2316 p->numa_pages_migrated += pages;
2317 if (flags & TNF_MIGRATE_FAIL)
2318 p->numa_faults_locality[2] += pages;
2320 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2321 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2322 p->numa_faults_locality[local] += pages;
2325 static void reset_ptenuma_scan(struct task_struct *p)
2328 * We only did a read acquisition of the mmap sem, so
2329 * p->mm->numa_scan_seq is written to without exclusive access
2330 * and the update is not guaranteed to be atomic. That's not
2331 * much of an issue though, since this is just used for
2332 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2333 * expensive, to avoid any form of compiler optimizations:
2335 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2336 p->mm->numa_scan_offset = 0;
2340 * The expensive part of numa migration is done from task_work context.
2341 * Triggered from task_tick_numa().
2343 void task_numa_work(struct callback_head *work)
2345 unsigned long migrate, next_scan, now = jiffies;
2346 struct task_struct *p = current;
2347 struct mm_struct *mm = p->mm;
2348 struct vm_area_struct *vma;
2349 unsigned long start, end;
2350 unsigned long nr_pte_updates = 0;
2351 long pages, virtpages;
2353 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2355 work->next = work; /* protect against double add */
2357 * Who cares about NUMA placement when they're dying.
2359 * NOTE: make sure not to dereference p->mm before this check,
2360 * exit_task_work() happens _after_ exit_mm() so we could be called
2361 * without p->mm even though we still had it when we enqueued this
2364 if (p->flags & PF_EXITING)
2367 if (!mm->numa_next_scan) {
2368 mm->numa_next_scan = now +
2369 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2373 * Enforce maximal scan/migration frequency..
2375 migrate = mm->numa_next_scan;
2376 if (time_before(now, migrate))
2379 if (p->numa_scan_period == 0) {
2380 p->numa_scan_period_max = task_scan_max(p);
2381 p->numa_scan_period = task_scan_min(p);
2384 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2385 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2389 * Delay this task enough that another task of this mm will likely win
2390 * the next time around.
2392 p->node_stamp += 2 * TICK_NSEC;
2394 start = mm->numa_scan_offset;
2395 pages = sysctl_numa_balancing_scan_size;
2396 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2397 virtpages = pages * 8; /* Scan up to this much virtual space */
2402 if (!down_read_trylock(&mm->mmap_sem))
2404 vma = find_vma(mm, start);
2406 reset_ptenuma_scan(p);
2410 for (; vma; vma = vma->vm_next) {
2411 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2412 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2417 * Shared library pages mapped by multiple processes are not
2418 * migrated as it is expected they are cache replicated. Avoid
2419 * hinting faults in read-only file-backed mappings or the vdso
2420 * as migrating the pages will be of marginal benefit.
2423 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2427 * Skip inaccessible VMAs to avoid any confusion between
2428 * PROT_NONE and NUMA hinting ptes
2430 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2434 start = max(start, vma->vm_start);
2435 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2436 end = min(end, vma->vm_end);
2437 nr_pte_updates = change_prot_numa(vma, start, end);
2440 * Try to scan sysctl_numa_balancing_size worth of
2441 * hpages that have at least one present PTE that
2442 * is not already pte-numa. If the VMA contains
2443 * areas that are unused or already full of prot_numa
2444 * PTEs, scan up to virtpages, to skip through those
2448 pages -= (end - start) >> PAGE_SHIFT;
2449 virtpages -= (end - start) >> PAGE_SHIFT;
2452 if (pages <= 0 || virtpages <= 0)
2456 } while (end != vma->vm_end);
2461 * It is possible to reach the end of the VMA list but the last few
2462 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2463 * would find the !migratable VMA on the next scan but not reset the
2464 * scanner to the start so check it now.
2467 mm->numa_scan_offset = start;
2469 reset_ptenuma_scan(p);
2470 up_read(&mm->mmap_sem);
2474 * Drive the periodic memory faults..
2476 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2478 struct callback_head *work = &curr->numa_work;
2482 * We don't care about NUMA placement if we don't have memory.
2484 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2488 * Using runtime rather than walltime has the dual advantage that
2489 * we (mostly) drive the selection from busy threads and that the
2490 * task needs to have done some actual work before we bother with
2493 now = curr->se.sum_exec_runtime;
2494 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2496 if (now > curr->node_stamp + period) {
2497 if (!curr->node_stamp)
2498 curr->numa_scan_period = task_scan_min(curr);
2499 curr->node_stamp += period;
2501 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2502 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2503 task_work_add(curr, work, true);
2508 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2512 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2516 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2519 #endif /* CONFIG_NUMA_BALANCING */
2522 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2524 update_load_add(&cfs_rq->load, se->load.weight);
2525 if (!parent_entity(se))
2526 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2528 if (entity_is_task(se)) {
2529 struct rq *rq = rq_of(cfs_rq);
2531 account_numa_enqueue(rq, task_of(se));
2532 list_add(&se->group_node, &rq->cfs_tasks);
2535 cfs_rq->nr_running++;
2539 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2541 update_load_sub(&cfs_rq->load, se->load.weight);
2542 if (!parent_entity(se))
2543 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2544 if (entity_is_task(se)) {
2545 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2546 list_del_init(&se->group_node);
2548 cfs_rq->nr_running--;
2551 #ifdef CONFIG_FAIR_GROUP_SCHED
2553 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2555 long tg_weight, load, shares;
2558 * This really should be: cfs_rq->avg.load_avg, but instead we use
2559 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2560 * the shares for small weight interactive tasks.
2562 load = scale_load_down(cfs_rq->load.weight);
2564 tg_weight = atomic_long_read(&tg->load_avg);
2566 /* Ensure tg_weight >= load */
2567 tg_weight -= cfs_rq->tg_load_avg_contrib;
2570 shares = (tg->shares * load);
2572 shares /= tg_weight;
2574 if (shares < MIN_SHARES)
2575 shares = MIN_SHARES;
2576 if (shares > tg->shares)
2577 shares = tg->shares;
2581 # else /* CONFIG_SMP */
2582 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2586 # endif /* CONFIG_SMP */
2588 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2589 unsigned long weight)
2592 /* commit outstanding execution time */
2593 if (cfs_rq->curr == se)
2594 update_curr(cfs_rq);
2595 account_entity_dequeue(cfs_rq, se);
2598 update_load_set(&se->load, weight);
2601 account_entity_enqueue(cfs_rq, se);
2604 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2606 static void update_cfs_shares(struct sched_entity *se)
2608 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2609 struct task_group *tg;
2615 if (throttled_hierarchy(cfs_rq))
2621 if (likely(se->load.weight == tg->shares))
2624 shares = calc_cfs_shares(cfs_rq, tg);
2626 reweight_entity(cfs_rq_of(se), se, shares);
2629 #else /* CONFIG_FAIR_GROUP_SCHED */
2630 static inline void update_cfs_shares(struct sched_entity *se)
2633 #endif /* CONFIG_FAIR_GROUP_SCHED */
2636 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2637 static const u32 runnable_avg_yN_inv[] = {
2638 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2639 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2640 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2641 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2642 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2643 0x85aac367, 0x82cd8698,
2647 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2648 * over-estimates when re-combining.
2650 static const u32 runnable_avg_yN_sum[] = {
2651 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2652 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2653 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2658 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2660 static __always_inline u64 decay_load(u64 val, u64 n)
2662 unsigned int local_n;
2666 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2669 /* after bounds checking we can collapse to 32-bit */
2673 * As y^PERIOD = 1/2, we can combine
2674 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2675 * With a look-up table which covers y^n (n<PERIOD)
2677 * To achieve constant time decay_load.
2679 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2680 val >>= local_n / LOAD_AVG_PERIOD;
2681 local_n %= LOAD_AVG_PERIOD;
2684 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2689 * For updates fully spanning n periods, the contribution to runnable
2690 * average will be: \Sum 1024*y^n
2692 * We can compute this reasonably efficiently by combining:
2693 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2695 static u32 __compute_runnable_contrib(u64 n)
2699 if (likely(n <= LOAD_AVG_PERIOD))
2700 return runnable_avg_yN_sum[n];
2701 else if (unlikely(n >= LOAD_AVG_MAX_N))
2702 return LOAD_AVG_MAX;
2704 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2706 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2707 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2709 n -= LOAD_AVG_PERIOD;
2710 } while (n > LOAD_AVG_PERIOD);
2712 contrib = decay_load(contrib, n);
2713 return contrib + runnable_avg_yN_sum[n];
2716 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2717 #error "load tracking assumes 2^10 as unit"
2720 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2723 * We can represent the historical contribution to runnable average as the
2724 * coefficients of a geometric series. To do this we sub-divide our runnable
2725 * history into segments of approximately 1ms (1024us); label the segment that
2726 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2728 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2730 * (now) (~1ms ago) (~2ms ago)
2732 * Let u_i denote the fraction of p_i that the entity was runnable.
2734 * We then designate the fractions u_i as our co-efficients, yielding the
2735 * following representation of historical load:
2736 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2738 * We choose y based on the with of a reasonably scheduling period, fixing:
2741 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2742 * approximately half as much as the contribution to load within the last ms
2745 * When a period "rolls over" and we have new u_0`, multiplying the previous
2746 * sum again by y is sufficient to update:
2747 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2748 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2750 static __always_inline int
2751 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2752 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2754 u64 delta, scaled_delta, periods;
2756 unsigned int delta_w, scaled_delta_w, decayed = 0;
2757 unsigned long scale_freq, scale_cpu;
2759 delta = now - sa->last_update_time;
2761 * This should only happen when time goes backwards, which it
2762 * unfortunately does during sched clock init when we swap over to TSC.
2764 if ((s64)delta < 0) {
2765 sa->last_update_time = now;
2770 * Use 1024ns as the unit of measurement since it's a reasonable
2771 * approximation of 1us and fast to compute.
2776 sa->last_update_time = now;
2778 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2779 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2780 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2782 /* delta_w is the amount already accumulated against our next period */
2783 delta_w = sa->period_contrib;
2784 if (delta + delta_w >= 1024) {
2787 /* how much left for next period will start over, we don't know yet */
2788 sa->period_contrib = 0;
2791 * Now that we know we're crossing a period boundary, figure
2792 * out how much from delta we need to complete the current
2793 * period and accrue it.
2795 delta_w = 1024 - delta_w;
2796 scaled_delta_w = cap_scale(delta_w, scale_freq);
2798 sa->load_sum += weight * scaled_delta_w;
2800 cfs_rq->runnable_load_sum +=
2801 weight * scaled_delta_w;
2805 sa->util_sum += scaled_delta_w * scale_cpu;
2809 /* Figure out how many additional periods this update spans */
2810 periods = delta / 1024;
2813 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2815 cfs_rq->runnable_load_sum =
2816 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2818 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2820 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2821 contrib = __compute_runnable_contrib(periods);
2822 contrib = cap_scale(contrib, scale_freq);
2824 sa->load_sum += weight * contrib;
2826 cfs_rq->runnable_load_sum += weight * contrib;
2829 sa->util_sum += contrib * scale_cpu;
2832 /* Remainder of delta accrued against u_0` */
2833 scaled_delta = cap_scale(delta, scale_freq);
2835 sa->load_sum += weight * scaled_delta;
2837 cfs_rq->runnable_load_sum += weight * scaled_delta;
2840 sa->util_sum += scaled_delta * scale_cpu;
2842 sa->period_contrib += delta;
2845 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2847 cfs_rq->runnable_load_avg =
2848 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2850 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2857 * Signed add and clamp on underflow.
2859 * Explicitly do a load-store to ensure the intermediate value never hits
2860 * memory. This allows lockless observations without ever seeing the negative
2863 #define add_positive(_ptr, _val) do { \
2864 typeof(_ptr) ptr = (_ptr); \
2865 typeof(_val) val = (_val); \
2866 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2870 if (val < 0 && res > var) \
2873 WRITE_ONCE(*ptr, res); \
2876 #ifdef CONFIG_FAIR_GROUP_SCHED
2878 * update_tg_load_avg - update the tg's load avg
2879 * @cfs_rq: the cfs_rq whose avg changed
2880 * @force: update regardless of how small the difference
2882 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2883 * However, because tg->load_avg is a global value there are performance
2886 * In order to avoid having to look at the other cfs_rq's, we use a
2887 * differential update where we store the last value we propagated. This in
2888 * turn allows skipping updates if the differential is 'small'.
2890 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2891 * done) and effective_load() (which is not done because it is too costly).
2893 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2895 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2898 * No need to update load_avg for root_task_group as it is not used.
2900 if (cfs_rq->tg == &root_task_group)
2903 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2904 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2905 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2910 * Called within set_task_rq() right before setting a task's cpu. The
2911 * caller only guarantees p->pi_lock is held; no other assumptions,
2912 * including the state of rq->lock, should be made.
2914 void set_task_rq_fair(struct sched_entity *se,
2915 struct cfs_rq *prev, struct cfs_rq *next)
2917 if (!sched_feat(ATTACH_AGE_LOAD))
2921 * We are supposed to update the task to "current" time, then its up to
2922 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2923 * getting what current time is, so simply throw away the out-of-date
2924 * time. This will result in the wakee task is less decayed, but giving
2925 * the wakee more load sounds not bad.
2927 if (se->avg.last_update_time && prev) {
2928 u64 p_last_update_time;
2929 u64 n_last_update_time;
2931 #ifndef CONFIG_64BIT
2932 u64 p_last_update_time_copy;
2933 u64 n_last_update_time_copy;
2936 p_last_update_time_copy = prev->load_last_update_time_copy;
2937 n_last_update_time_copy = next->load_last_update_time_copy;
2941 p_last_update_time = prev->avg.last_update_time;
2942 n_last_update_time = next->avg.last_update_time;
2944 } while (p_last_update_time != p_last_update_time_copy ||
2945 n_last_update_time != n_last_update_time_copy);
2947 p_last_update_time = prev->avg.last_update_time;
2948 n_last_update_time = next->avg.last_update_time;
2950 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2951 &se->avg, 0, 0, NULL);
2952 se->avg.last_update_time = n_last_update_time;
2956 /* Take into account change of utilization of a child task group */
2958 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
2960 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2961 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
2963 /* Nothing to update */
2967 /* Set new sched_entity's utilization */
2968 se->avg.util_avg = gcfs_rq->avg.util_avg;
2969 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
2971 /* Update parent cfs_rq utilization */
2972 add_positive(&cfs_rq->avg.util_avg, delta);
2973 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
2976 /* Take into account change of load of a child task group */
2978 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
2980 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2981 long delta, load = gcfs_rq->avg.load_avg;
2984 * If the load of group cfs_rq is null, the load of the
2985 * sched_entity will also be null so we can skip the formula
2990 /* Get tg's load and ensure tg_load > 0 */
2991 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
2993 /* Ensure tg_load >= load and updated with current load*/
2994 tg_load -= gcfs_rq->tg_load_avg_contrib;
2998 * We need to compute a correction term in the case that the
2999 * task group is consuming more CPU than a task of equal
3000 * weight. A task with a weight equals to tg->shares will have
3001 * a load less or equal to scale_load_down(tg->shares).
3002 * Similarly, the sched_entities that represent the task group
3003 * at parent level, can't have a load higher than
3004 * scale_load_down(tg->shares). And the Sum of sched_entities'
3005 * load must be <= scale_load_down(tg->shares).
3007 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3008 /* scale gcfs_rq's load into tg's shares*/
3009 load *= scale_load_down(gcfs_rq->tg->shares);
3014 delta = load - se->avg.load_avg;
3016 /* Nothing to update */
3020 /* Set new sched_entity's load */
3021 se->avg.load_avg = load;
3022 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3024 /* Update parent cfs_rq load */
3025 add_positive(&cfs_rq->avg.load_avg, delta);
3026 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3029 * If the sched_entity is already enqueued, we also have to update the
3030 * runnable load avg.
3033 /* Update parent cfs_rq runnable_load_avg */
3034 add_positive(&cfs_rq->runnable_load_avg, delta);
3035 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3039 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3041 cfs_rq->propagate_avg = 1;
3044 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3046 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3048 if (!cfs_rq->propagate_avg)
3051 cfs_rq->propagate_avg = 0;
3055 /* Update task and its cfs_rq load average */
3056 static inline int propagate_entity_load_avg(struct sched_entity *se)
3058 struct cfs_rq *cfs_rq;
3060 if (entity_is_task(se))
3063 if (!test_and_clear_tg_cfs_propagate(se))
3066 cfs_rq = cfs_rq_of(se);
3068 set_tg_cfs_propagate(cfs_rq);
3070 update_tg_cfs_util(cfs_rq, se);
3071 update_tg_cfs_load(cfs_rq, se);
3076 #else /* CONFIG_FAIR_GROUP_SCHED */
3078 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3080 static inline int propagate_entity_load_avg(struct sched_entity *se)
3085 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3087 #endif /* CONFIG_FAIR_GROUP_SCHED */
3089 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3091 if (&this_rq()->cfs == cfs_rq) {
3093 * There are a few boundary cases this might miss but it should
3094 * get called often enough that that should (hopefully) not be
3095 * a real problem -- added to that it only calls on the local
3096 * CPU, so if we enqueue remotely we'll miss an update, but
3097 * the next tick/schedule should update.
3099 * It will not get called when we go idle, because the idle
3100 * thread is a different class (!fair), nor will the utilization
3101 * number include things like RT tasks.
3103 * As is, the util number is not freq-invariant (we'd have to
3104 * implement arch_scale_freq_capacity() for that).
3108 cpufreq_update_util(rq_of(cfs_rq), 0);
3112 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
3115 * Unsigned subtract and clamp on underflow.
3117 * Explicitly do a load-store to ensure the intermediate value never hits
3118 * memory. This allows lockless observations without ever seeing the negative
3121 #define sub_positive(_ptr, _val) do { \
3122 typeof(_ptr) ptr = (_ptr); \
3123 typeof(*ptr) val = (_val); \
3124 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3128 WRITE_ONCE(*ptr, res); \
3132 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3133 * @now: current time, as per cfs_rq_clock_task()
3134 * @cfs_rq: cfs_rq to update
3135 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3137 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3138 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3139 * post_init_entity_util_avg().
3141 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3143 * Returns true if the load decayed or we removed load.
3145 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3146 * call update_tg_load_avg() when this function returns true.
3149 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3151 struct sched_avg *sa = &cfs_rq->avg;
3152 int decayed, removed = 0, removed_util = 0;
3154 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3155 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3156 sub_positive(&sa->load_avg, r);
3157 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3159 set_tg_cfs_propagate(cfs_rq);
3162 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3163 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3164 sub_positive(&sa->util_avg, r);
3165 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3167 set_tg_cfs_propagate(cfs_rq);
3170 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3171 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3173 #ifndef CONFIG_64BIT
3175 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3178 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
3179 if (cfs_rq == &rq_of(cfs_rq)->cfs)
3180 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
3182 if (update_freq && (decayed || removed_util))
3183 cfs_rq_util_change(cfs_rq);
3185 return decayed || removed;
3189 * Optional action to be done while updating the load average
3191 #define UPDATE_TG 0x1
3192 #define SKIP_AGE_LOAD 0x2
3194 /* Update task and its cfs_rq load average */
3195 static inline void update_load_avg(struct sched_entity *se, int flags)
3197 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3198 u64 now = cfs_rq_clock_task(cfs_rq);
3199 int cpu = cpu_of(rq_of(cfs_rq));
3204 * Track task load average for carrying it to new CPU after migrated, and
3205 * track group sched_entity load average for task_h_load calc in migration
3207 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
3208 __update_load_avg(now, cpu, &se->avg,
3209 se->on_rq * scale_load_down(se->load.weight),
3210 cfs_rq->curr == se, NULL);
3213 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3214 decayed |= propagate_entity_load_avg(se);
3216 if (decayed && (flags & UPDATE_TG))
3217 update_tg_load_avg(cfs_rq, 0);
3219 if (entity_is_task(se)) {
3220 #ifdef CONFIG_SCHED_WALT
3221 ptr = (void *)&(task_of(se)->ravg);
3223 trace_sched_load_avg_task(task_of(se), &se->avg, ptr);
3228 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3229 * @cfs_rq: cfs_rq to attach to
3230 * @se: sched_entity to attach
3232 * Must call update_cfs_rq_load_avg() before this, since we rely on
3233 * cfs_rq->avg.last_update_time being current.
3235 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3237 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3238 cfs_rq->avg.load_avg += se->avg.load_avg;
3239 cfs_rq->avg.load_sum += se->avg.load_sum;
3240 cfs_rq->avg.util_avg += se->avg.util_avg;
3241 cfs_rq->avg.util_sum += se->avg.util_sum;
3242 set_tg_cfs_propagate(cfs_rq);
3244 cfs_rq_util_change(cfs_rq);
3248 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3249 * @cfs_rq: cfs_rq to detach from
3250 * @se: sched_entity to detach
3252 * Must call update_cfs_rq_load_avg() before this, since we rely on
3253 * cfs_rq->avg.last_update_time being current.
3255 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3258 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3259 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3260 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3261 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3262 set_tg_cfs_propagate(cfs_rq);
3264 cfs_rq_util_change(cfs_rq);
3267 /* Add the load generated by se into cfs_rq's load average */
3269 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3271 struct sched_avg *sa = &se->avg;
3273 cfs_rq->runnable_load_avg += sa->load_avg;
3274 cfs_rq->runnable_load_sum += sa->load_sum;
3276 if (!sa->last_update_time) {
3277 attach_entity_load_avg(cfs_rq, se);
3278 update_tg_load_avg(cfs_rq, 0);
3282 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3284 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3286 cfs_rq->runnable_load_avg =
3287 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3288 cfs_rq->runnable_load_sum =
3289 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3292 #ifndef CONFIG_64BIT
3293 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3295 u64 last_update_time_copy;
3296 u64 last_update_time;
3299 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3301 last_update_time = cfs_rq->avg.last_update_time;
3302 } while (last_update_time != last_update_time_copy);
3304 return last_update_time;
3307 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3309 return cfs_rq->avg.last_update_time;
3314 * Synchronize entity load avg of dequeued entity without locking
3317 void sync_entity_load_avg(struct sched_entity *se)
3319 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3320 u64 last_update_time;
3322 last_update_time = cfs_rq_last_update_time(cfs_rq);
3323 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3327 * Task first catches up with cfs_rq, and then subtract
3328 * itself from the cfs_rq (task must be off the queue now).
3330 void remove_entity_load_avg(struct sched_entity *se)
3332 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3335 * tasks cannot exit without having gone through wake_up_new_task() ->
3336 * post_init_entity_util_avg() which will have added things to the
3337 * cfs_rq, so we can remove unconditionally.
3339 * Similarly for groups, they will have passed through
3340 * post_init_entity_util_avg() before unregister_sched_fair_group()
3344 sync_entity_load_avg(se);
3345 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3346 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3350 * Update the rq's load with the elapsed running time before entering
3351 * idle. if the last scheduled task is not a CFS task, idle_enter will
3352 * be the only way to update the runnable statistic.
3354 void idle_enter_fair(struct rq *this_rq)
3359 * Update the rq's load with the elapsed idle time before a task is
3360 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
3361 * be the only way to update the runnable statistic.
3363 void idle_exit_fair(struct rq *this_rq)
3367 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3369 return cfs_rq->runnable_load_avg;
3372 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3374 return cfs_rq->avg.load_avg;
3377 static int idle_balance(struct rq *this_rq);
3379 #else /* CONFIG_SMP */
3382 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3387 #define UPDATE_TG 0x0
3388 #define SKIP_AGE_LOAD 0x0
3390 static inline void update_load_avg(struct sched_entity *se, int not_used1){}
3392 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3394 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3395 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3398 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3400 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3402 static inline int idle_balance(struct rq *rq)
3407 #endif /* CONFIG_SMP */
3409 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3411 #ifdef CONFIG_SCHEDSTATS
3412 struct task_struct *tsk = NULL;
3414 if (entity_is_task(se))
3417 if (se->statistics.sleep_start) {
3418 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3423 if (unlikely(delta > se->statistics.sleep_max))
3424 se->statistics.sleep_max = delta;
3426 se->statistics.sleep_start = 0;
3427 se->statistics.sum_sleep_runtime += delta;
3430 account_scheduler_latency(tsk, delta >> 10, 1);
3431 trace_sched_stat_sleep(tsk, delta);
3434 if (se->statistics.block_start) {
3435 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3440 if (unlikely(delta > se->statistics.block_max))
3441 se->statistics.block_max = delta;
3443 se->statistics.block_start = 0;
3444 se->statistics.sum_sleep_runtime += delta;
3447 if (tsk->in_iowait) {
3448 se->statistics.iowait_sum += delta;
3449 se->statistics.iowait_count++;
3450 trace_sched_stat_iowait(tsk, delta);
3453 trace_sched_stat_blocked(tsk, delta);
3454 trace_sched_blocked_reason(tsk);
3457 * Blocking time is in units of nanosecs, so shift by
3458 * 20 to get a milliseconds-range estimation of the
3459 * amount of time that the task spent sleeping:
3461 if (unlikely(prof_on == SLEEP_PROFILING)) {
3462 profile_hits(SLEEP_PROFILING,
3463 (void *)get_wchan(tsk),
3466 account_scheduler_latency(tsk, delta >> 10, 0);
3472 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3474 #ifdef CONFIG_SCHED_DEBUG
3475 s64 d = se->vruntime - cfs_rq->min_vruntime;
3480 if (d > 3*sysctl_sched_latency)
3481 schedstat_inc(cfs_rq, nr_spread_over);
3486 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3488 u64 vruntime = cfs_rq->min_vruntime;
3491 * The 'current' period is already promised to the current tasks,
3492 * however the extra weight of the new task will slow them down a
3493 * little, place the new task so that it fits in the slot that
3494 * stays open at the end.
3496 if (initial && sched_feat(START_DEBIT))
3497 vruntime += sched_vslice(cfs_rq, se);
3499 /* sleeps up to a single latency don't count. */
3501 unsigned long thresh = sysctl_sched_latency;
3504 * Halve their sleep time's effect, to allow
3505 * for a gentler effect of sleepers:
3507 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3513 /* ensure we never gain time by being placed backwards. */
3514 se->vruntime = max_vruntime(se->vruntime, vruntime);
3517 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3520 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3523 * Update the normalized vruntime before updating min_vruntime
3524 * through calling update_curr().
3526 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3527 se->vruntime += cfs_rq->min_vruntime;
3530 * Update run-time statistics of the 'current'.
3532 update_curr(cfs_rq);
3533 update_load_avg(se, UPDATE_TG);
3534 enqueue_entity_load_avg(cfs_rq, se);
3535 update_cfs_shares(se);
3536 account_entity_enqueue(cfs_rq, se);
3538 if (flags & ENQUEUE_WAKEUP) {
3539 place_entity(cfs_rq, se, 0);
3540 enqueue_sleeper(cfs_rq, se);
3543 update_stats_enqueue(cfs_rq, se);
3544 check_spread(cfs_rq, se);
3545 if (se != cfs_rq->curr)
3546 __enqueue_entity(cfs_rq, se);
3549 if (cfs_rq->nr_running == 1) {
3550 list_add_leaf_cfs_rq(cfs_rq);
3551 check_enqueue_throttle(cfs_rq);
3555 static void __clear_buddies_last(struct sched_entity *se)
3557 for_each_sched_entity(se) {
3558 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3559 if (cfs_rq->last != se)
3562 cfs_rq->last = NULL;
3566 static void __clear_buddies_next(struct sched_entity *se)
3568 for_each_sched_entity(se) {
3569 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3570 if (cfs_rq->next != se)
3573 cfs_rq->next = NULL;
3577 static void __clear_buddies_skip(struct sched_entity *se)
3579 for_each_sched_entity(se) {
3580 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3581 if (cfs_rq->skip != se)
3584 cfs_rq->skip = NULL;
3588 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3590 if (cfs_rq->last == se)
3591 __clear_buddies_last(se);
3593 if (cfs_rq->next == se)
3594 __clear_buddies_next(se);
3596 if (cfs_rq->skip == se)
3597 __clear_buddies_skip(se);
3600 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3603 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3606 * Update run-time statistics of the 'current'.
3608 update_curr(cfs_rq);
3611 * When dequeuing a sched_entity, we must:
3612 * - Update loads to have both entity and cfs_rq synced with now.
3613 * - Substract its load from the cfs_rq->runnable_avg.
3614 * - Substract its previous weight from cfs_rq->load.weight.
3615 * - For group entity, update its weight to reflect the new share
3616 * of its group cfs_rq.
3618 update_load_avg(se, UPDATE_TG);
3619 dequeue_entity_load_avg(cfs_rq, se);
3621 update_stats_dequeue(cfs_rq, se);
3622 if (flags & DEQUEUE_SLEEP) {
3623 #ifdef CONFIG_SCHEDSTATS
3624 if (entity_is_task(se)) {
3625 struct task_struct *tsk = task_of(se);
3627 if (tsk->state & TASK_INTERRUPTIBLE)
3628 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3629 if (tsk->state & TASK_UNINTERRUPTIBLE)
3630 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3635 clear_buddies(cfs_rq, se);
3637 if (se != cfs_rq->curr)
3638 __dequeue_entity(cfs_rq, se);
3640 account_entity_dequeue(cfs_rq, se);
3643 * Normalize the entity after updating the min_vruntime because the
3644 * update can refer to the ->curr item and we need to reflect this
3645 * movement in our normalized position.
3647 if (!(flags & DEQUEUE_SLEEP))
3648 se->vruntime -= cfs_rq->min_vruntime;
3650 /* return excess runtime on last dequeue */
3651 return_cfs_rq_runtime(cfs_rq);
3653 update_min_vruntime(cfs_rq);
3654 update_cfs_shares(se);
3658 * Preempt the current task with a newly woken task if needed:
3661 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3663 unsigned long ideal_runtime, delta_exec;
3664 struct sched_entity *se;
3667 ideal_runtime = sched_slice(cfs_rq, curr);
3668 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3669 if (delta_exec > ideal_runtime) {
3670 resched_curr(rq_of(cfs_rq));
3672 * The current task ran long enough, ensure it doesn't get
3673 * re-elected due to buddy favours.
3675 clear_buddies(cfs_rq, curr);
3680 * Ensure that a task that missed wakeup preemption by a
3681 * narrow margin doesn't have to wait for a full slice.
3682 * This also mitigates buddy induced latencies under load.
3684 if (delta_exec < sysctl_sched_min_granularity)
3687 se = __pick_first_entity(cfs_rq);
3688 delta = curr->vruntime - se->vruntime;
3693 if (delta > ideal_runtime)
3694 resched_curr(rq_of(cfs_rq));
3698 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3700 /* 'current' is not kept within the tree. */
3703 * Any task has to be enqueued before it get to execute on
3704 * a CPU. So account for the time it spent waiting on the
3707 update_stats_wait_end(cfs_rq, se);
3708 __dequeue_entity(cfs_rq, se);
3709 update_load_avg(se, UPDATE_TG);
3712 update_stats_curr_start(cfs_rq, se);
3714 #ifdef CONFIG_SCHEDSTATS
3716 * Track our maximum slice length, if the CPU's load is at
3717 * least twice that of our own weight (i.e. dont track it
3718 * when there are only lesser-weight tasks around):
3720 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3721 se->statistics.slice_max = max(se->statistics.slice_max,
3722 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3725 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3729 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3732 * Pick the next process, keeping these things in mind, in this order:
3733 * 1) keep things fair between processes/task groups
3734 * 2) pick the "next" process, since someone really wants that to run
3735 * 3) pick the "last" process, for cache locality
3736 * 4) do not run the "skip" process, if something else is available
3738 static struct sched_entity *
3739 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3741 struct sched_entity *left = __pick_first_entity(cfs_rq);
3742 struct sched_entity *se;
3745 * If curr is set we have to see if its left of the leftmost entity
3746 * still in the tree, provided there was anything in the tree at all.
3748 if (!left || (curr && entity_before(curr, left)))
3751 se = left; /* ideally we run the leftmost entity */
3754 * Avoid running the skip buddy, if running something else can
3755 * be done without getting too unfair.
3757 if (cfs_rq->skip == se) {
3758 struct sched_entity *second;
3761 second = __pick_first_entity(cfs_rq);
3763 second = __pick_next_entity(se);
3764 if (!second || (curr && entity_before(curr, second)))
3768 if (second && wakeup_preempt_entity(second, left) < 1)
3773 * Prefer last buddy, try to return the CPU to a preempted task.
3775 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3779 * Someone really wants this to run. If it's not unfair, run it.
3781 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3784 clear_buddies(cfs_rq, se);
3789 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3791 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3794 * If still on the runqueue then deactivate_task()
3795 * was not called and update_curr() has to be done:
3798 update_curr(cfs_rq);
3800 /* throttle cfs_rqs exceeding runtime */
3801 check_cfs_rq_runtime(cfs_rq);
3803 check_spread(cfs_rq, prev);
3805 update_stats_wait_start(cfs_rq, prev);
3806 /* Put 'current' back into the tree. */
3807 __enqueue_entity(cfs_rq, prev);
3808 /* in !on_rq case, update occurred at dequeue */
3809 update_load_avg(prev, 0);
3811 cfs_rq->curr = NULL;
3815 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3818 * Update run-time statistics of the 'current'.
3820 update_curr(cfs_rq);
3823 * Ensure that runnable average is periodically updated.
3825 update_load_avg(curr, UPDATE_TG);
3826 update_cfs_shares(curr);
3828 #ifdef CONFIG_SCHED_HRTICK
3830 * queued ticks are scheduled to match the slice, so don't bother
3831 * validating it and just reschedule.
3834 resched_curr(rq_of(cfs_rq));
3838 * don't let the period tick interfere with the hrtick preemption
3840 if (!sched_feat(DOUBLE_TICK) &&
3841 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3845 if (cfs_rq->nr_running > 1)
3846 check_preempt_tick(cfs_rq, curr);
3850 /**************************************************
3851 * CFS bandwidth control machinery
3854 #ifdef CONFIG_CFS_BANDWIDTH
3856 #ifdef HAVE_JUMP_LABEL
3857 static struct static_key __cfs_bandwidth_used;
3859 static inline bool cfs_bandwidth_used(void)
3861 return static_key_false(&__cfs_bandwidth_used);
3864 void cfs_bandwidth_usage_inc(void)
3866 static_key_slow_inc(&__cfs_bandwidth_used);
3869 void cfs_bandwidth_usage_dec(void)
3871 static_key_slow_dec(&__cfs_bandwidth_used);
3873 #else /* HAVE_JUMP_LABEL */
3874 static bool cfs_bandwidth_used(void)
3879 void cfs_bandwidth_usage_inc(void) {}
3880 void cfs_bandwidth_usage_dec(void) {}
3881 #endif /* HAVE_JUMP_LABEL */
3884 * default period for cfs group bandwidth.
3885 * default: 0.1s, units: nanoseconds
3887 static inline u64 default_cfs_period(void)
3889 return 100000000ULL;
3892 static inline u64 sched_cfs_bandwidth_slice(void)
3894 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3898 * Replenish runtime according to assigned quota and update expiration time.
3899 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3900 * additional synchronization around rq->lock.
3902 * requires cfs_b->lock
3904 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3908 if (cfs_b->quota == RUNTIME_INF)
3911 now = sched_clock_cpu(smp_processor_id());
3912 cfs_b->runtime = cfs_b->quota;
3913 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3916 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3918 return &tg->cfs_bandwidth;
3921 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3922 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3924 if (unlikely(cfs_rq->throttle_count))
3925 return cfs_rq->throttled_clock_task;
3927 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3930 /* returns 0 on failure to allocate runtime */
3931 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3933 struct task_group *tg = cfs_rq->tg;
3934 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3935 u64 amount = 0, min_amount, expires;
3937 /* note: this is a positive sum as runtime_remaining <= 0 */
3938 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3940 raw_spin_lock(&cfs_b->lock);
3941 if (cfs_b->quota == RUNTIME_INF)
3942 amount = min_amount;
3944 start_cfs_bandwidth(cfs_b);
3946 if (cfs_b->runtime > 0) {
3947 amount = min(cfs_b->runtime, min_amount);
3948 cfs_b->runtime -= amount;
3952 expires = cfs_b->runtime_expires;
3953 raw_spin_unlock(&cfs_b->lock);
3955 cfs_rq->runtime_remaining += amount;
3957 * we may have advanced our local expiration to account for allowed
3958 * spread between our sched_clock and the one on which runtime was
3961 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3962 cfs_rq->runtime_expires = expires;
3964 return cfs_rq->runtime_remaining > 0;
3968 * Note: This depends on the synchronization provided by sched_clock and the
3969 * fact that rq->clock snapshots this value.
3971 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3973 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3975 /* if the deadline is ahead of our clock, nothing to do */
3976 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3979 if (cfs_rq->runtime_remaining < 0)
3983 * If the local deadline has passed we have to consider the
3984 * possibility that our sched_clock is 'fast' and the global deadline
3985 * has not truly expired.
3987 * Fortunately we can check determine whether this the case by checking
3988 * whether the global deadline has advanced. It is valid to compare
3989 * cfs_b->runtime_expires without any locks since we only care about
3990 * exact equality, so a partial write will still work.
3993 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3994 /* extend local deadline, drift is bounded above by 2 ticks */
3995 cfs_rq->runtime_expires += TICK_NSEC;
3997 /* global deadline is ahead, expiration has passed */
3998 cfs_rq->runtime_remaining = 0;
4002 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4004 /* dock delta_exec before expiring quota (as it could span periods) */
4005 cfs_rq->runtime_remaining -= delta_exec;
4006 expire_cfs_rq_runtime(cfs_rq);
4008 if (likely(cfs_rq->runtime_remaining > 0))
4012 * if we're unable to extend our runtime we resched so that the active
4013 * hierarchy can be throttled
4015 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4016 resched_curr(rq_of(cfs_rq));
4019 static __always_inline
4020 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4022 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4025 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4028 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4030 return cfs_bandwidth_used() && cfs_rq->throttled;
4033 /* check whether cfs_rq, or any parent, is throttled */
4034 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4036 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4040 * Ensure that neither of the group entities corresponding to src_cpu or
4041 * dest_cpu are members of a throttled hierarchy when performing group
4042 * load-balance operations.
4044 static inline int throttled_lb_pair(struct task_group *tg,
4045 int src_cpu, int dest_cpu)
4047 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4049 src_cfs_rq = tg->cfs_rq[src_cpu];
4050 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4052 return throttled_hierarchy(src_cfs_rq) ||
4053 throttled_hierarchy(dest_cfs_rq);
4056 /* updated child weight may affect parent so we have to do this bottom up */
4057 static int tg_unthrottle_up(struct task_group *tg, void *data)
4059 struct rq *rq = data;
4060 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4062 cfs_rq->throttle_count--;
4064 if (!cfs_rq->throttle_count) {
4065 /* adjust cfs_rq_clock_task() */
4066 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4067 cfs_rq->throttled_clock_task;
4074 static int tg_throttle_down(struct task_group *tg, void *data)
4076 struct rq *rq = data;
4077 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4079 /* group is entering throttled state, stop time */
4080 if (!cfs_rq->throttle_count)
4081 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4082 cfs_rq->throttle_count++;
4087 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4089 struct rq *rq = rq_of(cfs_rq);
4090 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4091 struct sched_entity *se;
4092 long task_delta, dequeue = 1;
4095 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4097 /* freeze hierarchy runnable averages while throttled */
4099 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4102 task_delta = cfs_rq->h_nr_running;
4103 for_each_sched_entity(se) {
4104 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4105 /* throttled entity or throttle-on-deactivate */
4110 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4111 qcfs_rq->h_nr_running -= task_delta;
4113 if (qcfs_rq->load.weight)
4118 sub_nr_running(rq, task_delta);
4120 cfs_rq->throttled = 1;
4121 cfs_rq->throttled_clock = rq_clock(rq);
4122 raw_spin_lock(&cfs_b->lock);
4123 empty = list_empty(&cfs_b->throttled_cfs_rq);
4126 * Add to the _head_ of the list, so that an already-started
4127 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4128 * not running add to the tail so that later runqueues don't get starved.
4130 if (cfs_b->distribute_running)
4131 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4133 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4136 * If we're the first throttled task, make sure the bandwidth
4140 start_cfs_bandwidth(cfs_b);
4142 raw_spin_unlock(&cfs_b->lock);
4145 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4147 struct rq *rq = rq_of(cfs_rq);
4148 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4149 struct sched_entity *se;
4153 se = cfs_rq->tg->se[cpu_of(rq)];
4155 cfs_rq->throttled = 0;
4157 update_rq_clock(rq);
4159 raw_spin_lock(&cfs_b->lock);
4160 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4161 list_del_rcu(&cfs_rq->throttled_list);
4162 raw_spin_unlock(&cfs_b->lock);
4164 /* update hierarchical throttle state */
4165 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4167 if (!cfs_rq->load.weight)
4170 task_delta = cfs_rq->h_nr_running;
4171 for_each_sched_entity(se) {
4175 cfs_rq = cfs_rq_of(se);
4177 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4178 cfs_rq->h_nr_running += task_delta;
4180 if (cfs_rq_throttled(cfs_rq))
4185 add_nr_running(rq, task_delta);
4187 /* determine whether we need to wake up potentially idle cpu */
4188 if (rq->curr == rq->idle && rq->cfs.nr_running)
4192 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4193 u64 remaining, u64 expires)
4195 struct cfs_rq *cfs_rq;
4197 u64 starting_runtime = remaining;
4200 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4202 struct rq *rq = rq_of(cfs_rq);
4204 raw_spin_lock(&rq->lock);
4205 if (!cfs_rq_throttled(cfs_rq))
4208 runtime = -cfs_rq->runtime_remaining + 1;
4209 if (runtime > remaining)
4210 runtime = remaining;
4211 remaining -= runtime;
4213 cfs_rq->runtime_remaining += runtime;
4214 cfs_rq->runtime_expires = expires;
4216 /* we check whether we're throttled above */
4217 if (cfs_rq->runtime_remaining > 0)
4218 unthrottle_cfs_rq(cfs_rq);
4221 raw_spin_unlock(&rq->lock);
4228 return starting_runtime - remaining;
4232 * Responsible for refilling a task_group's bandwidth and unthrottling its
4233 * cfs_rqs as appropriate. If there has been no activity within the last
4234 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4235 * used to track this state.
4237 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4239 u64 runtime, runtime_expires;
4242 /* no need to continue the timer with no bandwidth constraint */
4243 if (cfs_b->quota == RUNTIME_INF)
4244 goto out_deactivate;
4246 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4247 cfs_b->nr_periods += overrun;
4250 * idle depends on !throttled (for the case of a large deficit), and if
4251 * we're going inactive then everything else can be deferred
4253 if (cfs_b->idle && !throttled)
4254 goto out_deactivate;
4256 __refill_cfs_bandwidth_runtime(cfs_b);
4259 /* mark as potentially idle for the upcoming period */
4264 /* account preceding periods in which throttling occurred */
4265 cfs_b->nr_throttled += overrun;
4267 runtime_expires = cfs_b->runtime_expires;
4270 * This check is repeated as we are holding onto the new bandwidth while
4271 * we unthrottle. This can potentially race with an unthrottled group
4272 * trying to acquire new bandwidth from the global pool. This can result
4273 * in us over-using our runtime if it is all used during this loop, but
4274 * only by limited amounts in that extreme case.
4276 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4277 runtime = cfs_b->runtime;
4278 cfs_b->distribute_running = 1;
4279 raw_spin_unlock(&cfs_b->lock);
4280 /* we can't nest cfs_b->lock while distributing bandwidth */
4281 runtime = distribute_cfs_runtime(cfs_b, runtime,
4283 raw_spin_lock(&cfs_b->lock);
4285 cfs_b->distribute_running = 0;
4286 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4288 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4292 * While we are ensured activity in the period following an
4293 * unthrottle, this also covers the case in which the new bandwidth is
4294 * insufficient to cover the existing bandwidth deficit. (Forcing the
4295 * timer to remain active while there are any throttled entities.)
4305 /* a cfs_rq won't donate quota below this amount */
4306 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4307 /* minimum remaining period time to redistribute slack quota */
4308 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4309 /* how long we wait to gather additional slack before distributing */
4310 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4313 * Are we near the end of the current quota period?
4315 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4316 * hrtimer base being cleared by hrtimer_start. In the case of
4317 * migrate_hrtimers, base is never cleared, so we are fine.
4319 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4321 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4324 /* if the call-back is running a quota refresh is already occurring */
4325 if (hrtimer_callback_running(refresh_timer))
4328 /* is a quota refresh about to occur? */
4329 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4330 if (remaining < min_expire)
4336 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4338 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4340 /* if there's a quota refresh soon don't bother with slack */
4341 if (runtime_refresh_within(cfs_b, min_left))
4344 hrtimer_start(&cfs_b->slack_timer,
4345 ns_to_ktime(cfs_bandwidth_slack_period),
4349 /* we know any runtime found here is valid as update_curr() precedes return */
4350 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4352 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4353 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4355 if (slack_runtime <= 0)
4358 raw_spin_lock(&cfs_b->lock);
4359 if (cfs_b->quota != RUNTIME_INF &&
4360 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4361 cfs_b->runtime += slack_runtime;
4363 /* we are under rq->lock, defer unthrottling using a timer */
4364 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4365 !list_empty(&cfs_b->throttled_cfs_rq))
4366 start_cfs_slack_bandwidth(cfs_b);
4368 raw_spin_unlock(&cfs_b->lock);
4370 /* even if it's not valid for return we don't want to try again */
4371 cfs_rq->runtime_remaining -= slack_runtime;
4374 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4376 if (!cfs_bandwidth_used())
4379 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4382 __return_cfs_rq_runtime(cfs_rq);
4386 * This is done with a timer (instead of inline with bandwidth return) since
4387 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4389 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4391 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4394 /* confirm we're still not at a refresh boundary */
4395 raw_spin_lock(&cfs_b->lock);
4396 if (cfs_b->distribute_running) {
4397 raw_spin_unlock(&cfs_b->lock);
4401 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4402 raw_spin_unlock(&cfs_b->lock);
4406 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4407 runtime = cfs_b->runtime;
4409 expires = cfs_b->runtime_expires;
4411 cfs_b->distribute_running = 1;
4413 raw_spin_unlock(&cfs_b->lock);
4418 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4420 raw_spin_lock(&cfs_b->lock);
4421 if (expires == cfs_b->runtime_expires)
4422 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4423 cfs_b->distribute_running = 0;
4424 raw_spin_unlock(&cfs_b->lock);
4428 * When a group wakes up we want to make sure that its quota is not already
4429 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4430 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4432 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4434 if (!cfs_bandwidth_used())
4437 /* Synchronize hierarchical throttle counter: */
4438 if (unlikely(!cfs_rq->throttle_uptodate)) {
4439 struct rq *rq = rq_of(cfs_rq);
4440 struct cfs_rq *pcfs_rq;
4441 struct task_group *tg;
4443 cfs_rq->throttle_uptodate = 1;
4445 /* Get closest up-to-date node, because leaves go first: */
4446 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4447 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4448 if (pcfs_rq->throttle_uptodate)
4452 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4453 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4457 /* an active group must be handled by the update_curr()->put() path */
4458 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4461 /* ensure the group is not already throttled */
4462 if (cfs_rq_throttled(cfs_rq))
4465 /* update runtime allocation */
4466 account_cfs_rq_runtime(cfs_rq, 0);
4467 if (cfs_rq->runtime_remaining <= 0)
4468 throttle_cfs_rq(cfs_rq);
4471 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4472 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4474 if (!cfs_bandwidth_used())
4477 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4481 * it's possible for a throttled entity to be forced into a running
4482 * state (e.g. set_curr_task), in this case we're finished.
4484 if (cfs_rq_throttled(cfs_rq))
4487 throttle_cfs_rq(cfs_rq);
4491 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4493 struct cfs_bandwidth *cfs_b =
4494 container_of(timer, struct cfs_bandwidth, slack_timer);
4496 do_sched_cfs_slack_timer(cfs_b);
4498 return HRTIMER_NORESTART;
4501 extern const u64 max_cfs_quota_period;
4503 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4505 struct cfs_bandwidth *cfs_b =
4506 container_of(timer, struct cfs_bandwidth, period_timer);
4511 raw_spin_lock(&cfs_b->lock);
4513 overrun = hrtimer_forward_now(timer, cfs_b->period);
4518 u64 new, old = ktime_to_ns(cfs_b->period);
4520 new = (old * 147) / 128; /* ~115% */
4521 new = min(new, max_cfs_quota_period);
4523 cfs_b->period = ns_to_ktime(new);
4525 /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
4526 cfs_b->quota *= new;
4527 cfs_b->quota = div64_u64(cfs_b->quota, old);
4529 pr_warn_ratelimited(
4530 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
4532 div_u64(new, NSEC_PER_USEC),
4533 div_u64(cfs_b->quota, NSEC_PER_USEC));
4535 /* reset count so we don't come right back in here */
4539 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4542 cfs_b->period_active = 0;
4543 raw_spin_unlock(&cfs_b->lock);
4545 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4548 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4550 raw_spin_lock_init(&cfs_b->lock);
4552 cfs_b->quota = RUNTIME_INF;
4553 cfs_b->period = ns_to_ktime(default_cfs_period());
4555 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4556 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4557 cfs_b->period_timer.function = sched_cfs_period_timer;
4558 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4559 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4560 cfs_b->distribute_running = 0;
4563 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4565 cfs_rq->runtime_enabled = 0;
4566 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4569 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4571 lockdep_assert_held(&cfs_b->lock);
4573 if (!cfs_b->period_active) {
4574 cfs_b->period_active = 1;
4575 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4576 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4580 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4582 /* init_cfs_bandwidth() was not called */
4583 if (!cfs_b->throttled_cfs_rq.next)
4586 hrtimer_cancel(&cfs_b->period_timer);
4587 hrtimer_cancel(&cfs_b->slack_timer);
4590 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4592 struct cfs_rq *cfs_rq;
4594 for_each_leaf_cfs_rq(rq, cfs_rq) {
4595 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4597 raw_spin_lock(&cfs_b->lock);
4598 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4599 raw_spin_unlock(&cfs_b->lock);
4603 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4605 struct cfs_rq *cfs_rq;
4607 for_each_leaf_cfs_rq(rq, cfs_rq) {
4608 if (!cfs_rq->runtime_enabled)
4612 * clock_task is not advancing so we just need to make sure
4613 * there's some valid quota amount
4615 cfs_rq->runtime_remaining = 1;
4617 * Offline rq is schedulable till cpu is completely disabled
4618 * in take_cpu_down(), so we prevent new cfs throttling here.
4620 cfs_rq->runtime_enabled = 0;
4622 if (cfs_rq_throttled(cfs_rq))
4623 unthrottle_cfs_rq(cfs_rq);
4627 #else /* CONFIG_CFS_BANDWIDTH */
4628 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4630 return rq_clock_task(rq_of(cfs_rq));
4633 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4634 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4635 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4636 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4638 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4643 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4648 static inline int throttled_lb_pair(struct task_group *tg,
4649 int src_cpu, int dest_cpu)
4654 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4656 #ifdef CONFIG_FAIR_GROUP_SCHED
4657 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4660 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4664 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4665 static inline void update_runtime_enabled(struct rq *rq) {}
4666 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4668 #endif /* CONFIG_CFS_BANDWIDTH */
4670 /**************************************************
4671 * CFS operations on tasks:
4674 #ifdef CONFIG_SCHED_HRTICK
4675 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4677 struct sched_entity *se = &p->se;
4678 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4680 WARN_ON(task_rq(p) != rq);
4682 if (cfs_rq->nr_running > 1) {
4683 u64 slice = sched_slice(cfs_rq, se);
4684 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4685 s64 delta = slice - ran;
4692 hrtick_start(rq, delta);
4697 * called from enqueue/dequeue and updates the hrtick when the
4698 * current task is from our class and nr_running is low enough
4701 static void hrtick_update(struct rq *rq)
4703 struct task_struct *curr = rq->curr;
4705 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4708 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4709 hrtick_start_fair(rq, curr);
4711 #else /* !CONFIG_SCHED_HRTICK */
4713 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4717 static inline void hrtick_update(struct rq *rq)
4723 static bool __cpu_overutilized(int cpu, int delta);
4724 static bool cpu_overutilized(int cpu);
4725 unsigned long boosted_cpu_util(int cpu);
4727 #define boosted_cpu_util(cpu) cpu_util_freq(cpu)
4731 * The enqueue_task method is called before nr_running is
4732 * increased. Here we update the fair scheduling stats and
4733 * then put the task into the rbtree:
4736 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4738 struct cfs_rq *cfs_rq;
4739 struct sched_entity *se = &p->se;
4741 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4745 * If in_iowait is set, the code below may not trigger any cpufreq
4746 * utilization updates, so do it here explicitly with the IOWAIT flag
4750 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4752 for_each_sched_entity(se) {
4755 cfs_rq = cfs_rq_of(se);
4756 enqueue_entity(cfs_rq, se, flags);
4759 * end evaluation on encountering a throttled cfs_rq
4761 * note: in the case of encountering a throttled cfs_rq we will
4762 * post the final h_nr_running increment below.
4764 if (cfs_rq_throttled(cfs_rq))
4766 cfs_rq->h_nr_running++;
4767 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4769 flags = ENQUEUE_WAKEUP;
4772 for_each_sched_entity(se) {
4773 cfs_rq = cfs_rq_of(se);
4774 cfs_rq->h_nr_running++;
4775 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4777 if (cfs_rq_throttled(cfs_rq))
4780 update_load_avg(se, UPDATE_TG);
4781 update_cfs_shares(se);
4785 add_nr_running(rq, 1);
4790 * Update SchedTune accounting.
4792 * We do it before updating the CPU capacity to ensure the
4793 * boost value of the current task is accounted for in the
4794 * selection of the OPP.
4796 * We do it also in the case where we enqueue a throttled task;
4797 * we could argue that a throttled task should not boost a CPU,
4799 * a) properly implementing CPU boosting considering throttled
4800 * tasks will increase a lot the complexity of the solution
4801 * b) it's not easy to quantify the benefits introduced by
4802 * such a more complex solution.
4803 * Thus, for the time being we go for the simple solution and boost
4804 * also for throttled RQs.
4806 schedtune_enqueue_task(p, cpu_of(rq));
4809 walt_inc_cumulative_runnable_avg(rq, p);
4810 if (!task_new && !rq->rd->overutilized &&
4811 cpu_overutilized(rq->cpu)) {
4812 rq->rd->overutilized = true;
4813 trace_sched_overutilized(true);
4817 #endif /* CONFIG_SMP */
4821 static void set_next_buddy(struct sched_entity *se);
4824 * The dequeue_task method is called before nr_running is
4825 * decreased. We remove the task from the rbtree and
4826 * update the fair scheduling stats:
4828 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4830 struct cfs_rq *cfs_rq;
4831 struct sched_entity *se = &p->se;
4832 int task_sleep = flags & DEQUEUE_SLEEP;
4834 for_each_sched_entity(se) {
4835 cfs_rq = cfs_rq_of(se);
4836 dequeue_entity(cfs_rq, se, flags);
4839 * end evaluation on encountering a throttled cfs_rq
4841 * note: in the case of encountering a throttled cfs_rq we will
4842 * post the final h_nr_running decrement below.
4844 if (cfs_rq_throttled(cfs_rq))
4846 cfs_rq->h_nr_running--;
4847 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4849 /* Don't dequeue parent if it has other entities besides us */
4850 if (cfs_rq->load.weight) {
4851 /* Avoid re-evaluating load for this entity: */
4852 se = parent_entity(se);
4854 * Bias pick_next to pick a task from this cfs_rq, as
4855 * p is sleeping when it is within its sched_slice.
4857 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4861 flags |= DEQUEUE_SLEEP;
4864 for_each_sched_entity(se) {
4865 cfs_rq = cfs_rq_of(se);
4866 cfs_rq->h_nr_running--;
4867 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4869 if (cfs_rq_throttled(cfs_rq))
4872 update_load_avg(se, UPDATE_TG);
4873 update_cfs_shares(se);
4877 sub_nr_running(rq, 1);
4882 * Update SchedTune accounting
4884 * We do it before updating the CPU capacity to ensure the
4885 * boost value of the current task is accounted for in the
4886 * selection of the OPP.
4888 schedtune_dequeue_task(p, cpu_of(rq));
4891 walt_dec_cumulative_runnable_avg(rq, p);
4892 #endif /* CONFIG_SMP */
4900 * per rq 'load' arrray crap; XXX kill this.
4904 * The exact cpuload at various idx values, calculated at every tick would be
4905 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4907 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4908 * on nth tick when cpu may be busy, then we have:
4909 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4910 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4912 * decay_load_missed() below does efficient calculation of
4913 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4914 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4916 * The calculation is approximated on a 128 point scale.
4917 * degrade_zero_ticks is the number of ticks after which load at any
4918 * particular idx is approximated to be zero.
4919 * degrade_factor is a precomputed table, a row for each load idx.
4920 * Each column corresponds to degradation factor for a power of two ticks,
4921 * based on 128 point scale.
4923 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4924 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4926 * With this power of 2 load factors, we can degrade the load n times
4927 * by looking at 1 bits in n and doing as many mult/shift instead of
4928 * n mult/shifts needed by the exact degradation.
4930 #define DEGRADE_SHIFT 7
4931 static const unsigned char
4932 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4933 static const unsigned char
4934 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4935 {0, 0, 0, 0, 0, 0, 0, 0},
4936 {64, 32, 8, 0, 0, 0, 0, 0},
4937 {96, 72, 40, 12, 1, 0, 0},
4938 {112, 98, 75, 43, 15, 1, 0},
4939 {120, 112, 98, 76, 45, 16, 2} };
4942 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4943 * would be when CPU is idle and so we just decay the old load without
4944 * adding any new load.
4946 static unsigned long
4947 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4951 if (!missed_updates)
4954 if (missed_updates >= degrade_zero_ticks[idx])
4958 return load >> missed_updates;
4960 while (missed_updates) {
4961 if (missed_updates % 2)
4962 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4964 missed_updates >>= 1;
4971 * Update rq->cpu_load[] statistics. This function is usually called every
4972 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4973 * every tick. We fix it up based on jiffies.
4975 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4976 unsigned long pending_updates)
4980 this_rq->nr_load_updates++;
4982 /* Update our load: */
4983 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4984 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4985 unsigned long old_load, new_load;
4987 /* scale is effectively 1 << i now, and >> i divides by scale */
4989 old_load = this_rq->cpu_load[i];
4990 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4991 new_load = this_load;
4993 * Round up the averaging division if load is increasing. This
4994 * prevents us from getting stuck on 9 if the load is 10, for
4997 if (new_load > old_load)
4998 new_load += scale - 1;
5000 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5003 sched_avg_update(this_rq);
5006 /* Used instead of source_load when we know the type == 0 */
5007 static unsigned long weighted_cpuload(const int cpu)
5009 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
5012 #ifdef CONFIG_NO_HZ_COMMON
5014 * There is no sane way to deal with nohz on smp when using jiffies because the
5015 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5016 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5018 * Therefore we cannot use the delta approach from the regular tick since that
5019 * would seriously skew the load calculation. However we'll make do for those
5020 * updates happening while idle (nohz_idle_balance) or coming out of idle
5021 * (tick_nohz_idle_exit).
5023 * This means we might still be one tick off for nohz periods.
5027 * Called from nohz_idle_balance() to update the load ratings before doing the
5030 static void update_idle_cpu_load(struct rq *this_rq)
5032 unsigned long curr_jiffies = READ_ONCE(jiffies);
5033 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5034 unsigned long pending_updates;
5037 * bail if there's load or we're actually up-to-date.
5039 if (load || curr_jiffies == this_rq->last_load_update_tick)
5042 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5043 this_rq->last_load_update_tick = curr_jiffies;
5045 __update_cpu_load(this_rq, load, pending_updates);
5049 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
5051 void update_cpu_load_nohz(void)
5053 struct rq *this_rq = this_rq();
5054 unsigned long curr_jiffies = READ_ONCE(jiffies);
5055 unsigned long pending_updates;
5057 if (curr_jiffies == this_rq->last_load_update_tick)
5060 raw_spin_lock(&this_rq->lock);
5061 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5062 if (pending_updates) {
5063 this_rq->last_load_update_tick = curr_jiffies;
5065 * We were idle, this means load 0, the current load might be
5066 * !0 due to remote wakeups and the sort.
5068 __update_cpu_load(this_rq, 0, pending_updates);
5070 raw_spin_unlock(&this_rq->lock);
5072 #endif /* CONFIG_NO_HZ */
5075 * Called from scheduler_tick()
5077 void update_cpu_load_active(struct rq *this_rq)
5079 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5081 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
5083 this_rq->last_load_update_tick = jiffies;
5084 __update_cpu_load(this_rq, load, 1);
5088 * Return a low guess at the load of a migration-source cpu weighted
5089 * according to the scheduling class and "nice" value.
5091 * We want to under-estimate the load of migration sources, to
5092 * balance conservatively.
5094 static unsigned long source_load(int cpu, int type)
5096 struct rq *rq = cpu_rq(cpu);
5097 unsigned long total = weighted_cpuload(cpu);
5099 if (type == 0 || !sched_feat(LB_BIAS))
5102 return min(rq->cpu_load[type-1], total);
5106 * Return a high guess at the load of a migration-target cpu weighted
5107 * according to the scheduling class and "nice" value.
5109 static unsigned long target_load(int cpu, int type)
5111 struct rq *rq = cpu_rq(cpu);
5112 unsigned long total = weighted_cpuload(cpu);
5114 if (type == 0 || !sched_feat(LB_BIAS))
5117 return max(rq->cpu_load[type-1], total);
5121 static unsigned long cpu_avg_load_per_task(int cpu)
5123 struct rq *rq = cpu_rq(cpu);
5124 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5125 unsigned long load_avg = weighted_cpuload(cpu);
5128 return load_avg / nr_running;
5133 static void record_wakee(struct task_struct *p)
5136 * Rough decay (wiping) for cost saving, don't worry
5137 * about the boundary, really active task won't care
5140 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5141 current->wakee_flips >>= 1;
5142 current->wakee_flip_decay_ts = jiffies;
5145 if (current->last_wakee != p) {
5146 current->last_wakee = p;
5147 current->wakee_flips++;
5151 static void task_waking_fair(struct task_struct *p)
5153 struct sched_entity *se = &p->se;
5154 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5157 #ifndef CONFIG_64BIT
5158 u64 min_vruntime_copy;
5161 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5163 min_vruntime = cfs_rq->min_vruntime;
5164 } while (min_vruntime != min_vruntime_copy);
5166 min_vruntime = cfs_rq->min_vruntime;
5169 se->vruntime -= min_vruntime;
5173 #ifdef CONFIG_FAIR_GROUP_SCHED
5175 * effective_load() calculates the load change as seen from the root_task_group
5177 * Adding load to a group doesn't make a group heavier, but can cause movement
5178 * of group shares between cpus. Assuming the shares were perfectly aligned one
5179 * can calculate the shift in shares.
5181 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5182 * on this @cpu and results in a total addition (subtraction) of @wg to the
5183 * total group weight.
5185 * Given a runqueue weight distribution (rw_i) we can compute a shares
5186 * distribution (s_i) using:
5188 * s_i = rw_i / \Sum rw_j (1)
5190 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5191 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5192 * shares distribution (s_i):
5194 * rw_i = { 2, 4, 1, 0 }
5195 * s_i = { 2/7, 4/7, 1/7, 0 }
5197 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5198 * task used to run on and the CPU the waker is running on), we need to
5199 * compute the effect of waking a task on either CPU and, in case of a sync
5200 * wakeup, compute the effect of the current task going to sleep.
5202 * So for a change of @wl to the local @cpu with an overall group weight change
5203 * of @wl we can compute the new shares distribution (s'_i) using:
5205 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5207 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5208 * differences in waking a task to CPU 0. The additional task changes the
5209 * weight and shares distributions like:
5211 * rw'_i = { 3, 4, 1, 0 }
5212 * s'_i = { 3/8, 4/8, 1/8, 0 }
5214 * We can then compute the difference in effective weight by using:
5216 * dw_i = S * (s'_i - s_i) (3)
5218 * Where 'S' is the group weight as seen by its parent.
5220 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5221 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5222 * 4/7) times the weight of the group.
5224 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5226 struct sched_entity *se = tg->se[cpu];
5228 if (!tg->parent) /* the trivial, non-cgroup case */
5231 for_each_sched_entity(se) {
5232 struct cfs_rq *cfs_rq = se->my_q;
5233 long W, w = cfs_rq_load_avg(cfs_rq);
5238 * W = @wg + \Sum rw_j
5240 W = wg + atomic_long_read(&tg->load_avg);
5242 /* Ensure \Sum rw_j >= rw_i */
5243 W -= cfs_rq->tg_load_avg_contrib;
5252 * wl = S * s'_i; see (2)
5255 wl = (w * (long)tg->shares) / W;
5260 * Per the above, wl is the new se->load.weight value; since
5261 * those are clipped to [MIN_SHARES, ...) do so now. See
5262 * calc_cfs_shares().
5264 if (wl < MIN_SHARES)
5268 * wl = dw_i = S * (s'_i - s_i); see (3)
5270 wl -= se->avg.load_avg;
5273 * Recursively apply this logic to all parent groups to compute
5274 * the final effective load change on the root group. Since
5275 * only the @tg group gets extra weight, all parent groups can
5276 * only redistribute existing shares. @wl is the shift in shares
5277 * resulting from this level per the above.
5286 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5294 * Returns the current capacity of cpu after applying both
5295 * cpu and freq scaling.
5297 unsigned long capacity_curr_of(int cpu)
5299 return cpu_rq(cpu)->cpu_capacity_orig *
5300 arch_scale_freq_capacity(NULL, cpu)
5301 >> SCHED_CAPACITY_SHIFT;
5304 static inline bool energy_aware(void)
5306 return sched_feat(ENERGY_AWARE);
5310 struct sched_group *sg_top;
5311 struct sched_group *sg_cap;
5319 struct task_struct *task;
5333 static int cpu_util_wake(int cpu, struct task_struct *p);
5336 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
5337 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE], which is useful for
5338 * energy calculations.
5340 * Since util is a scale-invariant utilization defined as:
5342 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
5344 * the normalized util can be found using the specific capacity.
5346 * capacity = capacity_orig * curr_freq/max_freq
5348 * norm_util = running_time/time ~ util/capacity
5350 static unsigned long __cpu_norm_util(unsigned long util, unsigned long capacity)
5352 if (util >= capacity)
5353 return SCHED_CAPACITY_SCALE;
5355 return (util << SCHED_CAPACITY_SHIFT)/capacity;
5358 static unsigned long group_max_util(struct energy_env *eenv)
5360 unsigned long max_util = 0;
5364 for_each_cpu(cpu, sched_group_cpus(eenv->sg_cap)) {
5365 util = cpu_util_wake(cpu, eenv->task);
5368 * If we are looking at the target CPU specified by the eenv,
5369 * then we should add the (estimated) utilization of the task
5370 * assuming we will wake it up on that CPU.
5372 if (unlikely(cpu == eenv->trg_cpu))
5373 util += eenv->util_delta;
5375 max_util = max(max_util, util);
5382 * group_norm_util() returns the approximated group util relative to it's
5383 * current capacity (busy ratio), in the range [0..SCHED_LOAD_SCALE], for use
5384 * in energy calculations.
5386 * Since task executions may or may not overlap in time in the group the true
5387 * normalized util is between MAX(cpu_norm_util(i)) and SUM(cpu_norm_util(i))
5388 * when iterating over all CPUs in the group.
5389 * The latter estimate is used as it leads to a more pessimistic energy
5390 * estimate (more busy).
5393 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
5395 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
5396 unsigned long util, util_sum = 0;
5399 for_each_cpu(cpu, sched_group_cpus(sg)) {
5400 util = cpu_util_wake(cpu, eenv->task);
5403 * If we are looking at the target CPU specified by the eenv,
5404 * then we should add the (estimated) utilization of the task
5405 * assuming we will wake it up on that CPU.
5407 if (unlikely(cpu == eenv->trg_cpu))
5408 util += eenv->util_delta;
5410 util_sum += __cpu_norm_util(util, capacity);
5413 return min_t(unsigned long, util_sum, SCHED_CAPACITY_SCALE);
5416 static int find_new_capacity(struct energy_env *eenv,
5417 const struct sched_group_energy * const sge)
5419 int idx, max_idx = sge->nr_cap_states - 1;
5420 unsigned long util = group_max_util(eenv);
5422 /* default is max_cap if we don't find a match */
5423 eenv->cap_idx = max_idx;
5425 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5426 if (sge->cap_states[idx].cap >= util) {
5427 eenv->cap_idx = idx;
5432 return eenv->cap_idx;
5435 static int group_idle_state(struct energy_env *eenv, struct sched_group *sg)
5437 int i, state = INT_MAX;
5438 int src_in_grp, dst_in_grp;
5441 /* Find the shallowest idle state in the sched group. */
5442 for_each_cpu(i, sched_group_cpus(sg))
5443 state = min(state, idle_get_state_idx(cpu_rq(i)));
5445 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5448 src_in_grp = cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg));
5449 dst_in_grp = cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg));
5450 if (src_in_grp == dst_in_grp) {
5451 /* both CPUs under consideration are in the same group or not in
5452 * either group, migration should leave idle state the same.
5458 * Try to estimate if a deeper idle state is
5459 * achievable when we move the task.
5461 for_each_cpu(i, sched_group_cpus(sg)) {
5462 grp_util += cpu_util_wake(i, eenv->task);
5463 if (unlikely(i == eenv->trg_cpu))
5464 grp_util += eenv->util_delta;
5468 ((long)sg->sgc->max_capacity * (int)sg->group_weight)) {
5469 /* after moving, this group is at most partly
5470 * occupied, so it should have some idle time.
5472 int max_idle_state_idx = sg->sge->nr_idle_states - 2;
5473 int new_state = grp_util * max_idle_state_idx;
5475 /* group will have no util, use lowest state */
5476 new_state = max_idle_state_idx + 1;
5478 /* for partially idle, linearly map util to idle
5479 * states, excluding the lowest one. This does not
5480 * correspond to the state we expect to enter in
5481 * reality, but an indication of what might happen.
5483 new_state = min(max_idle_state_idx, (int)
5484 (new_state / sg->sgc->max_capacity));
5485 new_state = max_idle_state_idx - new_state;
5489 /* After moving, the group will be fully occupied
5490 * so assume it will not be idle at all.
5499 * sched_group_energy(): Computes the absolute energy consumption of cpus
5500 * belonging to the sched_group including shared resources shared only by
5501 * members of the group. Iterates over all cpus in the hierarchy below the
5502 * sched_group starting from the bottom working it's way up before going to
5503 * the next cpu until all cpus are covered at all levels. The current
5504 * implementation is likely to gather the same util statistics multiple times.
5505 * This can probably be done in a faster but more complex way.
5506 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5508 static int sched_group_energy(struct energy_env *eenv)
5510 struct cpumask visit_cpus;
5511 u64 total_energy = 0;
5514 WARN_ON(!eenv->sg_top->sge);
5516 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5517 /* If a cpu is hotplugged in while we are in this function,
5518 * it does not appear in the existing visit_cpus mask
5519 * which came from the sched_group pointer of the
5520 * sched_domain pointed at by sd_ea for either the prev
5521 * or next cpu and was dereferenced in __energy_diff.
5522 * Since we will dereference sd_scs later as we iterate
5523 * through the CPUs we expect to visit, new CPUs can
5524 * be present which are not in the visit_cpus mask.
5525 * Guard this with cpu_count.
5527 cpu_count = cpumask_weight(&visit_cpus);
5529 while (!cpumask_empty(&visit_cpus)) {
5530 struct sched_group *sg_shared_cap = NULL;
5531 int cpu = cpumask_first(&visit_cpus);
5532 struct sched_domain *sd;
5535 * Is the group utilization affected by cpus outside this
5537 * This sd may have groups with cpus which were not present
5538 * when we took visit_cpus.
5540 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5542 if (sd && sd->parent)
5543 sg_shared_cap = sd->parent->groups;
5545 for_each_domain(cpu, sd) {
5546 struct sched_group *sg = sd->groups;
5548 /* Has this sched_domain already been visited? */
5549 if (sd->child && group_first_cpu(sg) != cpu)
5553 unsigned long group_util;
5554 int sg_busy_energy, sg_idle_energy;
5555 int cap_idx, idle_idx;
5557 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5558 eenv->sg_cap = sg_shared_cap;
5562 cap_idx = find_new_capacity(eenv, sg->sge);
5564 if (sg->group_weight == 1) {
5565 /* Remove capacity of src CPU (before task move) */
5566 if (eenv->trg_cpu == eenv->src_cpu &&
5567 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5568 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5569 eenv->cap.delta -= eenv->cap.before;
5571 /* Add capacity of dst CPU (after task move) */
5572 if (eenv->trg_cpu == eenv->dst_cpu &&
5573 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5574 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5575 eenv->cap.delta += eenv->cap.after;
5579 idle_idx = group_idle_state(eenv, sg);
5580 group_util = group_norm_util(eenv, sg);
5582 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power);
5583 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5584 * sg->sge->idle_states[idle_idx].power);
5586 total_energy += sg_busy_energy + sg_idle_energy;
5590 * cpu_count here is the number of
5591 * cpus we expect to visit in this
5592 * calculation. If we race against
5593 * hotplug, we can have extra cpus
5594 * added to the groups we are
5595 * iterating which do not appear in
5596 * the visit_cpus mask. In that case
5597 * we are not able to calculate energy
5598 * without restarting so we will bail
5599 * out and use prev_cpu this time.
5603 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5607 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5610 } while (sg = sg->next, sg != sd->groups);
5614 * If we raced with hotplug and got an sd NULL-pointer;
5615 * returning a wrong energy estimation is better than
5616 * entering an infinite loop.
5617 * Specifically: If a cpu is unplugged after we took
5618 * the visit_cpus mask, it no longer has an sd_scs
5619 * pointer, so when we dereference it, we get NULL.
5621 if (cpumask_test_cpu(cpu, &visit_cpus))
5624 cpumask_clear_cpu(cpu, &visit_cpus);
5628 eenv->energy = total_energy >> SCHED_CAPACITY_SHIFT;
5632 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5634 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5637 static inline unsigned long task_util(struct task_struct *p);
5640 * energy_diff(): Estimate the energy impact of changing the utilization
5641 * distribution. eenv specifies the change: utilisation amount, source, and
5642 * destination cpu. Source or destination cpu may be -1 in which case the
5643 * utilization is removed from or added to the system (e.g. task wake-up). If
5644 * both are specified, the utilization is migrated.
5646 static inline int __energy_diff(struct energy_env *eenv)
5648 struct sched_domain *sd;
5649 struct sched_group *sg;
5650 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5653 struct energy_env eenv_before = {
5654 .util_delta = task_util(eenv->task),
5655 .src_cpu = eenv->src_cpu,
5656 .dst_cpu = eenv->dst_cpu,
5657 .trg_cpu = eenv->src_cpu,
5658 .nrg = { 0, 0, 0, 0},
5663 if (eenv->src_cpu == eenv->dst_cpu)
5666 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5667 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5670 return 0; /* Error */
5675 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5676 eenv_before.sg_top = eenv->sg_top = sg;
5678 if (sched_group_energy(&eenv_before))
5679 return 0; /* Invalid result abort */
5680 energy_before += eenv_before.energy;
5682 /* Keep track of SRC cpu (before) capacity */
5683 eenv->cap.before = eenv_before.cap.before;
5684 eenv->cap.delta = eenv_before.cap.delta;
5686 if (sched_group_energy(eenv))
5687 return 0; /* Invalid result abort */
5688 energy_after += eenv->energy;
5690 } while (sg = sg->next, sg != sd->groups);
5692 eenv->nrg.before = energy_before;
5693 eenv->nrg.after = energy_after;
5694 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5696 #ifndef CONFIG_SCHED_TUNE
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 * Dead-zone margin preventing too many migrations.
5707 margin = eenv->nrg.before >> 6; /* ~1.56% */
5709 diff = eenv->nrg.after - eenv->nrg.before;
5711 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5713 return eenv->nrg.diff;
5716 #ifdef CONFIG_SCHED_TUNE
5718 struct target_nrg schedtune_target_nrg;
5720 #ifdef CONFIG_CGROUP_SCHEDTUNE
5721 extern bool schedtune_initialized;
5722 #endif /* CONFIG_CGROUP_SCHEDTUNE */
5725 * System energy normalization
5726 * Returns the normalized value, in the range [0..SCHED_CAPACITY_SCALE],
5727 * corresponding to the specified energy variation.
5730 normalize_energy(int energy_diff)
5734 #ifdef CONFIG_CGROUP_SCHEDTUNE
5735 /* during early setup, we don't know the extents */
5736 if (unlikely(!schedtune_initialized))
5737 return energy_diff < 0 ? -1 : 1 ;
5738 #endif /* CONFIG_CGROUP_SCHEDTUNE */
5740 #ifdef CONFIG_SCHED_DEBUG
5744 /* Check for boundaries */
5745 max_delta = schedtune_target_nrg.max_power;
5746 max_delta -= schedtune_target_nrg.min_power;
5747 WARN_ON(abs(energy_diff) >= max_delta);
5751 /* Do scaling using positive numbers to increase the range */
5752 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5754 /* Scale by energy magnitude */
5755 normalized_nrg <<= SCHED_CAPACITY_SHIFT;
5757 /* Normalize on max energy for target platform */
5758 normalized_nrg = reciprocal_divide(
5759 normalized_nrg, schedtune_target_nrg.rdiv);
5761 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5765 energy_diff(struct energy_env *eenv)
5767 int boost = schedtune_task_boost(eenv->task);
5770 /* Conpute "absolute" energy diff */
5771 __energy_diff(eenv);
5773 /* Return energy diff when boost margin is 0 */
5775 trace_sched_energy_diff(eenv->task,
5776 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5777 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5778 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5779 0, -eenv->nrg.diff);
5780 return eenv->nrg.diff;
5783 /* Compute normalized energy diff */
5784 nrg_delta = normalize_energy(eenv->nrg.diff);
5785 eenv->nrg.delta = nrg_delta;
5787 eenv->payoff = schedtune_accept_deltas(
5792 trace_sched_energy_diff(eenv->task,
5793 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5794 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5795 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5796 eenv->nrg.delta, eenv->payoff);
5799 * When SchedTune is enabled, the energy_diff() function will return
5800 * the computed energy payoff value. Since the energy_diff() return
5801 * value is expected to be negative by its callers, this evaluation
5802 * function return a negative value each time the evaluation return a
5803 * positive payoff, which is the condition for the acceptance of
5804 * a scheduling decision
5806 return -eenv->payoff;
5808 #else /* CONFIG_SCHED_TUNE */
5809 #define energy_diff(eenv) __energy_diff(eenv)
5813 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5814 * A waker of many should wake a different task than the one last awakened
5815 * at a frequency roughly N times higher than one of its wakees. In order
5816 * to determine whether we should let the load spread vs consolodating to
5817 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5818 * partner, and a factor of lls_size higher frequency in the other. With
5819 * both conditions met, we can be relatively sure that the relationship is
5820 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5821 * being client/server, worker/dispatcher, interrupt source or whatever is
5822 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5824 static int wake_wide(struct task_struct *p, int sibling_count_hint)
5826 unsigned int master = current->wakee_flips;
5827 unsigned int slave = p->wakee_flips;
5828 int llc_size = this_cpu_read(sd_llc_size);
5830 if (sibling_count_hint >= llc_size)
5834 swap(master, slave);
5835 if (slave < llc_size || master < slave * llc_size)
5840 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5841 int prev_cpu, int sync)
5843 s64 this_load, load;
5844 s64 this_eff_load, prev_eff_load;
5846 struct task_group *tg;
5847 unsigned long weight;
5851 this_cpu = smp_processor_id();
5852 load = source_load(prev_cpu, idx);
5853 this_load = target_load(this_cpu, idx);
5856 * If sync wakeup then subtract the (maximum possible)
5857 * effect of the currently running task from the load
5858 * of the current CPU:
5861 tg = task_group(current);
5862 weight = current->se.avg.load_avg;
5864 this_load += effective_load(tg, this_cpu, -weight, -weight);
5865 load += effective_load(tg, prev_cpu, 0, -weight);
5869 weight = p->se.avg.load_avg;
5872 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5873 * due to the sync cause above having dropped this_load to 0, we'll
5874 * always have an imbalance, but there's really nothing you can do
5875 * about that, so that's good too.
5877 * Otherwise check if either cpus are near enough in load to allow this
5878 * task to be woken on this_cpu.
5880 this_eff_load = 100;
5881 this_eff_load *= capacity_of(prev_cpu);
5883 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5884 prev_eff_load *= capacity_of(this_cpu);
5886 if (this_load > 0) {
5887 this_eff_load *= this_load +
5888 effective_load(tg, this_cpu, weight, weight);
5890 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5893 balanced = this_eff_load <= prev_eff_load;
5895 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5900 schedstat_inc(sd, ttwu_move_affine);
5901 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5906 static inline unsigned long task_util(struct task_struct *p)
5908 #ifdef CONFIG_SCHED_WALT
5909 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5910 unsigned long demand = p->ravg.demand;
5911 return (demand << 10) / walt_ravg_window;
5914 return p->se.avg.util_avg;
5917 static inline unsigned long boosted_task_util(struct task_struct *task);
5919 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5921 unsigned long capacity = capacity_of(cpu);
5923 util += boosted_task_util(p);
5925 return (capacity * 1024) > (util * capacity_margin);
5928 static inline bool task_fits_max(struct task_struct *p, int cpu)
5930 unsigned long capacity = capacity_of(cpu);
5931 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5933 if (capacity == max_capacity)
5936 if (capacity * capacity_margin > max_capacity * 1024)
5939 return __task_fits(p, cpu, 0);
5942 static bool __cpu_overutilized(int cpu, int delta)
5944 return (capacity_of(cpu) * 1024) < ((cpu_util(cpu) + delta) * capacity_margin);
5947 static bool cpu_overutilized(int cpu)
5949 return __cpu_overutilized(cpu, 0);
5952 #ifdef CONFIG_SCHED_TUNE
5954 struct reciprocal_value schedtune_spc_rdiv;
5957 schedtune_margin(unsigned long signal, long boost)
5959 long long margin = 0;
5962 * Signal proportional compensation (SPC)
5964 * The Boost (B) value is used to compute a Margin (M) which is
5965 * proportional to the complement of the original Signal (S):
5966 * M = B * (SCHED_CAPACITY_SCALE - S)
5967 * The obtained M could be used by the caller to "boost" S.
5970 margin = SCHED_CAPACITY_SCALE - signal;
5973 margin = -signal * boost;
5975 margin = reciprocal_divide(margin, schedtune_spc_rdiv);
5983 schedtune_cpu_margin(unsigned long util, int cpu)
5985 int boost = schedtune_cpu_boost(cpu);
5990 return schedtune_margin(util, boost);
5994 schedtune_task_margin(struct task_struct *task)
5996 int boost = schedtune_task_boost(task);
6003 util = task_util(task);
6004 margin = schedtune_margin(util, boost);
6009 #else /* CONFIG_SCHED_TUNE */
6012 schedtune_cpu_margin(unsigned long util, int cpu)
6018 schedtune_task_margin(struct task_struct *task)
6023 #endif /* CONFIG_SCHED_TUNE */
6026 boosted_cpu_util(int cpu)
6028 unsigned long util = cpu_util_freq(cpu);
6029 long margin = schedtune_cpu_margin(util, cpu);
6031 trace_sched_boost_cpu(cpu, util, margin);
6033 return util + margin;
6036 static inline unsigned long
6037 boosted_task_util(struct task_struct *task)
6039 unsigned long util = task_util(task);
6040 long margin = schedtune_task_margin(task);
6042 trace_sched_boost_task(task, util, margin);
6044 return util + margin;
6047 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
6049 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
6053 * find_idlest_group finds and returns the least busy CPU group within the
6056 * Assumes p is allowed on at least one CPU in sd.
6058 static struct sched_group *
6059 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
6060 int this_cpu, int sd_flag)
6062 struct sched_group *idlest = NULL, *group = sd->groups;
6063 struct sched_group *most_spare_sg = NULL;
6064 unsigned long min_load = ULONG_MAX, this_load = ULONG_MAX;
6065 unsigned long most_spare = 0, this_spare = 0;
6066 int load_idx = sd->forkexec_idx;
6067 int imbalance = 100 + (sd->imbalance_pct-100)/2;
6069 if (sd_flag & SD_BALANCE_WAKE)
6070 load_idx = sd->wake_idx;
6073 unsigned long load, avg_load, spare_cap, max_spare_cap;
6077 /* Skip over this group if it has no CPUs allowed */
6078 if (!cpumask_intersects(sched_group_cpus(group),
6079 tsk_cpus_allowed(p)))
6082 local_group = cpumask_test_cpu(this_cpu,
6083 sched_group_cpus(group));
6086 * Tally up the load of all CPUs in the group and find
6087 * the group containing the CPU with most spare capacity.
6092 for_each_cpu(i, sched_group_cpus(group)) {
6093 /* Bias balancing toward cpus of our domain */
6095 load = source_load(i, load_idx);
6097 load = target_load(i, load_idx);
6101 spare_cap = capacity_spare_wake(i, p);
6103 if (spare_cap > max_spare_cap)
6104 max_spare_cap = spare_cap;
6107 /* Adjust by relative CPU capacity of the group */
6108 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
6111 this_load = avg_load;
6112 this_spare = max_spare_cap;
6114 if (avg_load < min_load) {
6115 min_load = avg_load;
6119 if (most_spare < max_spare_cap) {
6120 most_spare = max_spare_cap;
6121 most_spare_sg = group;
6124 } while (group = group->next, group != sd->groups);
6127 * The cross-over point between using spare capacity or least load
6128 * is too conservative for high utilization tasks on partially
6129 * utilized systems if we require spare_capacity > task_util(p),
6130 * so we allow for some task stuffing by using
6131 * spare_capacity > task_util(p)/2.
6133 * Spare capacity can't be used for fork because the utilization has
6134 * not been set yet, we must first select a rq to compute the initial
6137 if (sd_flag & SD_BALANCE_FORK)
6140 if (this_spare > task_util(p) / 2 &&
6141 imbalance*this_spare > 100*most_spare)
6143 else if (most_spare > task_util(p) / 2)
6144 return most_spare_sg;
6147 if (!idlest || 100*this_load < imbalance*min_load)
6153 * find_idlest_group_cpu - find the idlest cpu among the cpus in group.
6156 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6158 unsigned long load, min_load = ULONG_MAX;
6159 unsigned int min_exit_latency = UINT_MAX;
6160 u64 latest_idle_timestamp = 0;
6161 int least_loaded_cpu = this_cpu;
6162 int shallowest_idle_cpu = -1;
6165 /* Check if we have any choice: */
6166 if (group->group_weight == 1)
6167 return cpumask_first(sched_group_cpus(group));
6169 /* Traverse only the allowed CPUs */
6170 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
6172 struct rq *rq = cpu_rq(i);
6173 struct cpuidle_state *idle = idle_get_state(rq);
6174 if (idle && idle->exit_latency < min_exit_latency) {
6176 * We give priority to a CPU whose idle state
6177 * has the smallest exit latency irrespective
6178 * of any idle timestamp.
6180 min_exit_latency = idle->exit_latency;
6181 latest_idle_timestamp = rq->idle_stamp;
6182 shallowest_idle_cpu = i;
6183 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6184 rq->idle_stamp > latest_idle_timestamp) {
6186 * If equal or no active idle state, then
6187 * the most recently idled CPU might have
6190 latest_idle_timestamp = rq->idle_stamp;
6191 shallowest_idle_cpu = i;
6193 } else if (shallowest_idle_cpu == -1) {
6194 load = weighted_cpuload(i);
6195 if (load < min_load || (load == min_load && i == this_cpu)) {
6197 least_loaded_cpu = i;
6202 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6205 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6206 int cpu, int prev_cpu, int sd_flag)
6209 int wu = sd_flag & SD_BALANCE_WAKE;
6213 schedstat_inc(p, se.statistics.nr_wakeups_cas_attempts);
6214 schedstat_inc(this_rq(), eas_stats.cas_attempts);
6217 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6221 struct sched_group *group;
6222 struct sched_domain *tmp;
6226 schedstat_inc(sd, eas_stats.cas_attempts);
6228 if (!(sd->flags & sd_flag)) {
6233 group = find_idlest_group(sd, p, cpu, sd_flag);
6239 new_cpu = find_idlest_group_cpu(group, p, cpu);
6240 if (new_cpu == cpu) {
6241 /* Now try balancing at a lower domain level of cpu */
6246 /* Now try balancing at a lower domain level of new_cpu */
6247 cpu = cas_cpu = new_cpu;
6248 weight = sd->span_weight;
6250 for_each_domain(cpu, tmp) {
6251 if (weight <= tmp->span_weight)
6253 if (tmp->flags & sd_flag)
6256 /* while loop will break here if sd == NULL */
6259 if (wu && (cas_cpu >= 0)) {
6260 schedstat_inc(p, se.statistics.nr_wakeups_cas_count);
6261 schedstat_inc(this_rq(), eas_stats.cas_count);
6268 * Try and locate an idle CPU in the sched_domain.
6270 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6272 struct sched_domain *sd;
6273 struct sched_group *sg;
6274 int best_idle_cpu = -1;
6275 int best_idle_cstate = INT_MAX;
6276 unsigned long best_idle_capacity = ULONG_MAX;
6278 schedstat_inc(p, se.statistics.nr_wakeups_sis_attempts);
6279 schedstat_inc(this_rq(), eas_stats.sis_attempts);
6281 if (!sysctl_sched_cstate_aware) {
6282 if (idle_cpu(target)) {
6283 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle);
6284 schedstat_inc(this_rq(), eas_stats.sis_idle);
6289 * If the prevous cpu is cache affine and idle, don't be stupid.
6291 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) {
6292 schedstat_inc(p, se.statistics.nr_wakeups_sis_cache_affine);
6293 schedstat_inc(this_rq(), eas_stats.sis_cache_affine);
6299 * Otherwise, iterate the domains and find an elegible idle cpu.
6301 sd = rcu_dereference(per_cpu(sd_llc, target));
6302 for_each_lower_domain(sd) {
6306 if (!cpumask_intersects(sched_group_cpus(sg),
6307 tsk_cpus_allowed(p)))
6310 if (sysctl_sched_cstate_aware) {
6311 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6312 int idle_idx = idle_get_state_idx(cpu_rq(i));
6313 unsigned long new_usage = boosted_task_util(p);
6314 unsigned long capacity_orig = capacity_orig_of(i);
6316 if (new_usage > capacity_orig || !idle_cpu(i))
6319 if (i == target && new_usage <= capacity_curr_of(target)) {
6320 schedstat_inc(p, se.statistics.nr_wakeups_sis_suff_cap);
6321 schedstat_inc(this_rq(), eas_stats.sis_suff_cap);
6322 schedstat_inc(sd, eas_stats.sis_suff_cap);
6326 if (idle_idx < best_idle_cstate &&
6327 capacity_orig <= best_idle_capacity) {
6329 best_idle_cstate = idle_idx;
6330 best_idle_capacity = capacity_orig;
6334 for_each_cpu(i, sched_group_cpus(sg)) {
6335 if (i == target || !idle_cpu(i))
6339 target = cpumask_first_and(sched_group_cpus(sg),
6340 tsk_cpus_allowed(p));
6341 schedstat_inc(p, se.statistics.nr_wakeups_sis_idle_cpu);
6342 schedstat_inc(this_rq(), eas_stats.sis_idle_cpu);
6343 schedstat_inc(sd, eas_stats.sis_idle_cpu);
6348 } while (sg != sd->groups);
6351 if (best_idle_cpu >= 0)
6352 target = best_idle_cpu;
6355 schedstat_inc(p, se.statistics.nr_wakeups_sis_count);
6356 schedstat_inc(this_rq(), eas_stats.sis_count);
6362 * cpu_util_wake: Compute cpu utilization with any contributions from
6363 * the waking task p removed. check_for_migration() looks for a better CPU of
6364 * rq->curr. For that case we should return cpu util with contributions from
6365 * currently running task p removed.
6367 static int cpu_util_wake(int cpu, struct task_struct *p)
6369 unsigned long util, capacity;
6371 #ifdef CONFIG_SCHED_WALT
6373 * WALT does not decay idle tasks in the same manner
6374 * as PELT, so it makes little sense to subtract task
6375 * utilization from cpu utilization. Instead just use
6376 * cpu_util for this case.
6378 if (!walt_disabled && sysctl_sched_use_walt_cpu_util &&
6379 p->state == TASK_WAKING)
6380 return cpu_util(cpu);
6382 /* Task has no contribution or is new */
6383 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6384 return cpu_util(cpu);
6386 capacity = capacity_orig_of(cpu);
6387 util = max_t(long, cpu_util(cpu) - task_util(p), 0);
6389 return (util >= capacity) ? capacity : util;
6392 static int start_cpu(bool boosted)
6394 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6396 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
6399 static inline int find_best_target(struct task_struct *p, int *backup_cpu,
6400 bool boosted, bool prefer_idle)
6402 unsigned long best_idle_min_cap_orig = ULONG_MAX;
6403 unsigned long min_util = boosted_task_util(p);
6404 unsigned long target_capacity = ULONG_MAX;
6405 unsigned long min_wake_util = ULONG_MAX;
6406 unsigned long target_max_spare_cap = 0;
6407 unsigned long best_active_util = ULONG_MAX;
6408 int best_idle_cstate = INT_MAX;
6409 struct sched_domain *sd;
6410 struct sched_group *sg;
6411 int best_active_cpu = -1;
6412 int best_idle_cpu = -1;
6413 int target_cpu = -1;
6418 schedstat_inc(p, se.statistics.nr_wakeups_fbt_attempts);
6419 schedstat_inc(this_rq(), eas_stats.fbt_attempts);
6421 /* Find start CPU based on boost value */
6422 cpu = start_cpu(boosted);
6424 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_cpu);
6425 schedstat_inc(this_rq(), eas_stats.fbt_no_cpu);
6429 /* Find SD for the start CPU */
6430 sd = rcu_dereference(per_cpu(sd_ea, cpu));
6432 schedstat_inc(p, se.statistics.nr_wakeups_fbt_no_sd);
6433 schedstat_inc(this_rq(), eas_stats.fbt_no_sd);
6437 /* Scan CPUs in all SDs */
6440 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
6441 unsigned long capacity_curr = capacity_curr_of(i);
6442 unsigned long capacity_orig = capacity_orig_of(i);
6443 unsigned long wake_util, new_util;
6448 if (walt_cpu_high_irqload(i))
6452 * p's blocked utilization is still accounted for on prev_cpu
6453 * so prev_cpu will receive a negative bias due to the double
6454 * accounting. However, the blocked utilization may be zero.
6456 wake_util = cpu_util_wake(i, p);
6457 new_util = wake_util + task_util(p);
6460 * Ensure minimum capacity to grant the required boost.
6461 * The target CPU can be already at a capacity level higher
6462 * than the one required to boost the task.
6464 new_util = max(min_util, new_util);
6465 if (new_util > capacity_orig)
6469 * Case A) Latency sensitive tasks
6471 * Unconditionally favoring tasks that prefer idle CPU to
6475 * - an idle CPU, whatever its idle_state is, since
6476 * the first CPUs we explore are more likely to be
6477 * reserved for latency sensitive tasks.
6478 * - a non idle CPU where the task fits in its current
6479 * capacity and has the maximum spare capacity.
6480 * - a non idle CPU with lower contention from other
6481 * tasks and running at the lowest possible OPP.
6483 * The last two goals tries to favor a non idle CPU
6484 * where the task can run as if it is "almost alone".
6485 * A maximum spare capacity CPU is favoured since
6486 * the task already fits into that CPU's capacity
6487 * without waiting for an OPP chance.
6489 * The following code path is the only one in the CPUs
6490 * exploration loop which is always used by
6491 * prefer_idle tasks. It exits the loop with wither a
6492 * best_active_cpu or a target_cpu which should
6493 * represent an optimal choice for latency sensitive
6499 * Case A.1: IDLE CPU
6500 * Return the first IDLE CPU we find.
6503 schedstat_inc(p, se.statistics.nr_wakeups_fbt_pref_idle);
6504 schedstat_inc(this_rq(), eas_stats.fbt_pref_idle);
6506 trace_sched_find_best_target(p,
6507 prefer_idle, min_util,
6509 best_active_cpu, i);
6515 * Case A.2: Target ACTIVE CPU
6516 * Favor CPUs with max spare capacity.
6518 if ((capacity_curr > new_util) &&
6519 (capacity_orig - new_util > target_max_spare_cap)) {
6520 target_max_spare_cap = capacity_orig - new_util;
6524 if (target_cpu != -1)
6529 * Case A.3: Backup ACTIVE CPU
6531 * - lower utilization due to other tasks
6532 * - lower utilization with the task in
6534 if (wake_util > min_wake_util)
6536 if (new_util > best_active_util)
6538 min_wake_util = wake_util;
6539 best_active_util = new_util;
6540 best_active_cpu = i;
6547 * For non latency sensitive tasks, skip CPUs that
6548 * will be overutilized by moving the task there.
6550 * The goal here is to remain in EAS mode as long as
6551 * possible at least for !prefer_idle tasks.
6553 if ((new_util * capacity_margin) >
6554 (capacity_orig * SCHED_CAPACITY_SCALE))
6558 * Case B) Non latency sensitive tasks on IDLE CPUs.
6560 * Find an optimal backup IDLE CPU for non latency
6564 * - minimizing the capacity_orig,
6565 * i.e. preferring LITTLE CPUs
6566 * - favoring shallowest idle states
6567 * i.e. avoid to wakeup deep-idle CPUs
6569 * The following code path is used by non latency
6570 * sensitive tasks if IDLE CPUs are available. If at
6571 * least one of such CPUs are available it sets the
6572 * best_idle_cpu to the most suitable idle CPU to be
6575 * If idle CPUs are available, favour these CPUs to
6576 * improve performances by spreading tasks.
6577 * Indeed, the energy_diff() computed by the caller
6578 * will take care to ensure the minimization of energy
6579 * consumptions without affecting performance.
6582 int idle_idx = idle_get_state_idx(cpu_rq(i));
6584 /* Select idle CPU with lower cap_orig */
6585 if (capacity_orig > best_idle_min_cap_orig)
6589 * Skip CPUs in deeper idle state, but only
6590 * if they are also less energy efficient.
6591 * IOW, prefer a deep IDLE LITTLE CPU vs a
6592 * shallow idle big CPU.
6594 if (sysctl_sched_cstate_aware &&
6595 best_idle_cstate <= idle_idx)
6598 /* Keep track of best idle CPU */
6599 best_idle_min_cap_orig = capacity_orig;
6600 best_idle_cstate = idle_idx;
6606 * Case C) Non latency sensitive tasks on ACTIVE CPUs.
6608 * Pack tasks in the most energy efficient capacities.
6610 * This task packing strategy prefers more energy
6611 * efficient CPUs (i.e. pack on smaller maximum
6612 * capacity CPUs) while also trying to spread tasks to
6613 * run them all at the lower OPP.
6615 * This assumes for example that it's more energy
6616 * efficient to run two tasks on two CPUs at a lower
6617 * OPP than packing both on a single CPU but running
6618 * that CPU at an higher OPP.
6620 * Thus, this case keep track of the CPU with the
6621 * smallest maximum capacity and highest spare maximum
6625 /* Favor CPUs with smaller capacity */
6626 if (capacity_orig > target_capacity)
6629 /* Favor CPUs with maximum spare capacity */
6630 if ((capacity_orig - new_util) < target_max_spare_cap)
6633 target_max_spare_cap = capacity_orig - new_util;
6634 target_capacity = capacity_orig;
6638 } while (sg = sg->next, sg != sd->groups);
6641 * For non latency sensitive tasks, cases B and C in the previous loop,
6642 * we pick the best IDLE CPU only if we was not able to find a target
6645 * Policies priorities:
6647 * - prefer_idle tasks:
6649 * a) IDLE CPU available, we return immediately
6650 * b) ACTIVE CPU where task fits and has the bigger maximum spare
6651 * capacity (i.e. target_cpu)
6652 * c) ACTIVE CPU with less contention due to other tasks
6653 * (i.e. best_active_cpu)
6655 * - NON prefer_idle tasks:
6657 * a) ACTIVE CPU: target_cpu
6658 * b) IDLE CPU: best_idle_cpu
6660 if (target_cpu == -1)
6661 target_cpu = prefer_idle
6665 *backup_cpu = prefer_idle
6669 trace_sched_find_best_target(p, prefer_idle, min_util, cpu,
6670 best_idle_cpu, best_active_cpu,
6673 schedstat_inc(p, se.statistics.nr_wakeups_fbt_count);
6674 schedstat_inc(this_rq(), eas_stats.fbt_count);
6680 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6681 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6683 * In that case WAKE_AFFINE doesn't make sense and we'll let
6684 * BALANCE_WAKE sort things out.
6686 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6688 long min_cap, max_cap;
6690 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6691 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
6693 /* Minimum capacity is close to max, no need to abort wake_affine */
6694 if (max_cap - min_cap < max_cap >> 3)
6697 /* Bring task utilization in sync with prev_cpu */
6698 sync_entity_load_avg(&p->se);
6700 return min_cap * 1024 < task_util(p) * capacity_margin;
6703 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
6705 struct sched_domain *sd;
6706 int target_cpu = prev_cpu, tmp_target, tmp_backup;
6707 bool boosted, prefer_idle;
6709 schedstat_inc(p, se.statistics.nr_wakeups_secb_attempts);
6710 schedstat_inc(this_rq(), eas_stats.secb_attempts);
6712 if (sysctl_sched_sync_hint_enable && sync) {
6713 int cpu = smp_processor_id();
6715 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6716 schedstat_inc(p, se.statistics.nr_wakeups_secb_sync);
6717 schedstat_inc(this_rq(), eas_stats.secb_sync);
6723 #ifdef CONFIG_CGROUP_SCHEDTUNE
6724 boosted = schedtune_task_boost(p) > 0;
6725 prefer_idle = schedtune_prefer_idle(p) > 0;
6727 boosted = get_sysctl_sched_cfs_boost() > 0;
6731 sync_entity_load_avg(&p->se);
6733 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
6734 /* Find a cpu with sufficient capacity */
6735 tmp_target = find_best_target(p, &tmp_backup, boosted, prefer_idle);
6739 if (tmp_target >= 0) {
6740 target_cpu = tmp_target;
6741 if ((boosted || prefer_idle) && idle_cpu(target_cpu)) {
6742 schedstat_inc(p, se.statistics.nr_wakeups_secb_idle_bt);
6743 schedstat_inc(this_rq(), eas_stats.secb_idle_bt);
6748 if (target_cpu != prev_cpu) {
6750 struct energy_env eenv = {
6751 .util_delta = task_util(p),
6752 .src_cpu = prev_cpu,
6753 .dst_cpu = target_cpu,
6755 .trg_cpu = target_cpu,
6759 #ifdef CONFIG_SCHED_WALT
6760 if (!walt_disabled && sysctl_sched_use_walt_cpu_util &&
6761 p->state == TASK_WAKING)
6762 delta = task_util(p);
6764 /* Not enough spare capacity on previous cpu */
6765 if (__cpu_overutilized(prev_cpu, delta)) {
6766 schedstat_inc(p, se.statistics.nr_wakeups_secb_insuff_cap);
6767 schedstat_inc(this_rq(), eas_stats.secb_insuff_cap);
6771 if (energy_diff(&eenv) >= 0) {
6772 /* No energy saving for target_cpu, try backup */
6773 target_cpu = tmp_backup;
6774 eenv.dst_cpu = target_cpu;
6775 eenv.trg_cpu = target_cpu;
6776 if (tmp_backup < 0 ||
6777 tmp_backup == prev_cpu ||
6778 energy_diff(&eenv) >= 0) {
6779 schedstat_inc(p, se.statistics.nr_wakeups_secb_no_nrg_sav);
6780 schedstat_inc(this_rq(), eas_stats.secb_no_nrg_sav);
6781 target_cpu = prev_cpu;
6786 schedstat_inc(p, se.statistics.nr_wakeups_secb_nrg_sav);
6787 schedstat_inc(this_rq(), eas_stats.secb_nrg_sav);
6791 schedstat_inc(p, se.statistics.nr_wakeups_secb_count);
6792 schedstat_inc(this_rq(), eas_stats.secb_count);
6801 * select_task_rq_fair: Select target runqueue for the waking task in domains
6802 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6803 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6805 * Balances load by selecting the idlest cpu in the idlest group, or under
6806 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6808 * Returns the target cpu number.
6810 * preempt must be disabled.
6813 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags,
6814 int sibling_count_hint)
6816 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6817 int cpu = smp_processor_id();
6818 int new_cpu = prev_cpu;
6819 int want_affine = 0;
6820 int sync = wake_flags & WF_SYNC;
6822 if (sd_flag & SD_BALANCE_WAKE) {
6824 want_affine = !wake_wide(p, sibling_count_hint) &&
6825 !wake_cap(p, cpu, prev_cpu) &&
6826 cpumask_test_cpu(cpu, &p->cpus_allowed);
6829 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
6830 return select_energy_cpu_brute(p, prev_cpu, sync);
6833 for_each_domain(cpu, tmp) {
6834 if (!(tmp->flags & SD_LOAD_BALANCE))
6838 * If both cpu and prev_cpu are part of this domain,
6839 * cpu is a valid SD_WAKE_AFFINE target.
6841 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6842 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6847 if (tmp->flags & sd_flag)
6849 else if (!want_affine)
6854 sd = NULL; /* Prefer wake_affine over balance flags */
6855 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6859 if (sd && !(sd_flag & SD_BALANCE_FORK)) {
6861 * We're going to need the task's util for capacity_spare_wake
6862 * in find_idlest_group. Sync it up to prev_cpu's
6865 sync_entity_load_avg(&p->se);
6869 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6870 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6873 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6881 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6882 * cfs_rq_of(p) references at time of call are still valid and identify the
6883 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6884 * other assumptions, including the state of rq->lock, should be made.
6886 static void migrate_task_rq_fair(struct task_struct *p)
6889 * We are supposed to update the task to "current" time, then its up to date
6890 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6891 * what current time is, so simply throw away the out-of-date time. This
6892 * will result in the wakee task is less decayed, but giving the wakee more
6893 * load sounds not bad.
6895 remove_entity_load_avg(&p->se);
6897 /* Tell new CPU we are migrated */
6898 p->se.avg.last_update_time = 0;
6900 /* We have migrated, no longer consider this task hot */
6901 p->se.exec_start = 0;
6904 static void task_dead_fair(struct task_struct *p)
6906 remove_entity_load_avg(&p->se);
6909 #define task_fits_max(p, cpu) true
6910 #endif /* CONFIG_SMP */
6912 static unsigned long
6913 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6915 unsigned long gran = sysctl_sched_wakeup_granularity;
6918 * Since its curr running now, convert the gran from real-time
6919 * to virtual-time in his units.
6921 * By using 'se' instead of 'curr' we penalize light tasks, so
6922 * they get preempted easier. That is, if 'se' < 'curr' then
6923 * the resulting gran will be larger, therefore penalizing the
6924 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6925 * be smaller, again penalizing the lighter task.
6927 * This is especially important for buddies when the leftmost
6928 * task is higher priority than the buddy.
6930 return calc_delta_fair(gran, se);
6934 * Should 'se' preempt 'curr'.
6948 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6950 s64 gran, vdiff = curr->vruntime - se->vruntime;
6955 gran = wakeup_gran(curr, se);
6962 static void set_last_buddy(struct sched_entity *se)
6964 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6967 for_each_sched_entity(se)
6968 cfs_rq_of(se)->last = se;
6971 static void set_next_buddy(struct sched_entity *se)
6973 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6976 for_each_sched_entity(se)
6977 cfs_rq_of(se)->next = se;
6980 static void set_skip_buddy(struct sched_entity *se)
6982 for_each_sched_entity(se)
6983 cfs_rq_of(se)->skip = se;
6987 * Preempt the current task with a newly woken task if needed:
6989 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6991 struct task_struct *curr = rq->curr;
6992 struct sched_entity *se = &curr->se, *pse = &p->se;
6993 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6994 int scale = cfs_rq->nr_running >= sched_nr_latency;
6995 int next_buddy_marked = 0;
6997 if (unlikely(se == pse))
7001 * This is possible from callers such as attach_tasks(), in which we
7002 * unconditionally check_prempt_curr() after an enqueue (which may have
7003 * lead to a throttle). This both saves work and prevents false
7004 * next-buddy nomination below.
7006 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7009 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7010 set_next_buddy(pse);
7011 next_buddy_marked = 1;
7015 * We can come here with TIF_NEED_RESCHED already set from new task
7018 * Note: this also catches the edge-case of curr being in a throttled
7019 * group (e.g. via set_curr_task), since update_curr() (in the
7020 * enqueue of curr) will have resulted in resched being set. This
7021 * prevents us from potentially nominating it as a false LAST_BUDDY
7024 if (test_tsk_need_resched(curr))
7027 /* Idle tasks are by definition preempted by non-idle tasks. */
7028 if (unlikely(curr->policy == SCHED_IDLE) &&
7029 likely(p->policy != SCHED_IDLE))
7033 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7034 * is driven by the tick):
7036 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7039 find_matching_se(&se, &pse);
7040 update_curr(cfs_rq_of(se));
7042 if (wakeup_preempt_entity(se, pse) == 1) {
7044 * Bias pick_next to pick the sched entity that is
7045 * triggering this preemption.
7047 if (!next_buddy_marked)
7048 set_next_buddy(pse);
7057 * Only set the backward buddy when the current task is still
7058 * on the rq. This can happen when a wakeup gets interleaved
7059 * with schedule on the ->pre_schedule() or idle_balance()
7060 * point, either of which can * drop the rq lock.
7062 * Also, during early boot the idle thread is in the fair class,
7063 * for obvious reasons its a bad idea to schedule back to it.
7065 if (unlikely(!se->on_rq || curr == rq->idle))
7068 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7072 static struct task_struct *
7073 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
7075 struct cfs_rq *cfs_rq = &rq->cfs;
7076 struct sched_entity *se;
7077 struct task_struct *p;
7081 #ifdef CONFIG_FAIR_GROUP_SCHED
7082 if (!cfs_rq->nr_running)
7085 if (prev->sched_class != &fair_sched_class)
7089 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7090 * likely that a next task is from the same cgroup as the current.
7092 * Therefore attempt to avoid putting and setting the entire cgroup
7093 * hierarchy, only change the part that actually changes.
7097 struct sched_entity *curr = cfs_rq->curr;
7100 * Since we got here without doing put_prev_entity() we also
7101 * have to consider cfs_rq->curr. If it is still a runnable
7102 * entity, update_curr() will update its vruntime, otherwise
7103 * forget we've ever seen it.
7107 update_curr(cfs_rq);
7112 * This call to check_cfs_rq_runtime() will do the
7113 * throttle and dequeue its entity in the parent(s).
7114 * Therefore the 'simple' nr_running test will indeed
7117 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7121 se = pick_next_entity(cfs_rq, curr);
7122 cfs_rq = group_cfs_rq(se);
7128 * Since we haven't yet done put_prev_entity and if the selected task
7129 * is a different task than we started out with, try and touch the
7130 * least amount of cfs_rqs.
7133 struct sched_entity *pse = &prev->se;
7135 while (!(cfs_rq = is_same_group(se, pse))) {
7136 int se_depth = se->depth;
7137 int pse_depth = pse->depth;
7139 if (se_depth <= pse_depth) {
7140 put_prev_entity(cfs_rq_of(pse), pse);
7141 pse = parent_entity(pse);
7143 if (se_depth >= pse_depth) {
7144 set_next_entity(cfs_rq_of(se), se);
7145 se = parent_entity(se);
7149 put_prev_entity(cfs_rq, pse);
7150 set_next_entity(cfs_rq, se);
7153 if (hrtick_enabled(rq))
7154 hrtick_start_fair(rq, p);
7156 rq->misfit_task = !task_fits_max(p, rq->cpu);
7163 if (!cfs_rq->nr_running)
7166 put_prev_task(rq, prev);
7169 se = pick_next_entity(cfs_rq, NULL);
7170 set_next_entity(cfs_rq, se);
7171 cfs_rq = group_cfs_rq(se);
7176 if (hrtick_enabled(rq))
7177 hrtick_start_fair(rq, p);
7179 rq->misfit_task = !task_fits_max(p, rq->cpu);
7184 rq->misfit_task = 0;
7186 * This is OK, because current is on_cpu, which avoids it being picked
7187 * for load-balance and preemption/IRQs are still disabled avoiding
7188 * further scheduler activity on it and we're being very careful to
7189 * re-start the picking loop.
7191 lockdep_unpin_lock(&rq->lock);
7192 new_tasks = idle_balance(rq);
7193 lockdep_pin_lock(&rq->lock);
7195 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7196 * possible for any higher priority task to appear. In that case we
7197 * must re-start the pick_next_entity() loop.
7209 * Account for a descheduled task:
7211 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7213 struct sched_entity *se = &prev->se;
7214 struct cfs_rq *cfs_rq;
7216 for_each_sched_entity(se) {
7217 cfs_rq = cfs_rq_of(se);
7218 put_prev_entity(cfs_rq, se);
7223 * sched_yield() is very simple
7225 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7227 static void yield_task_fair(struct rq *rq)
7229 struct task_struct *curr = rq->curr;
7230 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7231 struct sched_entity *se = &curr->se;
7234 * Are we the only task in the tree?
7236 if (unlikely(rq->nr_running == 1))
7239 clear_buddies(cfs_rq, se);
7241 if (curr->policy != SCHED_BATCH) {
7242 update_rq_clock(rq);
7244 * Update run-time statistics of the 'current'.
7246 update_curr(cfs_rq);
7248 * Tell update_rq_clock() that we've just updated,
7249 * so we don't do microscopic update in schedule()
7250 * and double the fastpath cost.
7252 rq_clock_skip_update(rq, true);
7258 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7260 struct sched_entity *se = &p->se;
7262 /* throttled hierarchies are not runnable */
7263 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7266 /* Tell the scheduler that we'd really like pse to run next. */
7269 yield_task_fair(rq);
7275 /**************************************************
7276 * Fair scheduling class load-balancing methods.
7280 * The purpose of load-balancing is to achieve the same basic fairness the
7281 * per-cpu scheduler provides, namely provide a proportional amount of compute
7282 * time to each task. This is expressed in the following equation:
7284 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7286 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
7287 * W_i,0 is defined as:
7289 * W_i,0 = \Sum_j w_i,j (2)
7291 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
7292 * is derived from the nice value as per prio_to_weight[].
7294 * The weight average is an exponential decay average of the instantaneous
7297 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7299 * C_i is the compute capacity of cpu i, typically it is the
7300 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7301 * can also include other factors [XXX].
7303 * To achieve this balance we define a measure of imbalance which follows
7304 * directly from (1):
7306 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7308 * We them move tasks around to minimize the imbalance. In the continuous
7309 * function space it is obvious this converges, in the discrete case we get
7310 * a few fun cases generally called infeasible weight scenarios.
7313 * - infeasible weights;
7314 * - local vs global optima in the discrete case. ]
7319 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7320 * for all i,j solution, we create a tree of cpus that follows the hardware
7321 * topology where each level pairs two lower groups (or better). This results
7322 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
7323 * tree to only the first of the previous level and we decrease the frequency
7324 * of load-balance at each level inv. proportional to the number of cpus in
7330 * \Sum { --- * --- * 2^i } = O(n) (5)
7332 * `- size of each group
7333 * | | `- number of cpus doing load-balance
7335 * `- sum over all levels
7337 * Coupled with a limit on how many tasks we can migrate every balance pass,
7338 * this makes (5) the runtime complexity of the balancer.
7340 * An important property here is that each CPU is still (indirectly) connected
7341 * to every other cpu in at most O(log n) steps:
7343 * The adjacency matrix of the resulting graph is given by:
7346 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7349 * And you'll find that:
7351 * A^(log_2 n)_i,j != 0 for all i,j (7)
7353 * Showing there's indeed a path between every cpu in at most O(log n) steps.
7354 * The task movement gives a factor of O(m), giving a convergence complexity
7357 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7362 * In order to avoid CPUs going idle while there's still work to do, new idle
7363 * balancing is more aggressive and has the newly idle cpu iterate up the domain
7364 * tree itself instead of relying on other CPUs to bring it work.
7366 * This adds some complexity to both (5) and (8) but it reduces the total idle
7374 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7377 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7382 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7384 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
7386 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7389 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7390 * rewrite all of this once again.]
7393 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7395 enum fbq_type { regular, remote, all };
7404 #define LBF_ALL_PINNED 0x01
7405 #define LBF_NEED_BREAK 0x02
7406 #define LBF_DST_PINNED 0x04
7407 #define LBF_SOME_PINNED 0x08
7410 struct sched_domain *sd;
7418 struct cpumask *dst_grpmask;
7420 enum cpu_idle_type idle;
7422 unsigned int src_grp_nr_running;
7423 /* The set of CPUs under consideration for load-balancing */
7424 struct cpumask *cpus;
7429 unsigned int loop_break;
7430 unsigned int loop_max;
7432 enum fbq_type fbq_type;
7433 enum group_type busiest_group_type;
7434 struct list_head tasks;
7438 * Is this task likely cache-hot:
7440 static int task_hot(struct task_struct *p, struct lb_env *env)
7444 lockdep_assert_held(&env->src_rq->lock);
7446 if (p->sched_class != &fair_sched_class)
7449 if (unlikely(p->policy == SCHED_IDLE))
7453 * Buddy candidates are cache hot:
7455 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7456 (&p->se == cfs_rq_of(&p->se)->next ||
7457 &p->se == cfs_rq_of(&p->se)->last))
7460 if (sysctl_sched_migration_cost == -1)
7462 if (sysctl_sched_migration_cost == 0)
7465 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7467 return delta < (s64)sysctl_sched_migration_cost;
7470 #ifdef CONFIG_NUMA_BALANCING
7472 * Returns 1, if task migration degrades locality
7473 * Returns 0, if task migration improves locality i.e migration preferred.
7474 * Returns -1, if task migration is not affected by locality.
7476 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7478 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7479 unsigned long src_faults, dst_faults;
7480 int src_nid, dst_nid;
7482 if (!static_branch_likely(&sched_numa_balancing))
7485 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7488 src_nid = cpu_to_node(env->src_cpu);
7489 dst_nid = cpu_to_node(env->dst_cpu);
7491 if (src_nid == dst_nid)
7494 /* Migrating away from the preferred node is always bad. */
7495 if (src_nid == p->numa_preferred_nid) {
7496 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7502 /* Encourage migration to the preferred node. */
7503 if (dst_nid == p->numa_preferred_nid)
7507 src_faults = group_faults(p, src_nid);
7508 dst_faults = group_faults(p, dst_nid);
7510 src_faults = task_faults(p, src_nid);
7511 dst_faults = task_faults(p, dst_nid);
7514 return dst_faults < src_faults;
7518 static inline int migrate_degrades_locality(struct task_struct *p,
7526 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7529 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7533 lockdep_assert_held(&env->src_rq->lock);
7536 * We do not migrate tasks that are:
7537 * 1) throttled_lb_pair, or
7538 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7539 * 3) running (obviously), or
7540 * 4) are cache-hot on their current CPU.
7542 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7545 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
7548 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
7550 env->flags |= LBF_SOME_PINNED;
7553 * Remember if this task can be migrated to any other cpu in
7554 * our sched_group. We may want to revisit it if we couldn't
7555 * meet load balance goals by pulling other tasks on src_cpu.
7557 * Also avoid computing new_dst_cpu if we have already computed
7558 * one in current iteration.
7560 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
7563 /* Prevent to re-select dst_cpu via env's cpus */
7564 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7565 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
7566 env->flags |= LBF_DST_PINNED;
7567 env->new_dst_cpu = cpu;
7575 /* Record that we found atleast one task that could run on dst_cpu */
7576 env->flags &= ~LBF_ALL_PINNED;
7578 if (task_running(env->src_rq, p)) {
7579 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
7584 * Aggressive migration if:
7585 * 1) destination numa is preferred
7586 * 2) task is cache cold, or
7587 * 3) too many balance attempts have failed.
7589 tsk_cache_hot = migrate_degrades_locality(p, env);
7590 if (tsk_cache_hot == -1)
7591 tsk_cache_hot = task_hot(p, env);
7593 if (tsk_cache_hot <= 0 ||
7594 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7595 if (tsk_cache_hot == 1) {
7596 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
7597 schedstat_inc(p, se.statistics.nr_forced_migrations);
7602 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
7607 * detach_task() -- detach the task for the migration specified in env
7609 static void detach_task(struct task_struct *p, struct lb_env *env)
7611 lockdep_assert_held(&env->src_rq->lock);
7613 deactivate_task(env->src_rq, p, 0);
7614 p->on_rq = TASK_ON_RQ_MIGRATING;
7615 double_lock_balance(env->src_rq, env->dst_rq);
7616 set_task_cpu(p, env->dst_cpu);
7617 double_unlock_balance(env->src_rq, env->dst_rq);
7621 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7622 * part of active balancing operations within "domain".
7624 * Returns a task if successful and NULL otherwise.
7626 static struct task_struct *detach_one_task(struct lb_env *env)
7628 struct task_struct *p, *n;
7630 lockdep_assert_held(&env->src_rq->lock);
7632 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
7633 if (!can_migrate_task(p, env))
7636 detach_task(p, env);
7639 * Right now, this is only the second place where
7640 * lb_gained[env->idle] is updated (other is detach_tasks)
7641 * so we can safely collect stats here rather than
7642 * inside detach_tasks().
7644 schedstat_inc(env->sd, lb_gained[env->idle]);
7650 static const unsigned int sched_nr_migrate_break = 32;
7653 * detach_tasks() -- tries to detach up to imbalance weighted load from
7654 * busiest_rq, as part of a balancing operation within domain "sd".
7656 * Returns number of detached tasks if successful and 0 otherwise.
7658 static int detach_tasks(struct lb_env *env)
7660 struct list_head *tasks = &env->src_rq->cfs_tasks;
7661 struct task_struct *p;
7665 lockdep_assert_held(&env->src_rq->lock);
7667 if (env->imbalance <= 0)
7670 while (!list_empty(tasks)) {
7672 * We don't want to steal all, otherwise we may be treated likewise,
7673 * which could at worst lead to a livelock crash.
7675 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7678 p = list_first_entry(tasks, struct task_struct, se.group_node);
7681 /* We've more or less seen every task there is, call it quits */
7682 if (env->loop > env->loop_max)
7685 /* take a breather every nr_migrate tasks */
7686 if (env->loop > env->loop_break) {
7687 env->loop_break += sched_nr_migrate_break;
7688 env->flags |= LBF_NEED_BREAK;
7692 if (!can_migrate_task(p, env))
7695 load = task_h_load(p);
7697 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7700 if ((load / 2) > env->imbalance)
7703 detach_task(p, env);
7704 list_add(&p->se.group_node, &env->tasks);
7707 env->imbalance -= load;
7709 #ifdef CONFIG_PREEMPT
7711 * NEWIDLE balancing is a source of latency, so preemptible
7712 * kernels will stop after the first task is detached to minimize
7713 * the critical section.
7715 if (env->idle == CPU_NEWLY_IDLE)
7720 * We only want to steal up to the prescribed amount of
7723 if (env->imbalance <= 0)
7728 list_move_tail(&p->se.group_node, tasks);
7732 * Right now, this is one of only two places we collect this stat
7733 * so we can safely collect detach_one_task() stats here rather
7734 * than inside detach_one_task().
7736 schedstat_add(env->sd, lb_gained[env->idle], detached);
7742 * attach_task() -- attach the task detached by detach_task() to its new rq.
7744 static void attach_task(struct rq *rq, struct task_struct *p)
7746 lockdep_assert_held(&rq->lock);
7748 BUG_ON(task_rq(p) != rq);
7749 p->on_rq = TASK_ON_RQ_QUEUED;
7750 activate_task(rq, p, 0);
7751 check_preempt_curr(rq, p, 0);
7755 * attach_one_task() -- attaches the task returned from detach_one_task() to
7758 static void attach_one_task(struct rq *rq, struct task_struct *p)
7760 raw_spin_lock(&rq->lock);
7762 raw_spin_unlock(&rq->lock);
7766 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7769 static void attach_tasks(struct lb_env *env)
7771 struct list_head *tasks = &env->tasks;
7772 struct task_struct *p;
7774 raw_spin_lock(&env->dst_rq->lock);
7776 while (!list_empty(tasks)) {
7777 p = list_first_entry(tasks, struct task_struct, se.group_node);
7778 list_del_init(&p->se.group_node);
7780 attach_task(env->dst_rq, p);
7783 raw_spin_unlock(&env->dst_rq->lock);
7786 #ifdef CONFIG_FAIR_GROUP_SCHED
7787 static void update_blocked_averages(int cpu)
7789 struct rq *rq = cpu_rq(cpu);
7790 struct cfs_rq *cfs_rq;
7791 unsigned long flags;
7793 raw_spin_lock_irqsave(&rq->lock, flags);
7794 update_rq_clock(rq);
7797 * Iterates the task_group tree in a bottom up fashion, see
7798 * list_add_leaf_cfs_rq() for details.
7800 for_each_leaf_cfs_rq(rq, cfs_rq) {
7801 /* throttled entities do not contribute to load */
7802 if (throttled_hierarchy(cfs_rq))
7805 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
7807 update_tg_load_avg(cfs_rq, 0);
7809 /* Propagate pending load changes to the parent */
7810 if (cfs_rq->tg->se[cpu])
7811 update_load_avg(cfs_rq->tg->se[cpu], 0);
7813 raw_spin_unlock_irqrestore(&rq->lock, flags);
7817 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7818 * This needs to be done in a top-down fashion because the load of a child
7819 * group is a fraction of its parents load.
7821 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7823 struct rq *rq = rq_of(cfs_rq);
7824 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7825 unsigned long now = jiffies;
7828 if (cfs_rq->last_h_load_update == now)
7831 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7832 for_each_sched_entity(se) {
7833 cfs_rq = cfs_rq_of(se);
7834 WRITE_ONCE(cfs_rq->h_load_next, se);
7835 if (cfs_rq->last_h_load_update == now)
7840 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7841 cfs_rq->last_h_load_update = now;
7844 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7845 load = cfs_rq->h_load;
7846 load = div64_ul(load * se->avg.load_avg,
7847 cfs_rq_load_avg(cfs_rq) + 1);
7848 cfs_rq = group_cfs_rq(se);
7849 cfs_rq->h_load = load;
7850 cfs_rq->last_h_load_update = now;
7854 static unsigned long task_h_load(struct task_struct *p)
7856 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7858 update_cfs_rq_h_load(cfs_rq);
7859 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7860 cfs_rq_load_avg(cfs_rq) + 1);
7863 static inline void update_blocked_averages(int cpu)
7865 struct rq *rq = cpu_rq(cpu);
7866 struct cfs_rq *cfs_rq = &rq->cfs;
7867 unsigned long flags;
7869 raw_spin_lock_irqsave(&rq->lock, flags);
7870 update_rq_clock(rq);
7871 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7872 raw_spin_unlock_irqrestore(&rq->lock, flags);
7875 static unsigned long task_h_load(struct task_struct *p)
7877 return p->se.avg.load_avg;
7881 /********** Helpers for find_busiest_group ************************/
7884 * sg_lb_stats - stats of a sched_group required for load_balancing
7886 struct sg_lb_stats {
7887 unsigned long avg_load; /*Avg load across the CPUs of the group */
7888 unsigned long group_load; /* Total load over the CPUs of the group */
7889 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7890 unsigned long load_per_task;
7891 unsigned long group_capacity;
7892 unsigned long group_util; /* Total utilization of the group */
7893 unsigned int sum_nr_running; /* Nr tasks running in the group */
7894 unsigned int idle_cpus;
7895 unsigned int group_weight;
7896 enum group_type group_type;
7897 int group_no_capacity;
7898 int group_misfit_task; /* A cpu has a task too big for its capacity */
7899 #ifdef CONFIG_NUMA_BALANCING
7900 unsigned int nr_numa_running;
7901 unsigned int nr_preferred_running;
7906 * sd_lb_stats - Structure to store the statistics of a sched_domain
7907 * during load balancing.
7909 struct sd_lb_stats {
7910 struct sched_group *busiest; /* Busiest group in this sd */
7911 struct sched_group *local; /* Local group in this sd */
7912 unsigned long total_load; /* Total load of all groups in sd */
7913 unsigned long total_capacity; /* Total capacity of all groups in sd */
7914 unsigned long avg_load; /* Average load across all groups in sd */
7916 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7917 struct sg_lb_stats local_stat; /* Statistics of the local group */
7920 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7923 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7924 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7925 * We must however clear busiest_stat::avg_load because
7926 * update_sd_pick_busiest() reads this before assignment.
7928 *sds = (struct sd_lb_stats){
7932 .total_capacity = 0UL,
7935 .sum_nr_running = 0,
7936 .group_type = group_other,
7942 * get_sd_load_idx - Obtain the load index for a given sched domain.
7943 * @sd: The sched_domain whose load_idx is to be obtained.
7944 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7946 * Return: The load index.
7948 static inline int get_sd_load_idx(struct sched_domain *sd,
7949 enum cpu_idle_type idle)
7955 load_idx = sd->busy_idx;
7958 case CPU_NEWLY_IDLE:
7959 load_idx = sd->newidle_idx;
7962 load_idx = sd->idle_idx;
7969 static unsigned long scale_rt_capacity(int cpu)
7971 struct rq *rq = cpu_rq(cpu);
7972 u64 total, used, age_stamp, avg;
7976 * Since we're reading these variables without serialization make sure
7977 * we read them once before doing sanity checks on them.
7979 age_stamp = READ_ONCE(rq->age_stamp);
7980 avg = READ_ONCE(rq->rt_avg);
7981 delta = __rq_clock_broken(rq) - age_stamp;
7983 if (unlikely(delta < 0))
7986 total = sched_avg_period() + delta;
7988 used = div_u64(avg, total);
7991 * deadline bandwidth is defined at system level so we must
7992 * weight this bandwidth with the max capacity of the system.
7993 * As a reminder, avg_bw is 20bits width and
7994 * scale_cpu_capacity is 10 bits width
7996 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7998 if (likely(used < SCHED_CAPACITY_SCALE))
7999 return SCHED_CAPACITY_SCALE - used;
8004 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
8006 raw_spin_lock_init(&mcc->lock);
8011 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8013 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
8014 struct sched_group *sdg = sd->groups;
8015 struct max_cpu_capacity *mcc;
8016 unsigned long max_capacity;
8018 unsigned long flags;
8020 cpu_rq(cpu)->cpu_capacity_orig = capacity;
8022 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
8024 raw_spin_lock_irqsave(&mcc->lock, flags);
8025 max_capacity = mcc->val;
8026 max_cap_cpu = mcc->cpu;
8028 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
8029 (max_capacity < capacity)) {
8030 mcc->val = capacity;
8032 #ifdef CONFIG_SCHED_DEBUG
8033 raw_spin_unlock_irqrestore(&mcc->lock, flags);
8034 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
8039 raw_spin_unlock_irqrestore(&mcc->lock, flags);
8041 skip_unlock: __attribute__ ((unused));
8042 capacity *= scale_rt_capacity(cpu);
8043 capacity >>= SCHED_CAPACITY_SHIFT;
8048 cpu_rq(cpu)->cpu_capacity = capacity;
8049 sdg->sgc->capacity = capacity;
8050 sdg->sgc->max_capacity = capacity;
8051 sdg->sgc->min_capacity = capacity;
8054 void update_group_capacity(struct sched_domain *sd, int cpu)
8056 struct sched_domain *child = sd->child;
8057 struct sched_group *group, *sdg = sd->groups;
8058 unsigned long capacity, max_capacity, min_capacity;
8059 unsigned long interval;
8061 interval = msecs_to_jiffies(sd->balance_interval);
8062 interval = clamp(interval, 1UL, max_load_balance_interval);
8063 sdg->sgc->next_update = jiffies + interval;
8066 update_cpu_capacity(sd, cpu);
8072 min_capacity = ULONG_MAX;
8074 if (child->flags & SD_OVERLAP) {
8076 * SD_OVERLAP domains cannot assume that child groups
8077 * span the current group.
8080 for_each_cpu(cpu, sched_group_cpus(sdg)) {
8081 struct sched_group_capacity *sgc;
8082 struct rq *rq = cpu_rq(cpu);
8085 * build_sched_domains() -> init_sched_groups_capacity()
8086 * gets here before we've attached the domains to the
8089 * Use capacity_of(), which is set irrespective of domains
8090 * in update_cpu_capacity().
8092 * This avoids capacity from being 0 and
8093 * causing divide-by-zero issues on boot.
8095 if (unlikely(!rq->sd)) {
8096 capacity += capacity_of(cpu);
8098 sgc = rq->sd->groups->sgc;
8099 capacity += sgc->capacity;
8102 max_capacity = max(capacity, max_capacity);
8103 min_capacity = min(capacity, min_capacity);
8107 * !SD_OVERLAP domains can assume that child groups
8108 * span the current group.
8111 group = child->groups;
8113 struct sched_group_capacity *sgc = group->sgc;
8115 capacity += sgc->capacity;
8116 max_capacity = max(sgc->max_capacity, max_capacity);
8117 min_capacity = min(sgc->min_capacity, min_capacity);
8118 group = group->next;
8119 } while (group != child->groups);
8122 sdg->sgc->capacity = capacity;
8123 sdg->sgc->max_capacity = max_capacity;
8124 sdg->sgc->min_capacity = min_capacity;
8128 * Check whether the capacity of the rq has been noticeably reduced by side
8129 * activity. The imbalance_pct is used for the threshold.
8130 * Return true is the capacity is reduced
8133 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8135 return ((rq->cpu_capacity * sd->imbalance_pct) <
8136 (rq->cpu_capacity_orig * 100));
8140 * Group imbalance indicates (and tries to solve) the problem where balancing
8141 * groups is inadequate due to tsk_cpus_allowed() constraints.
8143 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
8144 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
8147 * { 0 1 2 3 } { 4 5 6 7 }
8150 * If we were to balance group-wise we'd place two tasks in the first group and
8151 * two tasks in the second group. Clearly this is undesired as it will overload
8152 * cpu 3 and leave one of the cpus in the second group unused.
8154 * The current solution to this issue is detecting the skew in the first group
8155 * by noticing the lower domain failed to reach balance and had difficulty
8156 * moving tasks due to affinity constraints.
8158 * When this is so detected; this group becomes a candidate for busiest; see
8159 * update_sd_pick_busiest(). And calculate_imbalance() and
8160 * find_busiest_group() avoid some of the usual balance conditions to allow it
8161 * to create an effective group imbalance.
8163 * This is a somewhat tricky proposition since the next run might not find the
8164 * group imbalance and decide the groups need to be balanced again. A most
8165 * subtle and fragile situation.
8168 static inline int sg_imbalanced(struct sched_group *group)
8170 return group->sgc->imbalance;
8174 * group_has_capacity returns true if the group has spare capacity that could
8175 * be used by some tasks.
8176 * We consider that a group has spare capacity if the * number of task is
8177 * smaller than the number of CPUs or if the utilization is lower than the
8178 * available capacity for CFS tasks.
8179 * For the latter, we use a threshold to stabilize the state, to take into
8180 * account the variance of the tasks' load and to return true if the available
8181 * capacity in meaningful for the load balancer.
8182 * As an example, an available capacity of 1% can appear but it doesn't make
8183 * any benefit for the load balance.
8186 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8188 if (sgs->sum_nr_running < sgs->group_weight)
8191 if ((sgs->group_capacity * 100) >
8192 (sgs->group_util * env->sd->imbalance_pct))
8199 * group_is_overloaded returns true if the group has more tasks than it can
8201 * group_is_overloaded is not equals to !group_has_capacity because a group
8202 * with the exact right number of tasks, has no more spare capacity but is not
8203 * overloaded so both group_has_capacity and group_is_overloaded return
8207 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8209 if (sgs->sum_nr_running <= sgs->group_weight)
8212 if ((sgs->group_capacity * 100) <
8213 (sgs->group_util * env->sd->imbalance_pct))
8221 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
8222 * per-cpu capacity than sched_group ref.
8225 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8227 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
8228 ref->sgc->max_capacity;
8232 group_type group_classify(struct sched_group *group,
8233 struct sg_lb_stats *sgs)
8235 if (sgs->group_no_capacity)
8236 return group_overloaded;
8238 if (sg_imbalanced(group))
8239 return group_imbalanced;
8241 if (sgs->group_misfit_task)
8242 return group_misfit_task;
8247 #ifdef CONFIG_NO_HZ_COMMON
8249 * idle load balancing data
8250 * - used by the nohz balance, but we want it available here
8251 * so that we can see which CPUs have no tick.
8254 cpumask_var_t idle_cpus_mask;
8256 unsigned long next_balance; /* in jiffy units */
8257 } nohz ____cacheline_aligned;
8259 static inline void update_cpu_stats_if_tickless(struct rq *rq)
8261 /* only called from update_sg_lb_stats when irqs are disabled */
8262 if (cpumask_test_cpu(rq->cpu, nohz.idle_cpus_mask)) {
8263 /* rate limit updates to once-per-jiffie at most */
8264 if (READ_ONCE(jiffies) <= rq->last_load_update_tick)
8267 raw_spin_lock(&rq->lock);
8268 update_rq_clock(rq);
8269 update_idle_cpu_load(rq);
8270 update_cfs_rq_load_avg(rq->clock_task, &rq->cfs, false);
8271 raw_spin_unlock(&rq->lock);
8276 static inline void update_cpu_stats_if_tickless(struct rq *rq) { }
8280 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8281 * @env: The load balancing environment.
8282 * @group: sched_group whose statistics are to be updated.
8283 * @load_idx: Load index of sched_domain of this_cpu for load calc.
8284 * @local_group: Does group contain this_cpu.
8285 * @sgs: variable to hold the statistics for this group.
8286 * @overload: Indicate more than one runnable task for any CPU.
8287 * @overutilized: Indicate overutilization for any CPU.
8289 static inline void update_sg_lb_stats(struct lb_env *env,
8290 struct sched_group *group, int load_idx,
8291 int local_group, struct sg_lb_stats *sgs,
8292 bool *overload, bool *overutilized)
8297 memset(sgs, 0, sizeof(*sgs));
8299 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8300 struct rq *rq = cpu_rq(i);
8302 /* if we are entering idle and there are CPUs with
8303 * their tick stopped, do an update for them
8305 if (env->idle == CPU_NEWLY_IDLE)
8306 update_cpu_stats_if_tickless(rq);
8308 /* Bias balancing toward cpus of our domain */
8310 load = target_load(i, load_idx);
8312 load = source_load(i, load_idx);
8314 sgs->group_load += load;
8315 sgs->group_util += cpu_util(i);
8316 sgs->sum_nr_running += rq->cfs.h_nr_running;
8318 nr_running = rq->nr_running;
8322 #ifdef CONFIG_NUMA_BALANCING
8323 sgs->nr_numa_running += rq->nr_numa_running;
8324 sgs->nr_preferred_running += rq->nr_preferred_running;
8326 sgs->sum_weighted_load += weighted_cpuload(i);
8328 * No need to call idle_cpu() if nr_running is not 0
8330 if (!nr_running && idle_cpu(i))
8333 if (cpu_overutilized(i)) {
8334 *overutilized = true;
8335 if (!sgs->group_misfit_task && rq->misfit_task)
8336 sgs->group_misfit_task = capacity_of(i);
8340 /* Adjust by relative CPU capacity of the group */
8341 sgs->group_capacity = group->sgc->capacity;
8342 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8344 if (sgs->sum_nr_running)
8345 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8347 sgs->group_weight = group->group_weight;
8349 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8350 sgs->group_type = group_classify(group, sgs);
8354 * update_sd_pick_busiest - return 1 on busiest group
8355 * @env: The load balancing environment.
8356 * @sds: sched_domain statistics
8357 * @sg: sched_group candidate to be checked for being the busiest
8358 * @sgs: sched_group statistics
8360 * Determine if @sg is a busier group than the previously selected
8363 * Return: %true if @sg is a busier group than the previously selected
8364 * busiest group. %false otherwise.
8366 static bool update_sd_pick_busiest(struct lb_env *env,
8367 struct sd_lb_stats *sds,
8368 struct sched_group *sg,
8369 struct sg_lb_stats *sgs)
8371 struct sg_lb_stats *busiest = &sds->busiest_stat;
8373 if (sgs->group_type > busiest->group_type)
8376 if (sgs->group_type < busiest->group_type)
8380 * Candidate sg doesn't face any serious load-balance problems
8381 * so don't pick it if the local sg is already filled up.
8383 if (sgs->group_type == group_other &&
8384 !group_has_capacity(env, &sds->local_stat))
8387 if (sgs->avg_load <= busiest->avg_load)
8390 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8394 * Candidate sg has no more than one task per CPU and
8395 * has higher per-CPU capacity. Migrating tasks to less
8396 * capable CPUs may harm throughput. Maximize throughput,
8397 * power/energy consequences are not considered.
8399 if (sgs->sum_nr_running <= sgs->group_weight &&
8400 group_smaller_cpu_capacity(sds->local, sg))
8404 /* This is the busiest node in its class. */
8405 if (!(env->sd->flags & SD_ASYM_PACKING))
8409 * ASYM_PACKING needs to move all the work to the lowest
8410 * numbered CPUs in the group, therefore mark all groups
8411 * higher than ourself as busy.
8413 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
8417 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
8424 #ifdef CONFIG_NUMA_BALANCING
8425 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8427 if (sgs->sum_nr_running > sgs->nr_numa_running)
8429 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8434 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8436 if (rq->nr_running > rq->nr_numa_running)
8438 if (rq->nr_running > rq->nr_preferred_running)
8443 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8448 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8452 #endif /* CONFIG_NUMA_BALANCING */
8454 #define lb_sd_parent(sd) \
8455 (sd->parent && sd->parent->groups != sd->parent->groups->next)
8458 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8459 * @env: The load balancing environment.
8460 * @sds: variable to hold the statistics for this sched_domain.
8462 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8464 struct sched_domain *child = env->sd->child;
8465 struct sched_group *sg = env->sd->groups;
8466 struct sg_lb_stats tmp_sgs;
8467 int load_idx, prefer_sibling = 0;
8468 bool overload = false, overutilized = false;
8470 if (child && child->flags & SD_PREFER_SIBLING)
8473 load_idx = get_sd_load_idx(env->sd, env->idle);
8476 struct sg_lb_stats *sgs = &tmp_sgs;
8479 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
8482 sgs = &sds->local_stat;
8484 if (env->idle != CPU_NEWLY_IDLE ||
8485 time_after_eq(jiffies, sg->sgc->next_update))
8486 update_group_capacity(env->sd, env->dst_cpu);
8489 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8490 &overload, &overutilized);
8496 * In case the child domain prefers tasks go to siblings
8497 * first, lower the sg capacity so that we'll try
8498 * and move all the excess tasks away. We lower the capacity
8499 * of a group only if the local group has the capacity to fit
8500 * these excess tasks. The extra check prevents the case where
8501 * you always pull from the heaviest group when it is already
8502 * under-utilized (possible with a large weight task outweighs
8503 * the tasks on the system).
8505 if (prefer_sibling && sds->local &&
8506 group_has_capacity(env, &sds->local_stat) &&
8507 (sgs->sum_nr_running > 1)) {
8508 sgs->group_no_capacity = 1;
8509 sgs->group_type = group_classify(sg, sgs);
8513 * Ignore task groups with misfit tasks if local group has no
8514 * capacity or if per-cpu capacity isn't higher.
8516 if (sgs->group_type == group_misfit_task &&
8517 (!group_has_capacity(env, &sds->local_stat) ||
8518 !group_smaller_cpu_capacity(sg, sds->local)))
8519 sgs->group_type = group_other;
8521 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8523 sds->busiest_stat = *sgs;
8527 /* Now, start updating sd_lb_stats */
8528 sds->total_load += sgs->group_load;
8529 sds->total_capacity += sgs->group_capacity;
8532 } while (sg != env->sd->groups);
8534 if (env->sd->flags & SD_NUMA)
8535 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8537 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
8539 if (!lb_sd_parent(env->sd)) {
8540 /* update overload indicator if we are at root domain */
8541 if (env->dst_rq->rd->overload != overload)
8542 env->dst_rq->rd->overload = overload;
8544 /* Update over-utilization (tipping point, U >= 0) indicator */
8545 if (env->dst_rq->rd->overutilized != overutilized) {
8546 env->dst_rq->rd->overutilized = overutilized;
8547 trace_sched_overutilized(overutilized);
8550 if (!env->dst_rq->rd->overutilized && overutilized) {
8551 env->dst_rq->rd->overutilized = true;
8552 trace_sched_overutilized(true);
8559 * check_asym_packing - Check to see if the group is packed into the
8562 * This is primarily intended to used at the sibling level. Some
8563 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8564 * case of POWER7, it can move to lower SMT modes only when higher
8565 * threads are idle. When in lower SMT modes, the threads will
8566 * perform better since they share less core resources. Hence when we
8567 * have idle threads, we want them to be the higher ones.
8569 * This packing function is run on idle threads. It checks to see if
8570 * the busiest CPU in this domain (core in the P7 case) has a higher
8571 * CPU number than the packing function is being run on. Here we are
8572 * assuming lower CPU number will be equivalent to lower a SMT thread
8575 * Return: 1 when packing is required and a task should be moved to
8576 * this CPU. The amount of the imbalance is returned in *imbalance.
8578 * @env: The load balancing environment.
8579 * @sds: Statistics of the sched_domain which is to be packed
8581 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8585 if (!(env->sd->flags & SD_ASYM_PACKING))
8591 busiest_cpu = group_first_cpu(sds->busiest);
8592 if (env->dst_cpu > busiest_cpu)
8595 env->imbalance = DIV_ROUND_CLOSEST(
8596 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8597 SCHED_CAPACITY_SCALE);
8603 * fix_small_imbalance - Calculate the minor imbalance that exists
8604 * amongst the groups of a sched_domain, during
8606 * @env: The load balancing environment.
8607 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8610 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8612 unsigned long tmp, capa_now = 0, capa_move = 0;
8613 unsigned int imbn = 2;
8614 unsigned long scaled_busy_load_per_task;
8615 struct sg_lb_stats *local, *busiest;
8617 local = &sds->local_stat;
8618 busiest = &sds->busiest_stat;
8620 if (!local->sum_nr_running)
8621 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8622 else if (busiest->load_per_task > local->load_per_task)
8625 scaled_busy_load_per_task =
8626 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8627 busiest->group_capacity;
8629 if (busiest->avg_load + scaled_busy_load_per_task >=
8630 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8631 env->imbalance = busiest->load_per_task;
8636 * OK, we don't have enough imbalance to justify moving tasks,
8637 * however we may be able to increase total CPU capacity used by
8641 capa_now += busiest->group_capacity *
8642 min(busiest->load_per_task, busiest->avg_load);
8643 capa_now += local->group_capacity *
8644 min(local->load_per_task, local->avg_load);
8645 capa_now /= SCHED_CAPACITY_SCALE;
8647 /* Amount of load we'd subtract */
8648 if (busiest->avg_load > scaled_busy_load_per_task) {
8649 capa_move += busiest->group_capacity *
8650 min(busiest->load_per_task,
8651 busiest->avg_load - scaled_busy_load_per_task);
8654 /* Amount of load we'd add */
8655 if (busiest->avg_load * busiest->group_capacity <
8656 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8657 tmp = (busiest->avg_load * busiest->group_capacity) /
8658 local->group_capacity;
8660 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8661 local->group_capacity;
8663 capa_move += local->group_capacity *
8664 min(local->load_per_task, local->avg_load + tmp);
8665 capa_move /= SCHED_CAPACITY_SCALE;
8667 /* Move if we gain throughput */
8668 if (capa_move > capa_now)
8669 env->imbalance = busiest->load_per_task;
8673 * calculate_imbalance - Calculate the amount of imbalance present within the
8674 * groups of a given sched_domain during load balance.
8675 * @env: load balance environment
8676 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8678 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8680 unsigned long max_pull, load_above_capacity = ~0UL;
8681 struct sg_lb_stats *local, *busiest;
8683 local = &sds->local_stat;
8684 busiest = &sds->busiest_stat;
8686 if (busiest->group_type == group_imbalanced) {
8688 * In the group_imb case we cannot rely on group-wide averages
8689 * to ensure cpu-load equilibrium, look at wider averages. XXX
8691 busiest->load_per_task =
8692 min(busiest->load_per_task, sds->avg_load);
8696 * In the presence of smp nice balancing, certain scenarios can have
8697 * max load less than avg load(as we skip the groups at or below
8698 * its cpu_capacity, while calculating max_load..)
8700 if (busiest->avg_load <= sds->avg_load ||
8701 local->avg_load >= sds->avg_load) {
8702 /* Misfitting tasks should be migrated in any case */
8703 if (busiest->group_type == group_misfit_task) {
8704 env->imbalance = busiest->group_misfit_task;
8709 * Busiest group is overloaded, local is not, use the spare
8710 * cycles to maximize throughput
8712 if (busiest->group_type == group_overloaded &&
8713 local->group_type <= group_misfit_task) {
8714 env->imbalance = busiest->load_per_task;
8719 return fix_small_imbalance(env, sds);
8723 * If there aren't any idle cpus, avoid creating some.
8725 if (busiest->group_type == group_overloaded &&
8726 local->group_type == group_overloaded) {
8727 load_above_capacity = busiest->sum_nr_running *
8729 if (load_above_capacity > busiest->group_capacity)
8730 load_above_capacity -= busiest->group_capacity;
8732 load_above_capacity = ~0UL;
8736 * We're trying to get all the cpus to the average_load, so we don't
8737 * want to push ourselves above the average load, nor do we wish to
8738 * reduce the max loaded cpu below the average load. At the same time,
8739 * we also don't want to reduce the group load below the group capacity
8740 * (so that we can implement power-savings policies etc). Thus we look
8741 * for the minimum possible imbalance.
8743 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8745 /* How much load to actually move to equalise the imbalance */
8746 env->imbalance = min(
8747 max_pull * busiest->group_capacity,
8748 (sds->avg_load - local->avg_load) * local->group_capacity
8749 ) / SCHED_CAPACITY_SCALE;
8751 /* Boost imbalance to allow misfit task to be balanced. */
8752 if (busiest->group_type == group_misfit_task)
8753 env->imbalance = max_t(long, env->imbalance,
8754 busiest->group_misfit_task);
8757 * if *imbalance is less than the average load per runnable task
8758 * there is no guarantee that any tasks will be moved so we'll have
8759 * a think about bumping its value to force at least one task to be
8762 if (env->imbalance < busiest->load_per_task)
8763 return fix_small_imbalance(env, sds);
8766 /******* find_busiest_group() helpers end here *********************/
8769 * find_busiest_group - Returns the busiest group within the sched_domain
8770 * if there is an imbalance. If there isn't an imbalance, and
8771 * the user has opted for power-savings, it returns a group whose
8772 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
8773 * such a group exists.
8775 * Also calculates the amount of weighted load which should be moved
8776 * to restore balance.
8778 * @env: The load balancing environment.
8780 * Return: - The busiest group if imbalance exists.
8781 * - If no imbalance and user has opted for power-savings balance,
8782 * return the least loaded group whose CPUs can be
8783 * put to idle by rebalancing its tasks onto our group.
8785 static struct sched_group *find_busiest_group(struct lb_env *env)
8787 struct sg_lb_stats *local, *busiest;
8788 struct sd_lb_stats sds;
8790 init_sd_lb_stats(&sds);
8793 * Compute the various statistics relavent for load balancing at
8796 update_sd_lb_stats(env, &sds);
8798 if (energy_aware() && !env->dst_rq->rd->overutilized)
8801 local = &sds.local_stat;
8802 busiest = &sds.busiest_stat;
8804 /* ASYM feature bypasses nice load balance check */
8805 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
8806 check_asym_packing(env, &sds))
8809 /* There is no busy sibling group to pull tasks from */
8810 if (!sds.busiest || busiest->sum_nr_running == 0)
8813 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8814 / sds.total_capacity;
8817 * If the busiest group is imbalanced the below checks don't
8818 * work because they assume all things are equal, which typically
8819 * isn't true due to cpus_allowed constraints and the like.
8821 if (busiest->group_type == group_imbalanced)
8825 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8826 * capacities from resulting in underutilization due to avg_load.
8828 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8829 busiest->group_no_capacity)
8832 /* Misfitting tasks should be dealt with regardless of the avg load */
8833 if (busiest->group_type == group_misfit_task) {
8838 * If the local group is busier than the selected busiest group
8839 * don't try and pull any tasks.
8841 if (local->avg_load >= busiest->avg_load)
8845 * Don't pull any tasks if this group is already above the domain
8848 if (local->avg_load >= sds.avg_load)
8851 if (env->idle == CPU_IDLE) {
8853 * This cpu is idle. If the busiest group is not overloaded
8854 * and there is no imbalance between this and busiest group
8855 * wrt idle cpus, it is balanced. The imbalance becomes
8856 * significant if the diff is greater than 1 otherwise we
8857 * might end up to just move the imbalance on another group
8859 if ((busiest->group_type != group_overloaded) &&
8860 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
8861 !group_smaller_cpu_capacity(sds.busiest, sds.local))
8865 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8866 * imbalance_pct to be conservative.
8868 if (100 * busiest->avg_load <=
8869 env->sd->imbalance_pct * local->avg_load)
8874 env->busiest_group_type = busiest->group_type;
8875 /* Looks like there is an imbalance. Compute it */
8876 calculate_imbalance(env, &sds);
8885 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8887 static struct rq *find_busiest_queue(struct lb_env *env,
8888 struct sched_group *group)
8890 struct rq *busiest = NULL, *rq;
8891 unsigned long busiest_load = 0, busiest_capacity = 1;
8894 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8895 unsigned long capacity, wl;
8899 rt = fbq_classify_rq(rq);
8902 * We classify groups/runqueues into three groups:
8903 * - regular: there are !numa tasks
8904 * - remote: there are numa tasks that run on the 'wrong' node
8905 * - all: there is no distinction
8907 * In order to avoid migrating ideally placed numa tasks,
8908 * ignore those when there's better options.
8910 * If we ignore the actual busiest queue to migrate another
8911 * task, the next balance pass can still reduce the busiest
8912 * queue by moving tasks around inside the node.
8914 * If we cannot move enough load due to this classification
8915 * the next pass will adjust the group classification and
8916 * allow migration of more tasks.
8918 * Both cases only affect the total convergence complexity.
8920 if (rt > env->fbq_type)
8923 capacity = capacity_of(i);
8925 wl = weighted_cpuload(i);
8928 * When comparing with imbalance, use weighted_cpuload()
8929 * which is not scaled with the cpu capacity.
8932 if (rq->nr_running == 1 && wl > env->imbalance &&
8933 !check_cpu_capacity(rq, env->sd) &&
8934 env->busiest_group_type != group_misfit_task)
8938 * For the load comparisons with the other cpu's, consider
8939 * the weighted_cpuload() scaled with the cpu capacity, so
8940 * that the load can be moved away from the cpu that is
8941 * potentially running at a lower capacity.
8943 * Thus we're looking for max(wl_i / capacity_i), crosswise
8944 * multiplication to rid ourselves of the division works out
8945 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8946 * our previous maximum.
8948 if (wl * busiest_capacity > busiest_load * capacity) {
8950 busiest_capacity = capacity;
8959 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8960 * so long as it is large enough.
8962 #define MAX_PINNED_INTERVAL 512
8964 /* Working cpumask for load_balance and load_balance_newidle. */
8965 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8967 static int need_active_balance(struct lb_env *env)
8969 struct sched_domain *sd = env->sd;
8971 if (env->idle == CPU_NEWLY_IDLE) {
8974 * ASYM_PACKING needs to force migrate tasks from busy but
8975 * higher numbered CPUs in order to pack all tasks in the
8976 * lowest numbered CPUs.
8978 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8983 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8984 * It's worth migrating the task if the src_cpu's capacity is reduced
8985 * because of other sched_class or IRQs if more capacity stays
8986 * available on dst_cpu.
8988 if ((env->idle != CPU_NOT_IDLE) &&
8989 (env->src_rq->cfs.h_nr_running == 1)) {
8990 if ((check_cpu_capacity(env->src_rq, sd)) &&
8991 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8995 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8996 ((capacity_orig_of(env->src_cpu) < capacity_orig_of(env->dst_cpu))) &&
8997 env->src_rq->cfs.h_nr_running == 1 &&
8998 cpu_overutilized(env->src_cpu) &&
8999 !cpu_overutilized(env->dst_cpu)) {
9003 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
9006 static int active_load_balance_cpu_stop(void *data);
9008 static int should_we_balance(struct lb_env *env)
9010 struct sched_group *sg = env->sd->groups;
9011 struct cpumask *sg_cpus, *sg_mask;
9012 int cpu, balance_cpu = -1;
9015 * In the newly idle case, we will allow all the cpu's
9016 * to do the newly idle load balance.
9018 if (env->idle == CPU_NEWLY_IDLE)
9021 sg_cpus = sched_group_cpus(sg);
9022 sg_mask = sched_group_mask(sg);
9023 /* Try to find first idle cpu */
9024 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
9025 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
9032 if (balance_cpu == -1)
9033 balance_cpu = group_balance_cpu(sg);
9036 * First idle cpu or the first cpu(busiest) in this sched group
9037 * is eligible for doing load balancing at this and above domains.
9039 return balance_cpu == env->dst_cpu;
9043 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9044 * tasks if there is an imbalance.
9046 static int load_balance(int this_cpu, struct rq *this_rq,
9047 struct sched_domain *sd, enum cpu_idle_type idle,
9048 int *continue_balancing)
9050 int ld_moved, cur_ld_moved, active_balance = 0;
9051 struct sched_domain *sd_parent = lb_sd_parent(sd) ? sd->parent : NULL;
9052 struct sched_group *group;
9054 unsigned long flags;
9055 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9057 struct lb_env env = {
9059 .dst_cpu = this_cpu,
9061 .dst_grpmask = sched_group_cpus(sd->groups),
9063 .loop_break = sched_nr_migrate_break,
9066 .tasks = LIST_HEAD_INIT(env.tasks),
9070 * For NEWLY_IDLE load_balancing, we don't need to consider
9071 * other cpus in our group
9073 if (idle == CPU_NEWLY_IDLE)
9074 env.dst_grpmask = NULL;
9076 cpumask_copy(cpus, cpu_active_mask);
9078 schedstat_inc(sd, lb_count[idle]);
9081 if (!should_we_balance(&env)) {
9082 *continue_balancing = 0;
9086 group = find_busiest_group(&env);
9088 schedstat_inc(sd, lb_nobusyg[idle]);
9092 busiest = find_busiest_queue(&env, group);
9094 schedstat_inc(sd, lb_nobusyq[idle]);
9098 BUG_ON(busiest == env.dst_rq);
9100 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
9102 env.src_cpu = busiest->cpu;
9103 env.src_rq = busiest;
9106 if (busiest->nr_running > 1) {
9108 * Attempt to move tasks. If find_busiest_group has found
9109 * an imbalance but busiest->nr_running <= 1, the group is
9110 * still unbalanced. ld_moved simply stays zero, so it is
9111 * correctly treated as an imbalance.
9113 env.flags |= LBF_ALL_PINNED;
9114 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9117 raw_spin_lock_irqsave(&busiest->lock, flags);
9118 update_rq_clock(busiest);
9121 * cur_ld_moved - load moved in current iteration
9122 * ld_moved - cumulative load moved across iterations
9124 cur_ld_moved = detach_tasks(&env);
9127 * We've detached some tasks from busiest_rq. Every
9128 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9129 * unlock busiest->lock, and we are able to be sure
9130 * that nobody can manipulate the tasks in parallel.
9131 * See task_rq_lock() family for the details.
9134 raw_spin_unlock(&busiest->lock);
9138 ld_moved += cur_ld_moved;
9141 local_irq_restore(flags);
9143 if (env.flags & LBF_NEED_BREAK) {
9144 env.flags &= ~LBF_NEED_BREAK;
9149 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9150 * us and move them to an alternate dst_cpu in our sched_group
9151 * where they can run. The upper limit on how many times we
9152 * iterate on same src_cpu is dependent on number of cpus in our
9155 * This changes load balance semantics a bit on who can move
9156 * load to a given_cpu. In addition to the given_cpu itself
9157 * (or a ilb_cpu acting on its behalf where given_cpu is
9158 * nohz-idle), we now have balance_cpu in a position to move
9159 * load to given_cpu. In rare situations, this may cause
9160 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9161 * _independently_ and at _same_ time to move some load to
9162 * given_cpu) causing exceess load to be moved to given_cpu.
9163 * This however should not happen so much in practice and
9164 * moreover subsequent load balance cycles should correct the
9165 * excess load moved.
9167 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9169 /* Prevent to re-select dst_cpu via env's cpus */
9170 cpumask_clear_cpu(env.dst_cpu, env.cpus);
9172 env.dst_rq = cpu_rq(env.new_dst_cpu);
9173 env.dst_cpu = env.new_dst_cpu;
9174 env.flags &= ~LBF_DST_PINNED;
9176 env.loop_break = sched_nr_migrate_break;
9179 * Go back to "more_balance" rather than "redo" since we
9180 * need to continue with same src_cpu.
9186 * We failed to reach balance because of affinity.
9189 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9191 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9192 *group_imbalance = 1;
9195 /* All tasks on this runqueue were pinned by CPU affinity */
9196 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9197 cpumask_clear_cpu(cpu_of(busiest), cpus);
9198 if (!cpumask_empty(cpus)) {
9200 env.loop_break = sched_nr_migrate_break;
9203 goto out_all_pinned;
9208 schedstat_inc(sd, lb_failed[idle]);
9210 * Increment the failure counter only on periodic balance.
9211 * We do not want newidle balance, which can be very
9212 * frequent, pollute the failure counter causing
9213 * excessive cache_hot migrations and active balances.
9215 if (idle != CPU_NEWLY_IDLE)
9216 if (env.src_grp_nr_running > 1)
9217 sd->nr_balance_failed++;
9219 if (need_active_balance(&env)) {
9220 raw_spin_lock_irqsave(&busiest->lock, flags);
9222 /* don't kick the active_load_balance_cpu_stop,
9223 * if the curr task on busiest cpu can't be
9226 if (!cpumask_test_cpu(this_cpu,
9227 tsk_cpus_allowed(busiest->curr))) {
9228 raw_spin_unlock_irqrestore(&busiest->lock,
9230 env.flags |= LBF_ALL_PINNED;
9231 goto out_one_pinned;
9235 * ->active_balance synchronizes accesses to
9236 * ->active_balance_work. Once set, it's cleared
9237 * only after active load balance is finished.
9239 if (!busiest->active_balance) {
9240 busiest->active_balance = 1;
9241 busiest->push_cpu = this_cpu;
9244 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9246 if (active_balance) {
9247 stop_one_cpu_nowait(cpu_of(busiest),
9248 active_load_balance_cpu_stop, busiest,
9249 &busiest->active_balance_work);
9253 * We've kicked active balancing, reset the failure
9256 sd->nr_balance_failed = sd->cache_nice_tries+1;
9259 sd->nr_balance_failed = 0;
9261 if (likely(!active_balance)) {
9262 /* We were unbalanced, so reset the balancing interval */
9263 sd->balance_interval = sd->min_interval;
9266 * If we've begun active balancing, start to back off. This
9267 * case may not be covered by the all_pinned logic if there
9268 * is only 1 task on the busy runqueue (because we don't call
9271 if (sd->balance_interval < sd->max_interval)
9272 sd->balance_interval *= 2;
9279 * We reach balance although we may have faced some affinity
9280 * constraints. Clear the imbalance flag if it was set.
9283 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9285 if (*group_imbalance)
9286 *group_imbalance = 0;
9291 * We reach balance because all tasks are pinned at this level so
9292 * we can't migrate them. Let the imbalance flag set so parent level
9293 * can try to migrate them.
9295 schedstat_inc(sd, lb_balanced[idle]);
9297 sd->nr_balance_failed = 0;
9300 /* tune up the balancing interval */
9301 if (((env.flags & LBF_ALL_PINNED) &&
9302 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9303 (sd->balance_interval < sd->max_interval))
9304 sd->balance_interval *= 2;
9311 static inline unsigned long
9312 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9314 unsigned long interval = sd->balance_interval;
9317 interval *= sd->busy_factor;
9319 /* scale ms to jiffies */
9320 interval = msecs_to_jiffies(interval);
9321 interval = clamp(interval, 1UL, max_load_balance_interval);
9327 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
9329 unsigned long interval, next;
9331 interval = get_sd_balance_interval(sd, cpu_busy);
9332 next = sd->last_balance + interval;
9334 if (time_after(*next_balance, next))
9335 *next_balance = next;
9339 * idle_balance is called by schedule() if this_cpu is about to become
9340 * idle. Attempts to pull tasks from other CPUs.
9342 static int idle_balance(struct rq *this_rq)
9344 unsigned long next_balance = jiffies + HZ;
9345 int this_cpu = this_rq->cpu;
9346 struct sched_domain *sd;
9347 int pulled_task = 0;
9350 idle_enter_fair(this_rq);
9353 * We must set idle_stamp _before_ calling idle_balance(), such that we
9354 * measure the duration of idle_balance() as idle time.
9356 this_rq->idle_stamp = rq_clock(this_rq);
9358 if (!energy_aware() &&
9359 (this_rq->avg_idle < sysctl_sched_migration_cost ||
9360 !this_rq->rd->overload)) {
9362 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9364 update_next_balance(sd, 0, &next_balance);
9370 raw_spin_unlock(&this_rq->lock);
9372 update_blocked_averages(this_cpu);
9374 for_each_domain(this_cpu, sd) {
9375 int continue_balancing = 1;
9376 u64 t0, domain_cost;
9378 if (!(sd->flags & SD_LOAD_BALANCE))
9381 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9382 update_next_balance(sd, 0, &next_balance);
9386 if (sd->flags & SD_BALANCE_NEWIDLE) {
9387 t0 = sched_clock_cpu(this_cpu);
9389 pulled_task = load_balance(this_cpu, this_rq,
9391 &continue_balancing);
9393 domain_cost = sched_clock_cpu(this_cpu) - t0;
9394 if (domain_cost > sd->max_newidle_lb_cost)
9395 sd->max_newidle_lb_cost = domain_cost;
9397 curr_cost += domain_cost;
9400 update_next_balance(sd, 0, &next_balance);
9403 * Stop searching for tasks to pull if there are
9404 * now runnable tasks on this rq.
9406 if (pulled_task || this_rq->nr_running > 0)
9411 raw_spin_lock(&this_rq->lock);
9413 if (curr_cost > this_rq->max_idle_balance_cost)
9414 this_rq->max_idle_balance_cost = curr_cost;
9417 * While browsing the domains, we released the rq lock, a task could
9418 * have been enqueued in the meantime. Since we're not going idle,
9419 * pretend we pulled a task.
9421 if (this_rq->cfs.h_nr_running && !pulled_task)
9425 /* Move the next balance forward */
9426 if (time_after(this_rq->next_balance, next_balance))
9427 this_rq->next_balance = next_balance;
9429 /* Is there a task of a high priority class? */
9430 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9434 idle_exit_fair(this_rq);
9435 this_rq->idle_stamp = 0;
9442 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
9443 * running tasks off the busiest CPU onto idle CPUs. It requires at
9444 * least 1 task to be running on each physical CPU where possible, and
9445 * avoids physical / logical imbalances.
9447 static int active_load_balance_cpu_stop(void *data)
9449 struct rq *busiest_rq = data;
9450 int busiest_cpu = cpu_of(busiest_rq);
9451 int target_cpu = busiest_rq->push_cpu;
9452 struct rq *target_rq = cpu_rq(target_cpu);
9453 struct sched_domain *sd = NULL;
9454 struct task_struct *p = NULL;
9455 struct task_struct *push_task = NULL;
9456 int push_task_detached = 0;
9457 struct lb_env env = {
9459 .dst_cpu = target_cpu,
9460 .dst_rq = target_rq,
9461 .src_cpu = busiest_rq->cpu,
9462 .src_rq = busiest_rq,
9466 raw_spin_lock_irq(&busiest_rq->lock);
9468 /* make sure the requested cpu hasn't gone down in the meantime */
9469 if (unlikely(busiest_cpu != smp_processor_id() ||
9470 !busiest_rq->active_balance))
9473 /* Is there any task to move? */
9474 if (busiest_rq->nr_running <= 1)
9478 * This condition is "impossible", if it occurs
9479 * we need to fix it. Originally reported by
9480 * Bjorn Helgaas on a 128-cpu setup.
9482 BUG_ON(busiest_rq == target_rq);
9484 push_task = busiest_rq->push_task;
9486 if (task_on_rq_queued(push_task) &&
9487 task_cpu(push_task) == busiest_cpu &&
9488 cpu_online(target_cpu)) {
9489 detach_task(push_task, &env);
9490 push_task_detached = 1;
9495 /* Search for an sd spanning us and the target CPU. */
9497 for_each_domain(target_cpu, sd) {
9498 if ((sd->flags & SD_LOAD_BALANCE) &&
9499 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9505 schedstat_inc(sd, alb_count);
9506 update_rq_clock(busiest_rq);
9508 p = detach_one_task(&env);
9510 schedstat_inc(sd, alb_pushed);
9512 schedstat_inc(sd, alb_failed);
9516 busiest_rq->active_balance = 0;
9519 busiest_rq->push_task = NULL;
9521 raw_spin_unlock(&busiest_rq->lock);
9524 if (push_task_detached)
9525 attach_one_task(target_rq, push_task);
9526 put_task_struct(push_task);
9530 attach_one_task(target_rq, p);
9537 static inline int on_null_domain(struct rq *rq)
9539 return unlikely(!rcu_dereference_sched(rq->sd));
9542 #ifdef CONFIG_NO_HZ_COMMON
9544 * idle load balancing details
9545 * - When one of the busy CPUs notice that there may be an idle rebalancing
9546 * needed, they will kick the idle load balancer, which then does idle
9547 * load balancing for all the idle CPUs.
9549 static inline int find_new_ilb(void)
9551 int ilb = cpumask_first(nohz.idle_cpus_mask);
9553 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9560 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9561 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9562 * CPU (if there is one).
9564 static void nohz_balancer_kick(void)
9568 nohz.next_balance++;
9570 ilb_cpu = find_new_ilb();
9572 if (ilb_cpu >= nr_cpu_ids)
9575 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9578 * Use smp_send_reschedule() instead of resched_cpu().
9579 * This way we generate a sched IPI on the target cpu which
9580 * is idle. And the softirq performing nohz idle load balance
9581 * will be run before returning from the IPI.
9583 smp_send_reschedule(ilb_cpu);
9587 static inline void nohz_balance_exit_idle(int cpu)
9589 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9591 * Completely isolated CPUs don't ever set, so we must test.
9593 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9594 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9595 atomic_dec(&nohz.nr_cpus);
9597 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9601 static inline void set_cpu_sd_state_busy(void)
9603 struct sched_domain *sd;
9604 int cpu = smp_processor_id();
9607 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9609 if (!sd || !sd->nohz_idle)
9613 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
9618 void set_cpu_sd_state_idle(void)
9620 struct sched_domain *sd;
9621 int cpu = smp_processor_id();
9624 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9626 if (!sd || sd->nohz_idle)
9630 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
9636 * This routine will record that the cpu is going idle with tick stopped.
9637 * This info will be used in performing idle load balancing in the future.
9639 void nohz_balance_enter_idle(int cpu)
9642 * If this cpu is going down, then nothing needs to be done.
9644 if (!cpu_active(cpu))
9647 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9651 * If we're a completely isolated CPU, we don't play.
9653 if (on_null_domain(cpu_rq(cpu)))
9656 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9657 atomic_inc(&nohz.nr_cpus);
9658 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9661 static int sched_ilb_notifier(struct notifier_block *nfb,
9662 unsigned long action, void *hcpu)
9664 switch (action & ~CPU_TASKS_FROZEN) {
9666 nohz_balance_exit_idle(smp_processor_id());
9674 static DEFINE_SPINLOCK(balancing);
9677 * Scale the max load_balance interval with the number of CPUs in the system.
9678 * This trades load-balance latency on larger machines for less cross talk.
9680 void update_max_interval(void)
9682 max_load_balance_interval = HZ*num_online_cpus()/10;
9686 * It checks each scheduling domain to see if it is due to be balanced,
9687 * and initiates a balancing operation if so.
9689 * Balancing parameters are set up in init_sched_domains.
9691 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9693 int continue_balancing = 1;
9695 unsigned long interval;
9696 struct sched_domain *sd;
9697 /* Earliest time when we have to do rebalance again */
9698 unsigned long next_balance = jiffies + 60*HZ;
9699 int update_next_balance = 0;
9700 int need_serialize, need_decay = 0;
9703 update_blocked_averages(cpu);
9706 for_each_domain(cpu, sd) {
9708 * Decay the newidle max times here because this is a regular
9709 * visit to all the domains. Decay ~1% per second.
9711 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9712 sd->max_newidle_lb_cost =
9713 (sd->max_newidle_lb_cost * 253) / 256;
9714 sd->next_decay_max_lb_cost = jiffies + HZ;
9717 max_cost += sd->max_newidle_lb_cost;
9719 if (!(sd->flags & SD_LOAD_BALANCE))
9723 * Stop the load balance at this level. There is another
9724 * CPU in our sched group which is doing load balancing more
9727 if (!continue_balancing) {
9733 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9735 need_serialize = sd->flags & SD_SERIALIZE;
9736 if (need_serialize) {
9737 if (!spin_trylock(&balancing))
9741 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9742 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9744 * The LBF_DST_PINNED logic could have changed
9745 * env->dst_cpu, so we can't know our idle
9746 * state even if we migrated tasks. Update it.
9748 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9750 sd->last_balance = jiffies;
9751 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9754 spin_unlock(&balancing);
9756 if (time_after(next_balance, sd->last_balance + interval)) {
9757 next_balance = sd->last_balance + interval;
9758 update_next_balance = 1;
9763 * Ensure the rq-wide value also decays but keep it at a
9764 * reasonable floor to avoid funnies with rq->avg_idle.
9766 rq->max_idle_balance_cost =
9767 max((u64)sysctl_sched_migration_cost, max_cost);
9772 * next_balance will be updated only when there is a need.
9773 * When the cpu is attached to null domain for ex, it will not be
9776 if (likely(update_next_balance)) {
9777 rq->next_balance = next_balance;
9779 #ifdef CONFIG_NO_HZ_COMMON
9781 * If this CPU has been elected to perform the nohz idle
9782 * balance. Other idle CPUs have already rebalanced with
9783 * nohz_idle_balance() and nohz.next_balance has been
9784 * updated accordingly. This CPU is now running the idle load
9785 * balance for itself and we need to update the
9786 * nohz.next_balance accordingly.
9788 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9789 nohz.next_balance = rq->next_balance;
9794 #ifdef CONFIG_NO_HZ_COMMON
9796 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9797 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9799 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9801 int this_cpu = this_rq->cpu;
9804 /* Earliest time when we have to do rebalance again */
9805 unsigned long next_balance = jiffies + 60*HZ;
9806 int update_next_balance = 0;
9808 if (idle != CPU_IDLE ||
9809 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9812 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9813 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9817 * If this cpu gets work to do, stop the load balancing
9818 * work being done for other cpus. Next load
9819 * balancing owner will pick it up.
9824 rq = cpu_rq(balance_cpu);
9827 * If time for next balance is due,
9830 if (time_after_eq(jiffies, rq->next_balance)) {
9831 raw_spin_lock_irq(&rq->lock);
9832 update_rq_clock(rq);
9833 update_idle_cpu_load(rq);
9834 raw_spin_unlock_irq(&rq->lock);
9835 rebalance_domains(rq, CPU_IDLE);
9838 if (time_after(next_balance, rq->next_balance)) {
9839 next_balance = rq->next_balance;
9840 update_next_balance = 1;
9845 * next_balance will be updated only when there is a need.
9846 * When the CPU is attached to null domain for ex, it will not be
9849 if (likely(update_next_balance))
9850 nohz.next_balance = next_balance;
9852 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9856 * Current heuristic for kicking the idle load balancer in the presence
9857 * of an idle cpu in the system.
9858 * - This rq has more than one task.
9859 * - This rq has at least one CFS task and the capacity of the CPU is
9860 * significantly reduced because of RT tasks or IRQs.
9861 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9862 * multiple busy cpu.
9863 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9864 * domain span are idle.
9866 static inline bool nohz_kick_needed(struct rq *rq)
9868 unsigned long now = jiffies;
9869 struct sched_domain *sd;
9870 struct sched_group_capacity *sgc;
9871 int nr_busy, cpu = rq->cpu;
9874 if (unlikely(rq->idle_balance))
9878 * We may be recently in ticked or tickless idle mode. At the first
9879 * busy tick after returning from idle, we will update the busy stats.
9881 set_cpu_sd_state_busy();
9882 nohz_balance_exit_idle(cpu);
9885 * None are in tickless mode and hence no need for NOHZ idle load
9888 if (likely(!atomic_read(&nohz.nr_cpus)))
9891 if (time_before(now, nohz.next_balance))
9894 if (rq->nr_running >= 2 &&
9895 (!energy_aware() || cpu_overutilized(cpu)))
9898 /* Do idle load balance if there have misfit task */
9900 return rq->misfit_task;
9903 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9905 sgc = sd->groups->sgc;
9906 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9915 sd = rcu_dereference(rq->sd);
9917 if ((rq->cfs.h_nr_running >= 1) &&
9918 check_cpu_capacity(rq, sd)) {
9924 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9925 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9926 sched_domain_span(sd)) < cpu)) {
9936 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9940 * run_rebalance_domains is triggered when needed from the scheduler tick.
9941 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9943 static void run_rebalance_domains(struct softirq_action *h)
9945 struct rq *this_rq = this_rq();
9946 enum cpu_idle_type idle = this_rq->idle_balance ?
9947 CPU_IDLE : CPU_NOT_IDLE;
9950 * If this cpu has a pending nohz_balance_kick, then do the
9951 * balancing on behalf of the other idle cpus whose ticks are
9952 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9953 * give the idle cpus a chance to load balance. Else we may
9954 * load balance only within the local sched_domain hierarchy
9955 * and abort nohz_idle_balance altogether if we pull some load.
9957 nohz_idle_balance(this_rq, idle);
9958 rebalance_domains(this_rq, idle);
9962 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9964 void trigger_load_balance(struct rq *rq)
9966 /* Don't need to rebalance while attached to NULL domain */
9967 if (unlikely(on_null_domain(rq)))
9970 if (time_after_eq(jiffies, rq->next_balance))
9971 raise_softirq(SCHED_SOFTIRQ);
9972 #ifdef CONFIG_NO_HZ_COMMON
9973 if (nohz_kick_needed(rq))
9974 nohz_balancer_kick();
9978 static void rq_online_fair(struct rq *rq)
9982 update_runtime_enabled(rq);
9985 static void rq_offline_fair(struct rq *rq)
9989 /* Ensure any throttled groups are reachable by pick_next_task */
9990 unthrottle_offline_cfs_rqs(rq);
9994 kick_active_balance(struct rq *rq, struct task_struct *p, int new_cpu)
9998 /* Invoke active balance to force migrate currently running task */
9999 raw_spin_lock(&rq->lock);
10000 if (!rq->active_balance) {
10001 rq->active_balance = 1;
10002 rq->push_cpu = new_cpu;
10003 get_task_struct(p);
10007 raw_spin_unlock(&rq->lock);
10012 void check_for_migration(struct rq *rq, struct task_struct *p)
10015 int active_balance;
10016 int cpu = task_cpu(p);
10018 if (energy_aware() && rq->misfit_task) {
10019 if (rq->curr->state != TASK_RUNNING ||
10020 rq->curr->nr_cpus_allowed == 1)
10023 new_cpu = select_energy_cpu_brute(p, cpu, 0);
10024 if (capacity_orig_of(new_cpu) > capacity_orig_of(cpu)) {
10025 active_balance = kick_active_balance(rq, p, new_cpu);
10026 if (active_balance)
10027 stop_one_cpu_nowait(cpu,
10028 active_load_balance_cpu_stop,
10029 rq, &rq->active_balance_work);
10034 #endif /* CONFIG_SMP */
10037 * scheduler tick hitting a task of our scheduling class:
10039 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10041 struct cfs_rq *cfs_rq;
10042 struct sched_entity *se = &curr->se;
10044 for_each_sched_entity(se) {
10045 cfs_rq = cfs_rq_of(se);
10046 entity_tick(cfs_rq, se, queued);
10049 if (static_branch_unlikely(&sched_numa_balancing))
10050 task_tick_numa(rq, curr);
10053 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
10054 rq->rd->overutilized = true;
10055 trace_sched_overutilized(true);
10058 rq->misfit_task = !task_fits_max(curr, rq->cpu);
10064 * called on fork with the child task as argument from the parent's context
10065 * - child not yet on the tasklist
10066 * - preemption disabled
10068 static void task_fork_fair(struct task_struct *p)
10070 struct cfs_rq *cfs_rq;
10071 struct sched_entity *se = &p->se, *curr;
10072 struct rq *rq = this_rq();
10074 raw_spin_lock(&rq->lock);
10075 update_rq_clock(rq);
10077 cfs_rq = task_cfs_rq(current);
10078 curr = cfs_rq->curr;
10080 update_curr(cfs_rq);
10081 se->vruntime = curr->vruntime;
10083 place_entity(cfs_rq, se, 1);
10085 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10087 * Upon rescheduling, sched_class::put_prev_task() will place
10088 * 'current' within the tree based on its new key value.
10090 swap(curr->vruntime, se->vruntime);
10094 se->vruntime -= cfs_rq->min_vruntime;
10095 raw_spin_unlock(&rq->lock);
10099 * Priority of the task has changed. Check to see if we preempt
10100 * the current task.
10103 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10105 if (!task_on_rq_queued(p))
10109 * Reschedule if we are currently running on this runqueue and
10110 * our priority decreased, or if we are not currently running on
10111 * this runqueue and our priority is higher than the current's
10113 if (rq->curr == p) {
10114 if (p->prio > oldprio)
10117 check_preempt_curr(rq, p, 0);
10120 static inline bool vruntime_normalized(struct task_struct *p)
10122 struct sched_entity *se = &p->se;
10125 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10126 * the dequeue_entity(.flags=0) will already have normalized the
10133 * When !on_rq, vruntime of the task has usually NOT been normalized.
10134 * But there are some cases where it has already been normalized:
10136 * - A forked child which is waiting for being woken up by
10137 * wake_up_new_task().
10138 * - A task which has been woken up by try_to_wake_up() and
10139 * waiting for actually being woken up by sched_ttwu_pending().
10141 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
10147 #ifdef CONFIG_FAIR_GROUP_SCHED
10149 * Propagate the changes of the sched_entity across the tg tree to make it
10150 * visible to the root
10152 static void propagate_entity_cfs_rq(struct sched_entity *se)
10154 struct cfs_rq *cfs_rq;
10156 /* Start to propagate at parent */
10159 for_each_sched_entity(se) {
10160 cfs_rq = cfs_rq_of(se);
10162 if (cfs_rq_throttled(cfs_rq))
10165 update_load_avg(se, UPDATE_TG);
10169 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10172 static void detach_entity_cfs_rq(struct sched_entity *se)
10174 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10176 /* Catch up with the cfs_rq and remove our load when we leave */
10177 update_load_avg(se, 0);
10178 detach_entity_load_avg(cfs_rq, se);
10179 update_tg_load_avg(cfs_rq, false);
10180 propagate_entity_cfs_rq(se);
10183 static void attach_entity_cfs_rq(struct sched_entity *se)
10185 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10187 #ifdef CONFIG_FAIR_GROUP_SCHED
10189 * Since the real-depth could have been changed (only FAIR
10190 * class maintain depth value), reset depth properly.
10192 se->depth = se->parent ? se->parent->depth + 1 : 0;
10195 /* Synchronize entity with its cfs_rq */
10196 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10197 attach_entity_load_avg(cfs_rq, se);
10198 update_tg_load_avg(cfs_rq, false);
10199 propagate_entity_cfs_rq(se);
10202 static void detach_task_cfs_rq(struct task_struct *p)
10204 struct sched_entity *se = &p->se;
10205 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10207 if (!vruntime_normalized(p)) {
10209 * Fix up our vruntime so that the current sleep doesn't
10210 * cause 'unlimited' sleep bonus.
10212 place_entity(cfs_rq, se, 0);
10213 se->vruntime -= cfs_rq->min_vruntime;
10216 detach_entity_cfs_rq(se);
10219 static void attach_task_cfs_rq(struct task_struct *p)
10221 struct sched_entity *se = &p->se;
10222 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10224 attach_entity_cfs_rq(se);
10226 if (!vruntime_normalized(p))
10227 se->vruntime += cfs_rq->min_vruntime;
10230 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10232 detach_task_cfs_rq(p);
10235 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10237 attach_task_cfs_rq(p);
10239 if (task_on_rq_queued(p)) {
10241 * We were most likely switched from sched_rt, so
10242 * kick off the schedule if running, otherwise just see
10243 * if we can still preempt the current task.
10248 check_preempt_curr(rq, p, 0);
10252 /* Account for a task changing its policy or group.
10254 * This routine is mostly called to set cfs_rq->curr field when a task
10255 * migrates between groups/classes.
10257 static void set_curr_task_fair(struct rq *rq)
10259 struct sched_entity *se = &rq->curr->se;
10261 for_each_sched_entity(se) {
10262 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10264 set_next_entity(cfs_rq, se);
10265 /* ensure bandwidth has been allocated on our new cfs_rq */
10266 account_cfs_rq_runtime(cfs_rq, 0);
10270 void init_cfs_rq(struct cfs_rq *cfs_rq)
10272 cfs_rq->tasks_timeline = RB_ROOT;
10273 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10274 #ifndef CONFIG_64BIT
10275 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10278 #ifdef CONFIG_FAIR_GROUP_SCHED
10279 cfs_rq->propagate_avg = 0;
10281 atomic_long_set(&cfs_rq->removed_load_avg, 0);
10282 atomic_long_set(&cfs_rq->removed_util_avg, 0);
10286 #ifdef CONFIG_FAIR_GROUP_SCHED
10287 static void task_set_group_fair(struct task_struct *p)
10289 struct sched_entity *se = &p->se;
10291 set_task_rq(p, task_cpu(p));
10292 se->depth = se->parent ? se->parent->depth + 1 : 0;
10295 static void task_move_group_fair(struct task_struct *p)
10297 detach_task_cfs_rq(p);
10298 set_task_rq(p, task_cpu(p));
10301 /* Tell se's cfs_rq has been changed -- migrated */
10302 p->se.avg.last_update_time = 0;
10304 attach_task_cfs_rq(p);
10307 static void task_change_group_fair(struct task_struct *p, int type)
10310 case TASK_SET_GROUP:
10311 task_set_group_fair(p);
10314 case TASK_MOVE_GROUP:
10315 task_move_group_fair(p);
10320 void free_fair_sched_group(struct task_group *tg)
10324 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10326 for_each_possible_cpu(i) {
10328 kfree(tg->cfs_rq[i]);
10337 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10339 struct sched_entity *se;
10340 struct cfs_rq *cfs_rq;
10344 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
10347 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
10351 tg->shares = NICE_0_LOAD;
10353 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10355 for_each_possible_cpu(i) {
10358 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10359 GFP_KERNEL, cpu_to_node(i));
10363 se = kzalloc_node(sizeof(struct sched_entity),
10364 GFP_KERNEL, cpu_to_node(i));
10368 init_cfs_rq(cfs_rq);
10369 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10370 init_entity_runnable_average(se);
10372 raw_spin_lock_irq(&rq->lock);
10373 post_init_entity_util_avg(se);
10374 raw_spin_unlock_irq(&rq->lock);
10385 void unregister_fair_sched_group(struct task_group *tg)
10387 unsigned long flags;
10391 for_each_possible_cpu(cpu) {
10393 remove_entity_load_avg(tg->se[cpu]);
10396 * Only empty task groups can be destroyed; so we can speculatively
10397 * check on_list without danger of it being re-added.
10399 if (!tg->cfs_rq[cpu]->on_list)
10404 raw_spin_lock_irqsave(&rq->lock, flags);
10405 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10406 raw_spin_unlock_irqrestore(&rq->lock, flags);
10410 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10411 struct sched_entity *se, int cpu,
10412 struct sched_entity *parent)
10414 struct rq *rq = cpu_rq(cpu);
10418 init_cfs_rq_runtime(cfs_rq);
10420 tg->cfs_rq[cpu] = cfs_rq;
10423 /* se could be NULL for root_task_group */
10428 se->cfs_rq = &rq->cfs;
10431 se->cfs_rq = parent->my_q;
10432 se->depth = parent->depth + 1;
10436 /* guarantee group entities always have weight */
10437 update_load_set(&se->load, NICE_0_LOAD);
10438 se->parent = parent;
10441 static DEFINE_MUTEX(shares_mutex);
10443 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10446 unsigned long flags;
10449 * We can't change the weight of the root cgroup.
10454 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10456 mutex_lock(&shares_mutex);
10457 if (tg->shares == shares)
10460 tg->shares = shares;
10461 for_each_possible_cpu(i) {
10462 struct rq *rq = cpu_rq(i);
10463 struct sched_entity *se;
10466 /* Propagate contribution to hierarchy */
10467 raw_spin_lock_irqsave(&rq->lock, flags);
10469 /* Possible calls to update_curr() need rq clock */
10470 update_rq_clock(rq);
10471 for_each_sched_entity(se) {
10472 update_load_avg(se, UPDATE_TG);
10473 update_cfs_shares(se);
10475 raw_spin_unlock_irqrestore(&rq->lock, flags);
10479 mutex_unlock(&shares_mutex);
10482 #else /* CONFIG_FAIR_GROUP_SCHED */
10484 void free_fair_sched_group(struct task_group *tg) { }
10486 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10491 void unregister_fair_sched_group(struct task_group *tg) { }
10493 #endif /* CONFIG_FAIR_GROUP_SCHED */
10496 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10498 struct sched_entity *se = &task->se;
10499 unsigned int rr_interval = 0;
10502 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10505 if (rq->cfs.load.weight)
10506 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10508 return rr_interval;
10512 * All the scheduling class methods:
10514 const struct sched_class fair_sched_class = {
10515 .next = &idle_sched_class,
10516 .enqueue_task = enqueue_task_fair,
10517 .dequeue_task = dequeue_task_fair,
10518 .yield_task = yield_task_fair,
10519 .yield_to_task = yield_to_task_fair,
10521 .check_preempt_curr = check_preempt_wakeup,
10523 .pick_next_task = pick_next_task_fair,
10524 .put_prev_task = put_prev_task_fair,
10527 .select_task_rq = select_task_rq_fair,
10528 .migrate_task_rq = migrate_task_rq_fair,
10530 .rq_online = rq_online_fair,
10531 .rq_offline = rq_offline_fair,
10533 .task_waking = task_waking_fair,
10534 .task_dead = task_dead_fair,
10535 .set_cpus_allowed = set_cpus_allowed_common,
10538 .set_curr_task = set_curr_task_fair,
10539 .task_tick = task_tick_fair,
10540 .task_fork = task_fork_fair,
10542 .prio_changed = prio_changed_fair,
10543 .switched_from = switched_from_fair,
10544 .switched_to = switched_to_fair,
10546 .get_rr_interval = get_rr_interval_fair,
10548 .update_curr = update_curr_fair,
10550 #ifdef CONFIG_FAIR_GROUP_SCHED
10551 .task_change_group = task_change_group_fair,
10555 #ifdef CONFIG_SCHED_DEBUG
10556 void print_cfs_stats(struct seq_file *m, int cpu)
10558 struct cfs_rq *cfs_rq;
10561 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
10562 print_cfs_rq(m, cpu, cfs_rq);
10566 #ifdef CONFIG_NUMA_BALANCING
10567 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10570 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10572 for_each_online_node(node) {
10573 if (p->numa_faults) {
10574 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10575 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10577 if (p->numa_group) {
10578 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10579 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10581 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10584 #endif /* CONFIG_NUMA_BALANCING */
10585 #endif /* CONFIG_SCHED_DEBUG */
10587 __init void init_sched_fair_class(void)
10590 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10592 #ifdef CONFIG_NO_HZ_COMMON
10593 nohz.next_balance = jiffies;
10594 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10595 cpu_notifier(sched_ilb_notifier, 0);