2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched/mm.h>
24 #include <linux/sched/topology.h>
26 #include <linux/latencytop.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpuidle.h>
29 #include <linux/slab.h>
30 #include <linux/profile.h>
31 #include <linux/interrupt.h>
32 #include <linux/mempolicy.h>
33 #include <linux/migrate.h>
34 #include <linux/task_work.h>
36 #include <trace/events/sched.h>
41 * Targeted preemption latency for CPU-bound tasks:
43 * NOTE: this latency value is not the same as the concept of
44 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
48 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
57 * The initial- and re-scaling of tunables is configurable
61 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
62 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
63 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
65 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
67 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
70 * Minimal preemption granularity for CPU-bound tasks:
72 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
74 unsigned int sysctl_sched_min_granularity = 750000ULL;
75 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
78 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
80 static unsigned int sched_nr_latency = 8;
83 * After fork, child runs first. If set to 0 (default) then
84 * parent will (try to) run first.
86 unsigned int sysctl_sched_child_runs_first __read_mostly;
89 * SCHED_OTHER wake-up granularity.
91 * This option delays the preemption effects of decoupled workloads
92 * and reduces their over-scheduling. Synchronous workloads will still
93 * have immediate wakeup/sleep latencies.
95 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
97 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
98 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
100 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
104 * For asym packing, by default the lower numbered cpu has higher priority.
106 int __weak arch_asym_cpu_priority(int cpu)
112 #ifdef CONFIG_CFS_BANDWIDTH
114 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
115 * each time a cfs_rq requests quota.
117 * Note: in the case that the slice exceeds the runtime remaining (either due
118 * to consumption or the quota being specified to be smaller than the slice)
119 * we will always only issue the remaining available time.
121 * (default: 5 msec, units: microseconds)
123 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
127 * The margin used when comparing utilization with CPU capacity:
128 * util * margin < capacity * 1024
132 unsigned int capacity_margin = 1280;
134 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
140 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
146 static inline void update_load_set(struct load_weight *lw, unsigned long w)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling) {
167 case SCHED_TUNABLESCALING_NONE:
170 case SCHED_TUNABLESCALING_LINEAR:
173 case SCHED_TUNABLESCALING_LOG:
175 factor = 1 + ilog2(cpus);
182 static void update_sysctl(void)
184 unsigned int factor = get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity);
189 SET_SYSCTL(sched_latency);
190 SET_SYSCTL(sched_wakeup_granularity);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight *lw)
206 if (likely(lw->inv_weight))
209 w = scale_load_down(lw->weight);
211 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
213 else if (unlikely(!w))
214 lw->inv_weight = WMULT_CONST;
216 lw->inv_weight = WMULT_CONST / w;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
233 u64 fact = scale_load_down(weight);
234 int shift = WMULT_SHIFT;
236 __update_inv_weight(lw);
238 if (unlikely(fact >> 32)) {
245 /* hint to use a 32x32->64 mul */
246 fact = (u64)(u32)fact * lw->inv_weight;
253 return mul_u64_u32_shr(delta_exec, fact, shift);
257 const struct sched_class fair_sched_class;
259 /**************************************************************
260 * CFS operations on generic schedulable entities:
263 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* cpu runqueue to which this cfs_rq is attached */
266 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
271 /* An entity is a task if it doesn't "own" a runqueue */
272 #define entity_is_task(se) (!se->my_q)
274 static inline struct task_struct *task_of(struct sched_entity *se)
276 SCHED_WARN_ON(!entity_is_task(se));
277 return container_of(se, struct task_struct, se);
280 /* Walk up scheduling entities hierarchy */
281 #define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
284 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
289 /* runqueue on which this entity is (to be) queued */
290 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
295 /* runqueue "owned" by this group */
296 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
301 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
303 if (!cfs_rq->on_list) {
304 struct rq *rq = rq_of(cfs_rq);
305 int cpu = cpu_of(rq);
307 * Ensure we either appear before our parent (if already
308 * enqueued) or force our parent to appear after us when it is
309 * enqueued. The fact that we always enqueue bottom-up
310 * reduces this to two cases and a special case for the root
311 * cfs_rq. Furthermore, it also means that we will always reset
312 * tmp_alone_branch either when the branch is connected
313 * to a tree or when we reach the beg of the tree
315 if (cfs_rq->tg->parent &&
316 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
318 * If parent is already on the list, we add the child
319 * just before. Thanks to circular linked property of
320 * the list, this means to put the child at the tail
321 * of the list that starts by parent.
323 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
324 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
326 * The branch is now connected to its tree so we can
327 * reset tmp_alone_branch to the beginning of the
330 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
331 } else if (!cfs_rq->tg->parent) {
333 * cfs rq without parent should be put
334 * at the tail of the list.
336 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
337 &rq->leaf_cfs_rq_list);
339 * We have reach the beg of a tree so we can reset
340 * tmp_alone_branch to the beginning of the list.
342 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the beg of the branch
348 * where we will add parent.
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
351 rq->tmp_alone_branch);
353 * update tmp_alone_branch to points to the new beg
356 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
363 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
365 if (cfs_rq->on_list) {
366 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
371 /* Iterate thr' all leaf cfs_rq's on a runqueue */
372 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
373 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
375 /* Do the two (enqueued) entities belong to the same group ? */
376 static inline struct cfs_rq *
377 is_same_group(struct sched_entity *se, struct sched_entity *pse)
379 if (se->cfs_rq == pse->cfs_rq)
385 static inline struct sched_entity *parent_entity(struct sched_entity *se)
391 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393 int se_depth, pse_depth;
396 * preemption test can be made between sibling entities who are in the
397 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
398 * both tasks until we find their ancestors who are siblings of common
402 /* First walk up until both entities are at same depth */
403 se_depth = (*se)->depth;
404 pse_depth = (*pse)->depth;
406 while (se_depth > pse_depth) {
408 *se = parent_entity(*se);
411 while (pse_depth > se_depth) {
413 *pse = parent_entity(*pse);
416 while (!is_same_group(*se, *pse)) {
417 *se = parent_entity(*se);
418 *pse = parent_entity(*pse);
422 #else /* !CONFIG_FAIR_GROUP_SCHED */
424 static inline struct task_struct *task_of(struct sched_entity *se)
426 return container_of(se, struct task_struct, se);
429 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
431 return container_of(cfs_rq, struct rq, cfs);
434 #define entity_is_task(se) 1
436 #define for_each_sched_entity(se) \
437 for (; se; se = NULL)
439 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
441 return &task_rq(p)->cfs;
444 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
446 struct task_struct *p = task_of(se);
447 struct rq *rq = task_rq(p);
452 /* runqueue "owned" by this group */
453 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
458 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
462 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
466 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
467 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
469 static inline struct sched_entity *parent_entity(struct sched_entity *se)
475 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
479 #endif /* CONFIG_FAIR_GROUP_SCHED */
481 static __always_inline
482 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
484 /**************************************************************
485 * Scheduling class tree data structure manipulation methods:
488 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
490 s64 delta = (s64)(vruntime - max_vruntime);
492 max_vruntime = vruntime;
497 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
499 s64 delta = (s64)(vruntime - min_vruntime);
501 min_vruntime = vruntime;
506 static inline int entity_before(struct sched_entity *a,
507 struct sched_entity *b)
509 return (s64)(a->vruntime - b->vruntime) < 0;
512 static void update_min_vruntime(struct cfs_rq *cfs_rq)
514 struct sched_entity *curr = cfs_rq->curr;
516 u64 vruntime = cfs_rq->min_vruntime;
520 vruntime = curr->vruntime;
525 if (cfs_rq->rb_leftmost) {
526 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
531 vruntime = se->vruntime;
533 vruntime = min_vruntime(vruntime, se->vruntime);
536 /* ensure we never gain time by being placed backwards. */
537 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
540 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
545 * Enqueue an entity into the rb-tree:
547 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
549 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
550 struct rb_node *parent = NULL;
551 struct sched_entity *entry;
555 * Find the right place in the rbtree:
559 entry = rb_entry(parent, struct sched_entity, run_node);
561 * We dont care about collisions. Nodes with
562 * the same key stay together.
564 if (entity_before(se, entry)) {
565 link = &parent->rb_left;
567 link = &parent->rb_right;
573 * Maintain a cache of leftmost tree entries (it is frequently
577 cfs_rq->rb_leftmost = &se->run_node;
579 rb_link_node(&se->run_node, parent, link);
580 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
583 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
585 if (cfs_rq->rb_leftmost == &se->run_node) {
586 struct rb_node *next_node;
588 next_node = rb_next(&se->run_node);
589 cfs_rq->rb_leftmost = next_node;
592 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
595 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
597 struct rb_node *left = cfs_rq->rb_leftmost;
602 return rb_entry(left, struct sched_entity, run_node);
605 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
607 struct rb_node *next = rb_next(&se->run_node);
612 return rb_entry(next, struct sched_entity, run_node);
615 #ifdef CONFIG_SCHED_DEBUG
616 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
618 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
623 return rb_entry(last, struct sched_entity, run_node);
626 /**************************************************************
627 * Scheduling class statistics methods:
630 int sched_proc_update_handler(struct ctl_table *table, int write,
631 void __user *buffer, size_t *lenp,
634 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
635 unsigned int factor = get_update_sysctl_factor();
640 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
641 sysctl_sched_min_granularity);
643 #define WRT_SYSCTL(name) \
644 (normalized_sysctl_##name = sysctl_##name / (factor))
645 WRT_SYSCTL(sched_min_granularity);
646 WRT_SYSCTL(sched_latency);
647 WRT_SYSCTL(sched_wakeup_granularity);
657 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
659 if (unlikely(se->load.weight != NICE_0_LOAD))
660 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
666 * The idea is to set a period in which each task runs once.
668 * When there are too many tasks (sched_nr_latency) we have to stretch
669 * this period because otherwise the slices get too small.
671 * p = (nr <= nl) ? l : l*nr/nl
673 static u64 __sched_period(unsigned long nr_running)
675 if (unlikely(nr_running > sched_nr_latency))
676 return nr_running * sysctl_sched_min_granularity;
678 return sysctl_sched_latency;
682 * We calculate the wall-time slice from the period by taking a part
683 * proportional to the weight.
687 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
689 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
691 for_each_sched_entity(se) {
692 struct load_weight *load;
693 struct load_weight lw;
695 cfs_rq = cfs_rq_of(se);
696 load = &cfs_rq->load;
698 if (unlikely(!se->on_rq)) {
701 update_load_add(&lw, se->load.weight);
704 slice = __calc_delta(slice, se->load.weight, load);
710 * We calculate the vruntime slice of a to-be-inserted task.
714 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
716 return calc_delta_fair(sched_slice(cfs_rq, se), se);
720 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
721 static unsigned long task_h_load(struct task_struct *p);
724 * We choose a half-life close to 1 scheduling period.
725 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
726 * dependent on this value.
728 #define LOAD_AVG_PERIOD 32
729 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
731 /* Give new sched_entity start runnable values to heavy its load in infant time */
732 void init_entity_runnable_average(struct sched_entity *se)
734 struct sched_avg *sa = &se->avg;
736 sa->last_update_time = 0;
738 * sched_avg's period_contrib should be strictly less then 1024, so
739 * we give it 1023 to make sure it is almost a period (1024us), and
740 * will definitely be update (after enqueue).
742 sa->period_contrib = 1023;
744 * Tasks are intialized with full load to be seen as heavy tasks until
745 * they get a chance to stabilize to their real load level.
746 * Group entities are intialized with zero load to reflect the fact that
747 * nothing has been attached to the task group yet.
749 if (entity_is_task(se))
750 sa->load_avg = scale_load_down(se->load.weight);
751 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
753 * At this point, util_avg won't be used in select_task_rq_fair anyway
757 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
760 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
761 static void attach_entity_cfs_rq(struct sched_entity *se);
764 * With new tasks being created, their initial util_avgs are extrapolated
765 * based on the cfs_rq's current util_avg:
767 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
769 * However, in many cases, the above util_avg does not give a desired
770 * value. Moreover, the sum of the util_avgs may be divergent, such
771 * as when the series is a harmonic series.
773 * To solve this problem, we also cap the util_avg of successive tasks to
774 * only 1/2 of the left utilization budget:
776 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
778 * where n denotes the nth task.
780 * For example, a simplest series from the beginning would be like:
782 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
783 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
785 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
786 * if util_avg > util_avg_cap.
788 void post_init_entity_util_avg(struct sched_entity *se)
790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
791 struct sched_avg *sa = &se->avg;
792 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
795 if (cfs_rq->avg.util_avg != 0) {
796 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
797 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
799 if (sa->util_avg > cap)
804 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
807 if (entity_is_task(se)) {
808 struct task_struct *p = task_of(se);
809 if (p->sched_class != &fair_sched_class) {
811 * For !fair tasks do:
813 update_cfs_rq_load_avg(now, cfs_rq, false);
814 attach_entity_load_avg(cfs_rq, se);
815 switched_from_fair(rq, p);
817 * such that the next switched_to_fair() has the
820 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
825 attach_entity_cfs_rq(se);
828 #else /* !CONFIG_SMP */
829 void init_entity_runnable_average(struct sched_entity *se)
832 void post_init_entity_util_avg(struct sched_entity *se)
835 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
838 #endif /* CONFIG_SMP */
841 * Update the current task's runtime statistics.
843 static void update_curr(struct cfs_rq *cfs_rq)
845 struct sched_entity *curr = cfs_rq->curr;
846 u64 now = rq_clock_task(rq_of(cfs_rq));
852 delta_exec = now - curr->exec_start;
853 if (unlikely((s64)delta_exec <= 0))
856 curr->exec_start = now;
858 schedstat_set(curr->statistics.exec_max,
859 max(delta_exec, curr->statistics.exec_max));
861 curr->sum_exec_runtime += delta_exec;
862 schedstat_add(cfs_rq->exec_clock, delta_exec);
864 curr->vruntime += calc_delta_fair(delta_exec, curr);
865 update_min_vruntime(cfs_rq);
867 if (entity_is_task(curr)) {
868 struct task_struct *curtask = task_of(curr);
870 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
871 cpuacct_charge(curtask, delta_exec);
872 account_group_exec_runtime(curtask, delta_exec);
875 account_cfs_rq_runtime(cfs_rq, delta_exec);
878 static void update_curr_fair(struct rq *rq)
880 update_curr(cfs_rq_of(&rq->curr->se));
884 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
886 u64 wait_start, prev_wait_start;
888 if (!schedstat_enabled())
891 wait_start = rq_clock(rq_of(cfs_rq));
892 prev_wait_start = schedstat_val(se->statistics.wait_start);
894 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
895 likely(wait_start > prev_wait_start))
896 wait_start -= prev_wait_start;
898 schedstat_set(se->statistics.wait_start, wait_start);
902 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
904 struct task_struct *p;
907 if (!schedstat_enabled())
910 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
912 if (entity_is_task(se)) {
914 if (task_on_rq_migrating(p)) {
916 * Preserve migrating task's wait time so wait_start
917 * time stamp can be adjusted to accumulate wait time
918 * prior to migration.
920 schedstat_set(se->statistics.wait_start, delta);
923 trace_sched_stat_wait(p, delta);
926 schedstat_set(se->statistics.wait_max,
927 max(schedstat_val(se->statistics.wait_max), delta));
928 schedstat_inc(se->statistics.wait_count);
929 schedstat_add(se->statistics.wait_sum, delta);
930 schedstat_set(se->statistics.wait_start, 0);
934 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
936 struct task_struct *tsk = NULL;
937 u64 sleep_start, block_start;
939 if (!schedstat_enabled())
942 sleep_start = schedstat_val(se->statistics.sleep_start);
943 block_start = schedstat_val(se->statistics.block_start);
945 if (entity_is_task(se))
949 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
954 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
955 schedstat_set(se->statistics.sleep_max, delta);
957 schedstat_set(se->statistics.sleep_start, 0);
958 schedstat_add(se->statistics.sum_sleep_runtime, delta);
961 account_scheduler_latency(tsk, delta >> 10, 1);
962 trace_sched_stat_sleep(tsk, delta);
966 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
971 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
972 schedstat_set(se->statistics.block_max, delta);
974 schedstat_set(se->statistics.block_start, 0);
975 schedstat_add(se->statistics.sum_sleep_runtime, delta);
978 if (tsk->in_iowait) {
979 schedstat_add(se->statistics.iowait_sum, delta);
980 schedstat_inc(se->statistics.iowait_count);
981 trace_sched_stat_iowait(tsk, delta);
984 trace_sched_stat_blocked(tsk, delta);
987 * Blocking time is in units of nanosecs, so shift by
988 * 20 to get a milliseconds-range estimation of the
989 * amount of time that the task spent sleeping:
991 if (unlikely(prof_on == SLEEP_PROFILING)) {
992 profile_hits(SLEEP_PROFILING,
993 (void *)get_wchan(tsk),
996 account_scheduler_latency(tsk, delta >> 10, 0);
1002 * Task is being enqueued - update stats:
1005 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1007 if (!schedstat_enabled())
1011 * Are we enqueueing a waiting task? (for current tasks
1012 * a dequeue/enqueue event is a NOP)
1014 if (se != cfs_rq->curr)
1015 update_stats_wait_start(cfs_rq, se);
1017 if (flags & ENQUEUE_WAKEUP)
1018 update_stats_enqueue_sleeper(cfs_rq, se);
1022 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1025 if (!schedstat_enabled())
1029 * Mark the end of the wait period if dequeueing a
1032 if (se != cfs_rq->curr)
1033 update_stats_wait_end(cfs_rq, se);
1035 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1036 struct task_struct *tsk = task_of(se);
1038 if (tsk->state & TASK_INTERRUPTIBLE)
1039 schedstat_set(se->statistics.sleep_start,
1040 rq_clock(rq_of(cfs_rq)));
1041 if (tsk->state & TASK_UNINTERRUPTIBLE)
1042 schedstat_set(se->statistics.block_start,
1043 rq_clock(rq_of(cfs_rq)));
1048 * We are picking a new current task - update its stats:
1051 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1054 * We are starting a new run period:
1056 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1059 /**************************************************
1060 * Scheduling class queueing methods:
1063 #ifdef CONFIG_NUMA_BALANCING
1065 * Approximate time to scan a full NUMA task in ms. The task scan period is
1066 * calculated based on the tasks virtual memory size and
1067 * numa_balancing_scan_size.
1069 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1070 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1072 /* Portion of address space to scan in MB */
1073 unsigned int sysctl_numa_balancing_scan_size = 256;
1075 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1076 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1078 static unsigned int task_nr_scan_windows(struct task_struct *p)
1080 unsigned long rss = 0;
1081 unsigned long nr_scan_pages;
1084 * Calculations based on RSS as non-present and empty pages are skipped
1085 * by the PTE scanner and NUMA hinting faults should be trapped based
1088 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1089 rss = get_mm_rss(p->mm);
1091 rss = nr_scan_pages;
1093 rss = round_up(rss, nr_scan_pages);
1094 return rss / nr_scan_pages;
1097 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1098 #define MAX_SCAN_WINDOW 2560
1100 static unsigned int task_scan_min(struct task_struct *p)
1102 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1103 unsigned int scan, floor;
1104 unsigned int windows = 1;
1106 if (scan_size < MAX_SCAN_WINDOW)
1107 windows = MAX_SCAN_WINDOW / scan_size;
1108 floor = 1000 / windows;
1110 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1111 return max_t(unsigned int, floor, scan);
1114 static unsigned int task_scan_max(struct task_struct *p)
1116 unsigned int smin = task_scan_min(p);
1119 /* Watch for min being lower than max due to floor calculations */
1120 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1121 return max(smin, smax);
1124 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1126 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1127 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1130 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1132 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1133 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1139 spinlock_t lock; /* nr_tasks, tasks */
1144 struct rcu_head rcu;
1145 unsigned long total_faults;
1146 unsigned long max_faults_cpu;
1148 * Faults_cpu is used to decide whether memory should move
1149 * towards the CPU. As a consequence, these stats are weighted
1150 * more by CPU use than by memory faults.
1152 unsigned long *faults_cpu;
1153 unsigned long faults[0];
1156 /* Shared or private faults. */
1157 #define NR_NUMA_HINT_FAULT_TYPES 2
1159 /* Memory and CPU locality */
1160 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1162 /* Averaged statistics, and temporary buffers. */
1163 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1165 pid_t task_numa_group_id(struct task_struct *p)
1167 return p->numa_group ? p->numa_group->gid : 0;
1171 * The averaged statistics, shared & private, memory & cpu,
1172 * occupy the first half of the array. The second half of the
1173 * array is for current counters, which are averaged into the
1174 * first set by task_numa_placement.
1176 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1178 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1181 static inline unsigned long task_faults(struct task_struct *p, int nid)
1183 if (!p->numa_faults)
1186 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1187 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1190 static inline unsigned long group_faults(struct task_struct *p, int nid)
1195 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1196 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1199 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1201 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1202 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1206 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1207 * considered part of a numa group's pseudo-interleaving set. Migrations
1208 * between these nodes are slowed down, to allow things to settle down.
1210 #define ACTIVE_NODE_FRACTION 3
1212 static bool numa_is_active_node(int nid, struct numa_group *ng)
1214 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1217 /* Handle placement on systems where not all nodes are directly connected. */
1218 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1219 int maxdist, bool task)
1221 unsigned long score = 0;
1225 * All nodes are directly connected, and the same distance
1226 * from each other. No need for fancy placement algorithms.
1228 if (sched_numa_topology_type == NUMA_DIRECT)
1232 * This code is called for each node, introducing N^2 complexity,
1233 * which should be ok given the number of nodes rarely exceeds 8.
1235 for_each_online_node(node) {
1236 unsigned long faults;
1237 int dist = node_distance(nid, node);
1240 * The furthest away nodes in the system are not interesting
1241 * for placement; nid was already counted.
1243 if (dist == sched_max_numa_distance || node == nid)
1247 * On systems with a backplane NUMA topology, compare groups
1248 * of nodes, and move tasks towards the group with the most
1249 * memory accesses. When comparing two nodes at distance
1250 * "hoplimit", only nodes closer by than "hoplimit" are part
1251 * of each group. Skip other nodes.
1253 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1257 /* Add up the faults from nearby nodes. */
1259 faults = task_faults(p, node);
1261 faults = group_faults(p, node);
1264 * On systems with a glueless mesh NUMA topology, there are
1265 * no fixed "groups of nodes". Instead, nodes that are not
1266 * directly connected bounce traffic through intermediate
1267 * nodes; a numa_group can occupy any set of nodes.
1268 * The further away a node is, the less the faults count.
1269 * This seems to result in good task placement.
1271 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1272 faults *= (sched_max_numa_distance - dist);
1273 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1283 * These return the fraction of accesses done by a particular task, or
1284 * task group, on a particular numa node. The group weight is given a
1285 * larger multiplier, in order to group tasks together that are almost
1286 * evenly spread out between numa nodes.
1288 static inline unsigned long task_weight(struct task_struct *p, int nid,
1291 unsigned long faults, total_faults;
1293 if (!p->numa_faults)
1296 total_faults = p->total_numa_faults;
1301 faults = task_faults(p, nid);
1302 faults += score_nearby_nodes(p, nid, dist, true);
1304 return 1000 * faults / total_faults;
1307 static inline unsigned long group_weight(struct task_struct *p, int nid,
1310 unsigned long faults, total_faults;
1315 total_faults = p->numa_group->total_faults;
1320 faults = group_faults(p, nid);
1321 faults += score_nearby_nodes(p, nid, dist, false);
1323 return 1000 * faults / total_faults;
1326 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1327 int src_nid, int dst_cpu)
1329 struct numa_group *ng = p->numa_group;
1330 int dst_nid = cpu_to_node(dst_cpu);
1331 int last_cpupid, this_cpupid;
1333 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1336 * Multi-stage node selection is used in conjunction with a periodic
1337 * migration fault to build a temporal task<->page relation. By using
1338 * a two-stage filter we remove short/unlikely relations.
1340 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1341 * a task's usage of a particular page (n_p) per total usage of this
1342 * page (n_t) (in a given time-span) to a probability.
1344 * Our periodic faults will sample this probability and getting the
1345 * same result twice in a row, given these samples are fully
1346 * independent, is then given by P(n)^2, provided our sample period
1347 * is sufficiently short compared to the usage pattern.
1349 * This quadric squishes small probabilities, making it less likely we
1350 * act on an unlikely task<->page relation.
1352 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1353 if (!cpupid_pid_unset(last_cpupid) &&
1354 cpupid_to_nid(last_cpupid) != dst_nid)
1357 /* Always allow migrate on private faults */
1358 if (cpupid_match_pid(p, last_cpupid))
1361 /* A shared fault, but p->numa_group has not been set up yet. */
1366 * Destination node is much more heavily used than the source
1367 * node? Allow migration.
1369 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1370 ACTIVE_NODE_FRACTION)
1374 * Distribute memory according to CPU & memory use on each node,
1375 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1377 * faults_cpu(dst) 3 faults_cpu(src)
1378 * --------------- * - > ---------------
1379 * faults_mem(dst) 4 faults_mem(src)
1381 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1382 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1385 static unsigned long weighted_cpuload(const int cpu);
1386 static unsigned long source_load(int cpu, int type);
1387 static unsigned long target_load(int cpu, int type);
1388 static unsigned long capacity_of(int cpu);
1389 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1391 /* Cached statistics for all CPUs within a node */
1393 unsigned long nr_running;
1396 /* Total compute capacity of CPUs on a node */
1397 unsigned long compute_capacity;
1399 /* Approximate capacity in terms of runnable tasks on a node */
1400 unsigned long task_capacity;
1401 int has_free_capacity;
1405 * XXX borrowed from update_sg_lb_stats
1407 static void update_numa_stats(struct numa_stats *ns, int nid)
1409 int smt, cpu, cpus = 0;
1410 unsigned long capacity;
1412 memset(ns, 0, sizeof(*ns));
1413 for_each_cpu(cpu, cpumask_of_node(nid)) {
1414 struct rq *rq = cpu_rq(cpu);
1416 ns->nr_running += rq->nr_running;
1417 ns->load += weighted_cpuload(cpu);
1418 ns->compute_capacity += capacity_of(cpu);
1424 * If we raced with hotplug and there are no CPUs left in our mask
1425 * the @ns structure is NULL'ed and task_numa_compare() will
1426 * not find this node attractive.
1428 * We'll either bail at !has_free_capacity, or we'll detect a huge
1429 * imbalance and bail there.
1434 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1435 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1436 capacity = cpus / smt; /* cores */
1438 ns->task_capacity = min_t(unsigned, capacity,
1439 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1440 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1443 struct task_numa_env {
1444 struct task_struct *p;
1446 int src_cpu, src_nid;
1447 int dst_cpu, dst_nid;
1449 struct numa_stats src_stats, dst_stats;
1454 struct task_struct *best_task;
1459 static void task_numa_assign(struct task_numa_env *env,
1460 struct task_struct *p, long imp)
1463 put_task_struct(env->best_task);
1468 env->best_imp = imp;
1469 env->best_cpu = env->dst_cpu;
1472 static bool load_too_imbalanced(long src_load, long dst_load,
1473 struct task_numa_env *env)
1476 long orig_src_load, orig_dst_load;
1477 long src_capacity, dst_capacity;
1480 * The load is corrected for the CPU capacity available on each node.
1483 * ------------ vs ---------
1484 * src_capacity dst_capacity
1486 src_capacity = env->src_stats.compute_capacity;
1487 dst_capacity = env->dst_stats.compute_capacity;
1489 /* We care about the slope of the imbalance, not the direction. */
1490 if (dst_load < src_load)
1491 swap(dst_load, src_load);
1493 /* Is the difference below the threshold? */
1494 imb = dst_load * src_capacity * 100 -
1495 src_load * dst_capacity * env->imbalance_pct;
1500 * The imbalance is above the allowed threshold.
1501 * Compare it with the old imbalance.
1503 orig_src_load = env->src_stats.load;
1504 orig_dst_load = env->dst_stats.load;
1506 if (orig_dst_load < orig_src_load)
1507 swap(orig_dst_load, orig_src_load);
1509 old_imb = orig_dst_load * src_capacity * 100 -
1510 orig_src_load * dst_capacity * env->imbalance_pct;
1512 /* Would this change make things worse? */
1513 return (imb > old_imb);
1517 * This checks if the overall compute and NUMA accesses of the system would
1518 * be improved if the source tasks was migrated to the target dst_cpu taking
1519 * into account that it might be best if task running on the dst_cpu should
1520 * be exchanged with the source task
1522 static void task_numa_compare(struct task_numa_env *env,
1523 long taskimp, long groupimp)
1525 struct rq *src_rq = cpu_rq(env->src_cpu);
1526 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1527 struct task_struct *cur;
1528 long src_load, dst_load;
1530 long imp = env->p->numa_group ? groupimp : taskimp;
1532 int dist = env->dist;
1535 cur = task_rcu_dereference(&dst_rq->curr);
1536 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1540 * Because we have preemption enabled we can get migrated around and
1541 * end try selecting ourselves (current == env->p) as a swap candidate.
1547 * "imp" is the fault differential for the source task between the
1548 * source and destination node. Calculate the total differential for
1549 * the source task and potential destination task. The more negative
1550 * the value is, the more rmeote accesses that would be expected to
1551 * be incurred if the tasks were swapped.
1554 /* Skip this swap candidate if cannot move to the source cpu */
1555 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1559 * If dst and source tasks are in the same NUMA group, or not
1560 * in any group then look only at task weights.
1562 if (cur->numa_group == env->p->numa_group) {
1563 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1564 task_weight(cur, env->dst_nid, dist);
1566 * Add some hysteresis to prevent swapping the
1567 * tasks within a group over tiny differences.
1569 if (cur->numa_group)
1573 * Compare the group weights. If a task is all by
1574 * itself (not part of a group), use the task weight
1577 if (cur->numa_group)
1578 imp += group_weight(cur, env->src_nid, dist) -
1579 group_weight(cur, env->dst_nid, dist);
1581 imp += task_weight(cur, env->src_nid, dist) -
1582 task_weight(cur, env->dst_nid, dist);
1586 if (imp <= env->best_imp && moveimp <= env->best_imp)
1590 /* Is there capacity at our destination? */
1591 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1592 !env->dst_stats.has_free_capacity)
1598 /* Balance doesn't matter much if we're running a task per cpu */
1599 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1600 dst_rq->nr_running == 1)
1604 * In the overloaded case, try and keep the load balanced.
1607 load = task_h_load(env->p);
1608 dst_load = env->dst_stats.load + load;
1609 src_load = env->src_stats.load - load;
1611 if (moveimp > imp && moveimp > env->best_imp) {
1613 * If the improvement from just moving env->p direction is
1614 * better than swapping tasks around, check if a move is
1615 * possible. Store a slightly smaller score than moveimp,
1616 * so an actually idle CPU will win.
1618 if (!load_too_imbalanced(src_load, dst_load, env)) {
1625 if (imp <= env->best_imp)
1629 load = task_h_load(cur);
1634 if (load_too_imbalanced(src_load, dst_load, env))
1638 * One idle CPU per node is evaluated for a task numa move.
1639 * Call select_idle_sibling to maybe find a better one.
1643 * select_idle_siblings() uses an per-cpu cpumask that
1644 * can be used from IRQ context.
1646 local_irq_disable();
1647 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1653 task_numa_assign(env, cur, imp);
1658 static void task_numa_find_cpu(struct task_numa_env *env,
1659 long taskimp, long groupimp)
1663 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1664 /* Skip this CPU if the source task cannot migrate */
1665 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1669 task_numa_compare(env, taskimp, groupimp);
1673 /* Only move tasks to a NUMA node less busy than the current node. */
1674 static bool numa_has_capacity(struct task_numa_env *env)
1676 struct numa_stats *src = &env->src_stats;
1677 struct numa_stats *dst = &env->dst_stats;
1679 if (src->has_free_capacity && !dst->has_free_capacity)
1683 * Only consider a task move if the source has a higher load
1684 * than the destination, corrected for CPU capacity on each node.
1686 * src->load dst->load
1687 * --------------------- vs ---------------------
1688 * src->compute_capacity dst->compute_capacity
1690 if (src->load * dst->compute_capacity * env->imbalance_pct >
1692 dst->load * src->compute_capacity * 100)
1698 static int task_numa_migrate(struct task_struct *p)
1700 struct task_numa_env env = {
1703 .src_cpu = task_cpu(p),
1704 .src_nid = task_node(p),
1706 .imbalance_pct = 112,
1712 struct sched_domain *sd;
1713 unsigned long taskweight, groupweight;
1715 long taskimp, groupimp;
1718 * Pick the lowest SD_NUMA domain, as that would have the smallest
1719 * imbalance and would be the first to start moving tasks about.
1721 * And we want to avoid any moving of tasks about, as that would create
1722 * random movement of tasks -- counter the numa conditions we're trying
1726 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1728 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1732 * Cpusets can break the scheduler domain tree into smaller
1733 * balance domains, some of which do not cross NUMA boundaries.
1734 * Tasks that are "trapped" in such domains cannot be migrated
1735 * elsewhere, so there is no point in (re)trying.
1737 if (unlikely(!sd)) {
1738 p->numa_preferred_nid = task_node(p);
1742 env.dst_nid = p->numa_preferred_nid;
1743 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1744 taskweight = task_weight(p, env.src_nid, dist);
1745 groupweight = group_weight(p, env.src_nid, dist);
1746 update_numa_stats(&env.src_stats, env.src_nid);
1747 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1748 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1749 update_numa_stats(&env.dst_stats, env.dst_nid);
1751 /* Try to find a spot on the preferred nid. */
1752 if (numa_has_capacity(&env))
1753 task_numa_find_cpu(&env, taskimp, groupimp);
1756 * Look at other nodes in these cases:
1757 * - there is no space available on the preferred_nid
1758 * - the task is part of a numa_group that is interleaved across
1759 * multiple NUMA nodes; in order to better consolidate the group,
1760 * we need to check other locations.
1762 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1763 for_each_online_node(nid) {
1764 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1767 dist = node_distance(env.src_nid, env.dst_nid);
1768 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1770 taskweight = task_weight(p, env.src_nid, dist);
1771 groupweight = group_weight(p, env.src_nid, dist);
1774 /* Only consider nodes where both task and groups benefit */
1775 taskimp = task_weight(p, nid, dist) - taskweight;
1776 groupimp = group_weight(p, nid, dist) - groupweight;
1777 if (taskimp < 0 && groupimp < 0)
1782 update_numa_stats(&env.dst_stats, env.dst_nid);
1783 if (numa_has_capacity(&env))
1784 task_numa_find_cpu(&env, taskimp, groupimp);
1789 * If the task is part of a workload that spans multiple NUMA nodes,
1790 * and is migrating into one of the workload's active nodes, remember
1791 * this node as the task's preferred numa node, so the workload can
1793 * A task that migrated to a second choice node will be better off
1794 * trying for a better one later. Do not set the preferred node here.
1796 if (p->numa_group) {
1797 struct numa_group *ng = p->numa_group;
1799 if (env.best_cpu == -1)
1804 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1805 sched_setnuma(p, env.dst_nid);
1808 /* No better CPU than the current one was found. */
1809 if (env.best_cpu == -1)
1813 * Reset the scan period if the task is being rescheduled on an
1814 * alternative node to recheck if the tasks is now properly placed.
1816 p->numa_scan_period = task_scan_min(p);
1818 if (env.best_task == NULL) {
1819 ret = migrate_task_to(p, env.best_cpu);
1821 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1825 ret = migrate_swap(p, env.best_task);
1827 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1828 put_task_struct(env.best_task);
1832 /* Attempt to migrate a task to a CPU on the preferred node. */
1833 static void numa_migrate_preferred(struct task_struct *p)
1835 unsigned long interval = HZ;
1837 /* This task has no NUMA fault statistics yet */
1838 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1841 /* Periodically retry migrating the task to the preferred node */
1842 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1843 p->numa_migrate_retry = jiffies + interval;
1845 /* Success if task is already running on preferred CPU */
1846 if (task_node(p) == p->numa_preferred_nid)
1849 /* Otherwise, try migrate to a CPU on the preferred node */
1850 task_numa_migrate(p);
1854 * Find out how many nodes on the workload is actively running on. Do this by
1855 * tracking the nodes from which NUMA hinting faults are triggered. This can
1856 * be different from the set of nodes where the workload's memory is currently
1859 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1861 unsigned long faults, max_faults = 0;
1862 int nid, active_nodes = 0;
1864 for_each_online_node(nid) {
1865 faults = group_faults_cpu(numa_group, nid);
1866 if (faults > max_faults)
1867 max_faults = faults;
1870 for_each_online_node(nid) {
1871 faults = group_faults_cpu(numa_group, nid);
1872 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1876 numa_group->max_faults_cpu = max_faults;
1877 numa_group->active_nodes = active_nodes;
1881 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1882 * increments. The more local the fault statistics are, the higher the scan
1883 * period will be for the next scan window. If local/(local+remote) ratio is
1884 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1885 * the scan period will decrease. Aim for 70% local accesses.
1887 #define NUMA_PERIOD_SLOTS 10
1888 #define NUMA_PERIOD_THRESHOLD 7
1891 * Increase the scan period (slow down scanning) if the majority of
1892 * our memory is already on our local node, or if the majority of
1893 * the page accesses are shared with other processes.
1894 * Otherwise, decrease the scan period.
1896 static void update_task_scan_period(struct task_struct *p,
1897 unsigned long shared, unsigned long private)
1899 unsigned int period_slot;
1903 unsigned long remote = p->numa_faults_locality[0];
1904 unsigned long local = p->numa_faults_locality[1];
1907 * If there were no record hinting faults then either the task is
1908 * completely idle or all activity is areas that are not of interest
1909 * to automatic numa balancing. Related to that, if there were failed
1910 * migration then it implies we are migrating too quickly or the local
1911 * node is overloaded. In either case, scan slower
1913 if (local + shared == 0 || p->numa_faults_locality[2]) {
1914 p->numa_scan_period = min(p->numa_scan_period_max,
1915 p->numa_scan_period << 1);
1917 p->mm->numa_next_scan = jiffies +
1918 msecs_to_jiffies(p->numa_scan_period);
1924 * Prepare to scale scan period relative to the current period.
1925 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1926 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1927 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1929 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1930 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1931 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1932 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1935 diff = slot * period_slot;
1937 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1940 * Scale scan rate increases based on sharing. There is an
1941 * inverse relationship between the degree of sharing and
1942 * the adjustment made to the scanning period. Broadly
1943 * speaking the intent is that there is little point
1944 * scanning faster if shared accesses dominate as it may
1945 * simply bounce migrations uselessly
1947 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1948 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1951 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1952 task_scan_min(p), task_scan_max(p));
1953 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1957 * Get the fraction of time the task has been running since the last
1958 * NUMA placement cycle. The scheduler keeps similar statistics, but
1959 * decays those on a 32ms period, which is orders of magnitude off
1960 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1961 * stats only if the task is so new there are no NUMA statistics yet.
1963 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1965 u64 runtime, delta, now;
1966 /* Use the start of this time slice to avoid calculations. */
1967 now = p->se.exec_start;
1968 runtime = p->se.sum_exec_runtime;
1970 if (p->last_task_numa_placement) {
1971 delta = runtime - p->last_sum_exec_runtime;
1972 *period = now - p->last_task_numa_placement;
1974 delta = p->se.avg.load_sum / p->se.load.weight;
1975 *period = LOAD_AVG_MAX;
1978 p->last_sum_exec_runtime = runtime;
1979 p->last_task_numa_placement = now;
1985 * Determine the preferred nid for a task in a numa_group. This needs to
1986 * be done in a way that produces consistent results with group_weight,
1987 * otherwise workloads might not converge.
1989 static int preferred_group_nid(struct task_struct *p, int nid)
1994 /* Direct connections between all NUMA nodes. */
1995 if (sched_numa_topology_type == NUMA_DIRECT)
1999 * On a system with glueless mesh NUMA topology, group_weight
2000 * scores nodes according to the number of NUMA hinting faults on
2001 * both the node itself, and on nearby nodes.
2003 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2004 unsigned long score, max_score = 0;
2005 int node, max_node = nid;
2007 dist = sched_max_numa_distance;
2009 for_each_online_node(node) {
2010 score = group_weight(p, node, dist);
2011 if (score > max_score) {
2020 * Finding the preferred nid in a system with NUMA backplane
2021 * interconnect topology is more involved. The goal is to locate
2022 * tasks from numa_groups near each other in the system, and
2023 * untangle workloads from different sides of the system. This requires
2024 * searching down the hierarchy of node groups, recursively searching
2025 * inside the highest scoring group of nodes. The nodemask tricks
2026 * keep the complexity of the search down.
2028 nodes = node_online_map;
2029 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2030 unsigned long max_faults = 0;
2031 nodemask_t max_group = NODE_MASK_NONE;
2034 /* Are there nodes at this distance from each other? */
2035 if (!find_numa_distance(dist))
2038 for_each_node_mask(a, nodes) {
2039 unsigned long faults = 0;
2040 nodemask_t this_group;
2041 nodes_clear(this_group);
2043 /* Sum group's NUMA faults; includes a==b case. */
2044 for_each_node_mask(b, nodes) {
2045 if (node_distance(a, b) < dist) {
2046 faults += group_faults(p, b);
2047 node_set(b, this_group);
2048 node_clear(b, nodes);
2052 /* Remember the top group. */
2053 if (faults > max_faults) {
2054 max_faults = faults;
2055 max_group = this_group;
2057 * subtle: at the smallest distance there is
2058 * just one node left in each "group", the
2059 * winner is the preferred nid.
2064 /* Next round, evaluate the nodes within max_group. */
2072 static void task_numa_placement(struct task_struct *p)
2074 int seq, nid, max_nid = -1, max_group_nid = -1;
2075 unsigned long max_faults = 0, max_group_faults = 0;
2076 unsigned long fault_types[2] = { 0, 0 };
2077 unsigned long total_faults;
2078 u64 runtime, period;
2079 spinlock_t *group_lock = NULL;
2082 * The p->mm->numa_scan_seq field gets updated without
2083 * exclusive access. Use READ_ONCE() here to ensure
2084 * that the field is read in a single access:
2086 seq = READ_ONCE(p->mm->numa_scan_seq);
2087 if (p->numa_scan_seq == seq)
2089 p->numa_scan_seq = seq;
2090 p->numa_scan_period_max = task_scan_max(p);
2092 total_faults = p->numa_faults_locality[0] +
2093 p->numa_faults_locality[1];
2094 runtime = numa_get_avg_runtime(p, &period);
2096 /* If the task is part of a group prevent parallel updates to group stats */
2097 if (p->numa_group) {
2098 group_lock = &p->numa_group->lock;
2099 spin_lock_irq(group_lock);
2102 /* Find the node with the highest number of faults */
2103 for_each_online_node(nid) {
2104 /* Keep track of the offsets in numa_faults array */
2105 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2106 unsigned long faults = 0, group_faults = 0;
2109 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2110 long diff, f_diff, f_weight;
2112 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2113 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2114 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2115 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2117 /* Decay existing window, copy faults since last scan */
2118 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2119 fault_types[priv] += p->numa_faults[membuf_idx];
2120 p->numa_faults[membuf_idx] = 0;
2123 * Normalize the faults_from, so all tasks in a group
2124 * count according to CPU use, instead of by the raw
2125 * number of faults. Tasks with little runtime have
2126 * little over-all impact on throughput, and thus their
2127 * faults are less important.
2129 f_weight = div64_u64(runtime << 16, period + 1);
2130 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2132 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2133 p->numa_faults[cpubuf_idx] = 0;
2135 p->numa_faults[mem_idx] += diff;
2136 p->numa_faults[cpu_idx] += f_diff;
2137 faults += p->numa_faults[mem_idx];
2138 p->total_numa_faults += diff;
2139 if (p->numa_group) {
2141 * safe because we can only change our own group
2143 * mem_idx represents the offset for a given
2144 * nid and priv in a specific region because it
2145 * is at the beginning of the numa_faults array.
2147 p->numa_group->faults[mem_idx] += diff;
2148 p->numa_group->faults_cpu[mem_idx] += f_diff;
2149 p->numa_group->total_faults += diff;
2150 group_faults += p->numa_group->faults[mem_idx];
2154 if (faults > max_faults) {
2155 max_faults = faults;
2159 if (group_faults > max_group_faults) {
2160 max_group_faults = group_faults;
2161 max_group_nid = nid;
2165 update_task_scan_period(p, fault_types[0], fault_types[1]);
2167 if (p->numa_group) {
2168 numa_group_count_active_nodes(p->numa_group);
2169 spin_unlock_irq(group_lock);
2170 max_nid = preferred_group_nid(p, max_group_nid);
2174 /* Set the new preferred node */
2175 if (max_nid != p->numa_preferred_nid)
2176 sched_setnuma(p, max_nid);
2178 if (task_node(p) != p->numa_preferred_nid)
2179 numa_migrate_preferred(p);
2183 static inline int get_numa_group(struct numa_group *grp)
2185 return atomic_inc_not_zero(&grp->refcount);
2188 static inline void put_numa_group(struct numa_group *grp)
2190 if (atomic_dec_and_test(&grp->refcount))
2191 kfree_rcu(grp, rcu);
2194 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2197 struct numa_group *grp, *my_grp;
2198 struct task_struct *tsk;
2200 int cpu = cpupid_to_cpu(cpupid);
2203 if (unlikely(!p->numa_group)) {
2204 unsigned int size = sizeof(struct numa_group) +
2205 4*nr_node_ids*sizeof(unsigned long);
2207 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2211 atomic_set(&grp->refcount, 1);
2212 grp->active_nodes = 1;
2213 grp->max_faults_cpu = 0;
2214 spin_lock_init(&grp->lock);
2216 /* Second half of the array tracks nids where faults happen */
2217 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2220 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2221 grp->faults[i] = p->numa_faults[i];
2223 grp->total_faults = p->total_numa_faults;
2226 rcu_assign_pointer(p->numa_group, grp);
2230 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2232 if (!cpupid_match_pid(tsk, cpupid))
2235 grp = rcu_dereference(tsk->numa_group);
2239 my_grp = p->numa_group;
2244 * Only join the other group if its bigger; if we're the bigger group,
2245 * the other task will join us.
2247 if (my_grp->nr_tasks > grp->nr_tasks)
2251 * Tie-break on the grp address.
2253 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2256 /* Always join threads in the same process. */
2257 if (tsk->mm == current->mm)
2260 /* Simple filter to avoid false positives due to PID collisions */
2261 if (flags & TNF_SHARED)
2264 /* Update priv based on whether false sharing was detected */
2267 if (join && !get_numa_group(grp))
2275 BUG_ON(irqs_disabled());
2276 double_lock_irq(&my_grp->lock, &grp->lock);
2278 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2279 my_grp->faults[i] -= p->numa_faults[i];
2280 grp->faults[i] += p->numa_faults[i];
2282 my_grp->total_faults -= p->total_numa_faults;
2283 grp->total_faults += p->total_numa_faults;
2288 spin_unlock(&my_grp->lock);
2289 spin_unlock_irq(&grp->lock);
2291 rcu_assign_pointer(p->numa_group, grp);
2293 put_numa_group(my_grp);
2301 void task_numa_free(struct task_struct *p)
2303 struct numa_group *grp = p->numa_group;
2304 void *numa_faults = p->numa_faults;
2305 unsigned long flags;
2309 spin_lock_irqsave(&grp->lock, flags);
2310 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2311 grp->faults[i] -= p->numa_faults[i];
2312 grp->total_faults -= p->total_numa_faults;
2315 spin_unlock_irqrestore(&grp->lock, flags);
2316 RCU_INIT_POINTER(p->numa_group, NULL);
2317 put_numa_group(grp);
2320 p->numa_faults = NULL;
2325 * Got a PROT_NONE fault for a page on @node.
2327 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2329 struct task_struct *p = current;
2330 bool migrated = flags & TNF_MIGRATED;
2331 int cpu_node = task_node(current);
2332 int local = !!(flags & TNF_FAULT_LOCAL);
2333 struct numa_group *ng;
2336 if (!static_branch_likely(&sched_numa_balancing))
2339 /* for example, ksmd faulting in a user's mm */
2343 /* Allocate buffer to track faults on a per-node basis */
2344 if (unlikely(!p->numa_faults)) {
2345 int size = sizeof(*p->numa_faults) *
2346 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2348 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2349 if (!p->numa_faults)
2352 p->total_numa_faults = 0;
2353 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2357 * First accesses are treated as private, otherwise consider accesses
2358 * to be private if the accessing pid has not changed
2360 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2363 priv = cpupid_match_pid(p, last_cpupid);
2364 if (!priv && !(flags & TNF_NO_GROUP))
2365 task_numa_group(p, last_cpupid, flags, &priv);
2369 * If a workload spans multiple NUMA nodes, a shared fault that
2370 * occurs wholly within the set of nodes that the workload is
2371 * actively using should be counted as local. This allows the
2372 * scan rate to slow down when a workload has settled down.
2375 if (!priv && !local && ng && ng->active_nodes > 1 &&
2376 numa_is_active_node(cpu_node, ng) &&
2377 numa_is_active_node(mem_node, ng))
2380 task_numa_placement(p);
2383 * Retry task to preferred node migration periodically, in case it
2384 * case it previously failed, or the scheduler moved us.
2386 if (time_after(jiffies, p->numa_migrate_retry))
2387 numa_migrate_preferred(p);
2390 p->numa_pages_migrated += pages;
2391 if (flags & TNF_MIGRATE_FAIL)
2392 p->numa_faults_locality[2] += pages;
2394 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2395 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2396 p->numa_faults_locality[local] += pages;
2399 static void reset_ptenuma_scan(struct task_struct *p)
2402 * We only did a read acquisition of the mmap sem, so
2403 * p->mm->numa_scan_seq is written to without exclusive access
2404 * and the update is not guaranteed to be atomic. That's not
2405 * much of an issue though, since this is just used for
2406 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2407 * expensive, to avoid any form of compiler optimizations:
2409 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2410 p->mm->numa_scan_offset = 0;
2414 * The expensive part of numa migration is done from task_work context.
2415 * Triggered from task_tick_numa().
2417 void task_numa_work(struct callback_head *work)
2419 unsigned long migrate, next_scan, now = jiffies;
2420 struct task_struct *p = current;
2421 struct mm_struct *mm = p->mm;
2422 u64 runtime = p->se.sum_exec_runtime;
2423 struct vm_area_struct *vma;
2424 unsigned long start, end;
2425 unsigned long nr_pte_updates = 0;
2426 long pages, virtpages;
2428 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2430 work->next = work; /* protect against double add */
2432 * Who cares about NUMA placement when they're dying.
2434 * NOTE: make sure not to dereference p->mm before this check,
2435 * exit_task_work() happens _after_ exit_mm() so we could be called
2436 * without p->mm even though we still had it when we enqueued this
2439 if (p->flags & PF_EXITING)
2442 if (!mm->numa_next_scan) {
2443 mm->numa_next_scan = now +
2444 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2448 * Enforce maximal scan/migration frequency..
2450 migrate = mm->numa_next_scan;
2451 if (time_before(now, migrate))
2454 if (p->numa_scan_period == 0) {
2455 p->numa_scan_period_max = task_scan_max(p);
2456 p->numa_scan_period = task_scan_min(p);
2459 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2460 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2464 * Delay this task enough that another task of this mm will likely win
2465 * the next time around.
2467 p->node_stamp += 2 * TICK_NSEC;
2469 start = mm->numa_scan_offset;
2470 pages = sysctl_numa_balancing_scan_size;
2471 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2472 virtpages = pages * 8; /* Scan up to this much virtual space */
2477 down_read(&mm->mmap_sem);
2478 vma = find_vma(mm, start);
2480 reset_ptenuma_scan(p);
2484 for (; vma; vma = vma->vm_next) {
2485 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2486 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2491 * Shared library pages mapped by multiple processes are not
2492 * migrated as it is expected they are cache replicated. Avoid
2493 * hinting faults in read-only file-backed mappings or the vdso
2494 * as migrating the pages will be of marginal benefit.
2497 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2501 * Skip inaccessible VMAs to avoid any confusion between
2502 * PROT_NONE and NUMA hinting ptes
2504 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2508 start = max(start, vma->vm_start);
2509 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2510 end = min(end, vma->vm_end);
2511 nr_pte_updates = change_prot_numa(vma, start, end);
2514 * Try to scan sysctl_numa_balancing_size worth of
2515 * hpages that have at least one present PTE that
2516 * is not already pte-numa. If the VMA contains
2517 * areas that are unused or already full of prot_numa
2518 * PTEs, scan up to virtpages, to skip through those
2522 pages -= (end - start) >> PAGE_SHIFT;
2523 virtpages -= (end - start) >> PAGE_SHIFT;
2526 if (pages <= 0 || virtpages <= 0)
2530 } while (end != vma->vm_end);
2535 * It is possible to reach the end of the VMA list but the last few
2536 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2537 * would find the !migratable VMA on the next scan but not reset the
2538 * scanner to the start so check it now.
2541 mm->numa_scan_offset = start;
2543 reset_ptenuma_scan(p);
2544 up_read(&mm->mmap_sem);
2547 * Make sure tasks use at least 32x as much time to run other code
2548 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2549 * Usually update_task_scan_period slows down scanning enough; on an
2550 * overloaded system we need to limit overhead on a per task basis.
2552 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2553 u64 diff = p->se.sum_exec_runtime - runtime;
2554 p->node_stamp += 32 * diff;
2559 * Drive the periodic memory faults..
2561 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2563 struct callback_head *work = &curr->numa_work;
2567 * We don't care about NUMA placement if we don't have memory.
2569 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2573 * Using runtime rather than walltime has the dual advantage that
2574 * we (mostly) drive the selection from busy threads and that the
2575 * task needs to have done some actual work before we bother with
2578 now = curr->se.sum_exec_runtime;
2579 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2581 if (now > curr->node_stamp + period) {
2582 if (!curr->node_stamp)
2583 curr->numa_scan_period = task_scan_min(curr);
2584 curr->node_stamp += period;
2586 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2587 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2588 task_work_add(curr, work, true);
2593 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2597 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2601 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2604 #endif /* CONFIG_NUMA_BALANCING */
2607 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2609 update_load_add(&cfs_rq->load, se->load.weight);
2610 if (!parent_entity(se))
2611 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2613 if (entity_is_task(se)) {
2614 struct rq *rq = rq_of(cfs_rq);
2616 account_numa_enqueue(rq, task_of(se));
2617 list_add(&se->group_node, &rq->cfs_tasks);
2620 cfs_rq->nr_running++;
2624 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2626 update_load_sub(&cfs_rq->load, se->load.weight);
2627 if (!parent_entity(se))
2628 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2630 if (entity_is_task(se)) {
2631 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2632 list_del_init(&se->group_node);
2635 cfs_rq->nr_running--;
2638 #ifdef CONFIG_FAIR_GROUP_SCHED
2640 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2642 long tg_weight, load, shares;
2645 * This really should be: cfs_rq->avg.load_avg, but instead we use
2646 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2647 * the shares for small weight interactive tasks.
2649 load = scale_load_down(cfs_rq->load.weight);
2651 tg_weight = atomic_long_read(&tg->load_avg);
2653 /* Ensure tg_weight >= load */
2654 tg_weight -= cfs_rq->tg_load_avg_contrib;
2657 shares = (tg->shares * load);
2659 shares /= tg_weight;
2662 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2663 * of a group with small tg->shares value. It is a floor value which is
2664 * assigned as a minimum load.weight to the sched_entity representing
2665 * the group on a CPU.
2667 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2668 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2669 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2670 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2673 if (shares < MIN_SHARES)
2674 shares = MIN_SHARES;
2675 if (shares > tg->shares)
2676 shares = tg->shares;
2680 # else /* CONFIG_SMP */
2681 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2685 # endif /* CONFIG_SMP */
2687 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2688 unsigned long weight)
2691 /* commit outstanding execution time */
2692 if (cfs_rq->curr == se)
2693 update_curr(cfs_rq);
2694 account_entity_dequeue(cfs_rq, se);
2697 update_load_set(&se->load, weight);
2700 account_entity_enqueue(cfs_rq, se);
2703 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2705 static void update_cfs_shares(struct sched_entity *se)
2707 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2708 struct task_group *tg;
2714 if (throttled_hierarchy(cfs_rq))
2720 if (likely(se->load.weight == tg->shares))
2723 shares = calc_cfs_shares(cfs_rq, tg);
2725 reweight_entity(cfs_rq_of(se), se, shares);
2728 #else /* CONFIG_FAIR_GROUP_SCHED */
2729 static inline void update_cfs_shares(struct sched_entity *se)
2732 #endif /* CONFIG_FAIR_GROUP_SCHED */
2735 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2736 static const u32 runnable_avg_yN_inv[] = {
2737 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2738 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2739 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2740 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2741 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2742 0x85aac367, 0x82cd8698,
2747 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2749 static u64 decay_load(u64 val, u64 n)
2751 unsigned int local_n;
2753 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2756 /* after bounds checking we can collapse to 32-bit */
2760 * As y^PERIOD = 1/2, we can combine
2761 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2762 * With a look-up table which covers y^n (n<PERIOD)
2764 * To achieve constant time decay_load.
2766 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2767 val >>= local_n / LOAD_AVG_PERIOD;
2768 local_n %= LOAD_AVG_PERIOD;
2771 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2775 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2777 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2782 c1 = decay_load((u64)d1, periods);
2786 * c2 = 1024 \Sum y^n
2790 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2793 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2795 return c1 + c2 + c3;
2798 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2801 * Accumulate the three separate parts of the sum; d1 the remainder
2802 * of the last (incomplete) period, d2 the span of full periods and d3
2803 * the remainder of the (incomplete) current period.
2808 * |<->|<----------------->|<--->|
2809 * ... |---x---|------| ... |------|-----x (now)
2812 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2815 * = u y^p + (Step 1)
2818 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2821 static __always_inline u32
2822 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2823 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2825 unsigned long scale_freq, scale_cpu;
2826 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2829 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2830 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2832 delta += sa->period_contrib;
2833 periods = delta / 1024; /* A period is 1024us (~1ms) */
2836 * Step 1: decay old *_sum if we crossed period boundaries.
2839 sa->load_sum = decay_load(sa->load_sum, periods);
2841 cfs_rq->runnable_load_sum =
2842 decay_load(cfs_rq->runnable_load_sum, periods);
2844 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2850 contrib = __accumulate_pelt_segments(periods,
2851 1024 - sa->period_contrib, delta);
2853 sa->period_contrib = delta;
2855 contrib = cap_scale(contrib, scale_freq);
2857 sa->load_sum += weight * contrib;
2859 cfs_rq->runnable_load_sum += weight * contrib;
2862 sa->util_sum += contrib * scale_cpu;
2868 * We can represent the historical contribution to runnable average as the
2869 * coefficients of a geometric series. To do this we sub-divide our runnable
2870 * history into segments of approximately 1ms (1024us); label the segment that
2871 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2873 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2875 * (now) (~1ms ago) (~2ms ago)
2877 * Let u_i denote the fraction of p_i that the entity was runnable.
2879 * We then designate the fractions u_i as our co-efficients, yielding the
2880 * following representation of historical load:
2881 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2883 * We choose y based on the with of a reasonably scheduling period, fixing:
2886 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2887 * approximately half as much as the contribution to load within the last ms
2890 * When a period "rolls over" and we have new u_0`, multiplying the previous
2891 * sum again by y is sufficient to update:
2892 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2893 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2895 static __always_inline int
2896 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2897 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2901 delta = now - sa->last_update_time;
2903 * This should only happen when time goes backwards, which it
2904 * unfortunately does during sched clock init when we swap over to TSC.
2906 if ((s64)delta < 0) {
2907 sa->last_update_time = now;
2912 * Use 1024ns as the unit of measurement since it's a reasonable
2913 * approximation of 1us and fast to compute.
2918 sa->last_update_time = now;
2921 * Now we know we crossed measurement unit boundaries. The *_avg
2922 * accrues by two steps:
2924 * Step 1: accumulate *_sum since last_update_time. If we haven't
2925 * crossed period boundaries, finish.
2927 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
2931 * Step 2: update *_avg.
2933 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2935 cfs_rq->runnable_load_avg =
2936 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2938 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2944 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
2946 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
2950 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
2952 return ___update_load_avg(now, cpu, &se->avg,
2953 se->on_rq * scale_load_down(se->load.weight),
2954 cfs_rq->curr == se, NULL);
2958 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
2960 return ___update_load_avg(now, cpu, &cfs_rq->avg,
2961 scale_load_down(cfs_rq->load.weight),
2962 cfs_rq->curr != NULL, cfs_rq);
2966 * Signed add and clamp on underflow.
2968 * Explicitly do a load-store to ensure the intermediate value never hits
2969 * memory. This allows lockless observations without ever seeing the negative
2972 #define add_positive(_ptr, _val) do { \
2973 typeof(_ptr) ptr = (_ptr); \
2974 typeof(_val) val = (_val); \
2975 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2979 if (val < 0 && res > var) \
2982 WRITE_ONCE(*ptr, res); \
2985 #ifdef CONFIG_FAIR_GROUP_SCHED
2987 * update_tg_load_avg - update the tg's load avg
2988 * @cfs_rq: the cfs_rq whose avg changed
2989 * @force: update regardless of how small the difference
2991 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2992 * However, because tg->load_avg is a global value there are performance
2995 * In order to avoid having to look at the other cfs_rq's, we use a
2996 * differential update where we store the last value we propagated. This in
2997 * turn allows skipping updates if the differential is 'small'.
2999 * Updating tg's load_avg is necessary before update_cfs_share() (which is
3000 * done) and effective_load() (which is not done because it is too costly).
3002 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3004 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3007 * No need to update load_avg for root_task_group as it is not used.
3009 if (cfs_rq->tg == &root_task_group)
3012 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3013 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3014 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3019 * Called within set_task_rq() right before setting a task's cpu. The
3020 * caller only guarantees p->pi_lock is held; no other assumptions,
3021 * including the state of rq->lock, should be made.
3023 void set_task_rq_fair(struct sched_entity *se,
3024 struct cfs_rq *prev, struct cfs_rq *next)
3026 u64 p_last_update_time;
3027 u64 n_last_update_time;
3029 if (!sched_feat(ATTACH_AGE_LOAD))
3033 * We are supposed to update the task to "current" time, then its up to
3034 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3035 * getting what current time is, so simply throw away the out-of-date
3036 * time. This will result in the wakee task is less decayed, but giving
3037 * the wakee more load sounds not bad.
3039 if (!(se->avg.last_update_time && prev))
3042 #ifndef CONFIG_64BIT
3044 u64 p_last_update_time_copy;
3045 u64 n_last_update_time_copy;
3048 p_last_update_time_copy = prev->load_last_update_time_copy;
3049 n_last_update_time_copy = next->load_last_update_time_copy;
3053 p_last_update_time = prev->avg.last_update_time;
3054 n_last_update_time = next->avg.last_update_time;
3056 } while (p_last_update_time != p_last_update_time_copy ||
3057 n_last_update_time != n_last_update_time_copy);
3060 p_last_update_time = prev->avg.last_update_time;
3061 n_last_update_time = next->avg.last_update_time;
3063 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3064 se->avg.last_update_time = n_last_update_time;
3067 /* Take into account change of utilization of a child task group */
3069 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3071 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3072 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3074 /* Nothing to update */
3078 /* Set new sched_entity's utilization */
3079 se->avg.util_avg = gcfs_rq->avg.util_avg;
3080 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3082 /* Update parent cfs_rq utilization */
3083 add_positive(&cfs_rq->avg.util_avg, delta);
3084 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3087 /* Take into account change of load of a child task group */
3089 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3091 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3092 long delta, load = gcfs_rq->avg.load_avg;
3095 * If the load of group cfs_rq is null, the load of the
3096 * sched_entity will also be null so we can skip the formula
3101 /* Get tg's load and ensure tg_load > 0 */
3102 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3104 /* Ensure tg_load >= load and updated with current load*/
3105 tg_load -= gcfs_rq->tg_load_avg_contrib;
3109 * We need to compute a correction term in the case that the
3110 * task group is consuming more CPU than a task of equal
3111 * weight. A task with a weight equals to tg->shares will have
3112 * a load less or equal to scale_load_down(tg->shares).
3113 * Similarly, the sched_entities that represent the task group
3114 * at parent level, can't have a load higher than
3115 * scale_load_down(tg->shares). And the Sum of sched_entities'
3116 * load must be <= scale_load_down(tg->shares).
3118 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3119 /* scale gcfs_rq's load into tg's shares*/
3120 load *= scale_load_down(gcfs_rq->tg->shares);
3125 delta = load - se->avg.load_avg;
3127 /* Nothing to update */
3131 /* Set new sched_entity's load */
3132 se->avg.load_avg = load;
3133 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3135 /* Update parent cfs_rq load */
3136 add_positive(&cfs_rq->avg.load_avg, delta);
3137 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3140 * If the sched_entity is already enqueued, we also have to update the
3141 * runnable load avg.
3144 /* Update parent cfs_rq runnable_load_avg */
3145 add_positive(&cfs_rq->runnable_load_avg, delta);
3146 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3150 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3152 cfs_rq->propagate_avg = 1;
3155 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3157 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3159 if (!cfs_rq->propagate_avg)
3162 cfs_rq->propagate_avg = 0;
3166 /* Update task and its cfs_rq load average */
3167 static inline int propagate_entity_load_avg(struct sched_entity *se)
3169 struct cfs_rq *cfs_rq;
3171 if (entity_is_task(se))
3174 if (!test_and_clear_tg_cfs_propagate(se))
3177 cfs_rq = cfs_rq_of(se);
3179 set_tg_cfs_propagate(cfs_rq);
3181 update_tg_cfs_util(cfs_rq, se);
3182 update_tg_cfs_load(cfs_rq, se);
3188 * Check if we need to update the load and the utilization of a blocked
3191 static inline bool skip_blocked_update(struct sched_entity *se)
3193 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3196 * If sched_entity still have not zero load or utilization, we have to
3199 if (se->avg.load_avg || se->avg.util_avg)
3203 * If there is a pending propagation, we have to update the load and
3204 * the utilization of the sched_entity:
3206 if (gcfs_rq->propagate_avg)
3210 * Otherwise, the load and the utilization of the sched_entity is
3211 * already zero and there is no pending propagation, so it will be a
3212 * waste of time to try to decay it:
3217 #else /* CONFIG_FAIR_GROUP_SCHED */
3219 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3221 static inline int propagate_entity_load_avg(struct sched_entity *se)
3226 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3228 #endif /* CONFIG_FAIR_GROUP_SCHED */
3230 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3232 if (&this_rq()->cfs == cfs_rq) {
3234 * There are a few boundary cases this might miss but it should
3235 * get called often enough that that should (hopefully) not be
3236 * a real problem -- added to that it only calls on the local
3237 * CPU, so if we enqueue remotely we'll miss an update, but
3238 * the next tick/schedule should update.
3240 * It will not get called when we go idle, because the idle
3241 * thread is a different class (!fair), nor will the utilization
3242 * number include things like RT tasks.
3244 * As is, the util number is not freq-invariant (we'd have to
3245 * implement arch_scale_freq_capacity() for that).
3249 cpufreq_update_util(rq_of(cfs_rq), 0);
3254 * Unsigned subtract and clamp on underflow.
3256 * Explicitly do a load-store to ensure the intermediate value never hits
3257 * memory. This allows lockless observations without ever seeing the negative
3260 #define sub_positive(_ptr, _val) do { \
3261 typeof(_ptr) ptr = (_ptr); \
3262 typeof(*ptr) val = (_val); \
3263 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3267 WRITE_ONCE(*ptr, res); \
3271 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3272 * @now: current time, as per cfs_rq_clock_task()
3273 * @cfs_rq: cfs_rq to update
3274 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3276 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3277 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3278 * post_init_entity_util_avg().
3280 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3282 * Returns true if the load decayed or we removed load.
3284 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3285 * call update_tg_load_avg() when this function returns true.
3288 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3290 struct sched_avg *sa = &cfs_rq->avg;
3291 int decayed, removed_load = 0, removed_util = 0;
3293 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3294 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3295 sub_positive(&sa->load_avg, r);
3296 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3298 set_tg_cfs_propagate(cfs_rq);
3301 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3302 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3303 sub_positive(&sa->util_avg, r);
3304 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3306 set_tg_cfs_propagate(cfs_rq);
3309 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3311 #ifndef CONFIG_64BIT
3313 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3316 if (update_freq && (decayed || removed_util))
3317 cfs_rq_util_change(cfs_rq);
3319 return decayed || removed_load;
3323 * Optional action to be done while updating the load average
3325 #define UPDATE_TG 0x1
3326 #define SKIP_AGE_LOAD 0x2
3328 /* Update task and its cfs_rq load average */
3329 static inline void update_load_avg(struct sched_entity *se, int flags)
3331 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3332 u64 now = cfs_rq_clock_task(cfs_rq);
3333 struct rq *rq = rq_of(cfs_rq);
3334 int cpu = cpu_of(rq);
3338 * Track task load average for carrying it to new CPU after migrated, and
3339 * track group sched_entity load average for task_h_load calc in migration
3341 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3342 __update_load_avg_se(now, cpu, cfs_rq, se);
3344 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3345 decayed |= propagate_entity_load_avg(se);
3347 if (decayed && (flags & UPDATE_TG))
3348 update_tg_load_avg(cfs_rq, 0);
3352 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3353 * @cfs_rq: cfs_rq to attach to
3354 * @se: sched_entity to attach
3356 * Must call update_cfs_rq_load_avg() before this, since we rely on
3357 * cfs_rq->avg.last_update_time being current.
3359 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3361 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3362 cfs_rq->avg.load_avg += se->avg.load_avg;
3363 cfs_rq->avg.load_sum += se->avg.load_sum;
3364 cfs_rq->avg.util_avg += se->avg.util_avg;
3365 cfs_rq->avg.util_sum += se->avg.util_sum;
3366 set_tg_cfs_propagate(cfs_rq);
3368 cfs_rq_util_change(cfs_rq);
3372 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3373 * @cfs_rq: cfs_rq to detach from
3374 * @se: sched_entity to detach
3376 * Must call update_cfs_rq_load_avg() before this, since we rely on
3377 * cfs_rq->avg.last_update_time being current.
3379 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3382 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3383 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3384 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3385 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3386 set_tg_cfs_propagate(cfs_rq);
3388 cfs_rq_util_change(cfs_rq);
3391 /* Add the load generated by se into cfs_rq's load average */
3393 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3395 struct sched_avg *sa = &se->avg;
3397 cfs_rq->runnable_load_avg += sa->load_avg;
3398 cfs_rq->runnable_load_sum += sa->load_sum;
3400 if (!sa->last_update_time) {
3401 attach_entity_load_avg(cfs_rq, se);
3402 update_tg_load_avg(cfs_rq, 0);
3406 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3408 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3410 cfs_rq->runnable_load_avg =
3411 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3412 cfs_rq->runnable_load_sum =
3413 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3416 #ifndef CONFIG_64BIT
3417 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3419 u64 last_update_time_copy;
3420 u64 last_update_time;
3423 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3425 last_update_time = cfs_rq->avg.last_update_time;
3426 } while (last_update_time != last_update_time_copy);
3428 return last_update_time;
3431 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3433 return cfs_rq->avg.last_update_time;
3438 * Synchronize entity load avg of dequeued entity without locking
3441 void sync_entity_load_avg(struct sched_entity *se)
3443 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3444 u64 last_update_time;
3446 last_update_time = cfs_rq_last_update_time(cfs_rq);
3447 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3451 * Task first catches up with cfs_rq, and then subtract
3452 * itself from the cfs_rq (task must be off the queue now).
3454 void remove_entity_load_avg(struct sched_entity *se)
3456 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3459 * tasks cannot exit without having gone through wake_up_new_task() ->
3460 * post_init_entity_util_avg() which will have added things to the
3461 * cfs_rq, so we can remove unconditionally.
3463 * Similarly for groups, they will have passed through
3464 * post_init_entity_util_avg() before unregister_sched_fair_group()
3468 sync_entity_load_avg(se);
3469 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3470 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3473 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3475 return cfs_rq->runnable_load_avg;
3478 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3480 return cfs_rq->avg.load_avg;
3483 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3485 #else /* CONFIG_SMP */
3488 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3493 #define UPDATE_TG 0x0
3494 #define SKIP_AGE_LOAD 0x0
3496 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3498 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3502 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3504 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3505 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3508 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3510 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3512 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3517 #endif /* CONFIG_SMP */
3519 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3521 #ifdef CONFIG_SCHED_DEBUG
3522 s64 d = se->vruntime - cfs_rq->min_vruntime;
3527 if (d > 3*sysctl_sched_latency)
3528 schedstat_inc(cfs_rq->nr_spread_over);
3533 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3535 u64 vruntime = cfs_rq->min_vruntime;
3538 * The 'current' period is already promised to the current tasks,
3539 * however the extra weight of the new task will slow them down a
3540 * little, place the new task so that it fits in the slot that
3541 * stays open at the end.
3543 if (initial && sched_feat(START_DEBIT))
3544 vruntime += sched_vslice(cfs_rq, se);
3546 /* sleeps up to a single latency don't count. */
3548 unsigned long thresh = sysctl_sched_latency;
3551 * Halve their sleep time's effect, to allow
3552 * for a gentler effect of sleepers:
3554 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3560 /* ensure we never gain time by being placed backwards. */
3561 se->vruntime = max_vruntime(se->vruntime, vruntime);
3564 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3566 static inline void check_schedstat_required(void)
3568 #ifdef CONFIG_SCHEDSTATS
3569 if (schedstat_enabled())
3572 /* Force schedstat enabled if a dependent tracepoint is active */
3573 if (trace_sched_stat_wait_enabled() ||
3574 trace_sched_stat_sleep_enabled() ||
3575 trace_sched_stat_iowait_enabled() ||
3576 trace_sched_stat_blocked_enabled() ||
3577 trace_sched_stat_runtime_enabled()) {
3578 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3579 "stat_blocked and stat_runtime require the "
3580 "kernel parameter schedstats=enabled or "
3581 "kernel.sched_schedstats=1\n");
3592 * update_min_vruntime()
3593 * vruntime -= min_vruntime
3597 * update_min_vruntime()
3598 * vruntime += min_vruntime
3600 * this way the vruntime transition between RQs is done when both
3601 * min_vruntime are up-to-date.
3605 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3606 * vruntime -= min_vruntime
3610 * update_min_vruntime()
3611 * vruntime += min_vruntime
3613 * this way we don't have the most up-to-date min_vruntime on the originating
3614 * CPU and an up-to-date min_vruntime on the destination CPU.
3618 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3620 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3621 bool curr = cfs_rq->curr == se;
3624 * If we're the current task, we must renormalise before calling
3628 se->vruntime += cfs_rq->min_vruntime;
3630 update_curr(cfs_rq);
3633 * Otherwise, renormalise after, such that we're placed at the current
3634 * moment in time, instead of some random moment in the past. Being
3635 * placed in the past could significantly boost this task to the
3636 * fairness detriment of existing tasks.
3638 if (renorm && !curr)
3639 se->vruntime += cfs_rq->min_vruntime;
3642 * When enqueuing a sched_entity, we must:
3643 * - Update loads to have both entity and cfs_rq synced with now.
3644 * - Add its load to cfs_rq->runnable_avg
3645 * - For group_entity, update its weight to reflect the new share of
3647 * - Add its new weight to cfs_rq->load.weight
3649 update_load_avg(se, UPDATE_TG);
3650 enqueue_entity_load_avg(cfs_rq, se);
3651 update_cfs_shares(se);
3652 account_entity_enqueue(cfs_rq, se);
3654 if (flags & ENQUEUE_WAKEUP)
3655 place_entity(cfs_rq, se, 0);
3657 check_schedstat_required();
3658 update_stats_enqueue(cfs_rq, se, flags);
3659 check_spread(cfs_rq, se);
3661 __enqueue_entity(cfs_rq, se);
3664 if (cfs_rq->nr_running == 1) {
3665 list_add_leaf_cfs_rq(cfs_rq);
3666 check_enqueue_throttle(cfs_rq);
3670 static void __clear_buddies_last(struct sched_entity *se)
3672 for_each_sched_entity(se) {
3673 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3674 if (cfs_rq->last != se)
3677 cfs_rq->last = NULL;
3681 static void __clear_buddies_next(struct sched_entity *se)
3683 for_each_sched_entity(se) {
3684 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3685 if (cfs_rq->next != se)
3688 cfs_rq->next = NULL;
3692 static void __clear_buddies_skip(struct sched_entity *se)
3694 for_each_sched_entity(se) {
3695 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3696 if (cfs_rq->skip != se)
3699 cfs_rq->skip = NULL;
3703 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3705 if (cfs_rq->last == se)
3706 __clear_buddies_last(se);
3708 if (cfs_rq->next == se)
3709 __clear_buddies_next(se);
3711 if (cfs_rq->skip == se)
3712 __clear_buddies_skip(se);
3715 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3718 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3721 * Update run-time statistics of the 'current'.
3723 update_curr(cfs_rq);
3726 * When dequeuing a sched_entity, we must:
3727 * - Update loads to have both entity and cfs_rq synced with now.
3728 * - Substract its load from the cfs_rq->runnable_avg.
3729 * - Substract its previous weight from cfs_rq->load.weight.
3730 * - For group entity, update its weight to reflect the new share
3731 * of its group cfs_rq.
3733 update_load_avg(se, UPDATE_TG);
3734 dequeue_entity_load_avg(cfs_rq, se);
3736 update_stats_dequeue(cfs_rq, se, flags);
3738 clear_buddies(cfs_rq, se);
3740 if (se != cfs_rq->curr)
3741 __dequeue_entity(cfs_rq, se);
3743 account_entity_dequeue(cfs_rq, se);
3746 * Normalize after update_curr(); which will also have moved
3747 * min_vruntime if @se is the one holding it back. But before doing
3748 * update_min_vruntime() again, which will discount @se's position and
3749 * can move min_vruntime forward still more.
3751 if (!(flags & DEQUEUE_SLEEP))
3752 se->vruntime -= cfs_rq->min_vruntime;
3754 /* return excess runtime on last dequeue */
3755 return_cfs_rq_runtime(cfs_rq);
3757 update_cfs_shares(se);
3760 * Now advance min_vruntime if @se was the entity holding it back,
3761 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3762 * put back on, and if we advance min_vruntime, we'll be placed back
3763 * further than we started -- ie. we'll be penalized.
3765 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3766 update_min_vruntime(cfs_rq);
3770 * Preempt the current task with a newly woken task if needed:
3773 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3775 unsigned long ideal_runtime, delta_exec;
3776 struct sched_entity *se;
3779 ideal_runtime = sched_slice(cfs_rq, curr);
3780 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3781 if (delta_exec > ideal_runtime) {
3782 resched_curr(rq_of(cfs_rq));
3784 * The current task ran long enough, ensure it doesn't get
3785 * re-elected due to buddy favours.
3787 clear_buddies(cfs_rq, curr);
3792 * Ensure that a task that missed wakeup preemption by a
3793 * narrow margin doesn't have to wait for a full slice.
3794 * This also mitigates buddy induced latencies under load.
3796 if (delta_exec < sysctl_sched_min_granularity)
3799 se = __pick_first_entity(cfs_rq);
3800 delta = curr->vruntime - se->vruntime;
3805 if (delta > ideal_runtime)
3806 resched_curr(rq_of(cfs_rq));
3810 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3812 /* 'current' is not kept within the tree. */
3815 * Any task has to be enqueued before it get to execute on
3816 * a CPU. So account for the time it spent waiting on the
3819 update_stats_wait_end(cfs_rq, se);
3820 __dequeue_entity(cfs_rq, se);
3821 update_load_avg(se, UPDATE_TG);
3824 update_stats_curr_start(cfs_rq, se);
3828 * Track our maximum slice length, if the CPU's load is at
3829 * least twice that of our own weight (i.e. dont track it
3830 * when there are only lesser-weight tasks around):
3832 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3833 schedstat_set(se->statistics.slice_max,
3834 max((u64)schedstat_val(se->statistics.slice_max),
3835 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3838 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3842 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3845 * Pick the next process, keeping these things in mind, in this order:
3846 * 1) keep things fair between processes/task groups
3847 * 2) pick the "next" process, since someone really wants that to run
3848 * 3) pick the "last" process, for cache locality
3849 * 4) do not run the "skip" process, if something else is available
3851 static struct sched_entity *
3852 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3854 struct sched_entity *left = __pick_first_entity(cfs_rq);
3855 struct sched_entity *se;
3858 * If curr is set we have to see if its left of the leftmost entity
3859 * still in the tree, provided there was anything in the tree at all.
3861 if (!left || (curr && entity_before(curr, left)))
3864 se = left; /* ideally we run the leftmost entity */
3867 * Avoid running the skip buddy, if running something else can
3868 * be done without getting too unfair.
3870 if (cfs_rq->skip == se) {
3871 struct sched_entity *second;
3874 second = __pick_first_entity(cfs_rq);
3876 second = __pick_next_entity(se);
3877 if (!second || (curr && entity_before(curr, second)))
3881 if (second && wakeup_preempt_entity(second, left) < 1)
3886 * Prefer last buddy, try to return the CPU to a preempted task.
3888 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3892 * Someone really wants this to run. If it's not unfair, run it.
3894 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3897 clear_buddies(cfs_rq, se);
3902 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3904 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3907 * If still on the runqueue then deactivate_task()
3908 * was not called and update_curr() has to be done:
3911 update_curr(cfs_rq);
3913 /* throttle cfs_rqs exceeding runtime */
3914 check_cfs_rq_runtime(cfs_rq);
3916 check_spread(cfs_rq, prev);
3919 update_stats_wait_start(cfs_rq, prev);
3920 /* Put 'current' back into the tree. */
3921 __enqueue_entity(cfs_rq, prev);
3922 /* in !on_rq case, update occurred at dequeue */
3923 update_load_avg(prev, 0);
3925 cfs_rq->curr = NULL;
3929 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3932 * Update run-time statistics of the 'current'.
3934 update_curr(cfs_rq);
3937 * Ensure that runnable average is periodically updated.
3939 update_load_avg(curr, UPDATE_TG);
3940 update_cfs_shares(curr);
3942 #ifdef CONFIG_SCHED_HRTICK
3944 * queued ticks are scheduled to match the slice, so don't bother
3945 * validating it and just reschedule.
3948 resched_curr(rq_of(cfs_rq));
3952 * don't let the period tick interfere with the hrtick preemption
3954 if (!sched_feat(DOUBLE_TICK) &&
3955 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3959 if (cfs_rq->nr_running > 1)
3960 check_preempt_tick(cfs_rq, curr);
3964 /**************************************************
3965 * CFS bandwidth control machinery
3968 #ifdef CONFIG_CFS_BANDWIDTH
3970 #ifdef HAVE_JUMP_LABEL
3971 static struct static_key __cfs_bandwidth_used;
3973 static inline bool cfs_bandwidth_used(void)
3975 return static_key_false(&__cfs_bandwidth_used);
3978 void cfs_bandwidth_usage_inc(void)
3980 static_key_slow_inc(&__cfs_bandwidth_used);
3983 void cfs_bandwidth_usage_dec(void)
3985 static_key_slow_dec(&__cfs_bandwidth_used);
3987 #else /* HAVE_JUMP_LABEL */
3988 static bool cfs_bandwidth_used(void)
3993 void cfs_bandwidth_usage_inc(void) {}
3994 void cfs_bandwidth_usage_dec(void) {}
3995 #endif /* HAVE_JUMP_LABEL */
3998 * default period for cfs group bandwidth.
3999 * default: 0.1s, units: nanoseconds
4001 static inline u64 default_cfs_period(void)
4003 return 100000000ULL;
4006 static inline u64 sched_cfs_bandwidth_slice(void)
4008 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4012 * Replenish runtime according to assigned quota and update expiration time.
4013 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4014 * additional synchronization around rq->lock.
4016 * requires cfs_b->lock
4018 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4022 if (cfs_b->quota == RUNTIME_INF)
4025 now = sched_clock_cpu(smp_processor_id());
4026 cfs_b->runtime = cfs_b->quota;
4027 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4030 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4032 return &tg->cfs_bandwidth;
4035 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4036 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4038 if (unlikely(cfs_rq->throttle_count))
4039 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4041 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4044 /* returns 0 on failure to allocate runtime */
4045 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4047 struct task_group *tg = cfs_rq->tg;
4048 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4049 u64 amount = 0, min_amount, expires;
4051 /* note: this is a positive sum as runtime_remaining <= 0 */
4052 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4054 raw_spin_lock(&cfs_b->lock);
4055 if (cfs_b->quota == RUNTIME_INF)
4056 amount = min_amount;
4058 start_cfs_bandwidth(cfs_b);
4060 if (cfs_b->runtime > 0) {
4061 amount = min(cfs_b->runtime, min_amount);
4062 cfs_b->runtime -= amount;
4066 expires = cfs_b->runtime_expires;
4067 raw_spin_unlock(&cfs_b->lock);
4069 cfs_rq->runtime_remaining += amount;
4071 * we may have advanced our local expiration to account for allowed
4072 * spread between our sched_clock and the one on which runtime was
4075 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4076 cfs_rq->runtime_expires = expires;
4078 return cfs_rq->runtime_remaining > 0;
4082 * Note: This depends on the synchronization provided by sched_clock and the
4083 * fact that rq->clock snapshots this value.
4085 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4087 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4089 /* if the deadline is ahead of our clock, nothing to do */
4090 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4093 if (cfs_rq->runtime_remaining < 0)
4097 * If the local deadline has passed we have to consider the
4098 * possibility that our sched_clock is 'fast' and the global deadline
4099 * has not truly expired.
4101 * Fortunately we can check determine whether this the case by checking
4102 * whether the global deadline has advanced. It is valid to compare
4103 * cfs_b->runtime_expires without any locks since we only care about
4104 * exact equality, so a partial write will still work.
4107 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4108 /* extend local deadline, drift is bounded above by 2 ticks */
4109 cfs_rq->runtime_expires += TICK_NSEC;
4111 /* global deadline is ahead, expiration has passed */
4112 cfs_rq->runtime_remaining = 0;
4116 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4118 /* dock delta_exec before expiring quota (as it could span periods) */
4119 cfs_rq->runtime_remaining -= delta_exec;
4120 expire_cfs_rq_runtime(cfs_rq);
4122 if (likely(cfs_rq->runtime_remaining > 0))
4126 * if we're unable to extend our runtime we resched so that the active
4127 * hierarchy can be throttled
4129 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4130 resched_curr(rq_of(cfs_rq));
4133 static __always_inline
4134 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4136 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4139 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4142 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4144 return cfs_bandwidth_used() && cfs_rq->throttled;
4147 /* check whether cfs_rq, or any parent, is throttled */
4148 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4150 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4154 * Ensure that neither of the group entities corresponding to src_cpu or
4155 * dest_cpu are members of a throttled hierarchy when performing group
4156 * load-balance operations.
4158 static inline int throttled_lb_pair(struct task_group *tg,
4159 int src_cpu, int dest_cpu)
4161 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4163 src_cfs_rq = tg->cfs_rq[src_cpu];
4164 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4166 return throttled_hierarchy(src_cfs_rq) ||
4167 throttled_hierarchy(dest_cfs_rq);
4170 /* updated child weight may affect parent so we have to do this bottom up */
4171 static int tg_unthrottle_up(struct task_group *tg, void *data)
4173 struct rq *rq = data;
4174 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4176 cfs_rq->throttle_count--;
4177 if (!cfs_rq->throttle_count) {
4178 /* adjust cfs_rq_clock_task() */
4179 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4180 cfs_rq->throttled_clock_task;
4186 static int tg_throttle_down(struct task_group *tg, void *data)
4188 struct rq *rq = data;
4189 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4191 /* group is entering throttled state, stop time */
4192 if (!cfs_rq->throttle_count)
4193 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4194 cfs_rq->throttle_count++;
4199 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4201 struct rq *rq = rq_of(cfs_rq);
4202 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4203 struct sched_entity *se;
4204 long task_delta, dequeue = 1;
4207 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4209 /* freeze hierarchy runnable averages while throttled */
4211 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4214 task_delta = cfs_rq->h_nr_running;
4215 for_each_sched_entity(se) {
4216 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4217 /* throttled entity or throttle-on-deactivate */
4222 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4223 qcfs_rq->h_nr_running -= task_delta;
4225 if (qcfs_rq->load.weight)
4230 sub_nr_running(rq, task_delta);
4232 cfs_rq->throttled = 1;
4233 cfs_rq->throttled_clock = rq_clock(rq);
4234 raw_spin_lock(&cfs_b->lock);
4235 empty = list_empty(&cfs_b->throttled_cfs_rq);
4238 * Add to the _head_ of the list, so that an already-started
4239 * distribute_cfs_runtime will not see us
4241 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4244 * If we're the first throttled task, make sure the bandwidth
4248 start_cfs_bandwidth(cfs_b);
4250 raw_spin_unlock(&cfs_b->lock);
4253 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4255 struct rq *rq = rq_of(cfs_rq);
4256 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4257 struct sched_entity *se;
4261 se = cfs_rq->tg->se[cpu_of(rq)];
4263 cfs_rq->throttled = 0;
4265 update_rq_clock(rq);
4267 raw_spin_lock(&cfs_b->lock);
4268 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4269 list_del_rcu(&cfs_rq->throttled_list);
4270 raw_spin_unlock(&cfs_b->lock);
4272 /* update hierarchical throttle state */
4273 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4275 if (!cfs_rq->load.weight)
4278 task_delta = cfs_rq->h_nr_running;
4279 for_each_sched_entity(se) {
4283 cfs_rq = cfs_rq_of(se);
4285 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4286 cfs_rq->h_nr_running += task_delta;
4288 if (cfs_rq_throttled(cfs_rq))
4293 add_nr_running(rq, task_delta);
4295 /* determine whether we need to wake up potentially idle cpu */
4296 if (rq->curr == rq->idle && rq->cfs.nr_running)
4300 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4301 u64 remaining, u64 expires)
4303 struct cfs_rq *cfs_rq;
4305 u64 starting_runtime = remaining;
4308 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4310 struct rq *rq = rq_of(cfs_rq);
4314 if (!cfs_rq_throttled(cfs_rq))
4317 runtime = -cfs_rq->runtime_remaining + 1;
4318 if (runtime > remaining)
4319 runtime = remaining;
4320 remaining -= runtime;
4322 cfs_rq->runtime_remaining += runtime;
4323 cfs_rq->runtime_expires = expires;
4325 /* we check whether we're throttled above */
4326 if (cfs_rq->runtime_remaining > 0)
4327 unthrottle_cfs_rq(cfs_rq);
4337 return starting_runtime - remaining;
4341 * Responsible for refilling a task_group's bandwidth and unthrottling its
4342 * cfs_rqs as appropriate. If there has been no activity within the last
4343 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4344 * used to track this state.
4346 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4348 u64 runtime, runtime_expires;
4351 /* no need to continue the timer with no bandwidth constraint */
4352 if (cfs_b->quota == RUNTIME_INF)
4353 goto out_deactivate;
4355 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4356 cfs_b->nr_periods += overrun;
4359 * idle depends on !throttled (for the case of a large deficit), and if
4360 * we're going inactive then everything else can be deferred
4362 if (cfs_b->idle && !throttled)
4363 goto out_deactivate;
4365 __refill_cfs_bandwidth_runtime(cfs_b);
4368 /* mark as potentially idle for the upcoming period */
4373 /* account preceding periods in which throttling occurred */
4374 cfs_b->nr_throttled += overrun;
4376 runtime_expires = cfs_b->runtime_expires;
4379 * This check is repeated as we are holding onto the new bandwidth while
4380 * we unthrottle. This can potentially race with an unthrottled group
4381 * trying to acquire new bandwidth from the global pool. This can result
4382 * in us over-using our runtime if it is all used during this loop, but
4383 * only by limited amounts in that extreme case.
4385 while (throttled && cfs_b->runtime > 0) {
4386 runtime = cfs_b->runtime;
4387 raw_spin_unlock(&cfs_b->lock);
4388 /* we can't nest cfs_b->lock while distributing bandwidth */
4389 runtime = distribute_cfs_runtime(cfs_b, runtime,
4391 raw_spin_lock(&cfs_b->lock);
4393 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4395 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4399 * While we are ensured activity in the period following an
4400 * unthrottle, this also covers the case in which the new bandwidth is
4401 * insufficient to cover the existing bandwidth deficit. (Forcing the
4402 * timer to remain active while there are any throttled entities.)
4412 /* a cfs_rq won't donate quota below this amount */
4413 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4414 /* minimum remaining period time to redistribute slack quota */
4415 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4416 /* how long we wait to gather additional slack before distributing */
4417 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4420 * Are we near the end of the current quota period?
4422 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4423 * hrtimer base being cleared by hrtimer_start. In the case of
4424 * migrate_hrtimers, base is never cleared, so we are fine.
4426 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4428 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4431 /* if the call-back is running a quota refresh is already occurring */
4432 if (hrtimer_callback_running(refresh_timer))
4435 /* is a quota refresh about to occur? */
4436 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4437 if (remaining < min_expire)
4443 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4445 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4447 /* if there's a quota refresh soon don't bother with slack */
4448 if (runtime_refresh_within(cfs_b, min_left))
4451 hrtimer_start(&cfs_b->slack_timer,
4452 ns_to_ktime(cfs_bandwidth_slack_period),
4456 /* we know any runtime found here is valid as update_curr() precedes return */
4457 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4459 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4460 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4462 if (slack_runtime <= 0)
4465 raw_spin_lock(&cfs_b->lock);
4466 if (cfs_b->quota != RUNTIME_INF &&
4467 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4468 cfs_b->runtime += slack_runtime;
4470 /* we are under rq->lock, defer unthrottling using a timer */
4471 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4472 !list_empty(&cfs_b->throttled_cfs_rq))
4473 start_cfs_slack_bandwidth(cfs_b);
4475 raw_spin_unlock(&cfs_b->lock);
4477 /* even if it's not valid for return we don't want to try again */
4478 cfs_rq->runtime_remaining -= slack_runtime;
4481 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4483 if (!cfs_bandwidth_used())
4486 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4489 __return_cfs_rq_runtime(cfs_rq);
4493 * This is done with a timer (instead of inline with bandwidth return) since
4494 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4496 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4498 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4501 /* confirm we're still not at a refresh boundary */
4502 raw_spin_lock(&cfs_b->lock);
4503 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4504 raw_spin_unlock(&cfs_b->lock);
4508 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4509 runtime = cfs_b->runtime;
4511 expires = cfs_b->runtime_expires;
4512 raw_spin_unlock(&cfs_b->lock);
4517 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4519 raw_spin_lock(&cfs_b->lock);
4520 if (expires == cfs_b->runtime_expires)
4521 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4522 raw_spin_unlock(&cfs_b->lock);
4526 * When a group wakes up we want to make sure that its quota is not already
4527 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4528 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4530 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4532 if (!cfs_bandwidth_used())
4535 /* an active group must be handled by the update_curr()->put() path */
4536 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4539 /* ensure the group is not already throttled */
4540 if (cfs_rq_throttled(cfs_rq))
4543 /* update runtime allocation */
4544 account_cfs_rq_runtime(cfs_rq, 0);
4545 if (cfs_rq->runtime_remaining <= 0)
4546 throttle_cfs_rq(cfs_rq);
4549 static void sync_throttle(struct task_group *tg, int cpu)
4551 struct cfs_rq *pcfs_rq, *cfs_rq;
4553 if (!cfs_bandwidth_used())
4559 cfs_rq = tg->cfs_rq[cpu];
4560 pcfs_rq = tg->parent->cfs_rq[cpu];
4562 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4563 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4566 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4567 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4569 if (!cfs_bandwidth_used())
4572 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4576 * it's possible for a throttled entity to be forced into a running
4577 * state (e.g. set_curr_task), in this case we're finished.
4579 if (cfs_rq_throttled(cfs_rq))
4582 throttle_cfs_rq(cfs_rq);
4586 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4588 struct cfs_bandwidth *cfs_b =
4589 container_of(timer, struct cfs_bandwidth, slack_timer);
4591 do_sched_cfs_slack_timer(cfs_b);
4593 return HRTIMER_NORESTART;
4596 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4598 struct cfs_bandwidth *cfs_b =
4599 container_of(timer, struct cfs_bandwidth, period_timer);
4603 raw_spin_lock(&cfs_b->lock);
4605 overrun = hrtimer_forward_now(timer, cfs_b->period);
4609 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4612 cfs_b->period_active = 0;
4613 raw_spin_unlock(&cfs_b->lock);
4615 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4618 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4620 raw_spin_lock_init(&cfs_b->lock);
4622 cfs_b->quota = RUNTIME_INF;
4623 cfs_b->period = ns_to_ktime(default_cfs_period());
4625 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4626 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4627 cfs_b->period_timer.function = sched_cfs_period_timer;
4628 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4629 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4632 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4634 cfs_rq->runtime_enabled = 0;
4635 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4638 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4640 lockdep_assert_held(&cfs_b->lock);
4642 if (!cfs_b->period_active) {
4643 cfs_b->period_active = 1;
4644 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4645 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4649 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4651 /* init_cfs_bandwidth() was not called */
4652 if (!cfs_b->throttled_cfs_rq.next)
4655 hrtimer_cancel(&cfs_b->period_timer);
4656 hrtimer_cancel(&cfs_b->slack_timer);
4659 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4661 struct cfs_rq *cfs_rq;
4663 for_each_leaf_cfs_rq(rq, cfs_rq) {
4664 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4666 raw_spin_lock(&cfs_b->lock);
4667 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4668 raw_spin_unlock(&cfs_b->lock);
4672 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4674 struct cfs_rq *cfs_rq;
4676 for_each_leaf_cfs_rq(rq, cfs_rq) {
4677 if (!cfs_rq->runtime_enabled)
4681 * clock_task is not advancing so we just need to make sure
4682 * there's some valid quota amount
4684 cfs_rq->runtime_remaining = 1;
4686 * Offline rq is schedulable till cpu is completely disabled
4687 * in take_cpu_down(), so we prevent new cfs throttling here.
4689 cfs_rq->runtime_enabled = 0;
4691 if (cfs_rq_throttled(cfs_rq))
4692 unthrottle_cfs_rq(cfs_rq);
4696 #else /* CONFIG_CFS_BANDWIDTH */
4697 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4699 return rq_clock_task(rq_of(cfs_rq));
4702 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4703 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4704 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4705 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4706 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4708 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4713 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4718 static inline int throttled_lb_pair(struct task_group *tg,
4719 int src_cpu, int dest_cpu)
4724 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4726 #ifdef CONFIG_FAIR_GROUP_SCHED
4727 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4730 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4734 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4735 static inline void update_runtime_enabled(struct rq *rq) {}
4736 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4738 #endif /* CONFIG_CFS_BANDWIDTH */
4740 /**************************************************
4741 * CFS operations on tasks:
4744 #ifdef CONFIG_SCHED_HRTICK
4745 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4747 struct sched_entity *se = &p->se;
4748 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4750 SCHED_WARN_ON(task_rq(p) != rq);
4752 if (rq->cfs.h_nr_running > 1) {
4753 u64 slice = sched_slice(cfs_rq, se);
4754 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4755 s64 delta = slice - ran;
4762 hrtick_start(rq, delta);
4767 * called from enqueue/dequeue and updates the hrtick when the
4768 * current task is from our class and nr_running is low enough
4771 static void hrtick_update(struct rq *rq)
4773 struct task_struct *curr = rq->curr;
4775 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4778 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4779 hrtick_start_fair(rq, curr);
4781 #else /* !CONFIG_SCHED_HRTICK */
4783 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4787 static inline void hrtick_update(struct rq *rq)
4793 * The enqueue_task method is called before nr_running is
4794 * increased. Here we update the fair scheduling stats and
4795 * then put the task into the rbtree:
4798 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4800 struct cfs_rq *cfs_rq;
4801 struct sched_entity *se = &p->se;
4804 * If in_iowait is set, the code below may not trigger any cpufreq
4805 * utilization updates, so do it here explicitly with the IOWAIT flag
4809 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4811 for_each_sched_entity(se) {
4814 cfs_rq = cfs_rq_of(se);
4815 enqueue_entity(cfs_rq, se, flags);
4818 * end evaluation on encountering a throttled cfs_rq
4820 * note: in the case of encountering a throttled cfs_rq we will
4821 * post the final h_nr_running increment below.
4823 if (cfs_rq_throttled(cfs_rq))
4825 cfs_rq->h_nr_running++;
4827 flags = ENQUEUE_WAKEUP;
4830 for_each_sched_entity(se) {
4831 cfs_rq = cfs_rq_of(se);
4832 cfs_rq->h_nr_running++;
4834 if (cfs_rq_throttled(cfs_rq))
4837 update_load_avg(se, UPDATE_TG);
4838 update_cfs_shares(se);
4842 add_nr_running(rq, 1);
4847 static void set_next_buddy(struct sched_entity *se);
4850 * The dequeue_task method is called before nr_running is
4851 * decreased. We remove the task from the rbtree and
4852 * update the fair scheduling stats:
4854 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4856 struct cfs_rq *cfs_rq;
4857 struct sched_entity *se = &p->se;
4858 int task_sleep = flags & DEQUEUE_SLEEP;
4860 for_each_sched_entity(se) {
4861 cfs_rq = cfs_rq_of(se);
4862 dequeue_entity(cfs_rq, se, flags);
4865 * end evaluation on encountering a throttled cfs_rq
4867 * note: in the case of encountering a throttled cfs_rq we will
4868 * post the final h_nr_running decrement below.
4870 if (cfs_rq_throttled(cfs_rq))
4872 cfs_rq->h_nr_running--;
4874 /* Don't dequeue parent if it has other entities besides us */
4875 if (cfs_rq->load.weight) {
4876 /* Avoid re-evaluating load for this entity: */
4877 se = parent_entity(se);
4879 * Bias pick_next to pick a task from this cfs_rq, as
4880 * p is sleeping when it is within its sched_slice.
4882 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4886 flags |= DEQUEUE_SLEEP;
4889 for_each_sched_entity(se) {
4890 cfs_rq = cfs_rq_of(se);
4891 cfs_rq->h_nr_running--;
4893 if (cfs_rq_throttled(cfs_rq))
4896 update_load_avg(se, UPDATE_TG);
4897 update_cfs_shares(se);
4901 sub_nr_running(rq, 1);
4908 /* Working cpumask for: load_balance, load_balance_newidle. */
4909 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4910 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4912 #ifdef CONFIG_NO_HZ_COMMON
4914 * per rq 'load' arrray crap; XXX kill this.
4918 * The exact cpuload calculated at every tick would be:
4920 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4922 * If a cpu misses updates for n ticks (as it was idle) and update gets
4923 * called on the n+1-th tick when cpu may be busy, then we have:
4925 * load_n = (1 - 1/2^i)^n * load_0
4926 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4928 * decay_load_missed() below does efficient calculation of
4930 * load' = (1 - 1/2^i)^n * load
4932 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4933 * This allows us to precompute the above in said factors, thereby allowing the
4934 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4935 * fixed_power_int())
4937 * The calculation is approximated on a 128 point scale.
4939 #define DEGRADE_SHIFT 7
4941 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4942 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4943 { 0, 0, 0, 0, 0, 0, 0, 0 },
4944 { 64, 32, 8, 0, 0, 0, 0, 0 },
4945 { 96, 72, 40, 12, 1, 0, 0, 0 },
4946 { 112, 98, 75, 43, 15, 1, 0, 0 },
4947 { 120, 112, 98, 76, 45, 16, 2, 0 }
4951 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4952 * would be when CPU is idle and so we just decay the old load without
4953 * adding any new load.
4955 static unsigned long
4956 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4960 if (!missed_updates)
4963 if (missed_updates >= degrade_zero_ticks[idx])
4967 return load >> missed_updates;
4969 while (missed_updates) {
4970 if (missed_updates % 2)
4971 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4973 missed_updates >>= 1;
4978 #endif /* CONFIG_NO_HZ_COMMON */
4981 * __cpu_load_update - update the rq->cpu_load[] statistics
4982 * @this_rq: The rq to update statistics for
4983 * @this_load: The current load
4984 * @pending_updates: The number of missed updates
4986 * Update rq->cpu_load[] statistics. This function is usually called every
4987 * scheduler tick (TICK_NSEC).
4989 * This function computes a decaying average:
4991 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4993 * Because of NOHZ it might not get called on every tick which gives need for
4994 * the @pending_updates argument.
4996 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4997 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4998 * = A * (A * load[i]_n-2 + B) + B
4999 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5000 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5001 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5002 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5003 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5005 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5006 * any change in load would have resulted in the tick being turned back on.
5008 * For regular NOHZ, this reduces to:
5010 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5012 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5015 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5016 unsigned long pending_updates)
5018 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5021 this_rq->nr_load_updates++;
5023 /* Update our load: */
5024 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5025 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5026 unsigned long old_load, new_load;
5028 /* scale is effectively 1 << i now, and >> i divides by scale */
5030 old_load = this_rq->cpu_load[i];
5031 #ifdef CONFIG_NO_HZ_COMMON
5032 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5033 if (tickless_load) {
5034 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5036 * old_load can never be a negative value because a
5037 * decayed tickless_load cannot be greater than the
5038 * original tickless_load.
5040 old_load += tickless_load;
5043 new_load = this_load;
5045 * Round up the averaging division if load is increasing. This
5046 * prevents us from getting stuck on 9 if the load is 10, for
5049 if (new_load > old_load)
5050 new_load += scale - 1;
5052 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5055 sched_avg_update(this_rq);
5058 /* Used instead of source_load when we know the type == 0 */
5059 static unsigned long weighted_cpuload(const int cpu)
5061 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
5064 #ifdef CONFIG_NO_HZ_COMMON
5066 * There is no sane way to deal with nohz on smp when using jiffies because the
5067 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5068 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5070 * Therefore we need to avoid the delta approach from the regular tick when
5071 * possible since that would seriously skew the load calculation. This is why we
5072 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5073 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5074 * loop exit, nohz_idle_balance, nohz full exit...)
5076 * This means we might still be one tick off for nohz periods.
5079 static void cpu_load_update_nohz(struct rq *this_rq,
5080 unsigned long curr_jiffies,
5083 unsigned long pending_updates;
5085 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5086 if (pending_updates) {
5087 this_rq->last_load_update_tick = curr_jiffies;
5089 * In the regular NOHZ case, we were idle, this means load 0.
5090 * In the NOHZ_FULL case, we were non-idle, we should consider
5091 * its weighted load.
5093 cpu_load_update(this_rq, load, pending_updates);
5098 * Called from nohz_idle_balance() to update the load ratings before doing the
5101 static void cpu_load_update_idle(struct rq *this_rq)
5104 * bail if there's load or we're actually up-to-date.
5106 if (weighted_cpuload(cpu_of(this_rq)))
5109 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5113 * Record CPU load on nohz entry so we know the tickless load to account
5114 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5115 * than other cpu_load[idx] but it should be fine as cpu_load readers
5116 * shouldn't rely into synchronized cpu_load[*] updates.
5118 void cpu_load_update_nohz_start(void)
5120 struct rq *this_rq = this_rq();
5123 * This is all lockless but should be fine. If weighted_cpuload changes
5124 * concurrently we'll exit nohz. And cpu_load write can race with
5125 * cpu_load_update_idle() but both updater would be writing the same.
5127 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
5131 * Account the tickless load in the end of a nohz frame.
5133 void cpu_load_update_nohz_stop(void)
5135 unsigned long curr_jiffies = READ_ONCE(jiffies);
5136 struct rq *this_rq = this_rq();
5140 if (curr_jiffies == this_rq->last_load_update_tick)
5143 load = weighted_cpuload(cpu_of(this_rq));
5144 rq_lock(this_rq, &rf);
5145 update_rq_clock(this_rq);
5146 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5147 rq_unlock(this_rq, &rf);
5149 #else /* !CONFIG_NO_HZ_COMMON */
5150 static inline void cpu_load_update_nohz(struct rq *this_rq,
5151 unsigned long curr_jiffies,
5152 unsigned long load) { }
5153 #endif /* CONFIG_NO_HZ_COMMON */
5155 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5157 #ifdef CONFIG_NO_HZ_COMMON
5158 /* See the mess around cpu_load_update_nohz(). */
5159 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5161 cpu_load_update(this_rq, load, 1);
5165 * Called from scheduler_tick()
5167 void cpu_load_update_active(struct rq *this_rq)
5169 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5171 if (tick_nohz_tick_stopped())
5172 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5174 cpu_load_update_periodic(this_rq, load);
5178 * Return a low guess at the load of a migration-source cpu weighted
5179 * according to the scheduling class and "nice" value.
5181 * We want to under-estimate the load of migration sources, to
5182 * balance conservatively.
5184 static unsigned long source_load(int cpu, int type)
5186 struct rq *rq = cpu_rq(cpu);
5187 unsigned long total = weighted_cpuload(cpu);
5189 if (type == 0 || !sched_feat(LB_BIAS))
5192 return min(rq->cpu_load[type-1], total);
5196 * Return a high guess at the load of a migration-target cpu weighted
5197 * according to the scheduling class and "nice" value.
5199 static unsigned long target_load(int cpu, int type)
5201 struct rq *rq = cpu_rq(cpu);
5202 unsigned long total = weighted_cpuload(cpu);
5204 if (type == 0 || !sched_feat(LB_BIAS))
5207 return max(rq->cpu_load[type-1], total);
5210 static unsigned long capacity_of(int cpu)
5212 return cpu_rq(cpu)->cpu_capacity;
5215 static unsigned long capacity_orig_of(int cpu)
5217 return cpu_rq(cpu)->cpu_capacity_orig;
5220 static unsigned long cpu_avg_load_per_task(int cpu)
5222 struct rq *rq = cpu_rq(cpu);
5223 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5224 unsigned long load_avg = weighted_cpuload(cpu);
5227 return load_avg / nr_running;
5232 #ifdef CONFIG_FAIR_GROUP_SCHED
5234 * effective_load() calculates the load change as seen from the root_task_group
5236 * Adding load to a group doesn't make a group heavier, but can cause movement
5237 * of group shares between cpus. Assuming the shares were perfectly aligned one
5238 * can calculate the shift in shares.
5240 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5241 * on this @cpu and results in a total addition (subtraction) of @wg to the
5242 * total group weight.
5244 * Given a runqueue weight distribution (rw_i) we can compute a shares
5245 * distribution (s_i) using:
5247 * s_i = rw_i / \Sum rw_j (1)
5249 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5250 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5251 * shares distribution (s_i):
5253 * rw_i = { 2, 4, 1, 0 }
5254 * s_i = { 2/7, 4/7, 1/7, 0 }
5256 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5257 * task used to run on and the CPU the waker is running on), we need to
5258 * compute the effect of waking a task on either CPU and, in case of a sync
5259 * wakeup, compute the effect of the current task going to sleep.
5261 * So for a change of @wl to the local @cpu with an overall group weight change
5262 * of @wl we can compute the new shares distribution (s'_i) using:
5264 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5266 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5267 * differences in waking a task to CPU 0. The additional task changes the
5268 * weight and shares distributions like:
5270 * rw'_i = { 3, 4, 1, 0 }
5271 * s'_i = { 3/8, 4/8, 1/8, 0 }
5273 * We can then compute the difference in effective weight by using:
5275 * dw_i = S * (s'_i - s_i) (3)
5277 * Where 'S' is the group weight as seen by its parent.
5279 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5280 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5281 * 4/7) times the weight of the group.
5283 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5285 struct sched_entity *se = tg->se[cpu];
5287 if (!tg->parent) /* the trivial, non-cgroup case */
5290 for_each_sched_entity(se) {
5291 struct cfs_rq *cfs_rq = se->my_q;
5292 long W, w = cfs_rq_load_avg(cfs_rq);
5297 * W = @wg + \Sum rw_j
5299 W = wg + atomic_long_read(&tg->load_avg);
5301 /* Ensure \Sum rw_j >= rw_i */
5302 W -= cfs_rq->tg_load_avg_contrib;
5311 * wl = S * s'_i; see (2)
5314 wl = (w * (long)scale_load_down(tg->shares)) / W;
5316 wl = scale_load_down(tg->shares);
5319 * Per the above, wl is the new se->load.weight value; since
5320 * those are clipped to [MIN_SHARES, ...) do so now. See
5321 * calc_cfs_shares().
5323 if (wl < MIN_SHARES)
5327 * wl = dw_i = S * (s'_i - s_i); see (3)
5329 wl -= se->avg.load_avg;
5332 * Recursively apply this logic to all parent groups to compute
5333 * the final effective load change on the root group. Since
5334 * only the @tg group gets extra weight, all parent groups can
5335 * only redistribute existing shares. @wl is the shift in shares
5336 * resulting from this level per the above.
5345 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5352 static void record_wakee(struct task_struct *p)
5355 * Only decay a single time; tasks that have less then 1 wakeup per
5356 * jiffy will not have built up many flips.
5358 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5359 current->wakee_flips >>= 1;
5360 current->wakee_flip_decay_ts = jiffies;
5363 if (current->last_wakee != p) {
5364 current->last_wakee = p;
5365 current->wakee_flips++;
5370 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5372 * A waker of many should wake a different task than the one last awakened
5373 * at a frequency roughly N times higher than one of its wakees.
5375 * In order to determine whether we should let the load spread vs consolidating
5376 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5377 * partner, and a factor of lls_size higher frequency in the other.
5379 * With both conditions met, we can be relatively sure that the relationship is
5380 * non-monogamous, with partner count exceeding socket size.
5382 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5383 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5386 static int wake_wide(struct task_struct *p)
5388 unsigned int master = current->wakee_flips;
5389 unsigned int slave = p->wakee_flips;
5390 int factor = this_cpu_read(sd_llc_size);
5393 swap(master, slave);
5394 if (slave < factor || master < slave * factor)
5399 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5400 int prev_cpu, int sync)
5402 s64 this_load, load;
5403 s64 this_eff_load, prev_eff_load;
5405 struct task_group *tg;
5406 unsigned long weight;
5410 this_cpu = smp_processor_id();
5411 load = source_load(prev_cpu, idx);
5412 this_load = target_load(this_cpu, idx);
5415 * If sync wakeup then subtract the (maximum possible)
5416 * effect of the currently running task from the load
5417 * of the current CPU:
5420 tg = task_group(current);
5421 weight = current->se.avg.load_avg;
5423 this_load += effective_load(tg, this_cpu, -weight, -weight);
5424 load += effective_load(tg, prev_cpu, 0, -weight);
5428 weight = p->se.avg.load_avg;
5431 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5432 * due to the sync cause above having dropped this_load to 0, we'll
5433 * always have an imbalance, but there's really nothing you can do
5434 * about that, so that's good too.
5436 * Otherwise check if either cpus are near enough in load to allow this
5437 * task to be woken on this_cpu.
5439 this_eff_load = 100;
5440 this_eff_load *= capacity_of(prev_cpu);
5442 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5443 prev_eff_load *= capacity_of(this_cpu);
5445 if (this_load > 0) {
5446 this_eff_load *= this_load +
5447 effective_load(tg, this_cpu, weight, weight);
5449 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5452 balanced = this_eff_load <= prev_eff_load;
5454 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5459 schedstat_inc(sd->ttwu_move_affine);
5460 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5465 static inline int task_util(struct task_struct *p);
5466 static int cpu_util_wake(int cpu, struct task_struct *p);
5468 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5470 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5474 * find_idlest_group finds and returns the least busy CPU group within the
5477 static struct sched_group *
5478 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5479 int this_cpu, int sd_flag)
5481 struct sched_group *idlest = NULL, *group = sd->groups;
5482 struct sched_group *most_spare_sg = NULL;
5483 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5484 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5485 unsigned long most_spare = 0, this_spare = 0;
5486 int load_idx = sd->forkexec_idx;
5487 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5488 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5489 (sd->imbalance_pct-100) / 100;
5491 if (sd_flag & SD_BALANCE_WAKE)
5492 load_idx = sd->wake_idx;
5495 unsigned long load, avg_load, runnable_load;
5496 unsigned long spare_cap, max_spare_cap;
5500 /* Skip over this group if it has no CPUs allowed */
5501 if (!cpumask_intersects(sched_group_cpus(group),
5505 local_group = cpumask_test_cpu(this_cpu,
5506 sched_group_cpus(group));
5509 * Tally up the load of all CPUs in the group and find
5510 * the group containing the CPU with most spare capacity.
5516 for_each_cpu(i, sched_group_cpus(group)) {
5517 /* Bias balancing toward cpus of our domain */
5519 load = source_load(i, load_idx);
5521 load = target_load(i, load_idx);
5523 runnable_load += load;
5525 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5527 spare_cap = capacity_spare_wake(i, p);
5529 if (spare_cap > max_spare_cap)
5530 max_spare_cap = spare_cap;
5533 /* Adjust by relative CPU capacity of the group */
5534 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5535 group->sgc->capacity;
5536 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5537 group->sgc->capacity;
5540 this_runnable_load = runnable_load;
5541 this_avg_load = avg_load;
5542 this_spare = max_spare_cap;
5544 if (min_runnable_load > (runnable_load + imbalance)) {
5546 * The runnable load is significantly smaller
5547 * so we can pick this new cpu
5549 min_runnable_load = runnable_load;
5550 min_avg_load = avg_load;
5552 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5553 (100*min_avg_load > imbalance_scale*avg_load)) {
5555 * The runnable loads are close so take the
5556 * blocked load into account through avg_load.
5558 min_avg_load = avg_load;
5562 if (most_spare < max_spare_cap) {
5563 most_spare = max_spare_cap;
5564 most_spare_sg = group;
5567 } while (group = group->next, group != sd->groups);
5570 * The cross-over point between using spare capacity or least load
5571 * is too conservative for high utilization tasks on partially
5572 * utilized systems if we require spare_capacity > task_util(p),
5573 * so we allow for some task stuffing by using
5574 * spare_capacity > task_util(p)/2.
5576 * Spare capacity can't be used for fork because the utilization has
5577 * not been set yet, we must first select a rq to compute the initial
5580 if (sd_flag & SD_BALANCE_FORK)
5583 if (this_spare > task_util(p) / 2 &&
5584 imbalance_scale*this_spare > 100*most_spare)
5587 if (most_spare > task_util(p) / 2)
5588 return most_spare_sg;
5594 if (min_runnable_load > (this_runnable_load + imbalance))
5597 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5598 (100*this_avg_load < imbalance_scale*min_avg_load))
5605 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5608 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5610 unsigned long load, min_load = ULONG_MAX;
5611 unsigned int min_exit_latency = UINT_MAX;
5612 u64 latest_idle_timestamp = 0;
5613 int least_loaded_cpu = this_cpu;
5614 int shallowest_idle_cpu = -1;
5617 /* Check if we have any choice: */
5618 if (group->group_weight == 1)
5619 return cpumask_first(sched_group_cpus(group));
5621 /* Traverse only the allowed CPUs */
5622 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
5624 struct rq *rq = cpu_rq(i);
5625 struct cpuidle_state *idle = idle_get_state(rq);
5626 if (idle && idle->exit_latency < min_exit_latency) {
5628 * We give priority to a CPU whose idle state
5629 * has the smallest exit latency irrespective
5630 * of any idle timestamp.
5632 min_exit_latency = idle->exit_latency;
5633 latest_idle_timestamp = rq->idle_stamp;
5634 shallowest_idle_cpu = i;
5635 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5636 rq->idle_stamp > latest_idle_timestamp) {
5638 * If equal or no active idle state, then
5639 * the most recently idled CPU might have
5642 latest_idle_timestamp = rq->idle_stamp;
5643 shallowest_idle_cpu = i;
5645 } else if (shallowest_idle_cpu == -1) {
5646 load = weighted_cpuload(i);
5647 if (load < min_load || (load == min_load && i == this_cpu)) {
5649 least_loaded_cpu = i;
5654 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5658 * Implement a for_each_cpu() variant that starts the scan at a given cpu
5659 * (@start), and wraps around.
5661 * This is used to scan for idle CPUs; such that not all CPUs looking for an
5662 * idle CPU find the same CPU. The down-side is that tasks tend to cycle
5663 * through the LLC domain.
5665 * Especially tbench is found sensitive to this.
5668 static int cpumask_next_wrap(int n, const struct cpumask *mask, int start, int *wrapped)
5673 next = find_next_bit(cpumask_bits(mask), nr_cpumask_bits, n+1);
5677 return nr_cpumask_bits;
5679 if (next >= nr_cpumask_bits) {
5689 #define for_each_cpu_wrap(cpu, mask, start, wrap) \
5690 for ((wrap) = 0, (cpu) = (start)-1; \
5691 (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)), \
5692 (cpu) < nr_cpumask_bits; )
5694 #ifdef CONFIG_SCHED_SMT
5696 static inline void set_idle_cores(int cpu, int val)
5698 struct sched_domain_shared *sds;
5700 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5702 WRITE_ONCE(sds->has_idle_cores, val);
5705 static inline bool test_idle_cores(int cpu, bool def)
5707 struct sched_domain_shared *sds;
5709 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5711 return READ_ONCE(sds->has_idle_cores);
5717 * Scans the local SMT mask to see if the entire core is idle, and records this
5718 * information in sd_llc_shared->has_idle_cores.
5720 * Since SMT siblings share all cache levels, inspecting this limited remote
5721 * state should be fairly cheap.
5723 void __update_idle_core(struct rq *rq)
5725 int core = cpu_of(rq);
5729 if (test_idle_cores(core, true))
5732 for_each_cpu(cpu, cpu_smt_mask(core)) {
5740 set_idle_cores(core, 1);
5746 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5747 * there are no idle cores left in the system; tracked through
5748 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5750 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5752 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5753 int core, cpu, wrap;
5755 if (!static_branch_likely(&sched_smt_present))
5758 if (!test_idle_cores(target, false))
5761 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5763 for_each_cpu_wrap(core, cpus, target, wrap) {
5766 for_each_cpu(cpu, cpu_smt_mask(core)) {
5767 cpumask_clear_cpu(cpu, cpus);
5777 * Failed to find an idle core; stop looking for one.
5779 set_idle_cores(target, 0);
5785 * Scan the local SMT mask for idle CPUs.
5787 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5791 if (!static_branch_likely(&sched_smt_present))
5794 for_each_cpu(cpu, cpu_smt_mask(target)) {
5795 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5804 #else /* CONFIG_SCHED_SMT */
5806 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5811 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5816 #endif /* CONFIG_SCHED_SMT */
5819 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5820 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5821 * average idle time for this rq (as found in rq->avg_idle).
5823 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5825 struct sched_domain *this_sd;
5826 u64 avg_cost, avg_idle = this_rq()->avg_idle;
5831 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5835 avg_cost = this_sd->avg_scan_cost;
5838 * Due to large variance we need a large fuzz factor; hackbench in
5839 * particularly is sensitive here.
5841 if (sched_feat(SIS_AVG_CPU) && (avg_idle / 512) < avg_cost)
5844 time = local_clock();
5846 for_each_cpu_wrap(cpu, sched_domain_span(sd), target, wrap) {
5847 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5853 time = local_clock() - time;
5854 cost = this_sd->avg_scan_cost;
5855 delta = (s64)(time - cost) / 8;
5856 this_sd->avg_scan_cost += delta;
5862 * Try and locate an idle core/thread in the LLC cache domain.
5864 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5866 struct sched_domain *sd;
5869 if (idle_cpu(target))
5873 * If the previous cpu is cache affine and idle, don't be stupid.
5875 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5878 sd = rcu_dereference(per_cpu(sd_llc, target));
5882 i = select_idle_core(p, sd, target);
5883 if ((unsigned)i < nr_cpumask_bits)
5886 i = select_idle_cpu(p, sd, target);
5887 if ((unsigned)i < nr_cpumask_bits)
5890 i = select_idle_smt(p, sd, target);
5891 if ((unsigned)i < nr_cpumask_bits)
5898 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5899 * tasks. The unit of the return value must be the one of capacity so we can
5900 * compare the utilization with the capacity of the CPU that is available for
5901 * CFS task (ie cpu_capacity).
5903 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5904 * recent utilization of currently non-runnable tasks on a CPU. It represents
5905 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5906 * capacity_orig is the cpu_capacity available at the highest frequency
5907 * (arch_scale_freq_capacity()).
5908 * The utilization of a CPU converges towards a sum equal to or less than the
5909 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5910 * the running time on this CPU scaled by capacity_curr.
5912 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5913 * higher than capacity_orig because of unfortunate rounding in
5914 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5915 * the average stabilizes with the new running time. We need to check that the
5916 * utilization stays within the range of [0..capacity_orig] and cap it if
5917 * necessary. Without utilization capping, a group could be seen as overloaded
5918 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5919 * available capacity. We allow utilization to overshoot capacity_curr (but not
5920 * capacity_orig) as it useful for predicting the capacity required after task
5921 * migrations (scheduler-driven DVFS).
5923 static int cpu_util(int cpu)
5925 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5926 unsigned long capacity = capacity_orig_of(cpu);
5928 return (util >= capacity) ? capacity : util;
5931 static inline int task_util(struct task_struct *p)
5933 return p->se.avg.util_avg;
5937 * cpu_util_wake: Compute cpu utilization with any contributions from
5938 * the waking task p removed.
5940 static int cpu_util_wake(int cpu, struct task_struct *p)
5942 unsigned long util, capacity;
5944 /* Task has no contribution or is new */
5945 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5946 return cpu_util(cpu);
5948 capacity = capacity_orig_of(cpu);
5949 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5951 return (util >= capacity) ? capacity : util;
5955 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5956 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5958 * In that case WAKE_AFFINE doesn't make sense and we'll let
5959 * BALANCE_WAKE sort things out.
5961 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5963 long min_cap, max_cap;
5965 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5966 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5968 /* Minimum capacity is close to max, no need to abort wake_affine */
5969 if (max_cap - min_cap < max_cap >> 3)
5972 /* Bring task utilization in sync with prev_cpu */
5973 sync_entity_load_avg(&p->se);
5975 return min_cap * 1024 < task_util(p) * capacity_margin;
5979 * select_task_rq_fair: Select target runqueue for the waking task in domains
5980 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5981 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5983 * Balances load by selecting the idlest cpu in the idlest group, or under
5984 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5986 * Returns the target cpu number.
5988 * preempt must be disabled.
5991 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5993 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5994 int cpu = smp_processor_id();
5995 int new_cpu = prev_cpu;
5996 int want_affine = 0;
5997 int sync = wake_flags & WF_SYNC;
5999 if (sd_flag & SD_BALANCE_WAKE) {
6001 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6002 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6006 for_each_domain(cpu, tmp) {
6007 if (!(tmp->flags & SD_LOAD_BALANCE))
6011 * If both cpu and prev_cpu are part of this domain,
6012 * cpu is a valid SD_WAKE_AFFINE target.
6014 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6015 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6020 if (tmp->flags & sd_flag)
6022 else if (!want_affine)
6027 sd = NULL; /* Prefer wake_affine over balance flags */
6028 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6033 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6034 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6037 struct sched_group *group;
6040 if (!(sd->flags & sd_flag)) {
6045 group = find_idlest_group(sd, p, cpu, sd_flag);
6051 new_cpu = find_idlest_cpu(group, p, cpu);
6052 if (new_cpu == -1 || new_cpu == cpu) {
6053 /* Now try balancing at a lower domain level of cpu */
6058 /* Now try balancing at a lower domain level of new_cpu */
6060 weight = sd->span_weight;
6062 for_each_domain(cpu, tmp) {
6063 if (weight <= tmp->span_weight)
6065 if (tmp->flags & sd_flag)
6068 /* while loop will break here if sd == NULL */
6076 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6077 * cfs_rq_of(p) references at time of call are still valid and identify the
6078 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6080 static void migrate_task_rq_fair(struct task_struct *p)
6083 * As blocked tasks retain absolute vruntime the migration needs to
6084 * deal with this by subtracting the old and adding the new
6085 * min_vruntime -- the latter is done by enqueue_entity() when placing
6086 * the task on the new runqueue.
6088 if (p->state == TASK_WAKING) {
6089 struct sched_entity *se = &p->se;
6090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6093 #ifndef CONFIG_64BIT
6094 u64 min_vruntime_copy;
6097 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6099 min_vruntime = cfs_rq->min_vruntime;
6100 } while (min_vruntime != min_vruntime_copy);
6102 min_vruntime = cfs_rq->min_vruntime;
6105 se->vruntime -= min_vruntime;
6109 * We are supposed to update the task to "current" time, then its up to date
6110 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6111 * what current time is, so simply throw away the out-of-date time. This
6112 * will result in the wakee task is less decayed, but giving the wakee more
6113 * load sounds not bad.
6115 remove_entity_load_avg(&p->se);
6117 /* Tell new CPU we are migrated */
6118 p->se.avg.last_update_time = 0;
6120 /* We have migrated, no longer consider this task hot */
6121 p->se.exec_start = 0;
6124 static void task_dead_fair(struct task_struct *p)
6126 remove_entity_load_avg(&p->se);
6128 #endif /* CONFIG_SMP */
6130 static unsigned long
6131 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6133 unsigned long gran = sysctl_sched_wakeup_granularity;
6136 * Since its curr running now, convert the gran from real-time
6137 * to virtual-time in his units.
6139 * By using 'se' instead of 'curr' we penalize light tasks, so
6140 * they get preempted easier. That is, if 'se' < 'curr' then
6141 * the resulting gran will be larger, therefore penalizing the
6142 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6143 * be smaller, again penalizing the lighter task.
6145 * This is especially important for buddies when the leftmost
6146 * task is higher priority than the buddy.
6148 return calc_delta_fair(gran, se);
6152 * Should 'se' preempt 'curr'.
6166 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6168 s64 gran, vdiff = curr->vruntime - se->vruntime;
6173 gran = wakeup_gran(curr, se);
6180 static void set_last_buddy(struct sched_entity *se)
6182 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6185 for_each_sched_entity(se)
6186 cfs_rq_of(se)->last = se;
6189 static void set_next_buddy(struct sched_entity *se)
6191 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6194 for_each_sched_entity(se)
6195 cfs_rq_of(se)->next = se;
6198 static void set_skip_buddy(struct sched_entity *se)
6200 for_each_sched_entity(se)
6201 cfs_rq_of(se)->skip = se;
6205 * Preempt the current task with a newly woken task if needed:
6207 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6209 struct task_struct *curr = rq->curr;
6210 struct sched_entity *se = &curr->se, *pse = &p->se;
6211 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6212 int scale = cfs_rq->nr_running >= sched_nr_latency;
6213 int next_buddy_marked = 0;
6215 if (unlikely(se == pse))
6219 * This is possible from callers such as attach_tasks(), in which we
6220 * unconditionally check_prempt_curr() after an enqueue (which may have
6221 * lead to a throttle). This both saves work and prevents false
6222 * next-buddy nomination below.
6224 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6227 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6228 set_next_buddy(pse);
6229 next_buddy_marked = 1;
6233 * We can come here with TIF_NEED_RESCHED already set from new task
6236 * Note: this also catches the edge-case of curr being in a throttled
6237 * group (e.g. via set_curr_task), since update_curr() (in the
6238 * enqueue of curr) will have resulted in resched being set. This
6239 * prevents us from potentially nominating it as a false LAST_BUDDY
6242 if (test_tsk_need_resched(curr))
6245 /* Idle tasks are by definition preempted by non-idle tasks. */
6246 if (unlikely(curr->policy == SCHED_IDLE) &&
6247 likely(p->policy != SCHED_IDLE))
6251 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6252 * is driven by the tick):
6254 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6257 find_matching_se(&se, &pse);
6258 update_curr(cfs_rq_of(se));
6260 if (wakeup_preempt_entity(se, pse) == 1) {
6262 * Bias pick_next to pick the sched entity that is
6263 * triggering this preemption.
6265 if (!next_buddy_marked)
6266 set_next_buddy(pse);
6275 * Only set the backward buddy when the current task is still
6276 * on the rq. This can happen when a wakeup gets interleaved
6277 * with schedule on the ->pre_schedule() or idle_balance()
6278 * point, either of which can * drop the rq lock.
6280 * Also, during early boot the idle thread is in the fair class,
6281 * for obvious reasons its a bad idea to schedule back to it.
6283 if (unlikely(!se->on_rq || curr == rq->idle))
6286 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6290 static struct task_struct *
6291 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6293 struct cfs_rq *cfs_rq = &rq->cfs;
6294 struct sched_entity *se;
6295 struct task_struct *p;
6299 #ifdef CONFIG_FAIR_GROUP_SCHED
6300 if (!cfs_rq->nr_running)
6303 if (prev->sched_class != &fair_sched_class)
6307 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6308 * likely that a next task is from the same cgroup as the current.
6310 * Therefore attempt to avoid putting and setting the entire cgroup
6311 * hierarchy, only change the part that actually changes.
6315 struct sched_entity *curr = cfs_rq->curr;
6318 * Since we got here without doing put_prev_entity() we also
6319 * have to consider cfs_rq->curr. If it is still a runnable
6320 * entity, update_curr() will update its vruntime, otherwise
6321 * forget we've ever seen it.
6325 update_curr(cfs_rq);
6330 * This call to check_cfs_rq_runtime() will do the
6331 * throttle and dequeue its entity in the parent(s).
6332 * Therefore the 'simple' nr_running test will indeed
6335 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6339 se = pick_next_entity(cfs_rq, curr);
6340 cfs_rq = group_cfs_rq(se);
6346 * Since we haven't yet done put_prev_entity and if the selected task
6347 * is a different task than we started out with, try and touch the
6348 * least amount of cfs_rqs.
6351 struct sched_entity *pse = &prev->se;
6353 while (!(cfs_rq = is_same_group(se, pse))) {
6354 int se_depth = se->depth;
6355 int pse_depth = pse->depth;
6357 if (se_depth <= pse_depth) {
6358 put_prev_entity(cfs_rq_of(pse), pse);
6359 pse = parent_entity(pse);
6361 if (se_depth >= pse_depth) {
6362 set_next_entity(cfs_rq_of(se), se);
6363 se = parent_entity(se);
6367 put_prev_entity(cfs_rq, pse);
6368 set_next_entity(cfs_rq, se);
6371 if (hrtick_enabled(rq))
6372 hrtick_start_fair(rq, p);
6379 if (!cfs_rq->nr_running)
6382 put_prev_task(rq, prev);
6385 se = pick_next_entity(cfs_rq, NULL);
6386 set_next_entity(cfs_rq, se);
6387 cfs_rq = group_cfs_rq(se);
6392 if (hrtick_enabled(rq))
6393 hrtick_start_fair(rq, p);
6398 new_tasks = idle_balance(rq, rf);
6401 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6402 * possible for any higher priority task to appear. In that case we
6403 * must re-start the pick_next_entity() loop.
6415 * Account for a descheduled task:
6417 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6419 struct sched_entity *se = &prev->se;
6420 struct cfs_rq *cfs_rq;
6422 for_each_sched_entity(se) {
6423 cfs_rq = cfs_rq_of(se);
6424 put_prev_entity(cfs_rq, se);
6429 * sched_yield() is very simple
6431 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6433 static void yield_task_fair(struct rq *rq)
6435 struct task_struct *curr = rq->curr;
6436 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6437 struct sched_entity *se = &curr->se;
6440 * Are we the only task in the tree?
6442 if (unlikely(rq->nr_running == 1))
6445 clear_buddies(cfs_rq, se);
6447 if (curr->policy != SCHED_BATCH) {
6448 update_rq_clock(rq);
6450 * Update run-time statistics of the 'current'.
6452 update_curr(cfs_rq);
6454 * Tell update_rq_clock() that we've just updated,
6455 * so we don't do microscopic update in schedule()
6456 * and double the fastpath cost.
6458 rq_clock_skip_update(rq, true);
6464 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6466 struct sched_entity *se = &p->se;
6468 /* throttled hierarchies are not runnable */
6469 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6472 /* Tell the scheduler that we'd really like pse to run next. */
6475 yield_task_fair(rq);
6481 /**************************************************
6482 * Fair scheduling class load-balancing methods.
6486 * The purpose of load-balancing is to achieve the same basic fairness the
6487 * per-cpu scheduler provides, namely provide a proportional amount of compute
6488 * time to each task. This is expressed in the following equation:
6490 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6492 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6493 * W_i,0 is defined as:
6495 * W_i,0 = \Sum_j w_i,j (2)
6497 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6498 * is derived from the nice value as per sched_prio_to_weight[].
6500 * The weight average is an exponential decay average of the instantaneous
6503 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6505 * C_i is the compute capacity of cpu i, typically it is the
6506 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6507 * can also include other factors [XXX].
6509 * To achieve this balance we define a measure of imbalance which follows
6510 * directly from (1):
6512 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6514 * We them move tasks around to minimize the imbalance. In the continuous
6515 * function space it is obvious this converges, in the discrete case we get
6516 * a few fun cases generally called infeasible weight scenarios.
6519 * - infeasible weights;
6520 * - local vs global optima in the discrete case. ]
6525 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6526 * for all i,j solution, we create a tree of cpus that follows the hardware
6527 * topology where each level pairs two lower groups (or better). This results
6528 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6529 * tree to only the first of the previous level and we decrease the frequency
6530 * of load-balance at each level inv. proportional to the number of cpus in
6536 * \Sum { --- * --- * 2^i } = O(n) (5)
6538 * `- size of each group
6539 * | | `- number of cpus doing load-balance
6541 * `- sum over all levels
6543 * Coupled with a limit on how many tasks we can migrate every balance pass,
6544 * this makes (5) the runtime complexity of the balancer.
6546 * An important property here is that each CPU is still (indirectly) connected
6547 * to every other cpu in at most O(log n) steps:
6549 * The adjacency matrix of the resulting graph is given by:
6552 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6555 * And you'll find that:
6557 * A^(log_2 n)_i,j != 0 for all i,j (7)
6559 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6560 * The task movement gives a factor of O(m), giving a convergence complexity
6563 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6568 * In order to avoid CPUs going idle while there's still work to do, new idle
6569 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6570 * tree itself instead of relying on other CPUs to bring it work.
6572 * This adds some complexity to both (5) and (8) but it reduces the total idle
6580 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6583 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6588 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6590 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6592 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6595 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6596 * rewrite all of this once again.]
6599 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6601 enum fbq_type { regular, remote, all };
6603 #define LBF_ALL_PINNED 0x01
6604 #define LBF_NEED_BREAK 0x02
6605 #define LBF_DST_PINNED 0x04
6606 #define LBF_SOME_PINNED 0x08
6609 struct sched_domain *sd;
6617 struct cpumask *dst_grpmask;
6619 enum cpu_idle_type idle;
6621 /* The set of CPUs under consideration for load-balancing */
6622 struct cpumask *cpus;
6627 unsigned int loop_break;
6628 unsigned int loop_max;
6630 enum fbq_type fbq_type;
6631 struct list_head tasks;
6635 * Is this task likely cache-hot:
6637 static int task_hot(struct task_struct *p, struct lb_env *env)
6641 lockdep_assert_held(&env->src_rq->lock);
6643 if (p->sched_class != &fair_sched_class)
6646 if (unlikely(p->policy == SCHED_IDLE))
6650 * Buddy candidates are cache hot:
6652 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6653 (&p->se == cfs_rq_of(&p->se)->next ||
6654 &p->se == cfs_rq_of(&p->se)->last))
6657 if (sysctl_sched_migration_cost == -1)
6659 if (sysctl_sched_migration_cost == 0)
6662 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6664 return delta < (s64)sysctl_sched_migration_cost;
6667 #ifdef CONFIG_NUMA_BALANCING
6669 * Returns 1, if task migration degrades locality
6670 * Returns 0, if task migration improves locality i.e migration preferred.
6671 * Returns -1, if task migration is not affected by locality.
6673 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6675 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6676 unsigned long src_faults, dst_faults;
6677 int src_nid, dst_nid;
6679 if (!static_branch_likely(&sched_numa_balancing))
6682 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6685 src_nid = cpu_to_node(env->src_cpu);
6686 dst_nid = cpu_to_node(env->dst_cpu);
6688 if (src_nid == dst_nid)
6691 /* Migrating away from the preferred node is always bad. */
6692 if (src_nid == p->numa_preferred_nid) {
6693 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6699 /* Encourage migration to the preferred node. */
6700 if (dst_nid == p->numa_preferred_nid)
6704 src_faults = group_faults(p, src_nid);
6705 dst_faults = group_faults(p, dst_nid);
6707 src_faults = task_faults(p, src_nid);
6708 dst_faults = task_faults(p, dst_nid);
6711 return dst_faults < src_faults;
6715 static inline int migrate_degrades_locality(struct task_struct *p,
6723 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6726 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6730 lockdep_assert_held(&env->src_rq->lock);
6733 * We do not migrate tasks that are:
6734 * 1) throttled_lb_pair, or
6735 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6736 * 3) running (obviously), or
6737 * 4) are cache-hot on their current CPU.
6739 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6742 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6745 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6747 env->flags |= LBF_SOME_PINNED;
6750 * Remember if this task can be migrated to any other cpu in
6751 * our sched_group. We may want to revisit it if we couldn't
6752 * meet load balance goals by pulling other tasks on src_cpu.
6754 * Also avoid computing new_dst_cpu if we have already computed
6755 * one in current iteration.
6757 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6760 /* Prevent to re-select dst_cpu via env's cpus */
6761 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6762 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6763 env->flags |= LBF_DST_PINNED;
6764 env->new_dst_cpu = cpu;
6772 /* Record that we found atleast one task that could run on dst_cpu */
6773 env->flags &= ~LBF_ALL_PINNED;
6775 if (task_running(env->src_rq, p)) {
6776 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6781 * Aggressive migration if:
6782 * 1) destination numa is preferred
6783 * 2) task is cache cold, or
6784 * 3) too many balance attempts have failed.
6786 tsk_cache_hot = migrate_degrades_locality(p, env);
6787 if (tsk_cache_hot == -1)
6788 tsk_cache_hot = task_hot(p, env);
6790 if (tsk_cache_hot <= 0 ||
6791 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6792 if (tsk_cache_hot == 1) {
6793 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6794 schedstat_inc(p->se.statistics.nr_forced_migrations);
6799 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6804 * detach_task() -- detach the task for the migration specified in env
6806 static void detach_task(struct task_struct *p, struct lb_env *env)
6808 lockdep_assert_held(&env->src_rq->lock);
6810 p->on_rq = TASK_ON_RQ_MIGRATING;
6811 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6812 set_task_cpu(p, env->dst_cpu);
6816 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6817 * part of active balancing operations within "domain".
6819 * Returns a task if successful and NULL otherwise.
6821 static struct task_struct *detach_one_task(struct lb_env *env)
6823 struct task_struct *p, *n;
6825 lockdep_assert_held(&env->src_rq->lock);
6827 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6828 if (!can_migrate_task(p, env))
6831 detach_task(p, env);
6834 * Right now, this is only the second place where
6835 * lb_gained[env->idle] is updated (other is detach_tasks)
6836 * so we can safely collect stats here rather than
6837 * inside detach_tasks().
6839 schedstat_inc(env->sd->lb_gained[env->idle]);
6845 static const unsigned int sched_nr_migrate_break = 32;
6848 * detach_tasks() -- tries to detach up to imbalance weighted load from
6849 * busiest_rq, as part of a balancing operation within domain "sd".
6851 * Returns number of detached tasks if successful and 0 otherwise.
6853 static int detach_tasks(struct lb_env *env)
6855 struct list_head *tasks = &env->src_rq->cfs_tasks;
6856 struct task_struct *p;
6860 lockdep_assert_held(&env->src_rq->lock);
6862 if (env->imbalance <= 0)
6865 while (!list_empty(tasks)) {
6867 * We don't want to steal all, otherwise we may be treated likewise,
6868 * which could at worst lead to a livelock crash.
6870 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6873 p = list_first_entry(tasks, struct task_struct, se.group_node);
6876 /* We've more or less seen every task there is, call it quits */
6877 if (env->loop > env->loop_max)
6880 /* take a breather every nr_migrate tasks */
6881 if (env->loop > env->loop_break) {
6882 env->loop_break += sched_nr_migrate_break;
6883 env->flags |= LBF_NEED_BREAK;
6887 if (!can_migrate_task(p, env))
6890 load = task_h_load(p);
6892 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6895 if ((load / 2) > env->imbalance)
6898 detach_task(p, env);
6899 list_add(&p->se.group_node, &env->tasks);
6902 env->imbalance -= load;
6904 #ifdef CONFIG_PREEMPT
6906 * NEWIDLE balancing is a source of latency, so preemptible
6907 * kernels will stop after the first task is detached to minimize
6908 * the critical section.
6910 if (env->idle == CPU_NEWLY_IDLE)
6915 * We only want to steal up to the prescribed amount of
6918 if (env->imbalance <= 0)
6923 list_move_tail(&p->se.group_node, tasks);
6927 * Right now, this is one of only two places we collect this stat
6928 * so we can safely collect detach_one_task() stats here rather
6929 * than inside detach_one_task().
6931 schedstat_add(env->sd->lb_gained[env->idle], detached);
6937 * attach_task() -- attach the task detached by detach_task() to its new rq.
6939 static void attach_task(struct rq *rq, struct task_struct *p)
6941 lockdep_assert_held(&rq->lock);
6943 BUG_ON(task_rq(p) != rq);
6944 activate_task(rq, p, ENQUEUE_NOCLOCK);
6945 p->on_rq = TASK_ON_RQ_QUEUED;
6946 check_preempt_curr(rq, p, 0);
6950 * attach_one_task() -- attaches the task returned from detach_one_task() to
6953 static void attach_one_task(struct rq *rq, struct task_struct *p)
6958 update_rq_clock(rq);
6964 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6967 static void attach_tasks(struct lb_env *env)
6969 struct list_head *tasks = &env->tasks;
6970 struct task_struct *p;
6973 rq_lock(env->dst_rq, &rf);
6974 update_rq_clock(env->dst_rq);
6976 while (!list_empty(tasks)) {
6977 p = list_first_entry(tasks, struct task_struct, se.group_node);
6978 list_del_init(&p->se.group_node);
6980 attach_task(env->dst_rq, p);
6983 rq_unlock(env->dst_rq, &rf);
6986 #ifdef CONFIG_FAIR_GROUP_SCHED
6987 static void update_blocked_averages(int cpu)
6989 struct rq *rq = cpu_rq(cpu);
6990 struct cfs_rq *cfs_rq;
6993 rq_lock_irqsave(rq, &rf);
6994 update_rq_clock(rq);
6997 * Iterates the task_group tree in a bottom up fashion, see
6998 * list_add_leaf_cfs_rq() for details.
7000 for_each_leaf_cfs_rq(rq, cfs_rq) {
7001 struct sched_entity *se;
7003 /* throttled entities do not contribute to load */
7004 if (throttled_hierarchy(cfs_rq))
7007 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
7008 update_tg_load_avg(cfs_rq, 0);
7010 /* Propagate pending load changes to the parent, if any: */
7011 se = cfs_rq->tg->se[cpu];
7012 if (se && !skip_blocked_update(se))
7013 update_load_avg(se, 0);
7015 rq_unlock_irqrestore(rq, &rf);
7019 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7020 * This needs to be done in a top-down fashion because the load of a child
7021 * group is a fraction of its parents load.
7023 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7025 struct rq *rq = rq_of(cfs_rq);
7026 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7027 unsigned long now = jiffies;
7030 if (cfs_rq->last_h_load_update == now)
7033 cfs_rq->h_load_next = NULL;
7034 for_each_sched_entity(se) {
7035 cfs_rq = cfs_rq_of(se);
7036 cfs_rq->h_load_next = se;
7037 if (cfs_rq->last_h_load_update == now)
7042 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7043 cfs_rq->last_h_load_update = now;
7046 while ((se = cfs_rq->h_load_next) != NULL) {
7047 load = cfs_rq->h_load;
7048 load = div64_ul(load * se->avg.load_avg,
7049 cfs_rq_load_avg(cfs_rq) + 1);
7050 cfs_rq = group_cfs_rq(se);
7051 cfs_rq->h_load = load;
7052 cfs_rq->last_h_load_update = now;
7056 static unsigned long task_h_load(struct task_struct *p)
7058 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7060 update_cfs_rq_h_load(cfs_rq);
7061 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7062 cfs_rq_load_avg(cfs_rq) + 1);
7065 static inline void update_blocked_averages(int cpu)
7067 struct rq *rq = cpu_rq(cpu);
7068 struct cfs_rq *cfs_rq = &rq->cfs;
7071 rq_lock_irqsave(rq, &rf);
7072 update_rq_clock(rq);
7073 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7074 rq_unlock_irqrestore(rq, &rf);
7077 static unsigned long task_h_load(struct task_struct *p)
7079 return p->se.avg.load_avg;
7083 /********** Helpers for find_busiest_group ************************/
7092 * sg_lb_stats - stats of a sched_group required for load_balancing
7094 struct sg_lb_stats {
7095 unsigned long avg_load; /*Avg load across the CPUs of the group */
7096 unsigned long group_load; /* Total load over the CPUs of the group */
7097 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7098 unsigned long load_per_task;
7099 unsigned long group_capacity;
7100 unsigned long group_util; /* Total utilization of the group */
7101 unsigned int sum_nr_running; /* Nr tasks running in the group */
7102 unsigned int idle_cpus;
7103 unsigned int group_weight;
7104 enum group_type group_type;
7105 int group_no_capacity;
7106 #ifdef CONFIG_NUMA_BALANCING
7107 unsigned int nr_numa_running;
7108 unsigned int nr_preferred_running;
7113 * sd_lb_stats - Structure to store the statistics of a sched_domain
7114 * during load balancing.
7116 struct sd_lb_stats {
7117 struct sched_group *busiest; /* Busiest group in this sd */
7118 struct sched_group *local; /* Local group in this sd */
7119 unsigned long total_load; /* Total load of all groups in sd */
7120 unsigned long total_capacity; /* Total capacity of all groups in sd */
7121 unsigned long avg_load; /* Average load across all groups in sd */
7123 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7124 struct sg_lb_stats local_stat; /* Statistics of the local group */
7127 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7130 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7131 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7132 * We must however clear busiest_stat::avg_load because
7133 * update_sd_pick_busiest() reads this before assignment.
7135 *sds = (struct sd_lb_stats){
7139 .total_capacity = 0UL,
7142 .sum_nr_running = 0,
7143 .group_type = group_other,
7149 * get_sd_load_idx - Obtain the load index for a given sched domain.
7150 * @sd: The sched_domain whose load_idx is to be obtained.
7151 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7153 * Return: The load index.
7155 static inline int get_sd_load_idx(struct sched_domain *sd,
7156 enum cpu_idle_type idle)
7162 load_idx = sd->busy_idx;
7165 case CPU_NEWLY_IDLE:
7166 load_idx = sd->newidle_idx;
7169 load_idx = sd->idle_idx;
7176 static unsigned long scale_rt_capacity(int cpu)
7178 struct rq *rq = cpu_rq(cpu);
7179 u64 total, used, age_stamp, avg;
7183 * Since we're reading these variables without serialization make sure
7184 * we read them once before doing sanity checks on them.
7186 age_stamp = READ_ONCE(rq->age_stamp);
7187 avg = READ_ONCE(rq->rt_avg);
7188 delta = __rq_clock_broken(rq) - age_stamp;
7190 if (unlikely(delta < 0))
7193 total = sched_avg_period() + delta;
7195 used = div_u64(avg, total);
7197 if (likely(used < SCHED_CAPACITY_SCALE))
7198 return SCHED_CAPACITY_SCALE - used;
7203 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7205 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7206 struct sched_group *sdg = sd->groups;
7208 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7210 capacity *= scale_rt_capacity(cpu);
7211 capacity >>= SCHED_CAPACITY_SHIFT;
7216 cpu_rq(cpu)->cpu_capacity = capacity;
7217 sdg->sgc->capacity = capacity;
7218 sdg->sgc->min_capacity = capacity;
7221 void update_group_capacity(struct sched_domain *sd, int cpu)
7223 struct sched_domain *child = sd->child;
7224 struct sched_group *group, *sdg = sd->groups;
7225 unsigned long capacity, min_capacity;
7226 unsigned long interval;
7228 interval = msecs_to_jiffies(sd->balance_interval);
7229 interval = clamp(interval, 1UL, max_load_balance_interval);
7230 sdg->sgc->next_update = jiffies + interval;
7233 update_cpu_capacity(sd, cpu);
7238 min_capacity = ULONG_MAX;
7240 if (child->flags & SD_OVERLAP) {
7242 * SD_OVERLAP domains cannot assume that child groups
7243 * span the current group.
7246 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7247 struct sched_group_capacity *sgc;
7248 struct rq *rq = cpu_rq(cpu);
7251 * build_sched_domains() -> init_sched_groups_capacity()
7252 * gets here before we've attached the domains to the
7255 * Use capacity_of(), which is set irrespective of domains
7256 * in update_cpu_capacity().
7258 * This avoids capacity from being 0 and
7259 * causing divide-by-zero issues on boot.
7261 if (unlikely(!rq->sd)) {
7262 capacity += capacity_of(cpu);
7264 sgc = rq->sd->groups->sgc;
7265 capacity += sgc->capacity;
7268 min_capacity = min(capacity, min_capacity);
7272 * !SD_OVERLAP domains can assume that child groups
7273 * span the current group.
7276 group = child->groups;
7278 struct sched_group_capacity *sgc = group->sgc;
7280 capacity += sgc->capacity;
7281 min_capacity = min(sgc->min_capacity, min_capacity);
7282 group = group->next;
7283 } while (group != child->groups);
7286 sdg->sgc->capacity = capacity;
7287 sdg->sgc->min_capacity = min_capacity;
7291 * Check whether the capacity of the rq has been noticeably reduced by side
7292 * activity. The imbalance_pct is used for the threshold.
7293 * Return true is the capacity is reduced
7296 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7298 return ((rq->cpu_capacity * sd->imbalance_pct) <
7299 (rq->cpu_capacity_orig * 100));
7303 * Group imbalance indicates (and tries to solve) the problem where balancing
7304 * groups is inadequate due to ->cpus_allowed constraints.
7306 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7307 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7310 * { 0 1 2 3 } { 4 5 6 7 }
7313 * If we were to balance group-wise we'd place two tasks in the first group and
7314 * two tasks in the second group. Clearly this is undesired as it will overload
7315 * cpu 3 and leave one of the cpus in the second group unused.
7317 * The current solution to this issue is detecting the skew in the first group
7318 * by noticing the lower domain failed to reach balance and had difficulty
7319 * moving tasks due to affinity constraints.
7321 * When this is so detected; this group becomes a candidate for busiest; see
7322 * update_sd_pick_busiest(). And calculate_imbalance() and
7323 * find_busiest_group() avoid some of the usual balance conditions to allow it
7324 * to create an effective group imbalance.
7326 * This is a somewhat tricky proposition since the next run might not find the
7327 * group imbalance and decide the groups need to be balanced again. A most
7328 * subtle and fragile situation.
7331 static inline int sg_imbalanced(struct sched_group *group)
7333 return group->sgc->imbalance;
7337 * group_has_capacity returns true if the group has spare capacity that could
7338 * be used by some tasks.
7339 * We consider that a group has spare capacity if the * number of task is
7340 * smaller than the number of CPUs or if the utilization is lower than the
7341 * available capacity for CFS tasks.
7342 * For the latter, we use a threshold to stabilize the state, to take into
7343 * account the variance of the tasks' load and to return true if the available
7344 * capacity in meaningful for the load balancer.
7345 * As an example, an available capacity of 1% can appear but it doesn't make
7346 * any benefit for the load balance.
7349 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7351 if (sgs->sum_nr_running < sgs->group_weight)
7354 if ((sgs->group_capacity * 100) >
7355 (sgs->group_util * env->sd->imbalance_pct))
7362 * group_is_overloaded returns true if the group has more tasks than it can
7364 * group_is_overloaded is not equals to !group_has_capacity because a group
7365 * with the exact right number of tasks, has no more spare capacity but is not
7366 * overloaded so both group_has_capacity and group_is_overloaded return
7370 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7372 if (sgs->sum_nr_running <= sgs->group_weight)
7375 if ((sgs->group_capacity * 100) <
7376 (sgs->group_util * env->sd->imbalance_pct))
7383 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7384 * per-CPU capacity than sched_group ref.
7387 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7389 return sg->sgc->min_capacity * capacity_margin <
7390 ref->sgc->min_capacity * 1024;
7394 group_type group_classify(struct sched_group *group,
7395 struct sg_lb_stats *sgs)
7397 if (sgs->group_no_capacity)
7398 return group_overloaded;
7400 if (sg_imbalanced(group))
7401 return group_imbalanced;
7407 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7408 * @env: The load balancing environment.
7409 * @group: sched_group whose statistics are to be updated.
7410 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7411 * @local_group: Does group contain this_cpu.
7412 * @sgs: variable to hold the statistics for this group.
7413 * @overload: Indicate more than one runnable task for any CPU.
7415 static inline void update_sg_lb_stats(struct lb_env *env,
7416 struct sched_group *group, int load_idx,
7417 int local_group, struct sg_lb_stats *sgs,
7423 memset(sgs, 0, sizeof(*sgs));
7425 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7426 struct rq *rq = cpu_rq(i);
7428 /* Bias balancing toward cpus of our domain */
7430 load = target_load(i, load_idx);
7432 load = source_load(i, load_idx);
7434 sgs->group_load += load;
7435 sgs->group_util += cpu_util(i);
7436 sgs->sum_nr_running += rq->cfs.h_nr_running;
7438 nr_running = rq->nr_running;
7442 #ifdef CONFIG_NUMA_BALANCING
7443 sgs->nr_numa_running += rq->nr_numa_running;
7444 sgs->nr_preferred_running += rq->nr_preferred_running;
7446 sgs->sum_weighted_load += weighted_cpuload(i);
7448 * No need to call idle_cpu() if nr_running is not 0
7450 if (!nr_running && idle_cpu(i))
7454 /* Adjust by relative CPU capacity of the group */
7455 sgs->group_capacity = group->sgc->capacity;
7456 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7458 if (sgs->sum_nr_running)
7459 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7461 sgs->group_weight = group->group_weight;
7463 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7464 sgs->group_type = group_classify(group, sgs);
7468 * update_sd_pick_busiest - return 1 on busiest group
7469 * @env: The load balancing environment.
7470 * @sds: sched_domain statistics
7471 * @sg: sched_group candidate to be checked for being the busiest
7472 * @sgs: sched_group statistics
7474 * Determine if @sg is a busier group than the previously selected
7477 * Return: %true if @sg is a busier group than the previously selected
7478 * busiest group. %false otherwise.
7480 static bool update_sd_pick_busiest(struct lb_env *env,
7481 struct sd_lb_stats *sds,
7482 struct sched_group *sg,
7483 struct sg_lb_stats *sgs)
7485 struct sg_lb_stats *busiest = &sds->busiest_stat;
7487 if (sgs->group_type > busiest->group_type)
7490 if (sgs->group_type < busiest->group_type)
7493 if (sgs->avg_load <= busiest->avg_load)
7496 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7500 * Candidate sg has no more than one task per CPU and
7501 * has higher per-CPU capacity. Migrating tasks to less
7502 * capable CPUs may harm throughput. Maximize throughput,
7503 * power/energy consequences are not considered.
7505 if (sgs->sum_nr_running <= sgs->group_weight &&
7506 group_smaller_cpu_capacity(sds->local, sg))
7510 /* This is the busiest node in its class. */
7511 if (!(env->sd->flags & SD_ASYM_PACKING))
7514 /* No ASYM_PACKING if target cpu is already busy */
7515 if (env->idle == CPU_NOT_IDLE)
7518 * ASYM_PACKING needs to move all the work to the highest
7519 * prority CPUs in the group, therefore mark all groups
7520 * of lower priority than ourself as busy.
7522 if (sgs->sum_nr_running &&
7523 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7527 /* Prefer to move from lowest priority cpu's work */
7528 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7529 sg->asym_prefer_cpu))
7536 #ifdef CONFIG_NUMA_BALANCING
7537 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7539 if (sgs->sum_nr_running > sgs->nr_numa_running)
7541 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7546 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7548 if (rq->nr_running > rq->nr_numa_running)
7550 if (rq->nr_running > rq->nr_preferred_running)
7555 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7560 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7564 #endif /* CONFIG_NUMA_BALANCING */
7567 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7568 * @env: The load balancing environment.
7569 * @sds: variable to hold the statistics for this sched_domain.
7571 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7573 struct sched_domain *child = env->sd->child;
7574 struct sched_group *sg = env->sd->groups;
7575 struct sg_lb_stats *local = &sds->local_stat;
7576 struct sg_lb_stats tmp_sgs;
7577 int load_idx, prefer_sibling = 0;
7578 bool overload = false;
7580 if (child && child->flags & SD_PREFER_SIBLING)
7583 load_idx = get_sd_load_idx(env->sd, env->idle);
7586 struct sg_lb_stats *sgs = &tmp_sgs;
7589 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7594 if (env->idle != CPU_NEWLY_IDLE ||
7595 time_after_eq(jiffies, sg->sgc->next_update))
7596 update_group_capacity(env->sd, env->dst_cpu);
7599 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7606 * In case the child domain prefers tasks go to siblings
7607 * first, lower the sg capacity so that we'll try
7608 * and move all the excess tasks away. We lower the capacity
7609 * of a group only if the local group has the capacity to fit
7610 * these excess tasks. The extra check prevents the case where
7611 * you always pull from the heaviest group when it is already
7612 * under-utilized (possible with a large weight task outweighs
7613 * the tasks on the system).
7615 if (prefer_sibling && sds->local &&
7616 group_has_capacity(env, local) &&
7617 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7618 sgs->group_no_capacity = 1;
7619 sgs->group_type = group_classify(sg, sgs);
7622 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7624 sds->busiest_stat = *sgs;
7628 /* Now, start updating sd_lb_stats */
7629 sds->total_load += sgs->group_load;
7630 sds->total_capacity += sgs->group_capacity;
7633 } while (sg != env->sd->groups);
7635 if (env->sd->flags & SD_NUMA)
7636 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7638 if (!env->sd->parent) {
7639 /* update overload indicator if we are at root domain */
7640 if (env->dst_rq->rd->overload != overload)
7641 env->dst_rq->rd->overload = overload;
7647 * check_asym_packing - Check to see if the group is packed into the
7650 * This is primarily intended to used at the sibling level. Some
7651 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7652 * case of POWER7, it can move to lower SMT modes only when higher
7653 * threads are idle. When in lower SMT modes, the threads will
7654 * perform better since they share less core resources. Hence when we
7655 * have idle threads, we want them to be the higher ones.
7657 * This packing function is run on idle threads. It checks to see if
7658 * the busiest CPU in this domain (core in the P7 case) has a higher
7659 * CPU number than the packing function is being run on. Here we are
7660 * assuming lower CPU number will be equivalent to lower a SMT thread
7663 * Return: 1 when packing is required and a task should be moved to
7664 * this CPU. The amount of the imbalance is returned in *imbalance.
7666 * @env: The load balancing environment.
7667 * @sds: Statistics of the sched_domain which is to be packed
7669 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7673 if (!(env->sd->flags & SD_ASYM_PACKING))
7676 if (env->idle == CPU_NOT_IDLE)
7682 busiest_cpu = sds->busiest->asym_prefer_cpu;
7683 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7686 env->imbalance = DIV_ROUND_CLOSEST(
7687 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7688 SCHED_CAPACITY_SCALE);
7694 * fix_small_imbalance - Calculate the minor imbalance that exists
7695 * amongst the groups of a sched_domain, during
7697 * @env: The load balancing environment.
7698 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7701 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7703 unsigned long tmp, capa_now = 0, capa_move = 0;
7704 unsigned int imbn = 2;
7705 unsigned long scaled_busy_load_per_task;
7706 struct sg_lb_stats *local, *busiest;
7708 local = &sds->local_stat;
7709 busiest = &sds->busiest_stat;
7711 if (!local->sum_nr_running)
7712 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7713 else if (busiest->load_per_task > local->load_per_task)
7716 scaled_busy_load_per_task =
7717 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7718 busiest->group_capacity;
7720 if (busiest->avg_load + scaled_busy_load_per_task >=
7721 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7722 env->imbalance = busiest->load_per_task;
7727 * OK, we don't have enough imbalance to justify moving tasks,
7728 * however we may be able to increase total CPU capacity used by
7732 capa_now += busiest->group_capacity *
7733 min(busiest->load_per_task, busiest->avg_load);
7734 capa_now += local->group_capacity *
7735 min(local->load_per_task, local->avg_load);
7736 capa_now /= SCHED_CAPACITY_SCALE;
7738 /* Amount of load we'd subtract */
7739 if (busiest->avg_load > scaled_busy_load_per_task) {
7740 capa_move += busiest->group_capacity *
7741 min(busiest->load_per_task,
7742 busiest->avg_load - scaled_busy_load_per_task);
7745 /* Amount of load we'd add */
7746 if (busiest->avg_load * busiest->group_capacity <
7747 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7748 tmp = (busiest->avg_load * busiest->group_capacity) /
7749 local->group_capacity;
7751 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7752 local->group_capacity;
7754 capa_move += local->group_capacity *
7755 min(local->load_per_task, local->avg_load + tmp);
7756 capa_move /= SCHED_CAPACITY_SCALE;
7758 /* Move if we gain throughput */
7759 if (capa_move > capa_now)
7760 env->imbalance = busiest->load_per_task;
7764 * calculate_imbalance - Calculate the amount of imbalance present within the
7765 * groups of a given sched_domain during load balance.
7766 * @env: load balance environment
7767 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7769 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7771 unsigned long max_pull, load_above_capacity = ~0UL;
7772 struct sg_lb_stats *local, *busiest;
7774 local = &sds->local_stat;
7775 busiest = &sds->busiest_stat;
7777 if (busiest->group_type == group_imbalanced) {
7779 * In the group_imb case we cannot rely on group-wide averages
7780 * to ensure cpu-load equilibrium, look at wider averages. XXX
7782 busiest->load_per_task =
7783 min(busiest->load_per_task, sds->avg_load);
7787 * Avg load of busiest sg can be less and avg load of local sg can
7788 * be greater than avg load across all sgs of sd because avg load
7789 * factors in sg capacity and sgs with smaller group_type are
7790 * skipped when updating the busiest sg:
7792 if (busiest->avg_load <= sds->avg_load ||
7793 local->avg_load >= sds->avg_load) {
7795 return fix_small_imbalance(env, sds);
7799 * If there aren't any idle cpus, avoid creating some.
7801 if (busiest->group_type == group_overloaded &&
7802 local->group_type == group_overloaded) {
7803 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7804 if (load_above_capacity > busiest->group_capacity) {
7805 load_above_capacity -= busiest->group_capacity;
7806 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7807 load_above_capacity /= busiest->group_capacity;
7809 load_above_capacity = ~0UL;
7813 * We're trying to get all the cpus to the average_load, so we don't
7814 * want to push ourselves above the average load, nor do we wish to
7815 * reduce the max loaded cpu below the average load. At the same time,
7816 * we also don't want to reduce the group load below the group
7817 * capacity. Thus we look for the minimum possible imbalance.
7819 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7821 /* How much load to actually move to equalise the imbalance */
7822 env->imbalance = min(
7823 max_pull * busiest->group_capacity,
7824 (sds->avg_load - local->avg_load) * local->group_capacity
7825 ) / SCHED_CAPACITY_SCALE;
7828 * if *imbalance is less than the average load per runnable task
7829 * there is no guarantee that any tasks will be moved so we'll have
7830 * a think about bumping its value to force at least one task to be
7833 if (env->imbalance < busiest->load_per_task)
7834 return fix_small_imbalance(env, sds);
7837 /******* find_busiest_group() helpers end here *********************/
7840 * find_busiest_group - Returns the busiest group within the sched_domain
7841 * if there is an imbalance.
7843 * Also calculates the amount of weighted load which should be moved
7844 * to restore balance.
7846 * @env: The load balancing environment.
7848 * Return: - The busiest group if imbalance exists.
7850 static struct sched_group *find_busiest_group(struct lb_env *env)
7852 struct sg_lb_stats *local, *busiest;
7853 struct sd_lb_stats sds;
7855 init_sd_lb_stats(&sds);
7858 * Compute the various statistics relavent for load balancing at
7861 update_sd_lb_stats(env, &sds);
7862 local = &sds.local_stat;
7863 busiest = &sds.busiest_stat;
7865 /* ASYM feature bypasses nice load balance check */
7866 if (check_asym_packing(env, &sds))
7869 /* There is no busy sibling group to pull tasks from */
7870 if (!sds.busiest || busiest->sum_nr_running == 0)
7873 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7874 / sds.total_capacity;
7877 * If the busiest group is imbalanced the below checks don't
7878 * work because they assume all things are equal, which typically
7879 * isn't true due to cpus_allowed constraints and the like.
7881 if (busiest->group_type == group_imbalanced)
7884 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7885 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7886 busiest->group_no_capacity)
7890 * If the local group is busier than the selected busiest group
7891 * don't try and pull any tasks.
7893 if (local->avg_load >= busiest->avg_load)
7897 * Don't pull any tasks if this group is already above the domain
7900 if (local->avg_load >= sds.avg_load)
7903 if (env->idle == CPU_IDLE) {
7905 * This cpu is idle. If the busiest group is not overloaded
7906 * and there is no imbalance between this and busiest group
7907 * wrt idle cpus, it is balanced. The imbalance becomes
7908 * significant if the diff is greater than 1 otherwise we
7909 * might end up to just move the imbalance on another group
7911 if ((busiest->group_type != group_overloaded) &&
7912 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7916 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7917 * imbalance_pct to be conservative.
7919 if (100 * busiest->avg_load <=
7920 env->sd->imbalance_pct * local->avg_load)
7925 /* Looks like there is an imbalance. Compute it */
7926 calculate_imbalance(env, &sds);
7935 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7937 static struct rq *find_busiest_queue(struct lb_env *env,
7938 struct sched_group *group)
7940 struct rq *busiest = NULL, *rq;
7941 unsigned long busiest_load = 0, busiest_capacity = 1;
7944 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7945 unsigned long capacity, wl;
7949 rt = fbq_classify_rq(rq);
7952 * We classify groups/runqueues into three groups:
7953 * - regular: there are !numa tasks
7954 * - remote: there are numa tasks that run on the 'wrong' node
7955 * - all: there is no distinction
7957 * In order to avoid migrating ideally placed numa tasks,
7958 * ignore those when there's better options.
7960 * If we ignore the actual busiest queue to migrate another
7961 * task, the next balance pass can still reduce the busiest
7962 * queue by moving tasks around inside the node.
7964 * If we cannot move enough load due to this classification
7965 * the next pass will adjust the group classification and
7966 * allow migration of more tasks.
7968 * Both cases only affect the total convergence complexity.
7970 if (rt > env->fbq_type)
7973 capacity = capacity_of(i);
7975 wl = weighted_cpuload(i);
7978 * When comparing with imbalance, use weighted_cpuload()
7979 * which is not scaled with the cpu capacity.
7982 if (rq->nr_running == 1 && wl > env->imbalance &&
7983 !check_cpu_capacity(rq, env->sd))
7987 * For the load comparisons with the other cpu's, consider
7988 * the weighted_cpuload() scaled with the cpu capacity, so
7989 * that the load can be moved away from the cpu that is
7990 * potentially running at a lower capacity.
7992 * Thus we're looking for max(wl_i / capacity_i), crosswise
7993 * multiplication to rid ourselves of the division works out
7994 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7995 * our previous maximum.
7997 if (wl * busiest_capacity > busiest_load * capacity) {
7999 busiest_capacity = capacity;
8008 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8009 * so long as it is large enough.
8011 #define MAX_PINNED_INTERVAL 512
8013 static int need_active_balance(struct lb_env *env)
8015 struct sched_domain *sd = env->sd;
8017 if (env->idle == CPU_NEWLY_IDLE) {
8020 * ASYM_PACKING needs to force migrate tasks from busy but
8021 * lower priority CPUs in order to pack all tasks in the
8022 * highest priority CPUs.
8024 if ((sd->flags & SD_ASYM_PACKING) &&
8025 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8030 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8031 * It's worth migrating the task if the src_cpu's capacity is reduced
8032 * because of other sched_class or IRQs if more capacity stays
8033 * available on dst_cpu.
8035 if ((env->idle != CPU_NOT_IDLE) &&
8036 (env->src_rq->cfs.h_nr_running == 1)) {
8037 if ((check_cpu_capacity(env->src_rq, sd)) &&
8038 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8042 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8045 static int active_load_balance_cpu_stop(void *data);
8047 static int should_we_balance(struct lb_env *env)
8049 struct sched_group *sg = env->sd->groups;
8050 struct cpumask *sg_cpus, *sg_mask;
8051 int cpu, balance_cpu = -1;
8054 * In the newly idle case, we will allow all the cpu's
8055 * to do the newly idle load balance.
8057 if (env->idle == CPU_NEWLY_IDLE)
8060 sg_cpus = sched_group_cpus(sg);
8061 sg_mask = sched_group_mask(sg);
8062 /* Try to find first idle cpu */
8063 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8064 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8071 if (balance_cpu == -1)
8072 balance_cpu = group_balance_cpu(sg);
8075 * First idle cpu or the first cpu(busiest) in this sched group
8076 * is eligible for doing load balancing at this and above domains.
8078 return balance_cpu == env->dst_cpu;
8082 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8083 * tasks if there is an imbalance.
8085 static int load_balance(int this_cpu, struct rq *this_rq,
8086 struct sched_domain *sd, enum cpu_idle_type idle,
8087 int *continue_balancing)
8089 int ld_moved, cur_ld_moved, active_balance = 0;
8090 struct sched_domain *sd_parent = sd->parent;
8091 struct sched_group *group;
8094 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8096 struct lb_env env = {
8098 .dst_cpu = this_cpu,
8100 .dst_grpmask = sched_group_cpus(sd->groups),
8102 .loop_break = sched_nr_migrate_break,
8105 .tasks = LIST_HEAD_INIT(env.tasks),
8109 * For NEWLY_IDLE load_balancing, we don't need to consider
8110 * other cpus in our group
8112 if (idle == CPU_NEWLY_IDLE)
8113 env.dst_grpmask = NULL;
8115 cpumask_copy(cpus, cpu_active_mask);
8117 schedstat_inc(sd->lb_count[idle]);
8120 if (!should_we_balance(&env)) {
8121 *continue_balancing = 0;
8125 group = find_busiest_group(&env);
8127 schedstat_inc(sd->lb_nobusyg[idle]);
8131 busiest = find_busiest_queue(&env, group);
8133 schedstat_inc(sd->lb_nobusyq[idle]);
8137 BUG_ON(busiest == env.dst_rq);
8139 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8141 env.src_cpu = busiest->cpu;
8142 env.src_rq = busiest;
8145 if (busiest->nr_running > 1) {
8147 * Attempt to move tasks. If find_busiest_group has found
8148 * an imbalance but busiest->nr_running <= 1, the group is
8149 * still unbalanced. ld_moved simply stays zero, so it is
8150 * correctly treated as an imbalance.
8152 env.flags |= LBF_ALL_PINNED;
8153 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8156 rq_lock_irqsave(busiest, &rf);
8157 update_rq_clock(busiest);
8160 * cur_ld_moved - load moved in current iteration
8161 * ld_moved - cumulative load moved across iterations
8163 cur_ld_moved = detach_tasks(&env);
8166 * We've detached some tasks from busiest_rq. Every
8167 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8168 * unlock busiest->lock, and we are able to be sure
8169 * that nobody can manipulate the tasks in parallel.
8170 * See task_rq_lock() family for the details.
8173 rq_unlock(busiest, &rf);
8177 ld_moved += cur_ld_moved;
8180 local_irq_restore(rf.flags);
8182 if (env.flags & LBF_NEED_BREAK) {
8183 env.flags &= ~LBF_NEED_BREAK;
8188 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8189 * us and move them to an alternate dst_cpu in our sched_group
8190 * where they can run. The upper limit on how many times we
8191 * iterate on same src_cpu is dependent on number of cpus in our
8194 * This changes load balance semantics a bit on who can move
8195 * load to a given_cpu. In addition to the given_cpu itself
8196 * (or a ilb_cpu acting on its behalf where given_cpu is
8197 * nohz-idle), we now have balance_cpu in a position to move
8198 * load to given_cpu. In rare situations, this may cause
8199 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8200 * _independently_ and at _same_ time to move some load to
8201 * given_cpu) causing exceess load to be moved to given_cpu.
8202 * This however should not happen so much in practice and
8203 * moreover subsequent load balance cycles should correct the
8204 * excess load moved.
8206 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8208 /* Prevent to re-select dst_cpu via env's cpus */
8209 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8211 env.dst_rq = cpu_rq(env.new_dst_cpu);
8212 env.dst_cpu = env.new_dst_cpu;
8213 env.flags &= ~LBF_DST_PINNED;
8215 env.loop_break = sched_nr_migrate_break;
8218 * Go back to "more_balance" rather than "redo" since we
8219 * need to continue with same src_cpu.
8225 * We failed to reach balance because of affinity.
8228 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8230 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8231 *group_imbalance = 1;
8234 /* All tasks on this runqueue were pinned by CPU affinity */
8235 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8236 cpumask_clear_cpu(cpu_of(busiest), cpus);
8237 if (!cpumask_empty(cpus)) {
8239 env.loop_break = sched_nr_migrate_break;
8242 goto out_all_pinned;
8247 schedstat_inc(sd->lb_failed[idle]);
8249 * Increment the failure counter only on periodic balance.
8250 * We do not want newidle balance, which can be very
8251 * frequent, pollute the failure counter causing
8252 * excessive cache_hot migrations and active balances.
8254 if (idle != CPU_NEWLY_IDLE)
8255 sd->nr_balance_failed++;
8257 if (need_active_balance(&env)) {
8258 unsigned long flags;
8260 raw_spin_lock_irqsave(&busiest->lock, flags);
8262 /* don't kick the active_load_balance_cpu_stop,
8263 * if the curr task on busiest cpu can't be
8266 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8267 raw_spin_unlock_irqrestore(&busiest->lock,
8269 env.flags |= LBF_ALL_PINNED;
8270 goto out_one_pinned;
8274 * ->active_balance synchronizes accesses to
8275 * ->active_balance_work. Once set, it's cleared
8276 * only after active load balance is finished.
8278 if (!busiest->active_balance) {
8279 busiest->active_balance = 1;
8280 busiest->push_cpu = this_cpu;
8283 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8285 if (active_balance) {
8286 stop_one_cpu_nowait(cpu_of(busiest),
8287 active_load_balance_cpu_stop, busiest,
8288 &busiest->active_balance_work);
8291 /* We've kicked active balancing, force task migration. */
8292 sd->nr_balance_failed = sd->cache_nice_tries+1;
8295 sd->nr_balance_failed = 0;
8297 if (likely(!active_balance)) {
8298 /* We were unbalanced, so reset the balancing interval */
8299 sd->balance_interval = sd->min_interval;
8302 * If we've begun active balancing, start to back off. This
8303 * case may not be covered by the all_pinned logic if there
8304 * is only 1 task on the busy runqueue (because we don't call
8307 if (sd->balance_interval < sd->max_interval)
8308 sd->balance_interval *= 2;
8315 * We reach balance although we may have faced some affinity
8316 * constraints. Clear the imbalance flag if it was set.
8319 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8321 if (*group_imbalance)
8322 *group_imbalance = 0;
8327 * We reach balance because all tasks are pinned at this level so
8328 * we can't migrate them. Let the imbalance flag set so parent level
8329 * can try to migrate them.
8331 schedstat_inc(sd->lb_balanced[idle]);
8333 sd->nr_balance_failed = 0;
8336 /* tune up the balancing interval */
8337 if (((env.flags & LBF_ALL_PINNED) &&
8338 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8339 (sd->balance_interval < sd->max_interval))
8340 sd->balance_interval *= 2;
8347 static inline unsigned long
8348 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8350 unsigned long interval = sd->balance_interval;
8353 interval *= sd->busy_factor;
8355 /* scale ms to jiffies */
8356 interval = msecs_to_jiffies(interval);
8357 interval = clamp(interval, 1UL, max_load_balance_interval);
8363 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8365 unsigned long interval, next;
8367 /* used by idle balance, so cpu_busy = 0 */
8368 interval = get_sd_balance_interval(sd, 0);
8369 next = sd->last_balance + interval;
8371 if (time_after(*next_balance, next))
8372 *next_balance = next;
8376 * idle_balance is called by schedule() if this_cpu is about to become
8377 * idle. Attempts to pull tasks from other CPUs.
8379 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8381 unsigned long next_balance = jiffies + HZ;
8382 int this_cpu = this_rq->cpu;
8383 struct sched_domain *sd;
8384 int pulled_task = 0;
8388 * We must set idle_stamp _before_ calling idle_balance(), such that we
8389 * measure the duration of idle_balance() as idle time.
8391 this_rq->idle_stamp = rq_clock(this_rq);
8394 * This is OK, because current is on_cpu, which avoids it being picked
8395 * for load-balance and preemption/IRQs are still disabled avoiding
8396 * further scheduler activity on it and we're being very careful to
8397 * re-start the picking loop.
8399 rq_unpin_lock(this_rq, rf);
8401 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8402 !this_rq->rd->overload) {
8404 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8406 update_next_balance(sd, &next_balance);
8412 raw_spin_unlock(&this_rq->lock);
8414 update_blocked_averages(this_cpu);
8416 for_each_domain(this_cpu, sd) {
8417 int continue_balancing = 1;
8418 u64 t0, domain_cost;
8420 if (!(sd->flags & SD_LOAD_BALANCE))
8423 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8424 update_next_balance(sd, &next_balance);
8428 if (sd->flags & SD_BALANCE_NEWIDLE) {
8429 t0 = sched_clock_cpu(this_cpu);
8431 pulled_task = load_balance(this_cpu, this_rq,
8433 &continue_balancing);
8435 domain_cost = sched_clock_cpu(this_cpu) - t0;
8436 if (domain_cost > sd->max_newidle_lb_cost)
8437 sd->max_newidle_lb_cost = domain_cost;
8439 curr_cost += domain_cost;
8442 update_next_balance(sd, &next_balance);
8445 * Stop searching for tasks to pull if there are
8446 * now runnable tasks on this rq.
8448 if (pulled_task || this_rq->nr_running > 0)
8453 raw_spin_lock(&this_rq->lock);
8455 if (curr_cost > this_rq->max_idle_balance_cost)
8456 this_rq->max_idle_balance_cost = curr_cost;
8459 * While browsing the domains, we released the rq lock, a task could
8460 * have been enqueued in the meantime. Since we're not going idle,
8461 * pretend we pulled a task.
8463 if (this_rq->cfs.h_nr_running && !pulled_task)
8467 /* Move the next balance forward */
8468 if (time_after(this_rq->next_balance, next_balance))
8469 this_rq->next_balance = next_balance;
8471 /* Is there a task of a high priority class? */
8472 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8476 this_rq->idle_stamp = 0;
8478 rq_repin_lock(this_rq, rf);
8484 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8485 * running tasks off the busiest CPU onto idle CPUs. It requires at
8486 * least 1 task to be running on each physical CPU where possible, and
8487 * avoids physical / logical imbalances.
8489 static int active_load_balance_cpu_stop(void *data)
8491 struct rq *busiest_rq = data;
8492 int busiest_cpu = cpu_of(busiest_rq);
8493 int target_cpu = busiest_rq->push_cpu;
8494 struct rq *target_rq = cpu_rq(target_cpu);
8495 struct sched_domain *sd;
8496 struct task_struct *p = NULL;
8499 rq_lock_irq(busiest_rq, &rf);
8501 /* make sure the requested cpu hasn't gone down in the meantime */
8502 if (unlikely(busiest_cpu != smp_processor_id() ||
8503 !busiest_rq->active_balance))
8506 /* Is there any task to move? */
8507 if (busiest_rq->nr_running <= 1)
8511 * This condition is "impossible", if it occurs
8512 * we need to fix it. Originally reported by
8513 * Bjorn Helgaas on a 128-cpu setup.
8515 BUG_ON(busiest_rq == target_rq);
8517 /* Search for an sd spanning us and the target CPU. */
8519 for_each_domain(target_cpu, sd) {
8520 if ((sd->flags & SD_LOAD_BALANCE) &&
8521 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8526 struct lb_env env = {
8528 .dst_cpu = target_cpu,
8529 .dst_rq = target_rq,
8530 .src_cpu = busiest_rq->cpu,
8531 .src_rq = busiest_rq,
8535 schedstat_inc(sd->alb_count);
8536 update_rq_clock(busiest_rq);
8538 p = detach_one_task(&env);
8540 schedstat_inc(sd->alb_pushed);
8541 /* Active balancing done, reset the failure counter. */
8542 sd->nr_balance_failed = 0;
8544 schedstat_inc(sd->alb_failed);
8549 busiest_rq->active_balance = 0;
8550 rq_unlock(busiest_rq, &rf);
8553 attach_one_task(target_rq, p);
8560 static inline int on_null_domain(struct rq *rq)
8562 return unlikely(!rcu_dereference_sched(rq->sd));
8565 #ifdef CONFIG_NO_HZ_COMMON
8567 * idle load balancing details
8568 * - When one of the busy CPUs notice that there may be an idle rebalancing
8569 * needed, they will kick the idle load balancer, which then does idle
8570 * load balancing for all the idle CPUs.
8573 cpumask_var_t idle_cpus_mask;
8575 unsigned long next_balance; /* in jiffy units */
8576 } nohz ____cacheline_aligned;
8578 static inline int find_new_ilb(void)
8580 int ilb = cpumask_first(nohz.idle_cpus_mask);
8582 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8589 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8590 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8591 * CPU (if there is one).
8593 static void nohz_balancer_kick(void)
8597 nohz.next_balance++;
8599 ilb_cpu = find_new_ilb();
8601 if (ilb_cpu >= nr_cpu_ids)
8604 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8607 * Use smp_send_reschedule() instead of resched_cpu().
8608 * This way we generate a sched IPI on the target cpu which
8609 * is idle. And the softirq performing nohz idle load balance
8610 * will be run before returning from the IPI.
8612 smp_send_reschedule(ilb_cpu);
8616 void nohz_balance_exit_idle(unsigned int cpu)
8618 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8620 * Completely isolated CPUs don't ever set, so we must test.
8622 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8623 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8624 atomic_dec(&nohz.nr_cpus);
8626 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8630 static inline void set_cpu_sd_state_busy(void)
8632 struct sched_domain *sd;
8633 int cpu = smp_processor_id();
8636 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8638 if (!sd || !sd->nohz_idle)
8642 atomic_inc(&sd->shared->nr_busy_cpus);
8647 void set_cpu_sd_state_idle(void)
8649 struct sched_domain *sd;
8650 int cpu = smp_processor_id();
8653 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8655 if (!sd || sd->nohz_idle)
8659 atomic_dec(&sd->shared->nr_busy_cpus);
8665 * This routine will record that the cpu is going idle with tick stopped.
8666 * This info will be used in performing idle load balancing in the future.
8668 void nohz_balance_enter_idle(int cpu)
8671 * If this cpu is going down, then nothing needs to be done.
8673 if (!cpu_active(cpu))
8676 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8680 * If we're a completely isolated CPU, we don't play.
8682 if (on_null_domain(cpu_rq(cpu)))
8685 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8686 atomic_inc(&nohz.nr_cpus);
8687 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8691 static DEFINE_SPINLOCK(balancing);
8694 * Scale the max load_balance interval with the number of CPUs in the system.
8695 * This trades load-balance latency on larger machines for less cross talk.
8697 void update_max_interval(void)
8699 max_load_balance_interval = HZ*num_online_cpus()/10;
8703 * It checks each scheduling domain to see if it is due to be balanced,
8704 * and initiates a balancing operation if so.
8706 * Balancing parameters are set up in init_sched_domains.
8708 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8710 int continue_balancing = 1;
8712 unsigned long interval;
8713 struct sched_domain *sd;
8714 /* Earliest time when we have to do rebalance again */
8715 unsigned long next_balance = jiffies + 60*HZ;
8716 int update_next_balance = 0;
8717 int need_serialize, need_decay = 0;
8720 update_blocked_averages(cpu);
8723 for_each_domain(cpu, sd) {
8725 * Decay the newidle max times here because this is a regular
8726 * visit to all the domains. Decay ~1% per second.
8728 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8729 sd->max_newidle_lb_cost =
8730 (sd->max_newidle_lb_cost * 253) / 256;
8731 sd->next_decay_max_lb_cost = jiffies + HZ;
8734 max_cost += sd->max_newidle_lb_cost;
8736 if (!(sd->flags & SD_LOAD_BALANCE))
8740 * Stop the load balance at this level. There is another
8741 * CPU in our sched group which is doing load balancing more
8744 if (!continue_balancing) {
8750 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8752 need_serialize = sd->flags & SD_SERIALIZE;
8753 if (need_serialize) {
8754 if (!spin_trylock(&balancing))
8758 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8759 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8761 * The LBF_DST_PINNED logic could have changed
8762 * env->dst_cpu, so we can't know our idle
8763 * state even if we migrated tasks. Update it.
8765 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8767 sd->last_balance = jiffies;
8768 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8771 spin_unlock(&balancing);
8773 if (time_after(next_balance, sd->last_balance + interval)) {
8774 next_balance = sd->last_balance + interval;
8775 update_next_balance = 1;
8780 * Ensure the rq-wide value also decays but keep it at a
8781 * reasonable floor to avoid funnies with rq->avg_idle.
8783 rq->max_idle_balance_cost =
8784 max((u64)sysctl_sched_migration_cost, max_cost);
8789 * next_balance will be updated only when there is a need.
8790 * When the cpu is attached to null domain for ex, it will not be
8793 if (likely(update_next_balance)) {
8794 rq->next_balance = next_balance;
8796 #ifdef CONFIG_NO_HZ_COMMON
8798 * If this CPU has been elected to perform the nohz idle
8799 * balance. Other idle CPUs have already rebalanced with
8800 * nohz_idle_balance() and nohz.next_balance has been
8801 * updated accordingly. This CPU is now running the idle load
8802 * balance for itself and we need to update the
8803 * nohz.next_balance accordingly.
8805 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8806 nohz.next_balance = rq->next_balance;
8811 #ifdef CONFIG_NO_HZ_COMMON
8813 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8814 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8816 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8818 int this_cpu = this_rq->cpu;
8821 /* Earliest time when we have to do rebalance again */
8822 unsigned long next_balance = jiffies + 60*HZ;
8823 int update_next_balance = 0;
8825 if (idle != CPU_IDLE ||
8826 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8829 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8830 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8834 * If this cpu gets work to do, stop the load balancing
8835 * work being done for other cpus. Next load
8836 * balancing owner will pick it up.
8841 rq = cpu_rq(balance_cpu);
8844 * If time for next balance is due,
8847 if (time_after_eq(jiffies, rq->next_balance)) {
8850 rq_lock_irq(rq, &rf);
8851 update_rq_clock(rq);
8852 cpu_load_update_idle(rq);
8853 rq_unlock_irq(rq, &rf);
8855 rebalance_domains(rq, CPU_IDLE);
8858 if (time_after(next_balance, rq->next_balance)) {
8859 next_balance = rq->next_balance;
8860 update_next_balance = 1;
8865 * next_balance will be updated only when there is a need.
8866 * When the CPU is attached to null domain for ex, it will not be
8869 if (likely(update_next_balance))
8870 nohz.next_balance = next_balance;
8872 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8876 * Current heuristic for kicking the idle load balancer in the presence
8877 * of an idle cpu in the system.
8878 * - This rq has more than one task.
8879 * - This rq has at least one CFS task and the capacity of the CPU is
8880 * significantly reduced because of RT tasks or IRQs.
8881 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8882 * multiple busy cpu.
8883 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8884 * domain span are idle.
8886 static inline bool nohz_kick_needed(struct rq *rq)
8888 unsigned long now = jiffies;
8889 struct sched_domain_shared *sds;
8890 struct sched_domain *sd;
8891 int nr_busy, i, cpu = rq->cpu;
8894 if (unlikely(rq->idle_balance))
8898 * We may be recently in ticked or tickless idle mode. At the first
8899 * busy tick after returning from idle, we will update the busy stats.
8901 set_cpu_sd_state_busy();
8902 nohz_balance_exit_idle(cpu);
8905 * None are in tickless mode and hence no need for NOHZ idle load
8908 if (likely(!atomic_read(&nohz.nr_cpus)))
8911 if (time_before(now, nohz.next_balance))
8914 if (rq->nr_running >= 2)
8918 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8921 * XXX: write a coherent comment on why we do this.
8922 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8924 nr_busy = atomic_read(&sds->nr_busy_cpus);
8932 sd = rcu_dereference(rq->sd);
8934 if ((rq->cfs.h_nr_running >= 1) &&
8935 check_cpu_capacity(rq, sd)) {
8941 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8943 for_each_cpu(i, sched_domain_span(sd)) {
8945 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
8948 if (sched_asym_prefer(i, cpu)) {
8959 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8963 * run_rebalance_domains is triggered when needed from the scheduler tick.
8964 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8966 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
8968 struct rq *this_rq = this_rq();
8969 enum cpu_idle_type idle = this_rq->idle_balance ?
8970 CPU_IDLE : CPU_NOT_IDLE;
8973 * If this cpu has a pending nohz_balance_kick, then do the
8974 * balancing on behalf of the other idle cpus whose ticks are
8975 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8976 * give the idle cpus a chance to load balance. Else we may
8977 * load balance only within the local sched_domain hierarchy
8978 * and abort nohz_idle_balance altogether if we pull some load.
8980 nohz_idle_balance(this_rq, idle);
8981 rebalance_domains(this_rq, idle);
8985 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8987 void trigger_load_balance(struct rq *rq)
8989 /* Don't need to rebalance while attached to NULL domain */
8990 if (unlikely(on_null_domain(rq)))
8993 if (time_after_eq(jiffies, rq->next_balance))
8994 raise_softirq(SCHED_SOFTIRQ);
8995 #ifdef CONFIG_NO_HZ_COMMON
8996 if (nohz_kick_needed(rq))
8997 nohz_balancer_kick();
9001 static void rq_online_fair(struct rq *rq)
9005 update_runtime_enabled(rq);
9008 static void rq_offline_fair(struct rq *rq)
9012 /* Ensure any throttled groups are reachable by pick_next_task */
9013 unthrottle_offline_cfs_rqs(rq);
9016 #endif /* CONFIG_SMP */
9019 * scheduler tick hitting a task of our scheduling class:
9021 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9023 struct cfs_rq *cfs_rq;
9024 struct sched_entity *se = &curr->se;
9026 for_each_sched_entity(se) {
9027 cfs_rq = cfs_rq_of(se);
9028 entity_tick(cfs_rq, se, queued);
9031 if (static_branch_unlikely(&sched_numa_balancing))
9032 task_tick_numa(rq, curr);
9036 * called on fork with the child task as argument from the parent's context
9037 * - child not yet on the tasklist
9038 * - preemption disabled
9040 static void task_fork_fair(struct task_struct *p)
9042 struct cfs_rq *cfs_rq;
9043 struct sched_entity *se = &p->se, *curr;
9044 struct rq *rq = this_rq();
9048 update_rq_clock(rq);
9050 cfs_rq = task_cfs_rq(current);
9051 curr = cfs_rq->curr;
9053 update_curr(cfs_rq);
9054 se->vruntime = curr->vruntime;
9056 place_entity(cfs_rq, se, 1);
9058 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9060 * Upon rescheduling, sched_class::put_prev_task() will place
9061 * 'current' within the tree based on its new key value.
9063 swap(curr->vruntime, se->vruntime);
9067 se->vruntime -= cfs_rq->min_vruntime;
9072 * Priority of the task has changed. Check to see if we preempt
9076 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9078 if (!task_on_rq_queued(p))
9082 * Reschedule if we are currently running on this runqueue and
9083 * our priority decreased, or if we are not currently running on
9084 * this runqueue and our priority is higher than the current's
9086 if (rq->curr == p) {
9087 if (p->prio > oldprio)
9090 check_preempt_curr(rq, p, 0);
9093 static inline bool vruntime_normalized(struct task_struct *p)
9095 struct sched_entity *se = &p->se;
9098 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9099 * the dequeue_entity(.flags=0) will already have normalized the
9106 * When !on_rq, vruntime of the task has usually NOT been normalized.
9107 * But there are some cases where it has already been normalized:
9109 * - A forked child which is waiting for being woken up by
9110 * wake_up_new_task().
9111 * - A task which has been woken up by try_to_wake_up() and
9112 * waiting for actually being woken up by sched_ttwu_pending().
9114 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9120 #ifdef CONFIG_FAIR_GROUP_SCHED
9122 * Propagate the changes of the sched_entity across the tg tree to make it
9123 * visible to the root
9125 static void propagate_entity_cfs_rq(struct sched_entity *se)
9127 struct cfs_rq *cfs_rq;
9129 /* Start to propagate at parent */
9132 for_each_sched_entity(se) {
9133 cfs_rq = cfs_rq_of(se);
9135 if (cfs_rq_throttled(cfs_rq))
9138 update_load_avg(se, UPDATE_TG);
9142 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9145 static void detach_entity_cfs_rq(struct sched_entity *se)
9147 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9149 /* Catch up with the cfs_rq and remove our load when we leave */
9150 update_load_avg(se, 0);
9151 detach_entity_load_avg(cfs_rq, se);
9152 update_tg_load_avg(cfs_rq, false);
9153 propagate_entity_cfs_rq(se);
9156 static void attach_entity_cfs_rq(struct sched_entity *se)
9158 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9160 #ifdef CONFIG_FAIR_GROUP_SCHED
9162 * Since the real-depth could have been changed (only FAIR
9163 * class maintain depth value), reset depth properly.
9165 se->depth = se->parent ? se->parent->depth + 1 : 0;
9168 /* Synchronize entity with its cfs_rq */
9169 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9170 attach_entity_load_avg(cfs_rq, se);
9171 update_tg_load_avg(cfs_rq, false);
9172 propagate_entity_cfs_rq(se);
9175 static void detach_task_cfs_rq(struct task_struct *p)
9177 struct sched_entity *se = &p->se;
9178 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9180 if (!vruntime_normalized(p)) {
9182 * Fix up our vruntime so that the current sleep doesn't
9183 * cause 'unlimited' sleep bonus.
9185 place_entity(cfs_rq, se, 0);
9186 se->vruntime -= cfs_rq->min_vruntime;
9189 detach_entity_cfs_rq(se);
9192 static void attach_task_cfs_rq(struct task_struct *p)
9194 struct sched_entity *se = &p->se;
9195 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9197 attach_entity_cfs_rq(se);
9199 if (!vruntime_normalized(p))
9200 se->vruntime += cfs_rq->min_vruntime;
9203 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9205 detach_task_cfs_rq(p);
9208 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9210 attach_task_cfs_rq(p);
9212 if (task_on_rq_queued(p)) {
9214 * We were most likely switched from sched_rt, so
9215 * kick off the schedule if running, otherwise just see
9216 * if we can still preempt the current task.
9221 check_preempt_curr(rq, p, 0);
9225 /* Account for a task changing its policy or group.
9227 * This routine is mostly called to set cfs_rq->curr field when a task
9228 * migrates between groups/classes.
9230 static void set_curr_task_fair(struct rq *rq)
9232 struct sched_entity *se = &rq->curr->se;
9234 for_each_sched_entity(se) {
9235 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9237 set_next_entity(cfs_rq, se);
9238 /* ensure bandwidth has been allocated on our new cfs_rq */
9239 account_cfs_rq_runtime(cfs_rq, 0);
9243 void init_cfs_rq(struct cfs_rq *cfs_rq)
9245 cfs_rq->tasks_timeline = RB_ROOT;
9246 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9247 #ifndef CONFIG_64BIT
9248 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9251 #ifdef CONFIG_FAIR_GROUP_SCHED
9252 cfs_rq->propagate_avg = 0;
9254 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9255 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9259 #ifdef CONFIG_FAIR_GROUP_SCHED
9260 static void task_set_group_fair(struct task_struct *p)
9262 struct sched_entity *se = &p->se;
9264 set_task_rq(p, task_cpu(p));
9265 se->depth = se->parent ? se->parent->depth + 1 : 0;
9268 static void task_move_group_fair(struct task_struct *p)
9270 detach_task_cfs_rq(p);
9271 set_task_rq(p, task_cpu(p));
9274 /* Tell se's cfs_rq has been changed -- migrated */
9275 p->se.avg.last_update_time = 0;
9277 attach_task_cfs_rq(p);
9280 static void task_change_group_fair(struct task_struct *p, int type)
9283 case TASK_SET_GROUP:
9284 task_set_group_fair(p);
9287 case TASK_MOVE_GROUP:
9288 task_move_group_fair(p);
9293 void free_fair_sched_group(struct task_group *tg)
9297 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9299 for_each_possible_cpu(i) {
9301 kfree(tg->cfs_rq[i]);
9310 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9312 struct sched_entity *se;
9313 struct cfs_rq *cfs_rq;
9316 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9319 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9323 tg->shares = NICE_0_LOAD;
9325 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9327 for_each_possible_cpu(i) {
9328 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9329 GFP_KERNEL, cpu_to_node(i));
9333 se = kzalloc_node(sizeof(struct sched_entity),
9334 GFP_KERNEL, cpu_to_node(i));
9338 init_cfs_rq(cfs_rq);
9339 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9340 init_entity_runnable_average(se);
9351 void online_fair_sched_group(struct task_group *tg)
9353 struct sched_entity *se;
9357 for_each_possible_cpu(i) {
9361 raw_spin_lock_irq(&rq->lock);
9362 update_rq_clock(rq);
9363 attach_entity_cfs_rq(se);
9364 sync_throttle(tg, i);
9365 raw_spin_unlock_irq(&rq->lock);
9369 void unregister_fair_sched_group(struct task_group *tg)
9371 unsigned long flags;
9375 for_each_possible_cpu(cpu) {
9377 remove_entity_load_avg(tg->se[cpu]);
9380 * Only empty task groups can be destroyed; so we can speculatively
9381 * check on_list without danger of it being re-added.
9383 if (!tg->cfs_rq[cpu]->on_list)
9388 raw_spin_lock_irqsave(&rq->lock, flags);
9389 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9390 raw_spin_unlock_irqrestore(&rq->lock, flags);
9394 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9395 struct sched_entity *se, int cpu,
9396 struct sched_entity *parent)
9398 struct rq *rq = cpu_rq(cpu);
9402 init_cfs_rq_runtime(cfs_rq);
9404 tg->cfs_rq[cpu] = cfs_rq;
9407 /* se could be NULL for root_task_group */
9412 se->cfs_rq = &rq->cfs;
9415 se->cfs_rq = parent->my_q;
9416 se->depth = parent->depth + 1;
9420 /* guarantee group entities always have weight */
9421 update_load_set(&se->load, NICE_0_LOAD);
9422 se->parent = parent;
9425 static DEFINE_MUTEX(shares_mutex);
9427 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9432 * We can't change the weight of the root cgroup.
9437 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9439 mutex_lock(&shares_mutex);
9440 if (tg->shares == shares)
9443 tg->shares = shares;
9444 for_each_possible_cpu(i) {
9445 struct rq *rq = cpu_rq(i);
9446 struct sched_entity *se = tg->se[i];
9449 /* Propagate contribution to hierarchy */
9450 rq_lock_irqsave(rq, &rf);
9451 update_rq_clock(rq);
9452 for_each_sched_entity(se) {
9453 update_load_avg(se, UPDATE_TG);
9454 update_cfs_shares(se);
9456 rq_unlock_irqrestore(rq, &rf);
9460 mutex_unlock(&shares_mutex);
9463 #else /* CONFIG_FAIR_GROUP_SCHED */
9465 void free_fair_sched_group(struct task_group *tg) { }
9467 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9472 void online_fair_sched_group(struct task_group *tg) { }
9474 void unregister_fair_sched_group(struct task_group *tg) { }
9476 #endif /* CONFIG_FAIR_GROUP_SCHED */
9479 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9481 struct sched_entity *se = &task->se;
9482 unsigned int rr_interval = 0;
9485 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9488 if (rq->cfs.load.weight)
9489 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9495 * All the scheduling class methods:
9497 const struct sched_class fair_sched_class = {
9498 .next = &idle_sched_class,
9499 .enqueue_task = enqueue_task_fair,
9500 .dequeue_task = dequeue_task_fair,
9501 .yield_task = yield_task_fair,
9502 .yield_to_task = yield_to_task_fair,
9504 .check_preempt_curr = check_preempt_wakeup,
9506 .pick_next_task = pick_next_task_fair,
9507 .put_prev_task = put_prev_task_fair,
9510 .select_task_rq = select_task_rq_fair,
9511 .migrate_task_rq = migrate_task_rq_fair,
9513 .rq_online = rq_online_fair,
9514 .rq_offline = rq_offline_fair,
9516 .task_dead = task_dead_fair,
9517 .set_cpus_allowed = set_cpus_allowed_common,
9520 .set_curr_task = set_curr_task_fair,
9521 .task_tick = task_tick_fair,
9522 .task_fork = task_fork_fair,
9524 .prio_changed = prio_changed_fair,
9525 .switched_from = switched_from_fair,
9526 .switched_to = switched_to_fair,
9528 .get_rr_interval = get_rr_interval_fair,
9530 .update_curr = update_curr_fair,
9532 #ifdef CONFIG_FAIR_GROUP_SCHED
9533 .task_change_group = task_change_group_fair,
9537 #ifdef CONFIG_SCHED_DEBUG
9538 void print_cfs_stats(struct seq_file *m, int cpu)
9540 struct cfs_rq *cfs_rq;
9543 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9544 print_cfs_rq(m, cpu, cfs_rq);
9548 #ifdef CONFIG_NUMA_BALANCING
9549 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9552 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9554 for_each_online_node(node) {
9555 if (p->numa_faults) {
9556 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9557 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9559 if (p->numa_group) {
9560 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9561 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9563 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9566 #endif /* CONFIG_NUMA_BALANCING */
9567 #endif /* CONFIG_SCHED_DEBUG */
9569 __init void init_sched_fair_class(void)
9572 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9574 #ifdef CONFIG_NO_HZ_COMMON
9575 nohz.next_balance = jiffies;
9576 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);