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_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
376 /* Do the two (enqueued) entities belong to the same group ? */
377 static inline struct cfs_rq *
378 is_same_group(struct sched_entity *se, struct sched_entity *pse)
380 if (se->cfs_rq == pse->cfs_rq)
386 static inline struct sched_entity *parent_entity(struct sched_entity *se)
392 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
394 int se_depth, pse_depth;
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
403 /* First walk up until both entities are at same depth */
404 se_depth = (*se)->depth;
405 pse_depth = (*pse)->depth;
407 while (se_depth > pse_depth) {
409 *se = parent_entity(*se);
412 while (pse_depth > se_depth) {
414 *pse = parent_entity(*pse);
417 while (!is_same_group(*se, *pse)) {
418 *se = parent_entity(*se);
419 *pse = parent_entity(*pse);
423 #else /* !CONFIG_FAIR_GROUP_SCHED */
425 static inline struct task_struct *task_of(struct sched_entity *se)
427 return container_of(se, struct task_struct, se);
430 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
432 return container_of(cfs_rq, struct rq, cfs);
435 #define entity_is_task(se) 1
437 #define for_each_sched_entity(se) \
438 for (; se; se = NULL)
440 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
442 return &task_rq(p)->cfs;
445 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
447 struct task_struct *p = task_of(se);
448 struct rq *rq = task_rq(p);
453 /* runqueue "owned" by this group */
454 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
459 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
463 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
467 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
468 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
470 static inline struct sched_entity *parent_entity(struct sched_entity *se)
476 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
480 #endif /* CONFIG_FAIR_GROUP_SCHED */
482 static __always_inline
483 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
485 /**************************************************************
486 * Scheduling class tree data structure manipulation methods:
489 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
491 s64 delta = (s64)(vruntime - max_vruntime);
493 max_vruntime = vruntime;
498 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
500 s64 delta = (s64)(vruntime - min_vruntime);
502 min_vruntime = vruntime;
507 static inline int entity_before(struct sched_entity *a,
508 struct sched_entity *b)
510 return (s64)(a->vruntime - b->vruntime) < 0;
513 static void update_min_vruntime(struct cfs_rq *cfs_rq)
515 struct sched_entity *curr = cfs_rq->curr;
517 u64 vruntime = cfs_rq->min_vruntime;
521 vruntime = curr->vruntime;
526 if (cfs_rq->rb_leftmost) {
527 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
532 vruntime = se->vruntime;
534 vruntime = min_vruntime(vruntime, se->vruntime);
537 /* ensure we never gain time by being placed backwards. */
538 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
541 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
546 * Enqueue an entity into the rb-tree:
548 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
550 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
551 struct rb_node *parent = NULL;
552 struct sched_entity *entry;
556 * Find the right place in the rbtree:
560 entry = rb_entry(parent, struct sched_entity, run_node);
562 * We dont care about collisions. Nodes with
563 * the same key stay together.
565 if (entity_before(se, entry)) {
566 link = &parent->rb_left;
568 link = &parent->rb_right;
574 * Maintain a cache of leftmost tree entries (it is frequently
578 cfs_rq->rb_leftmost = &se->run_node;
580 rb_link_node(&se->run_node, parent, link);
581 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
584 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
586 if (cfs_rq->rb_leftmost == &se->run_node) {
587 struct rb_node *next_node;
589 next_node = rb_next(&se->run_node);
590 cfs_rq->rb_leftmost = next_node;
593 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
596 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
598 struct rb_node *left = cfs_rq->rb_leftmost;
603 return rb_entry(left, struct sched_entity, run_node);
606 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
608 struct rb_node *next = rb_next(&se->run_node);
613 return rb_entry(next, struct sched_entity, run_node);
616 #ifdef CONFIG_SCHED_DEBUG
617 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
619 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
624 return rb_entry(last, struct sched_entity, run_node);
627 /**************************************************************
628 * Scheduling class statistics methods:
631 int sched_proc_update_handler(struct ctl_table *table, int write,
632 void __user *buffer, size_t *lenp,
635 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
636 unsigned int factor = get_update_sysctl_factor();
641 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
642 sysctl_sched_min_granularity);
644 #define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity);
647 WRT_SYSCTL(sched_latency);
648 WRT_SYSCTL(sched_wakeup_granularity);
658 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
660 if (unlikely(se->load.weight != NICE_0_LOAD))
661 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
667 * The idea is to set a period in which each task runs once.
669 * When there are too many tasks (sched_nr_latency) we have to stretch
670 * this period because otherwise the slices get too small.
672 * p = (nr <= nl) ? l : l*nr/nl
674 static u64 __sched_period(unsigned long nr_running)
676 if (unlikely(nr_running > sched_nr_latency))
677 return nr_running * sysctl_sched_min_granularity;
679 return sysctl_sched_latency;
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
688 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
692 for_each_sched_entity(se) {
693 struct load_weight *load;
694 struct load_weight lw;
696 cfs_rq = cfs_rq_of(se);
697 load = &cfs_rq->load;
699 if (unlikely(!se->on_rq)) {
702 update_load_add(&lw, se->load.weight);
705 slice = __calc_delta(slice, se->load.weight, load);
711 * We calculate the vruntime slice of a to-be-inserted task.
715 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
717 return calc_delta_fair(sched_slice(cfs_rq, se), se);
722 #include "sched-pelt.h"
724 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
725 static unsigned long task_h_load(struct task_struct *p);
727 /* Give new sched_entity start runnable values to heavy its load in infant time */
728 void init_entity_runnable_average(struct sched_entity *se)
730 struct sched_avg *sa = &se->avg;
732 sa->last_update_time = 0;
734 * sched_avg's period_contrib should be strictly less then 1024, so
735 * we give it 1023 to make sure it is almost a period (1024us), and
736 * will definitely be update (after enqueue).
738 sa->period_contrib = 1023;
740 * Tasks are intialized with full load to be seen as heavy tasks until
741 * they get a chance to stabilize to their real load level.
742 * Group entities are intialized with zero load to reflect the fact that
743 * nothing has been attached to the task group yet.
745 if (entity_is_task(se))
746 sa->load_avg = scale_load_down(se->load.weight);
747 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
749 * At this point, util_avg won't be used in select_task_rq_fair anyway
753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
756 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
757 static void attach_entity_cfs_rq(struct sched_entity *se);
760 * With new tasks being created, their initial util_avgs are extrapolated
761 * based on the cfs_rq's current util_avg:
763 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
765 * However, in many cases, the above util_avg does not give a desired
766 * value. Moreover, the sum of the util_avgs may be divergent, such
767 * as when the series is a harmonic series.
769 * To solve this problem, we also cap the util_avg of successive tasks to
770 * only 1/2 of the left utilization budget:
772 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
774 * where n denotes the nth task.
776 * For example, a simplest series from the beginning would be like:
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
784 void post_init_entity_util_avg(struct sched_entity *se)
786 struct cfs_rq *cfs_rq = cfs_rq_of(se);
787 struct sched_avg *sa = &se->avg;
788 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
791 if (cfs_rq->avg.util_avg != 0) {
792 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
793 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
795 if (sa->util_avg > cap)
800 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
803 if (entity_is_task(se)) {
804 struct task_struct *p = task_of(se);
805 if (p->sched_class != &fair_sched_class) {
807 * For !fair tasks do:
809 update_cfs_rq_load_avg(now, cfs_rq, false);
810 attach_entity_load_avg(cfs_rq, se);
811 switched_from_fair(rq, p);
813 * such that the next switched_to_fair() has the
816 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
821 attach_entity_cfs_rq(se);
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity *se)
828 void post_init_entity_util_avg(struct sched_entity *se)
831 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
834 #endif /* CONFIG_SMP */
837 * Update the current task's runtime statistics.
839 static void update_curr(struct cfs_rq *cfs_rq)
841 struct sched_entity *curr = cfs_rq->curr;
842 u64 now = rq_clock_task(rq_of(cfs_rq));
848 delta_exec = now - curr->exec_start;
849 if (unlikely((s64)delta_exec <= 0))
852 curr->exec_start = now;
854 schedstat_set(curr->statistics.exec_max,
855 max(delta_exec, curr->statistics.exec_max));
857 curr->sum_exec_runtime += delta_exec;
858 schedstat_add(cfs_rq->exec_clock, delta_exec);
860 curr->vruntime += calc_delta_fair(delta_exec, curr);
861 update_min_vruntime(cfs_rq);
863 if (entity_is_task(curr)) {
864 struct task_struct *curtask = task_of(curr);
866 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
867 cpuacct_charge(curtask, delta_exec);
868 account_group_exec_runtime(curtask, delta_exec);
871 account_cfs_rq_runtime(cfs_rq, delta_exec);
874 static void update_curr_fair(struct rq *rq)
876 update_curr(cfs_rq_of(&rq->curr->se));
880 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
882 u64 wait_start, prev_wait_start;
884 if (!schedstat_enabled())
887 wait_start = rq_clock(rq_of(cfs_rq));
888 prev_wait_start = schedstat_val(se->statistics.wait_start);
890 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
891 likely(wait_start > prev_wait_start))
892 wait_start -= prev_wait_start;
894 schedstat_set(se->statistics.wait_start, wait_start);
898 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
900 struct task_struct *p;
903 if (!schedstat_enabled())
906 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
908 if (entity_is_task(se)) {
910 if (task_on_rq_migrating(p)) {
912 * Preserve migrating task's wait time so wait_start
913 * time stamp can be adjusted to accumulate wait time
914 * prior to migration.
916 schedstat_set(se->statistics.wait_start, delta);
919 trace_sched_stat_wait(p, delta);
922 schedstat_set(se->statistics.wait_max,
923 max(schedstat_val(se->statistics.wait_max), delta));
924 schedstat_inc(se->statistics.wait_count);
925 schedstat_add(se->statistics.wait_sum, delta);
926 schedstat_set(se->statistics.wait_start, 0);
930 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
932 struct task_struct *tsk = NULL;
933 u64 sleep_start, block_start;
935 if (!schedstat_enabled())
938 sleep_start = schedstat_val(se->statistics.sleep_start);
939 block_start = schedstat_val(se->statistics.block_start);
941 if (entity_is_task(se))
945 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
950 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
951 schedstat_set(se->statistics.sleep_max, delta);
953 schedstat_set(se->statistics.sleep_start, 0);
954 schedstat_add(se->statistics.sum_sleep_runtime, delta);
957 account_scheduler_latency(tsk, delta >> 10, 1);
958 trace_sched_stat_sleep(tsk, delta);
962 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
967 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
968 schedstat_set(se->statistics.block_max, delta);
970 schedstat_set(se->statistics.block_start, 0);
971 schedstat_add(se->statistics.sum_sleep_runtime, delta);
974 if (tsk->in_iowait) {
975 schedstat_add(se->statistics.iowait_sum, delta);
976 schedstat_inc(se->statistics.iowait_count);
977 trace_sched_stat_iowait(tsk, delta);
980 trace_sched_stat_blocked(tsk, delta);
983 * Blocking time is in units of nanosecs, so shift by
984 * 20 to get a milliseconds-range estimation of the
985 * amount of time that the task spent sleeping:
987 if (unlikely(prof_on == SLEEP_PROFILING)) {
988 profile_hits(SLEEP_PROFILING,
989 (void *)get_wchan(tsk),
992 account_scheduler_latency(tsk, delta >> 10, 0);
998 * Task is being enqueued - update stats:
1001 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1003 if (!schedstat_enabled())
1007 * Are we enqueueing a waiting task? (for current tasks
1008 * a dequeue/enqueue event is a NOP)
1010 if (se != cfs_rq->curr)
1011 update_stats_wait_start(cfs_rq, se);
1013 if (flags & ENQUEUE_WAKEUP)
1014 update_stats_enqueue_sleeper(cfs_rq, se);
1018 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1021 if (!schedstat_enabled())
1025 * Mark the end of the wait period if dequeueing a
1028 if (se != cfs_rq->curr)
1029 update_stats_wait_end(cfs_rq, se);
1031 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1032 struct task_struct *tsk = task_of(se);
1034 if (tsk->state & TASK_INTERRUPTIBLE)
1035 schedstat_set(se->statistics.sleep_start,
1036 rq_clock(rq_of(cfs_rq)));
1037 if (tsk->state & TASK_UNINTERRUPTIBLE)
1038 schedstat_set(se->statistics.block_start,
1039 rq_clock(rq_of(cfs_rq)));
1044 * We are picking a new current task - update its stats:
1047 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1050 * We are starting a new run period:
1052 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1055 /**************************************************
1056 * Scheduling class queueing methods:
1059 #ifdef CONFIG_NUMA_BALANCING
1061 * Approximate time to scan a full NUMA task in ms. The task scan period is
1062 * calculated based on the tasks virtual memory size and
1063 * numa_balancing_scan_size.
1065 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1066 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1068 /* Portion of address space to scan in MB */
1069 unsigned int sysctl_numa_balancing_scan_size = 256;
1071 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1074 static unsigned int task_nr_scan_windows(struct task_struct *p)
1076 unsigned long rss = 0;
1077 unsigned long nr_scan_pages;
1080 * Calculations based on RSS as non-present and empty pages are skipped
1081 * by the PTE scanner and NUMA hinting faults should be trapped based
1084 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1085 rss = get_mm_rss(p->mm);
1087 rss = nr_scan_pages;
1089 rss = round_up(rss, nr_scan_pages);
1090 return rss / nr_scan_pages;
1093 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1094 #define MAX_SCAN_WINDOW 2560
1096 static unsigned int task_scan_min(struct task_struct *p)
1098 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1099 unsigned int scan, floor;
1100 unsigned int windows = 1;
1102 if (scan_size < MAX_SCAN_WINDOW)
1103 windows = MAX_SCAN_WINDOW / scan_size;
1104 floor = 1000 / windows;
1106 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1107 return max_t(unsigned int, floor, scan);
1110 static unsigned int task_scan_max(struct task_struct *p)
1112 unsigned int smin = task_scan_min(p);
1115 /* Watch for min being lower than max due to floor calculations */
1116 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1117 return max(smin, smax);
1120 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1122 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1123 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1126 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1128 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1129 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1135 spinlock_t lock; /* nr_tasks, tasks */
1140 struct rcu_head rcu;
1141 unsigned long total_faults;
1142 unsigned long max_faults_cpu;
1144 * Faults_cpu is used to decide whether memory should move
1145 * towards the CPU. As a consequence, these stats are weighted
1146 * more by CPU use than by memory faults.
1148 unsigned long *faults_cpu;
1149 unsigned long faults[0];
1152 /* Shared or private faults. */
1153 #define NR_NUMA_HINT_FAULT_TYPES 2
1155 /* Memory and CPU locality */
1156 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1158 /* Averaged statistics, and temporary buffers. */
1159 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1161 pid_t task_numa_group_id(struct task_struct *p)
1163 return p->numa_group ? p->numa_group->gid : 0;
1167 * The averaged statistics, shared & private, memory & cpu,
1168 * occupy the first half of the array. The second half of the
1169 * array is for current counters, which are averaged into the
1170 * first set by task_numa_placement.
1172 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1174 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1177 static inline unsigned long task_faults(struct task_struct *p, int nid)
1179 if (!p->numa_faults)
1182 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1183 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1186 static inline unsigned long group_faults(struct task_struct *p, int nid)
1191 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1192 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1195 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1197 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1198 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1202 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1203 * considered part of a numa group's pseudo-interleaving set. Migrations
1204 * between these nodes are slowed down, to allow things to settle down.
1206 #define ACTIVE_NODE_FRACTION 3
1208 static bool numa_is_active_node(int nid, struct numa_group *ng)
1210 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1213 /* Handle placement on systems where not all nodes are directly connected. */
1214 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1215 int maxdist, bool task)
1217 unsigned long score = 0;
1221 * All nodes are directly connected, and the same distance
1222 * from each other. No need for fancy placement algorithms.
1224 if (sched_numa_topology_type == NUMA_DIRECT)
1228 * This code is called for each node, introducing N^2 complexity,
1229 * which should be ok given the number of nodes rarely exceeds 8.
1231 for_each_online_node(node) {
1232 unsigned long faults;
1233 int dist = node_distance(nid, node);
1236 * The furthest away nodes in the system are not interesting
1237 * for placement; nid was already counted.
1239 if (dist == sched_max_numa_distance || node == nid)
1243 * On systems with a backplane NUMA topology, compare groups
1244 * of nodes, and move tasks towards the group with the most
1245 * memory accesses. When comparing two nodes at distance
1246 * "hoplimit", only nodes closer by than "hoplimit" are part
1247 * of each group. Skip other nodes.
1249 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1253 /* Add up the faults from nearby nodes. */
1255 faults = task_faults(p, node);
1257 faults = group_faults(p, node);
1260 * On systems with a glueless mesh NUMA topology, there are
1261 * no fixed "groups of nodes". Instead, nodes that are not
1262 * directly connected bounce traffic through intermediate
1263 * nodes; a numa_group can occupy any set of nodes.
1264 * The further away a node is, the less the faults count.
1265 * This seems to result in good task placement.
1267 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1268 faults *= (sched_max_numa_distance - dist);
1269 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1279 * These return the fraction of accesses done by a particular task, or
1280 * task group, on a particular numa node. The group weight is given a
1281 * larger multiplier, in order to group tasks together that are almost
1282 * evenly spread out between numa nodes.
1284 static inline unsigned long task_weight(struct task_struct *p, int nid,
1287 unsigned long faults, total_faults;
1289 if (!p->numa_faults)
1292 total_faults = p->total_numa_faults;
1297 faults = task_faults(p, nid);
1298 faults += score_nearby_nodes(p, nid, dist, true);
1300 return 1000 * faults / total_faults;
1303 static inline unsigned long group_weight(struct task_struct *p, int nid,
1306 unsigned long faults, total_faults;
1311 total_faults = p->numa_group->total_faults;
1316 faults = group_faults(p, nid);
1317 faults += score_nearby_nodes(p, nid, dist, false);
1319 return 1000 * faults / total_faults;
1322 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1323 int src_nid, int dst_cpu)
1325 struct numa_group *ng = p->numa_group;
1326 int dst_nid = cpu_to_node(dst_cpu);
1327 int last_cpupid, this_cpupid;
1329 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1332 * Multi-stage node selection is used in conjunction with a periodic
1333 * migration fault to build a temporal task<->page relation. By using
1334 * a two-stage filter we remove short/unlikely relations.
1336 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1337 * a task's usage of a particular page (n_p) per total usage of this
1338 * page (n_t) (in a given time-span) to a probability.
1340 * Our periodic faults will sample this probability and getting the
1341 * same result twice in a row, given these samples are fully
1342 * independent, is then given by P(n)^2, provided our sample period
1343 * is sufficiently short compared to the usage pattern.
1345 * This quadric squishes small probabilities, making it less likely we
1346 * act on an unlikely task<->page relation.
1348 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1349 if (!cpupid_pid_unset(last_cpupid) &&
1350 cpupid_to_nid(last_cpupid) != dst_nid)
1353 /* Always allow migrate on private faults */
1354 if (cpupid_match_pid(p, last_cpupid))
1357 /* A shared fault, but p->numa_group has not been set up yet. */
1362 * Destination node is much more heavily used than the source
1363 * node? Allow migration.
1365 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1366 ACTIVE_NODE_FRACTION)
1370 * Distribute memory according to CPU & memory use on each node,
1371 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1373 * faults_cpu(dst) 3 faults_cpu(src)
1374 * --------------- * - > ---------------
1375 * faults_mem(dst) 4 faults_mem(src)
1377 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1378 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1381 static unsigned long weighted_cpuload(const int cpu);
1382 static unsigned long source_load(int cpu, int type);
1383 static unsigned long target_load(int cpu, int type);
1384 static unsigned long capacity_of(int cpu);
1385 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1387 /* Cached statistics for all CPUs within a node */
1389 unsigned long nr_running;
1392 /* Total compute capacity of CPUs on a node */
1393 unsigned long compute_capacity;
1395 /* Approximate capacity in terms of runnable tasks on a node */
1396 unsigned long task_capacity;
1397 int has_free_capacity;
1401 * XXX borrowed from update_sg_lb_stats
1403 static void update_numa_stats(struct numa_stats *ns, int nid)
1405 int smt, cpu, cpus = 0;
1406 unsigned long capacity;
1408 memset(ns, 0, sizeof(*ns));
1409 for_each_cpu(cpu, cpumask_of_node(nid)) {
1410 struct rq *rq = cpu_rq(cpu);
1412 ns->nr_running += rq->nr_running;
1413 ns->load += weighted_cpuload(cpu);
1414 ns->compute_capacity += capacity_of(cpu);
1420 * If we raced with hotplug and there are no CPUs left in our mask
1421 * the @ns structure is NULL'ed and task_numa_compare() will
1422 * not find this node attractive.
1424 * We'll either bail at !has_free_capacity, or we'll detect a huge
1425 * imbalance and bail there.
1430 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1431 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1432 capacity = cpus / smt; /* cores */
1434 ns->task_capacity = min_t(unsigned, capacity,
1435 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1436 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1439 struct task_numa_env {
1440 struct task_struct *p;
1442 int src_cpu, src_nid;
1443 int dst_cpu, dst_nid;
1445 struct numa_stats src_stats, dst_stats;
1450 struct task_struct *best_task;
1455 static void task_numa_assign(struct task_numa_env *env,
1456 struct task_struct *p, long imp)
1459 put_task_struct(env->best_task);
1464 env->best_imp = imp;
1465 env->best_cpu = env->dst_cpu;
1468 static bool load_too_imbalanced(long src_load, long dst_load,
1469 struct task_numa_env *env)
1472 long orig_src_load, orig_dst_load;
1473 long src_capacity, dst_capacity;
1476 * The load is corrected for the CPU capacity available on each node.
1479 * ------------ vs ---------
1480 * src_capacity dst_capacity
1482 src_capacity = env->src_stats.compute_capacity;
1483 dst_capacity = env->dst_stats.compute_capacity;
1485 /* We care about the slope of the imbalance, not the direction. */
1486 if (dst_load < src_load)
1487 swap(dst_load, src_load);
1489 /* Is the difference below the threshold? */
1490 imb = dst_load * src_capacity * 100 -
1491 src_load * dst_capacity * env->imbalance_pct;
1496 * The imbalance is above the allowed threshold.
1497 * Compare it with the old imbalance.
1499 orig_src_load = env->src_stats.load;
1500 orig_dst_load = env->dst_stats.load;
1502 if (orig_dst_load < orig_src_load)
1503 swap(orig_dst_load, orig_src_load);
1505 old_imb = orig_dst_load * src_capacity * 100 -
1506 orig_src_load * dst_capacity * env->imbalance_pct;
1508 /* Would this change make things worse? */
1509 return (imb > old_imb);
1513 * This checks if the overall compute and NUMA accesses of the system would
1514 * be improved if the source tasks was migrated to the target dst_cpu taking
1515 * into account that it might be best if task running on the dst_cpu should
1516 * be exchanged with the source task
1518 static void task_numa_compare(struct task_numa_env *env,
1519 long taskimp, long groupimp)
1521 struct rq *src_rq = cpu_rq(env->src_cpu);
1522 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1523 struct task_struct *cur;
1524 long src_load, dst_load;
1526 long imp = env->p->numa_group ? groupimp : taskimp;
1528 int dist = env->dist;
1531 cur = task_rcu_dereference(&dst_rq->curr);
1532 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1536 * Because we have preemption enabled we can get migrated around and
1537 * end try selecting ourselves (current == env->p) as a swap candidate.
1543 * "imp" is the fault differential for the source task between the
1544 * source and destination node. Calculate the total differential for
1545 * the source task and potential destination task. The more negative
1546 * the value is, the more rmeote accesses that would be expected to
1547 * be incurred if the tasks were swapped.
1550 /* Skip this swap candidate if cannot move to the source cpu */
1551 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1555 * If dst and source tasks are in the same NUMA group, or not
1556 * in any group then look only at task weights.
1558 if (cur->numa_group == env->p->numa_group) {
1559 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1560 task_weight(cur, env->dst_nid, dist);
1562 * Add some hysteresis to prevent swapping the
1563 * tasks within a group over tiny differences.
1565 if (cur->numa_group)
1569 * Compare the group weights. If a task is all by
1570 * itself (not part of a group), use the task weight
1573 if (cur->numa_group)
1574 imp += group_weight(cur, env->src_nid, dist) -
1575 group_weight(cur, env->dst_nid, dist);
1577 imp += task_weight(cur, env->src_nid, dist) -
1578 task_weight(cur, env->dst_nid, dist);
1582 if (imp <= env->best_imp && moveimp <= env->best_imp)
1586 /* Is there capacity at our destination? */
1587 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1588 !env->dst_stats.has_free_capacity)
1594 /* Balance doesn't matter much if we're running a task per cpu */
1595 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1596 dst_rq->nr_running == 1)
1600 * In the overloaded case, try and keep the load balanced.
1603 load = task_h_load(env->p);
1604 dst_load = env->dst_stats.load + load;
1605 src_load = env->src_stats.load - load;
1607 if (moveimp > imp && moveimp > env->best_imp) {
1609 * If the improvement from just moving env->p direction is
1610 * better than swapping tasks around, check if a move is
1611 * possible. Store a slightly smaller score than moveimp,
1612 * so an actually idle CPU will win.
1614 if (!load_too_imbalanced(src_load, dst_load, env)) {
1621 if (imp <= env->best_imp)
1625 load = task_h_load(cur);
1630 if (load_too_imbalanced(src_load, dst_load, env))
1634 * One idle CPU per node is evaluated for a task numa move.
1635 * Call select_idle_sibling to maybe find a better one.
1639 * select_idle_siblings() uses an per-cpu cpumask that
1640 * can be used from IRQ context.
1642 local_irq_disable();
1643 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1649 task_numa_assign(env, cur, imp);
1654 static void task_numa_find_cpu(struct task_numa_env *env,
1655 long taskimp, long groupimp)
1659 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1660 /* Skip this CPU if the source task cannot migrate */
1661 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1665 task_numa_compare(env, taskimp, groupimp);
1669 /* Only move tasks to a NUMA node less busy than the current node. */
1670 static bool numa_has_capacity(struct task_numa_env *env)
1672 struct numa_stats *src = &env->src_stats;
1673 struct numa_stats *dst = &env->dst_stats;
1675 if (src->has_free_capacity && !dst->has_free_capacity)
1679 * Only consider a task move if the source has a higher load
1680 * than the destination, corrected for CPU capacity on each node.
1682 * src->load dst->load
1683 * --------------------- vs ---------------------
1684 * src->compute_capacity dst->compute_capacity
1686 if (src->load * dst->compute_capacity * env->imbalance_pct >
1688 dst->load * src->compute_capacity * 100)
1694 static int task_numa_migrate(struct task_struct *p)
1696 struct task_numa_env env = {
1699 .src_cpu = task_cpu(p),
1700 .src_nid = task_node(p),
1702 .imbalance_pct = 112,
1708 struct sched_domain *sd;
1709 unsigned long taskweight, groupweight;
1711 long taskimp, groupimp;
1714 * Pick the lowest SD_NUMA domain, as that would have the smallest
1715 * imbalance and would be the first to start moving tasks about.
1717 * And we want to avoid any moving of tasks about, as that would create
1718 * random movement of tasks -- counter the numa conditions we're trying
1722 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1724 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1728 * Cpusets can break the scheduler domain tree into smaller
1729 * balance domains, some of which do not cross NUMA boundaries.
1730 * Tasks that are "trapped" in such domains cannot be migrated
1731 * elsewhere, so there is no point in (re)trying.
1733 if (unlikely(!sd)) {
1734 p->numa_preferred_nid = task_node(p);
1738 env.dst_nid = p->numa_preferred_nid;
1739 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1740 taskweight = task_weight(p, env.src_nid, dist);
1741 groupweight = group_weight(p, env.src_nid, dist);
1742 update_numa_stats(&env.src_stats, env.src_nid);
1743 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1744 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1745 update_numa_stats(&env.dst_stats, env.dst_nid);
1747 /* Try to find a spot on the preferred nid. */
1748 if (numa_has_capacity(&env))
1749 task_numa_find_cpu(&env, taskimp, groupimp);
1752 * Look at other nodes in these cases:
1753 * - there is no space available on the preferred_nid
1754 * - the task is part of a numa_group that is interleaved across
1755 * multiple NUMA nodes; in order to better consolidate the group,
1756 * we need to check other locations.
1758 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1759 for_each_online_node(nid) {
1760 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1763 dist = node_distance(env.src_nid, env.dst_nid);
1764 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1766 taskweight = task_weight(p, env.src_nid, dist);
1767 groupweight = group_weight(p, env.src_nid, dist);
1770 /* Only consider nodes where both task and groups benefit */
1771 taskimp = task_weight(p, nid, dist) - taskweight;
1772 groupimp = group_weight(p, nid, dist) - groupweight;
1773 if (taskimp < 0 && groupimp < 0)
1778 update_numa_stats(&env.dst_stats, env.dst_nid);
1779 if (numa_has_capacity(&env))
1780 task_numa_find_cpu(&env, taskimp, groupimp);
1785 * If the task is part of a workload that spans multiple NUMA nodes,
1786 * and is migrating into one of the workload's active nodes, remember
1787 * this node as the task's preferred numa node, so the workload can
1789 * A task that migrated to a second choice node will be better off
1790 * trying for a better one later. Do not set the preferred node here.
1792 if (p->numa_group) {
1793 struct numa_group *ng = p->numa_group;
1795 if (env.best_cpu == -1)
1800 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1801 sched_setnuma(p, env.dst_nid);
1804 /* No better CPU than the current one was found. */
1805 if (env.best_cpu == -1)
1809 * Reset the scan period if the task is being rescheduled on an
1810 * alternative node to recheck if the tasks is now properly placed.
1812 p->numa_scan_period = task_scan_min(p);
1814 if (env.best_task == NULL) {
1815 ret = migrate_task_to(p, env.best_cpu);
1817 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1821 ret = migrate_swap(p, env.best_task);
1823 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1824 put_task_struct(env.best_task);
1828 /* Attempt to migrate a task to a CPU on the preferred node. */
1829 static void numa_migrate_preferred(struct task_struct *p)
1831 unsigned long interval = HZ;
1833 /* This task has no NUMA fault statistics yet */
1834 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1837 /* Periodically retry migrating the task to the preferred node */
1838 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1839 p->numa_migrate_retry = jiffies + interval;
1841 /* Success if task is already running on preferred CPU */
1842 if (task_node(p) == p->numa_preferred_nid)
1845 /* Otherwise, try migrate to a CPU on the preferred node */
1846 task_numa_migrate(p);
1850 * Find out how many nodes on the workload is actively running on. Do this by
1851 * tracking the nodes from which NUMA hinting faults are triggered. This can
1852 * be different from the set of nodes where the workload's memory is currently
1855 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1857 unsigned long faults, max_faults = 0;
1858 int nid, active_nodes = 0;
1860 for_each_online_node(nid) {
1861 faults = group_faults_cpu(numa_group, nid);
1862 if (faults > max_faults)
1863 max_faults = faults;
1866 for_each_online_node(nid) {
1867 faults = group_faults_cpu(numa_group, nid);
1868 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1872 numa_group->max_faults_cpu = max_faults;
1873 numa_group->active_nodes = active_nodes;
1877 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1878 * increments. The more local the fault statistics are, the higher the scan
1879 * period will be for the next scan window. If local/(local+remote) ratio is
1880 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1881 * the scan period will decrease. Aim for 70% local accesses.
1883 #define NUMA_PERIOD_SLOTS 10
1884 #define NUMA_PERIOD_THRESHOLD 7
1887 * Increase the scan period (slow down scanning) if the majority of
1888 * our memory is already on our local node, or if the majority of
1889 * the page accesses are shared with other processes.
1890 * Otherwise, decrease the scan period.
1892 static void update_task_scan_period(struct task_struct *p,
1893 unsigned long shared, unsigned long private)
1895 unsigned int period_slot;
1899 unsigned long remote = p->numa_faults_locality[0];
1900 unsigned long local = p->numa_faults_locality[1];
1903 * If there were no record hinting faults then either the task is
1904 * completely idle or all activity is areas that are not of interest
1905 * to automatic numa balancing. Related to that, if there were failed
1906 * migration then it implies we are migrating too quickly or the local
1907 * node is overloaded. In either case, scan slower
1909 if (local + shared == 0 || p->numa_faults_locality[2]) {
1910 p->numa_scan_period = min(p->numa_scan_period_max,
1911 p->numa_scan_period << 1);
1913 p->mm->numa_next_scan = jiffies +
1914 msecs_to_jiffies(p->numa_scan_period);
1920 * Prepare to scale scan period relative to the current period.
1921 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1922 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1923 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1925 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1926 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1927 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1928 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1931 diff = slot * period_slot;
1933 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1936 * Scale scan rate increases based on sharing. There is an
1937 * inverse relationship between the degree of sharing and
1938 * the adjustment made to the scanning period. Broadly
1939 * speaking the intent is that there is little point
1940 * scanning faster if shared accesses dominate as it may
1941 * simply bounce migrations uselessly
1943 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1944 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1947 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1948 task_scan_min(p), task_scan_max(p));
1949 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1953 * Get the fraction of time the task has been running since the last
1954 * NUMA placement cycle. The scheduler keeps similar statistics, but
1955 * decays those on a 32ms period, which is orders of magnitude off
1956 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1957 * stats only if the task is so new there are no NUMA statistics yet.
1959 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1961 u64 runtime, delta, now;
1962 /* Use the start of this time slice to avoid calculations. */
1963 now = p->se.exec_start;
1964 runtime = p->se.sum_exec_runtime;
1966 if (p->last_task_numa_placement) {
1967 delta = runtime - p->last_sum_exec_runtime;
1968 *period = now - p->last_task_numa_placement;
1970 delta = p->se.avg.load_sum / p->se.load.weight;
1971 *period = LOAD_AVG_MAX;
1974 p->last_sum_exec_runtime = runtime;
1975 p->last_task_numa_placement = now;
1981 * Determine the preferred nid for a task in a numa_group. This needs to
1982 * be done in a way that produces consistent results with group_weight,
1983 * otherwise workloads might not converge.
1985 static int preferred_group_nid(struct task_struct *p, int nid)
1990 /* Direct connections between all NUMA nodes. */
1991 if (sched_numa_topology_type == NUMA_DIRECT)
1995 * On a system with glueless mesh NUMA topology, group_weight
1996 * scores nodes according to the number of NUMA hinting faults on
1997 * both the node itself, and on nearby nodes.
1999 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2000 unsigned long score, max_score = 0;
2001 int node, max_node = nid;
2003 dist = sched_max_numa_distance;
2005 for_each_online_node(node) {
2006 score = group_weight(p, node, dist);
2007 if (score > max_score) {
2016 * Finding the preferred nid in a system with NUMA backplane
2017 * interconnect topology is more involved. The goal is to locate
2018 * tasks from numa_groups near each other in the system, and
2019 * untangle workloads from different sides of the system. This requires
2020 * searching down the hierarchy of node groups, recursively searching
2021 * inside the highest scoring group of nodes. The nodemask tricks
2022 * keep the complexity of the search down.
2024 nodes = node_online_map;
2025 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2026 unsigned long max_faults = 0;
2027 nodemask_t max_group = NODE_MASK_NONE;
2030 /* Are there nodes at this distance from each other? */
2031 if (!find_numa_distance(dist))
2034 for_each_node_mask(a, nodes) {
2035 unsigned long faults = 0;
2036 nodemask_t this_group;
2037 nodes_clear(this_group);
2039 /* Sum group's NUMA faults; includes a==b case. */
2040 for_each_node_mask(b, nodes) {
2041 if (node_distance(a, b) < dist) {
2042 faults += group_faults(p, b);
2043 node_set(b, this_group);
2044 node_clear(b, nodes);
2048 /* Remember the top group. */
2049 if (faults > max_faults) {
2050 max_faults = faults;
2051 max_group = this_group;
2053 * subtle: at the smallest distance there is
2054 * just one node left in each "group", the
2055 * winner is the preferred nid.
2060 /* Next round, evaluate the nodes within max_group. */
2068 static void task_numa_placement(struct task_struct *p)
2070 int seq, nid, max_nid = -1, max_group_nid = -1;
2071 unsigned long max_faults = 0, max_group_faults = 0;
2072 unsigned long fault_types[2] = { 0, 0 };
2073 unsigned long total_faults;
2074 u64 runtime, period;
2075 spinlock_t *group_lock = NULL;
2078 * The p->mm->numa_scan_seq field gets updated without
2079 * exclusive access. Use READ_ONCE() here to ensure
2080 * that the field is read in a single access:
2082 seq = READ_ONCE(p->mm->numa_scan_seq);
2083 if (p->numa_scan_seq == seq)
2085 p->numa_scan_seq = seq;
2086 p->numa_scan_period_max = task_scan_max(p);
2088 total_faults = p->numa_faults_locality[0] +
2089 p->numa_faults_locality[1];
2090 runtime = numa_get_avg_runtime(p, &period);
2092 /* If the task is part of a group prevent parallel updates to group stats */
2093 if (p->numa_group) {
2094 group_lock = &p->numa_group->lock;
2095 spin_lock_irq(group_lock);
2098 /* Find the node with the highest number of faults */
2099 for_each_online_node(nid) {
2100 /* Keep track of the offsets in numa_faults array */
2101 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2102 unsigned long faults = 0, group_faults = 0;
2105 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2106 long diff, f_diff, f_weight;
2108 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2109 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2110 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2111 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2113 /* Decay existing window, copy faults since last scan */
2114 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2115 fault_types[priv] += p->numa_faults[membuf_idx];
2116 p->numa_faults[membuf_idx] = 0;
2119 * Normalize the faults_from, so all tasks in a group
2120 * count according to CPU use, instead of by the raw
2121 * number of faults. Tasks with little runtime have
2122 * little over-all impact on throughput, and thus their
2123 * faults are less important.
2125 f_weight = div64_u64(runtime << 16, period + 1);
2126 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2128 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2129 p->numa_faults[cpubuf_idx] = 0;
2131 p->numa_faults[mem_idx] += diff;
2132 p->numa_faults[cpu_idx] += f_diff;
2133 faults += p->numa_faults[mem_idx];
2134 p->total_numa_faults += diff;
2135 if (p->numa_group) {
2137 * safe because we can only change our own group
2139 * mem_idx represents the offset for a given
2140 * nid and priv in a specific region because it
2141 * is at the beginning of the numa_faults array.
2143 p->numa_group->faults[mem_idx] += diff;
2144 p->numa_group->faults_cpu[mem_idx] += f_diff;
2145 p->numa_group->total_faults += diff;
2146 group_faults += p->numa_group->faults[mem_idx];
2150 if (faults > max_faults) {
2151 max_faults = faults;
2155 if (group_faults > max_group_faults) {
2156 max_group_faults = group_faults;
2157 max_group_nid = nid;
2161 update_task_scan_period(p, fault_types[0], fault_types[1]);
2163 if (p->numa_group) {
2164 numa_group_count_active_nodes(p->numa_group);
2165 spin_unlock_irq(group_lock);
2166 max_nid = preferred_group_nid(p, max_group_nid);
2170 /* Set the new preferred node */
2171 if (max_nid != p->numa_preferred_nid)
2172 sched_setnuma(p, max_nid);
2174 if (task_node(p) != p->numa_preferred_nid)
2175 numa_migrate_preferred(p);
2179 static inline int get_numa_group(struct numa_group *grp)
2181 return atomic_inc_not_zero(&grp->refcount);
2184 static inline void put_numa_group(struct numa_group *grp)
2186 if (atomic_dec_and_test(&grp->refcount))
2187 kfree_rcu(grp, rcu);
2190 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2193 struct numa_group *grp, *my_grp;
2194 struct task_struct *tsk;
2196 int cpu = cpupid_to_cpu(cpupid);
2199 if (unlikely(!p->numa_group)) {
2200 unsigned int size = sizeof(struct numa_group) +
2201 4*nr_node_ids*sizeof(unsigned long);
2203 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2207 atomic_set(&grp->refcount, 1);
2208 grp->active_nodes = 1;
2209 grp->max_faults_cpu = 0;
2210 spin_lock_init(&grp->lock);
2212 /* Second half of the array tracks nids where faults happen */
2213 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2216 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2217 grp->faults[i] = p->numa_faults[i];
2219 grp->total_faults = p->total_numa_faults;
2222 rcu_assign_pointer(p->numa_group, grp);
2226 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2228 if (!cpupid_match_pid(tsk, cpupid))
2231 grp = rcu_dereference(tsk->numa_group);
2235 my_grp = p->numa_group;
2240 * Only join the other group if its bigger; if we're the bigger group,
2241 * the other task will join us.
2243 if (my_grp->nr_tasks > grp->nr_tasks)
2247 * Tie-break on the grp address.
2249 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2252 /* Always join threads in the same process. */
2253 if (tsk->mm == current->mm)
2256 /* Simple filter to avoid false positives due to PID collisions */
2257 if (flags & TNF_SHARED)
2260 /* Update priv based on whether false sharing was detected */
2263 if (join && !get_numa_group(grp))
2271 BUG_ON(irqs_disabled());
2272 double_lock_irq(&my_grp->lock, &grp->lock);
2274 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2275 my_grp->faults[i] -= p->numa_faults[i];
2276 grp->faults[i] += p->numa_faults[i];
2278 my_grp->total_faults -= p->total_numa_faults;
2279 grp->total_faults += p->total_numa_faults;
2284 spin_unlock(&my_grp->lock);
2285 spin_unlock_irq(&grp->lock);
2287 rcu_assign_pointer(p->numa_group, grp);
2289 put_numa_group(my_grp);
2297 void task_numa_free(struct task_struct *p)
2299 struct numa_group *grp = p->numa_group;
2300 void *numa_faults = p->numa_faults;
2301 unsigned long flags;
2305 spin_lock_irqsave(&grp->lock, flags);
2306 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2307 grp->faults[i] -= p->numa_faults[i];
2308 grp->total_faults -= p->total_numa_faults;
2311 spin_unlock_irqrestore(&grp->lock, flags);
2312 RCU_INIT_POINTER(p->numa_group, NULL);
2313 put_numa_group(grp);
2316 p->numa_faults = NULL;
2321 * Got a PROT_NONE fault for a page on @node.
2323 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2325 struct task_struct *p = current;
2326 bool migrated = flags & TNF_MIGRATED;
2327 int cpu_node = task_node(current);
2328 int local = !!(flags & TNF_FAULT_LOCAL);
2329 struct numa_group *ng;
2332 if (!static_branch_likely(&sched_numa_balancing))
2335 /* for example, ksmd faulting in a user's mm */
2339 /* Allocate buffer to track faults on a per-node basis */
2340 if (unlikely(!p->numa_faults)) {
2341 int size = sizeof(*p->numa_faults) *
2342 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2344 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2345 if (!p->numa_faults)
2348 p->total_numa_faults = 0;
2349 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2353 * First accesses are treated as private, otherwise consider accesses
2354 * to be private if the accessing pid has not changed
2356 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2359 priv = cpupid_match_pid(p, last_cpupid);
2360 if (!priv && !(flags & TNF_NO_GROUP))
2361 task_numa_group(p, last_cpupid, flags, &priv);
2365 * If a workload spans multiple NUMA nodes, a shared fault that
2366 * occurs wholly within the set of nodes that the workload is
2367 * actively using should be counted as local. This allows the
2368 * scan rate to slow down when a workload has settled down.
2371 if (!priv && !local && ng && ng->active_nodes > 1 &&
2372 numa_is_active_node(cpu_node, ng) &&
2373 numa_is_active_node(mem_node, ng))
2376 task_numa_placement(p);
2379 * Retry task to preferred node migration periodically, in case it
2380 * case it previously failed, or the scheduler moved us.
2382 if (time_after(jiffies, p->numa_migrate_retry))
2383 numa_migrate_preferred(p);
2386 p->numa_pages_migrated += pages;
2387 if (flags & TNF_MIGRATE_FAIL)
2388 p->numa_faults_locality[2] += pages;
2390 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2391 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2392 p->numa_faults_locality[local] += pages;
2395 static void reset_ptenuma_scan(struct task_struct *p)
2398 * We only did a read acquisition of the mmap sem, so
2399 * p->mm->numa_scan_seq is written to without exclusive access
2400 * and the update is not guaranteed to be atomic. That's not
2401 * much of an issue though, since this is just used for
2402 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2403 * expensive, to avoid any form of compiler optimizations:
2405 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2406 p->mm->numa_scan_offset = 0;
2410 * The expensive part of numa migration is done from task_work context.
2411 * Triggered from task_tick_numa().
2413 void task_numa_work(struct callback_head *work)
2415 unsigned long migrate, next_scan, now = jiffies;
2416 struct task_struct *p = current;
2417 struct mm_struct *mm = p->mm;
2418 u64 runtime = p->se.sum_exec_runtime;
2419 struct vm_area_struct *vma;
2420 unsigned long start, end;
2421 unsigned long nr_pte_updates = 0;
2422 long pages, virtpages;
2424 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2426 work->next = work; /* protect against double add */
2428 * Who cares about NUMA placement when they're dying.
2430 * NOTE: make sure not to dereference p->mm before this check,
2431 * exit_task_work() happens _after_ exit_mm() so we could be called
2432 * without p->mm even though we still had it when we enqueued this
2435 if (p->flags & PF_EXITING)
2438 if (!mm->numa_next_scan) {
2439 mm->numa_next_scan = now +
2440 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2444 * Enforce maximal scan/migration frequency..
2446 migrate = mm->numa_next_scan;
2447 if (time_before(now, migrate))
2450 if (p->numa_scan_period == 0) {
2451 p->numa_scan_period_max = task_scan_max(p);
2452 p->numa_scan_period = task_scan_min(p);
2455 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2456 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2460 * Delay this task enough that another task of this mm will likely win
2461 * the next time around.
2463 p->node_stamp += 2 * TICK_NSEC;
2465 start = mm->numa_scan_offset;
2466 pages = sysctl_numa_balancing_scan_size;
2467 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2468 virtpages = pages * 8; /* Scan up to this much virtual space */
2473 if (!down_read_trylock(&mm->mmap_sem))
2475 vma = find_vma(mm, start);
2477 reset_ptenuma_scan(p);
2481 for (; vma; vma = vma->vm_next) {
2482 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2483 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2488 * Shared library pages mapped by multiple processes are not
2489 * migrated as it is expected they are cache replicated. Avoid
2490 * hinting faults in read-only file-backed mappings or the vdso
2491 * as migrating the pages will be of marginal benefit.
2494 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2498 * Skip inaccessible VMAs to avoid any confusion between
2499 * PROT_NONE and NUMA hinting ptes
2501 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2505 start = max(start, vma->vm_start);
2506 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2507 end = min(end, vma->vm_end);
2508 nr_pte_updates = change_prot_numa(vma, start, end);
2511 * Try to scan sysctl_numa_balancing_size worth of
2512 * hpages that have at least one present PTE that
2513 * is not already pte-numa. If the VMA contains
2514 * areas that are unused or already full of prot_numa
2515 * PTEs, scan up to virtpages, to skip through those
2519 pages -= (end - start) >> PAGE_SHIFT;
2520 virtpages -= (end - start) >> PAGE_SHIFT;
2523 if (pages <= 0 || virtpages <= 0)
2527 } while (end != vma->vm_end);
2532 * It is possible to reach the end of the VMA list but the last few
2533 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2534 * would find the !migratable VMA on the next scan but not reset the
2535 * scanner to the start so check it now.
2538 mm->numa_scan_offset = start;
2540 reset_ptenuma_scan(p);
2541 up_read(&mm->mmap_sem);
2544 * Make sure tasks use at least 32x as much time to run other code
2545 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2546 * Usually update_task_scan_period slows down scanning enough; on an
2547 * overloaded system we need to limit overhead on a per task basis.
2549 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2550 u64 diff = p->se.sum_exec_runtime - runtime;
2551 p->node_stamp += 32 * diff;
2556 * Drive the periodic memory faults..
2558 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2560 struct callback_head *work = &curr->numa_work;
2564 * We don't care about NUMA placement if we don't have memory.
2566 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2570 * Using runtime rather than walltime has the dual advantage that
2571 * we (mostly) drive the selection from busy threads and that the
2572 * task needs to have done some actual work before we bother with
2575 now = curr->se.sum_exec_runtime;
2576 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2578 if (now > curr->node_stamp + period) {
2579 if (!curr->node_stamp)
2580 curr->numa_scan_period = task_scan_min(curr);
2581 curr->node_stamp += period;
2583 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2584 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2585 task_work_add(curr, work, true);
2591 * Can a task be moved from prev_cpu to this_cpu without causing a load
2592 * imbalance that would trigger the load balancer?
2594 static inline bool numa_wake_affine(struct sched_domain *sd,
2595 struct task_struct *p, int this_cpu,
2596 int prev_cpu, int sync)
2598 struct numa_stats prev_load, this_load;
2599 s64 this_eff_load, prev_eff_load;
2601 update_numa_stats(&prev_load, cpu_to_node(prev_cpu));
2602 update_numa_stats(&this_load, cpu_to_node(this_cpu));
2605 * If sync wakeup then subtract the (maximum possible)
2606 * effect of the currently running task from the load
2607 * of the current CPU:
2610 unsigned long current_load = task_h_load(current);
2612 if (this_load.load > current_load)
2613 this_load.load -= current_load;
2619 * In low-load situations, where this_cpu's node is idle due to the
2620 * sync cause above having dropped this_load.load to 0, move the task.
2621 * Moving to an idle socket will not create a bad imbalance.
2623 * Otherwise check if the nodes are near enough in load to allow this
2624 * task to be woken on this_cpu's node.
2626 if (this_load.load > 0) {
2627 unsigned long task_load = task_h_load(p);
2629 this_eff_load = 100;
2630 this_eff_load *= prev_load.compute_capacity;
2632 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2633 prev_eff_load *= this_load.compute_capacity;
2635 this_eff_load *= this_load.load + task_load;
2636 prev_eff_load *= prev_load.load - task_load;
2638 return this_eff_load <= prev_eff_load;
2644 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2648 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2652 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2656 static inline bool numa_wake_affine(struct sched_domain *sd,
2657 struct task_struct *p, int this_cpu,
2658 int prev_cpu, int sync)
2662 #endif /* CONFIG_NUMA_BALANCING */
2665 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2667 update_load_add(&cfs_rq->load, se->load.weight);
2668 if (!parent_entity(se))
2669 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2671 if (entity_is_task(se)) {
2672 struct rq *rq = rq_of(cfs_rq);
2674 account_numa_enqueue(rq, task_of(se));
2675 list_add(&se->group_node, &rq->cfs_tasks);
2678 cfs_rq->nr_running++;
2682 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2684 update_load_sub(&cfs_rq->load, se->load.weight);
2685 if (!parent_entity(se))
2686 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2688 if (entity_is_task(se)) {
2689 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2690 list_del_init(&se->group_node);
2693 cfs_rq->nr_running--;
2696 #ifdef CONFIG_FAIR_GROUP_SCHED
2698 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2700 long tg_weight, load, shares;
2703 * This really should be: cfs_rq->avg.load_avg, but instead we use
2704 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2705 * the shares for small weight interactive tasks.
2707 load = scale_load_down(cfs_rq->load.weight);
2709 tg_weight = atomic_long_read(&tg->load_avg);
2711 /* Ensure tg_weight >= load */
2712 tg_weight -= cfs_rq->tg_load_avg_contrib;
2715 shares = (tg->shares * load);
2717 shares /= tg_weight;
2720 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2721 * of a group with small tg->shares value. It is a floor value which is
2722 * assigned as a minimum load.weight to the sched_entity representing
2723 * the group on a CPU.
2725 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2726 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2727 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2728 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2731 if (shares < MIN_SHARES)
2732 shares = MIN_SHARES;
2733 if (shares > tg->shares)
2734 shares = tg->shares;
2738 # else /* CONFIG_SMP */
2739 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2743 # endif /* CONFIG_SMP */
2745 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2746 unsigned long weight)
2749 /* commit outstanding execution time */
2750 if (cfs_rq->curr == se)
2751 update_curr(cfs_rq);
2752 account_entity_dequeue(cfs_rq, se);
2755 update_load_set(&se->load, weight);
2758 account_entity_enqueue(cfs_rq, se);
2761 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2763 static void update_cfs_shares(struct sched_entity *se)
2765 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2766 struct task_group *tg;
2772 if (throttled_hierarchy(cfs_rq))
2778 if (likely(se->load.weight == tg->shares))
2781 shares = calc_cfs_shares(cfs_rq, tg);
2783 reweight_entity(cfs_rq_of(se), se, shares);
2786 #else /* CONFIG_FAIR_GROUP_SCHED */
2787 static inline void update_cfs_shares(struct sched_entity *se)
2790 #endif /* CONFIG_FAIR_GROUP_SCHED */
2795 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2797 static u64 decay_load(u64 val, u64 n)
2799 unsigned int local_n;
2801 if (unlikely(n > LOAD_AVG_PERIOD * 63))
2804 /* after bounds checking we can collapse to 32-bit */
2808 * As y^PERIOD = 1/2, we can combine
2809 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2810 * With a look-up table which covers y^n (n<PERIOD)
2812 * To achieve constant time decay_load.
2814 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2815 val >>= local_n / LOAD_AVG_PERIOD;
2816 local_n %= LOAD_AVG_PERIOD;
2819 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2823 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
2825 u32 c1, c2, c3 = d3; /* y^0 == 1 */
2830 c1 = decay_load((u64)d1, periods);
2834 * c2 = 1024 \Sum y^n
2838 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
2841 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
2843 return c1 + c2 + c3;
2846 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2849 * Accumulate the three separate parts of the sum; d1 the remainder
2850 * of the last (incomplete) period, d2 the span of full periods and d3
2851 * the remainder of the (incomplete) current period.
2856 * |<->|<----------------->|<--->|
2857 * ... |---x---|------| ... |------|-----x (now)
2860 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2863 * = u y^p + (Step 1)
2866 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2869 static __always_inline u32
2870 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2871 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2873 unsigned long scale_freq, scale_cpu;
2874 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
2877 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2878 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2880 delta += sa->period_contrib;
2881 periods = delta / 1024; /* A period is 1024us (~1ms) */
2884 * Step 1: decay old *_sum if we crossed period boundaries.
2887 sa->load_sum = decay_load(sa->load_sum, periods);
2889 cfs_rq->runnable_load_sum =
2890 decay_load(cfs_rq->runnable_load_sum, periods);
2892 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
2898 contrib = __accumulate_pelt_segments(periods,
2899 1024 - sa->period_contrib, delta);
2901 sa->period_contrib = delta;
2903 contrib = cap_scale(contrib, scale_freq);
2905 sa->load_sum += weight * contrib;
2907 cfs_rq->runnable_load_sum += weight * contrib;
2910 sa->util_sum += contrib * scale_cpu;
2916 * We can represent the historical contribution to runnable average as the
2917 * coefficients of a geometric series. To do this we sub-divide our runnable
2918 * history into segments of approximately 1ms (1024us); label the segment that
2919 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2921 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2923 * (now) (~1ms ago) (~2ms ago)
2925 * Let u_i denote the fraction of p_i that the entity was runnable.
2927 * We then designate the fractions u_i as our co-efficients, yielding the
2928 * following representation of historical load:
2929 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2931 * We choose y based on the with of a reasonably scheduling period, fixing:
2934 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2935 * approximately half as much as the contribution to load within the last ms
2938 * When a period "rolls over" and we have new u_0`, multiplying the previous
2939 * sum again by y is sufficient to update:
2940 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2941 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2943 static __always_inline int
2944 ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2945 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2949 delta = now - sa->last_update_time;
2951 * This should only happen when time goes backwards, which it
2952 * unfortunately does during sched clock init when we swap over to TSC.
2954 if ((s64)delta < 0) {
2955 sa->last_update_time = now;
2960 * Use 1024ns as the unit of measurement since it's a reasonable
2961 * approximation of 1us and fast to compute.
2967 sa->last_update_time += delta << 10;
2970 * Now we know we crossed measurement unit boundaries. The *_avg
2971 * accrues by two steps:
2973 * Step 1: accumulate *_sum since last_update_time. If we haven't
2974 * crossed period boundaries, finish.
2976 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
2980 * Step 2: update *_avg.
2982 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2984 cfs_rq->runnable_load_avg =
2985 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
2987 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
2993 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
2995 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
2999 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3001 return ___update_load_avg(now, cpu, &se->avg,
3002 se->on_rq * scale_load_down(se->load.weight),
3003 cfs_rq->curr == se, NULL);
3007 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3009 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3010 scale_load_down(cfs_rq->load.weight),
3011 cfs_rq->curr != NULL, cfs_rq);
3015 * Signed add and clamp on underflow.
3017 * Explicitly do a load-store to ensure the intermediate value never hits
3018 * memory. This allows lockless observations without ever seeing the negative
3021 #define add_positive(_ptr, _val) do { \
3022 typeof(_ptr) ptr = (_ptr); \
3023 typeof(_val) val = (_val); \
3024 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3028 if (val < 0 && res > var) \
3031 WRITE_ONCE(*ptr, res); \
3034 #ifdef CONFIG_FAIR_GROUP_SCHED
3036 * update_tg_load_avg - update the tg's load avg
3037 * @cfs_rq: the cfs_rq whose avg changed
3038 * @force: update regardless of how small the difference
3040 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3041 * However, because tg->load_avg is a global value there are performance
3044 * In order to avoid having to look at the other cfs_rq's, we use a
3045 * differential update where we store the last value we propagated. This in
3046 * turn allows skipping updates if the differential is 'small'.
3048 * Updating tg's load_avg is necessary before update_cfs_share() (which is
3049 * done) and effective_load() (which is not done because it is too costly).
3051 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3053 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3056 * No need to update load_avg for root_task_group as it is not used.
3058 if (cfs_rq->tg == &root_task_group)
3061 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3062 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3063 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3068 * Called within set_task_rq() right before setting a task's cpu. The
3069 * caller only guarantees p->pi_lock is held; no other assumptions,
3070 * including the state of rq->lock, should be made.
3072 void set_task_rq_fair(struct sched_entity *se,
3073 struct cfs_rq *prev, struct cfs_rq *next)
3075 u64 p_last_update_time;
3076 u64 n_last_update_time;
3078 if (!sched_feat(ATTACH_AGE_LOAD))
3082 * We are supposed to update the task to "current" time, then its up to
3083 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3084 * getting what current time is, so simply throw away the out-of-date
3085 * time. This will result in the wakee task is less decayed, but giving
3086 * the wakee more load sounds not bad.
3088 if (!(se->avg.last_update_time && prev))
3091 #ifndef CONFIG_64BIT
3093 u64 p_last_update_time_copy;
3094 u64 n_last_update_time_copy;
3097 p_last_update_time_copy = prev->load_last_update_time_copy;
3098 n_last_update_time_copy = next->load_last_update_time_copy;
3102 p_last_update_time = prev->avg.last_update_time;
3103 n_last_update_time = next->avg.last_update_time;
3105 } while (p_last_update_time != p_last_update_time_copy ||
3106 n_last_update_time != n_last_update_time_copy);
3109 p_last_update_time = prev->avg.last_update_time;
3110 n_last_update_time = next->avg.last_update_time;
3112 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3113 se->avg.last_update_time = n_last_update_time;
3116 /* Take into account change of utilization of a child task group */
3118 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3120 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3121 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3123 /* Nothing to update */
3127 /* Set new sched_entity's utilization */
3128 se->avg.util_avg = gcfs_rq->avg.util_avg;
3129 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3131 /* Update parent cfs_rq utilization */
3132 add_positive(&cfs_rq->avg.util_avg, delta);
3133 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3136 /* Take into account change of load of a child task group */
3138 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3140 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3141 long delta, load = gcfs_rq->avg.load_avg;
3144 * If the load of group cfs_rq is null, the load of the
3145 * sched_entity will also be null so we can skip the formula
3150 /* Get tg's load and ensure tg_load > 0 */
3151 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3153 /* Ensure tg_load >= load and updated with current load*/
3154 tg_load -= gcfs_rq->tg_load_avg_contrib;
3158 * We need to compute a correction term in the case that the
3159 * task group is consuming more CPU than a task of equal
3160 * weight. A task with a weight equals to tg->shares will have
3161 * a load less or equal to scale_load_down(tg->shares).
3162 * Similarly, the sched_entities that represent the task group
3163 * at parent level, can't have a load higher than
3164 * scale_load_down(tg->shares). And the Sum of sched_entities'
3165 * load must be <= scale_load_down(tg->shares).
3167 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3168 /* scale gcfs_rq's load into tg's shares*/
3169 load *= scale_load_down(gcfs_rq->tg->shares);
3174 delta = load - se->avg.load_avg;
3176 /* Nothing to update */
3180 /* Set new sched_entity's load */
3181 se->avg.load_avg = load;
3182 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3184 /* Update parent cfs_rq load */
3185 add_positive(&cfs_rq->avg.load_avg, delta);
3186 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3189 * If the sched_entity is already enqueued, we also have to update the
3190 * runnable load avg.
3193 /* Update parent cfs_rq runnable_load_avg */
3194 add_positive(&cfs_rq->runnable_load_avg, delta);
3195 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3199 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3201 cfs_rq->propagate_avg = 1;
3204 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3206 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3208 if (!cfs_rq->propagate_avg)
3211 cfs_rq->propagate_avg = 0;
3215 /* Update task and its cfs_rq load average */
3216 static inline int propagate_entity_load_avg(struct sched_entity *se)
3218 struct cfs_rq *cfs_rq;
3220 if (entity_is_task(se))
3223 if (!test_and_clear_tg_cfs_propagate(se))
3226 cfs_rq = cfs_rq_of(se);
3228 set_tg_cfs_propagate(cfs_rq);
3230 update_tg_cfs_util(cfs_rq, se);
3231 update_tg_cfs_load(cfs_rq, se);
3237 * Check if we need to update the load and the utilization of a blocked
3240 static inline bool skip_blocked_update(struct sched_entity *se)
3242 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3245 * If sched_entity still have not zero load or utilization, we have to
3248 if (se->avg.load_avg || se->avg.util_avg)
3252 * If there is a pending propagation, we have to update the load and
3253 * the utilization of the sched_entity:
3255 if (gcfs_rq->propagate_avg)
3259 * Otherwise, the load and the utilization of the sched_entity is
3260 * already zero and there is no pending propagation, so it will be a
3261 * waste of time to try to decay it:
3266 #else /* CONFIG_FAIR_GROUP_SCHED */
3268 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3270 static inline int propagate_entity_load_avg(struct sched_entity *se)
3275 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3277 #endif /* CONFIG_FAIR_GROUP_SCHED */
3279 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3281 if (&this_rq()->cfs == cfs_rq) {
3283 * There are a few boundary cases this might miss but it should
3284 * get called often enough that that should (hopefully) not be
3285 * a real problem -- added to that it only calls on the local
3286 * CPU, so if we enqueue remotely we'll miss an update, but
3287 * the next tick/schedule should update.
3289 * It will not get called when we go idle, because the idle
3290 * thread is a different class (!fair), nor will the utilization
3291 * number include things like RT tasks.
3293 * As is, the util number is not freq-invariant (we'd have to
3294 * implement arch_scale_freq_capacity() for that).
3298 cpufreq_update_util(rq_of(cfs_rq), 0);
3303 * Unsigned subtract and clamp on underflow.
3305 * Explicitly do a load-store to ensure the intermediate value never hits
3306 * memory. This allows lockless observations without ever seeing the negative
3309 #define sub_positive(_ptr, _val) do { \
3310 typeof(_ptr) ptr = (_ptr); \
3311 typeof(*ptr) val = (_val); \
3312 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3316 WRITE_ONCE(*ptr, res); \
3320 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3321 * @now: current time, as per cfs_rq_clock_task()
3322 * @cfs_rq: cfs_rq to update
3323 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3325 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3326 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3327 * post_init_entity_util_avg().
3329 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3331 * Returns true if the load decayed or we removed load.
3333 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3334 * call update_tg_load_avg() when this function returns true.
3337 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3339 struct sched_avg *sa = &cfs_rq->avg;
3340 int decayed, removed_load = 0, removed_util = 0;
3342 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3343 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3344 sub_positive(&sa->load_avg, r);
3345 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3347 set_tg_cfs_propagate(cfs_rq);
3350 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3351 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3352 sub_positive(&sa->util_avg, r);
3353 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3355 set_tg_cfs_propagate(cfs_rq);
3358 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3360 #ifndef CONFIG_64BIT
3362 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3365 if (update_freq && (decayed || removed_util))
3366 cfs_rq_util_change(cfs_rq);
3368 return decayed || removed_load;
3372 * Optional action to be done while updating the load average
3374 #define UPDATE_TG 0x1
3375 #define SKIP_AGE_LOAD 0x2
3377 /* Update task and its cfs_rq load average */
3378 static inline void update_load_avg(struct sched_entity *se, int flags)
3380 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3381 u64 now = cfs_rq_clock_task(cfs_rq);
3382 struct rq *rq = rq_of(cfs_rq);
3383 int cpu = cpu_of(rq);
3387 * Track task load average for carrying it to new CPU after migrated, and
3388 * track group sched_entity load average for task_h_load calc in migration
3390 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3391 __update_load_avg_se(now, cpu, cfs_rq, se);
3393 decayed = update_cfs_rq_load_avg(now, cfs_rq, true);
3394 decayed |= propagate_entity_load_avg(se);
3396 if (decayed && (flags & UPDATE_TG))
3397 update_tg_load_avg(cfs_rq, 0);
3401 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3402 * @cfs_rq: cfs_rq to attach to
3403 * @se: sched_entity to attach
3405 * Must call update_cfs_rq_load_avg() before this, since we rely on
3406 * cfs_rq->avg.last_update_time being current.
3408 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3410 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3411 cfs_rq->avg.load_avg += se->avg.load_avg;
3412 cfs_rq->avg.load_sum += se->avg.load_sum;
3413 cfs_rq->avg.util_avg += se->avg.util_avg;
3414 cfs_rq->avg.util_sum += se->avg.util_sum;
3415 set_tg_cfs_propagate(cfs_rq);
3417 cfs_rq_util_change(cfs_rq);
3421 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3422 * @cfs_rq: cfs_rq to detach from
3423 * @se: sched_entity to detach
3425 * Must call update_cfs_rq_load_avg() before this, since we rely on
3426 * cfs_rq->avg.last_update_time being current.
3428 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3431 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3432 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3433 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3434 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3435 set_tg_cfs_propagate(cfs_rq);
3437 cfs_rq_util_change(cfs_rq);
3440 /* Add the load generated by se into cfs_rq's load average */
3442 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3444 struct sched_avg *sa = &se->avg;
3446 cfs_rq->runnable_load_avg += sa->load_avg;
3447 cfs_rq->runnable_load_sum += sa->load_sum;
3449 if (!sa->last_update_time) {
3450 attach_entity_load_avg(cfs_rq, se);
3451 update_tg_load_avg(cfs_rq, 0);
3455 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3457 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3459 cfs_rq->runnable_load_avg =
3460 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3461 cfs_rq->runnable_load_sum =
3462 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3465 #ifndef CONFIG_64BIT
3466 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3468 u64 last_update_time_copy;
3469 u64 last_update_time;
3472 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3474 last_update_time = cfs_rq->avg.last_update_time;
3475 } while (last_update_time != last_update_time_copy);
3477 return last_update_time;
3480 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3482 return cfs_rq->avg.last_update_time;
3487 * Synchronize entity load avg of dequeued entity without locking
3490 void sync_entity_load_avg(struct sched_entity *se)
3492 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3493 u64 last_update_time;
3495 last_update_time = cfs_rq_last_update_time(cfs_rq);
3496 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3500 * Task first catches up with cfs_rq, and then subtract
3501 * itself from the cfs_rq (task must be off the queue now).
3503 void remove_entity_load_avg(struct sched_entity *se)
3505 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3508 * tasks cannot exit without having gone through wake_up_new_task() ->
3509 * post_init_entity_util_avg() which will have added things to the
3510 * cfs_rq, so we can remove unconditionally.
3512 * Similarly for groups, they will have passed through
3513 * post_init_entity_util_avg() before unregister_sched_fair_group()
3517 sync_entity_load_avg(se);
3518 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3519 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3522 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3524 return cfs_rq->runnable_load_avg;
3527 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3529 return cfs_rq->avg.load_avg;
3532 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3534 #else /* CONFIG_SMP */
3537 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3542 #define UPDATE_TG 0x0
3543 #define SKIP_AGE_LOAD 0x0
3545 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3547 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3551 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3553 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3554 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3557 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3559 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3561 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3566 #endif /* CONFIG_SMP */
3568 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3570 #ifdef CONFIG_SCHED_DEBUG
3571 s64 d = se->vruntime - cfs_rq->min_vruntime;
3576 if (d > 3*sysctl_sched_latency)
3577 schedstat_inc(cfs_rq->nr_spread_over);
3582 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3584 u64 vruntime = cfs_rq->min_vruntime;
3587 * The 'current' period is already promised to the current tasks,
3588 * however the extra weight of the new task will slow them down a
3589 * little, place the new task so that it fits in the slot that
3590 * stays open at the end.
3592 if (initial && sched_feat(START_DEBIT))
3593 vruntime += sched_vslice(cfs_rq, se);
3595 /* sleeps up to a single latency don't count. */
3597 unsigned long thresh = sysctl_sched_latency;
3600 * Halve their sleep time's effect, to allow
3601 * for a gentler effect of sleepers:
3603 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3609 /* ensure we never gain time by being placed backwards. */
3610 se->vruntime = max_vruntime(se->vruntime, vruntime);
3613 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3615 static inline void check_schedstat_required(void)
3617 #ifdef CONFIG_SCHEDSTATS
3618 if (schedstat_enabled())
3621 /* Force schedstat enabled if a dependent tracepoint is active */
3622 if (trace_sched_stat_wait_enabled() ||
3623 trace_sched_stat_sleep_enabled() ||
3624 trace_sched_stat_iowait_enabled() ||
3625 trace_sched_stat_blocked_enabled() ||
3626 trace_sched_stat_runtime_enabled()) {
3627 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3628 "stat_blocked and stat_runtime require the "
3629 "kernel parameter schedstats=enable or "
3630 "kernel.sched_schedstats=1\n");
3641 * update_min_vruntime()
3642 * vruntime -= min_vruntime
3646 * update_min_vruntime()
3647 * vruntime += min_vruntime
3649 * this way the vruntime transition between RQs is done when both
3650 * min_vruntime are up-to-date.
3654 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3655 * vruntime -= min_vruntime
3659 * update_min_vruntime()
3660 * vruntime += min_vruntime
3662 * this way we don't have the most up-to-date min_vruntime on the originating
3663 * CPU and an up-to-date min_vruntime on the destination CPU.
3667 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3669 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3670 bool curr = cfs_rq->curr == se;
3673 * If we're the current task, we must renormalise before calling
3677 se->vruntime += cfs_rq->min_vruntime;
3679 update_curr(cfs_rq);
3682 * Otherwise, renormalise after, such that we're placed at the current
3683 * moment in time, instead of some random moment in the past. Being
3684 * placed in the past could significantly boost this task to the
3685 * fairness detriment of existing tasks.
3687 if (renorm && !curr)
3688 se->vruntime += cfs_rq->min_vruntime;
3691 * When enqueuing a sched_entity, we must:
3692 * - Update loads to have both entity and cfs_rq synced with now.
3693 * - Add its load to cfs_rq->runnable_avg
3694 * - For group_entity, update its weight to reflect the new share of
3696 * - Add its new weight to cfs_rq->load.weight
3698 update_load_avg(se, UPDATE_TG);
3699 enqueue_entity_load_avg(cfs_rq, se);
3700 update_cfs_shares(se);
3701 account_entity_enqueue(cfs_rq, se);
3703 if (flags & ENQUEUE_WAKEUP)
3704 place_entity(cfs_rq, se, 0);
3706 check_schedstat_required();
3707 update_stats_enqueue(cfs_rq, se, flags);
3708 check_spread(cfs_rq, se);
3710 __enqueue_entity(cfs_rq, se);
3713 if (cfs_rq->nr_running == 1) {
3714 list_add_leaf_cfs_rq(cfs_rq);
3715 check_enqueue_throttle(cfs_rq);
3719 static void __clear_buddies_last(struct sched_entity *se)
3721 for_each_sched_entity(se) {
3722 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3723 if (cfs_rq->last != se)
3726 cfs_rq->last = NULL;
3730 static void __clear_buddies_next(struct sched_entity *se)
3732 for_each_sched_entity(se) {
3733 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3734 if (cfs_rq->next != se)
3737 cfs_rq->next = NULL;
3741 static void __clear_buddies_skip(struct sched_entity *se)
3743 for_each_sched_entity(se) {
3744 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3745 if (cfs_rq->skip != se)
3748 cfs_rq->skip = NULL;
3752 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3754 if (cfs_rq->last == se)
3755 __clear_buddies_last(se);
3757 if (cfs_rq->next == se)
3758 __clear_buddies_next(se);
3760 if (cfs_rq->skip == se)
3761 __clear_buddies_skip(se);
3764 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3767 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3770 * Update run-time statistics of the 'current'.
3772 update_curr(cfs_rq);
3775 * When dequeuing a sched_entity, we must:
3776 * - Update loads to have both entity and cfs_rq synced with now.
3777 * - Substract its load from the cfs_rq->runnable_avg.
3778 * - Substract its previous weight from cfs_rq->load.weight.
3779 * - For group entity, update its weight to reflect the new share
3780 * of its group cfs_rq.
3782 update_load_avg(se, UPDATE_TG);
3783 dequeue_entity_load_avg(cfs_rq, se);
3785 update_stats_dequeue(cfs_rq, se, flags);
3787 clear_buddies(cfs_rq, se);
3789 if (se != cfs_rq->curr)
3790 __dequeue_entity(cfs_rq, se);
3792 account_entity_dequeue(cfs_rq, se);
3795 * Normalize after update_curr(); which will also have moved
3796 * min_vruntime if @se is the one holding it back. But before doing
3797 * update_min_vruntime() again, which will discount @se's position and
3798 * can move min_vruntime forward still more.
3800 if (!(flags & DEQUEUE_SLEEP))
3801 se->vruntime -= cfs_rq->min_vruntime;
3803 /* return excess runtime on last dequeue */
3804 return_cfs_rq_runtime(cfs_rq);
3806 update_cfs_shares(se);
3809 * Now advance min_vruntime if @se was the entity holding it back,
3810 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3811 * put back on, and if we advance min_vruntime, we'll be placed back
3812 * further than we started -- ie. we'll be penalized.
3814 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3815 update_min_vruntime(cfs_rq);
3819 * Preempt the current task with a newly woken task if needed:
3822 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3824 unsigned long ideal_runtime, delta_exec;
3825 struct sched_entity *se;
3828 ideal_runtime = sched_slice(cfs_rq, curr);
3829 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3830 if (delta_exec > ideal_runtime) {
3831 resched_curr(rq_of(cfs_rq));
3833 * The current task ran long enough, ensure it doesn't get
3834 * re-elected due to buddy favours.
3836 clear_buddies(cfs_rq, curr);
3841 * Ensure that a task that missed wakeup preemption by a
3842 * narrow margin doesn't have to wait for a full slice.
3843 * This also mitigates buddy induced latencies under load.
3845 if (delta_exec < sysctl_sched_min_granularity)
3848 se = __pick_first_entity(cfs_rq);
3849 delta = curr->vruntime - se->vruntime;
3854 if (delta > ideal_runtime)
3855 resched_curr(rq_of(cfs_rq));
3859 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3861 /* 'current' is not kept within the tree. */
3864 * Any task has to be enqueued before it get to execute on
3865 * a CPU. So account for the time it spent waiting on the
3868 update_stats_wait_end(cfs_rq, se);
3869 __dequeue_entity(cfs_rq, se);
3870 update_load_avg(se, UPDATE_TG);
3873 update_stats_curr_start(cfs_rq, se);
3877 * Track our maximum slice length, if the CPU's load is at
3878 * least twice that of our own weight (i.e. dont track it
3879 * when there are only lesser-weight tasks around):
3881 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3882 schedstat_set(se->statistics.slice_max,
3883 max((u64)schedstat_val(se->statistics.slice_max),
3884 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3887 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3891 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3894 * Pick the next process, keeping these things in mind, in this order:
3895 * 1) keep things fair between processes/task groups
3896 * 2) pick the "next" process, since someone really wants that to run
3897 * 3) pick the "last" process, for cache locality
3898 * 4) do not run the "skip" process, if something else is available
3900 static struct sched_entity *
3901 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3903 struct sched_entity *left = __pick_first_entity(cfs_rq);
3904 struct sched_entity *se;
3907 * If curr is set we have to see if its left of the leftmost entity
3908 * still in the tree, provided there was anything in the tree at all.
3910 if (!left || (curr && entity_before(curr, left)))
3913 se = left; /* ideally we run the leftmost entity */
3916 * Avoid running the skip buddy, if running something else can
3917 * be done without getting too unfair.
3919 if (cfs_rq->skip == se) {
3920 struct sched_entity *second;
3923 second = __pick_first_entity(cfs_rq);
3925 second = __pick_next_entity(se);
3926 if (!second || (curr && entity_before(curr, second)))
3930 if (second && wakeup_preempt_entity(second, left) < 1)
3935 * Prefer last buddy, try to return the CPU to a preempted task.
3937 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3941 * Someone really wants this to run. If it's not unfair, run it.
3943 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3946 clear_buddies(cfs_rq, se);
3951 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3953 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3956 * If still on the runqueue then deactivate_task()
3957 * was not called and update_curr() has to be done:
3960 update_curr(cfs_rq);
3962 /* throttle cfs_rqs exceeding runtime */
3963 check_cfs_rq_runtime(cfs_rq);
3965 check_spread(cfs_rq, prev);
3968 update_stats_wait_start(cfs_rq, prev);
3969 /* Put 'current' back into the tree. */
3970 __enqueue_entity(cfs_rq, prev);
3971 /* in !on_rq case, update occurred at dequeue */
3972 update_load_avg(prev, 0);
3974 cfs_rq->curr = NULL;
3978 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3981 * Update run-time statistics of the 'current'.
3983 update_curr(cfs_rq);
3986 * Ensure that runnable average is periodically updated.
3988 update_load_avg(curr, UPDATE_TG);
3989 update_cfs_shares(curr);
3991 #ifdef CONFIG_SCHED_HRTICK
3993 * queued ticks are scheduled to match the slice, so don't bother
3994 * validating it and just reschedule.
3997 resched_curr(rq_of(cfs_rq));
4001 * don't let the period tick interfere with the hrtick preemption
4003 if (!sched_feat(DOUBLE_TICK) &&
4004 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4008 if (cfs_rq->nr_running > 1)
4009 check_preempt_tick(cfs_rq, curr);
4013 /**************************************************
4014 * CFS bandwidth control machinery
4017 #ifdef CONFIG_CFS_BANDWIDTH
4019 #ifdef HAVE_JUMP_LABEL
4020 static struct static_key __cfs_bandwidth_used;
4022 static inline bool cfs_bandwidth_used(void)
4024 return static_key_false(&__cfs_bandwidth_used);
4027 void cfs_bandwidth_usage_inc(void)
4029 static_key_slow_inc(&__cfs_bandwidth_used);
4032 void cfs_bandwidth_usage_dec(void)
4034 static_key_slow_dec(&__cfs_bandwidth_used);
4036 #else /* HAVE_JUMP_LABEL */
4037 static bool cfs_bandwidth_used(void)
4042 void cfs_bandwidth_usage_inc(void) {}
4043 void cfs_bandwidth_usage_dec(void) {}
4044 #endif /* HAVE_JUMP_LABEL */
4047 * default period for cfs group bandwidth.
4048 * default: 0.1s, units: nanoseconds
4050 static inline u64 default_cfs_period(void)
4052 return 100000000ULL;
4055 static inline u64 sched_cfs_bandwidth_slice(void)
4057 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4061 * Replenish runtime according to assigned quota and update expiration time.
4062 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4063 * additional synchronization around rq->lock.
4065 * requires cfs_b->lock
4067 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4071 if (cfs_b->quota == RUNTIME_INF)
4074 now = sched_clock_cpu(smp_processor_id());
4075 cfs_b->runtime = cfs_b->quota;
4076 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4079 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4081 return &tg->cfs_bandwidth;
4084 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4085 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4087 if (unlikely(cfs_rq->throttle_count))
4088 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4090 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4093 /* returns 0 on failure to allocate runtime */
4094 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4096 struct task_group *tg = cfs_rq->tg;
4097 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4098 u64 amount = 0, min_amount, expires;
4100 /* note: this is a positive sum as runtime_remaining <= 0 */
4101 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4103 raw_spin_lock(&cfs_b->lock);
4104 if (cfs_b->quota == RUNTIME_INF)
4105 amount = min_amount;
4107 start_cfs_bandwidth(cfs_b);
4109 if (cfs_b->runtime > 0) {
4110 amount = min(cfs_b->runtime, min_amount);
4111 cfs_b->runtime -= amount;
4115 expires = cfs_b->runtime_expires;
4116 raw_spin_unlock(&cfs_b->lock);
4118 cfs_rq->runtime_remaining += amount;
4120 * we may have advanced our local expiration to account for allowed
4121 * spread between our sched_clock and the one on which runtime was
4124 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4125 cfs_rq->runtime_expires = expires;
4127 return cfs_rq->runtime_remaining > 0;
4131 * Note: This depends on the synchronization provided by sched_clock and the
4132 * fact that rq->clock snapshots this value.
4134 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4136 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4138 /* if the deadline is ahead of our clock, nothing to do */
4139 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4142 if (cfs_rq->runtime_remaining < 0)
4146 * If the local deadline has passed we have to consider the
4147 * possibility that our sched_clock is 'fast' and the global deadline
4148 * has not truly expired.
4150 * Fortunately we can check determine whether this the case by checking
4151 * whether the global deadline has advanced. It is valid to compare
4152 * cfs_b->runtime_expires without any locks since we only care about
4153 * exact equality, so a partial write will still work.
4156 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4157 /* extend local deadline, drift is bounded above by 2 ticks */
4158 cfs_rq->runtime_expires += TICK_NSEC;
4160 /* global deadline is ahead, expiration has passed */
4161 cfs_rq->runtime_remaining = 0;
4165 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4167 /* dock delta_exec before expiring quota (as it could span periods) */
4168 cfs_rq->runtime_remaining -= delta_exec;
4169 expire_cfs_rq_runtime(cfs_rq);
4171 if (likely(cfs_rq->runtime_remaining > 0))
4175 * if we're unable to extend our runtime we resched so that the active
4176 * hierarchy can be throttled
4178 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4179 resched_curr(rq_of(cfs_rq));
4182 static __always_inline
4183 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4185 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4188 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4191 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4193 return cfs_bandwidth_used() && cfs_rq->throttled;
4196 /* check whether cfs_rq, or any parent, is throttled */
4197 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4199 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4203 * Ensure that neither of the group entities corresponding to src_cpu or
4204 * dest_cpu are members of a throttled hierarchy when performing group
4205 * load-balance operations.
4207 static inline int throttled_lb_pair(struct task_group *tg,
4208 int src_cpu, int dest_cpu)
4210 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4212 src_cfs_rq = tg->cfs_rq[src_cpu];
4213 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4215 return throttled_hierarchy(src_cfs_rq) ||
4216 throttled_hierarchy(dest_cfs_rq);
4219 /* updated child weight may affect parent so we have to do this bottom up */
4220 static int tg_unthrottle_up(struct task_group *tg, void *data)
4222 struct rq *rq = data;
4223 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4225 cfs_rq->throttle_count--;
4226 if (!cfs_rq->throttle_count) {
4227 /* adjust cfs_rq_clock_task() */
4228 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4229 cfs_rq->throttled_clock_task;
4235 static int tg_throttle_down(struct task_group *tg, void *data)
4237 struct rq *rq = data;
4238 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4240 /* group is entering throttled state, stop time */
4241 if (!cfs_rq->throttle_count)
4242 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4243 cfs_rq->throttle_count++;
4248 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4250 struct rq *rq = rq_of(cfs_rq);
4251 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4252 struct sched_entity *se;
4253 long task_delta, dequeue = 1;
4256 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4258 /* freeze hierarchy runnable averages while throttled */
4260 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4263 task_delta = cfs_rq->h_nr_running;
4264 for_each_sched_entity(se) {
4265 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4266 /* throttled entity or throttle-on-deactivate */
4271 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4272 qcfs_rq->h_nr_running -= task_delta;
4274 if (qcfs_rq->load.weight)
4279 sub_nr_running(rq, task_delta);
4281 cfs_rq->throttled = 1;
4282 cfs_rq->throttled_clock = rq_clock(rq);
4283 raw_spin_lock(&cfs_b->lock);
4284 empty = list_empty(&cfs_b->throttled_cfs_rq);
4287 * Add to the _head_ of the list, so that an already-started
4288 * distribute_cfs_runtime will not see us
4290 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4293 * If we're the first throttled task, make sure the bandwidth
4297 start_cfs_bandwidth(cfs_b);
4299 raw_spin_unlock(&cfs_b->lock);
4302 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4304 struct rq *rq = rq_of(cfs_rq);
4305 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4306 struct sched_entity *se;
4310 se = cfs_rq->tg->se[cpu_of(rq)];
4312 cfs_rq->throttled = 0;
4314 update_rq_clock(rq);
4316 raw_spin_lock(&cfs_b->lock);
4317 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4318 list_del_rcu(&cfs_rq->throttled_list);
4319 raw_spin_unlock(&cfs_b->lock);
4321 /* update hierarchical throttle state */
4322 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4324 if (!cfs_rq->load.weight)
4327 task_delta = cfs_rq->h_nr_running;
4328 for_each_sched_entity(se) {
4332 cfs_rq = cfs_rq_of(se);
4334 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4335 cfs_rq->h_nr_running += task_delta;
4337 if (cfs_rq_throttled(cfs_rq))
4342 add_nr_running(rq, task_delta);
4344 /* determine whether we need to wake up potentially idle cpu */
4345 if (rq->curr == rq->idle && rq->cfs.nr_running)
4349 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4350 u64 remaining, u64 expires)
4352 struct cfs_rq *cfs_rq;
4354 u64 starting_runtime = remaining;
4357 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4359 struct rq *rq = rq_of(cfs_rq);
4363 if (!cfs_rq_throttled(cfs_rq))
4366 runtime = -cfs_rq->runtime_remaining + 1;
4367 if (runtime > remaining)
4368 runtime = remaining;
4369 remaining -= runtime;
4371 cfs_rq->runtime_remaining += runtime;
4372 cfs_rq->runtime_expires = expires;
4374 /* we check whether we're throttled above */
4375 if (cfs_rq->runtime_remaining > 0)
4376 unthrottle_cfs_rq(cfs_rq);
4386 return starting_runtime - remaining;
4390 * Responsible for refilling a task_group's bandwidth and unthrottling its
4391 * cfs_rqs as appropriate. If there has been no activity within the last
4392 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4393 * used to track this state.
4395 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4397 u64 runtime, runtime_expires;
4400 /* no need to continue the timer with no bandwidth constraint */
4401 if (cfs_b->quota == RUNTIME_INF)
4402 goto out_deactivate;
4404 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4405 cfs_b->nr_periods += overrun;
4408 * idle depends on !throttled (for the case of a large deficit), and if
4409 * we're going inactive then everything else can be deferred
4411 if (cfs_b->idle && !throttled)
4412 goto out_deactivate;
4414 __refill_cfs_bandwidth_runtime(cfs_b);
4417 /* mark as potentially idle for the upcoming period */
4422 /* account preceding periods in which throttling occurred */
4423 cfs_b->nr_throttled += overrun;
4425 runtime_expires = cfs_b->runtime_expires;
4428 * This check is repeated as we are holding onto the new bandwidth while
4429 * we unthrottle. This can potentially race with an unthrottled group
4430 * trying to acquire new bandwidth from the global pool. This can result
4431 * in us over-using our runtime if it is all used during this loop, but
4432 * only by limited amounts in that extreme case.
4434 while (throttled && cfs_b->runtime > 0) {
4435 runtime = cfs_b->runtime;
4436 raw_spin_unlock(&cfs_b->lock);
4437 /* we can't nest cfs_b->lock while distributing bandwidth */
4438 runtime = distribute_cfs_runtime(cfs_b, runtime,
4440 raw_spin_lock(&cfs_b->lock);
4442 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4444 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4448 * While we are ensured activity in the period following an
4449 * unthrottle, this also covers the case in which the new bandwidth is
4450 * insufficient to cover the existing bandwidth deficit. (Forcing the
4451 * timer to remain active while there are any throttled entities.)
4461 /* a cfs_rq won't donate quota below this amount */
4462 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4463 /* minimum remaining period time to redistribute slack quota */
4464 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4465 /* how long we wait to gather additional slack before distributing */
4466 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4469 * Are we near the end of the current quota period?
4471 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4472 * hrtimer base being cleared by hrtimer_start. In the case of
4473 * migrate_hrtimers, base is never cleared, so we are fine.
4475 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4477 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4480 /* if the call-back is running a quota refresh is already occurring */
4481 if (hrtimer_callback_running(refresh_timer))
4484 /* is a quota refresh about to occur? */
4485 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4486 if (remaining < min_expire)
4492 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4494 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4496 /* if there's a quota refresh soon don't bother with slack */
4497 if (runtime_refresh_within(cfs_b, min_left))
4500 hrtimer_start(&cfs_b->slack_timer,
4501 ns_to_ktime(cfs_bandwidth_slack_period),
4505 /* we know any runtime found here is valid as update_curr() precedes return */
4506 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4508 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4509 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4511 if (slack_runtime <= 0)
4514 raw_spin_lock(&cfs_b->lock);
4515 if (cfs_b->quota != RUNTIME_INF &&
4516 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4517 cfs_b->runtime += slack_runtime;
4519 /* we are under rq->lock, defer unthrottling using a timer */
4520 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4521 !list_empty(&cfs_b->throttled_cfs_rq))
4522 start_cfs_slack_bandwidth(cfs_b);
4524 raw_spin_unlock(&cfs_b->lock);
4526 /* even if it's not valid for return we don't want to try again */
4527 cfs_rq->runtime_remaining -= slack_runtime;
4530 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4532 if (!cfs_bandwidth_used())
4535 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4538 __return_cfs_rq_runtime(cfs_rq);
4542 * This is done with a timer (instead of inline with bandwidth return) since
4543 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4545 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4547 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4550 /* confirm we're still not at a refresh boundary */
4551 raw_spin_lock(&cfs_b->lock);
4552 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4553 raw_spin_unlock(&cfs_b->lock);
4557 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4558 runtime = cfs_b->runtime;
4560 expires = cfs_b->runtime_expires;
4561 raw_spin_unlock(&cfs_b->lock);
4566 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4568 raw_spin_lock(&cfs_b->lock);
4569 if (expires == cfs_b->runtime_expires)
4570 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4571 raw_spin_unlock(&cfs_b->lock);
4575 * When a group wakes up we want to make sure that its quota is not already
4576 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4577 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4579 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4581 if (!cfs_bandwidth_used())
4584 /* an active group must be handled by the update_curr()->put() path */
4585 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4588 /* ensure the group is not already throttled */
4589 if (cfs_rq_throttled(cfs_rq))
4592 /* update runtime allocation */
4593 account_cfs_rq_runtime(cfs_rq, 0);
4594 if (cfs_rq->runtime_remaining <= 0)
4595 throttle_cfs_rq(cfs_rq);
4598 static void sync_throttle(struct task_group *tg, int cpu)
4600 struct cfs_rq *pcfs_rq, *cfs_rq;
4602 if (!cfs_bandwidth_used())
4608 cfs_rq = tg->cfs_rq[cpu];
4609 pcfs_rq = tg->parent->cfs_rq[cpu];
4611 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4612 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4615 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4616 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4618 if (!cfs_bandwidth_used())
4621 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4625 * it's possible for a throttled entity to be forced into a running
4626 * state (e.g. set_curr_task), in this case we're finished.
4628 if (cfs_rq_throttled(cfs_rq))
4631 throttle_cfs_rq(cfs_rq);
4635 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4637 struct cfs_bandwidth *cfs_b =
4638 container_of(timer, struct cfs_bandwidth, slack_timer);
4640 do_sched_cfs_slack_timer(cfs_b);
4642 return HRTIMER_NORESTART;
4645 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4647 struct cfs_bandwidth *cfs_b =
4648 container_of(timer, struct cfs_bandwidth, period_timer);
4652 raw_spin_lock(&cfs_b->lock);
4654 overrun = hrtimer_forward_now(timer, cfs_b->period);
4658 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4661 cfs_b->period_active = 0;
4662 raw_spin_unlock(&cfs_b->lock);
4664 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4667 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4669 raw_spin_lock_init(&cfs_b->lock);
4671 cfs_b->quota = RUNTIME_INF;
4672 cfs_b->period = ns_to_ktime(default_cfs_period());
4674 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4675 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4676 cfs_b->period_timer.function = sched_cfs_period_timer;
4677 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4678 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4681 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4683 cfs_rq->runtime_enabled = 0;
4684 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4687 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4689 lockdep_assert_held(&cfs_b->lock);
4691 if (!cfs_b->period_active) {
4692 cfs_b->period_active = 1;
4693 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4694 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4698 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4700 /* init_cfs_bandwidth() was not called */
4701 if (!cfs_b->throttled_cfs_rq.next)
4704 hrtimer_cancel(&cfs_b->period_timer);
4705 hrtimer_cancel(&cfs_b->slack_timer);
4709 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4711 * The race is harmless, since modifying bandwidth settings of unhooked group
4712 * bits doesn't do much.
4715 /* cpu online calback */
4716 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4718 struct task_group *tg;
4720 lockdep_assert_held(&rq->lock);
4723 list_for_each_entry_rcu(tg, &task_groups, list) {
4724 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4725 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4727 raw_spin_lock(&cfs_b->lock);
4728 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4729 raw_spin_unlock(&cfs_b->lock);
4734 /* cpu offline callback */
4735 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4737 struct task_group *tg;
4739 lockdep_assert_held(&rq->lock);
4742 list_for_each_entry_rcu(tg, &task_groups, list) {
4743 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4745 if (!cfs_rq->runtime_enabled)
4749 * clock_task is not advancing so we just need to make sure
4750 * there's some valid quota amount
4752 cfs_rq->runtime_remaining = 1;
4754 * Offline rq is schedulable till cpu is completely disabled
4755 * in take_cpu_down(), so we prevent new cfs throttling here.
4757 cfs_rq->runtime_enabled = 0;
4759 if (cfs_rq_throttled(cfs_rq))
4760 unthrottle_cfs_rq(cfs_rq);
4765 #else /* CONFIG_CFS_BANDWIDTH */
4766 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4768 return rq_clock_task(rq_of(cfs_rq));
4771 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4772 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4773 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4774 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4775 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4777 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4782 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4787 static inline int throttled_lb_pair(struct task_group *tg,
4788 int src_cpu, int dest_cpu)
4793 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4795 #ifdef CONFIG_FAIR_GROUP_SCHED
4796 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4799 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4803 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4804 static inline void update_runtime_enabled(struct rq *rq) {}
4805 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4807 #endif /* CONFIG_CFS_BANDWIDTH */
4809 /**************************************************
4810 * CFS operations on tasks:
4813 #ifdef CONFIG_SCHED_HRTICK
4814 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4816 struct sched_entity *se = &p->se;
4817 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4819 SCHED_WARN_ON(task_rq(p) != rq);
4821 if (rq->cfs.h_nr_running > 1) {
4822 u64 slice = sched_slice(cfs_rq, se);
4823 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4824 s64 delta = slice - ran;
4831 hrtick_start(rq, delta);
4836 * called from enqueue/dequeue and updates the hrtick when the
4837 * current task is from our class and nr_running is low enough
4840 static void hrtick_update(struct rq *rq)
4842 struct task_struct *curr = rq->curr;
4844 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4847 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4848 hrtick_start_fair(rq, curr);
4850 #else /* !CONFIG_SCHED_HRTICK */
4852 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4856 static inline void hrtick_update(struct rq *rq)
4862 * The enqueue_task method is called before nr_running is
4863 * increased. Here we update the fair scheduling stats and
4864 * then put the task into the rbtree:
4867 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4869 struct cfs_rq *cfs_rq;
4870 struct sched_entity *se = &p->se;
4873 * If in_iowait is set, the code below may not trigger any cpufreq
4874 * utilization updates, so do it here explicitly with the IOWAIT flag
4878 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4880 for_each_sched_entity(se) {
4883 cfs_rq = cfs_rq_of(se);
4884 enqueue_entity(cfs_rq, se, flags);
4887 * end evaluation on encountering a throttled cfs_rq
4889 * note: in the case of encountering a throttled cfs_rq we will
4890 * post the final h_nr_running increment below.
4892 if (cfs_rq_throttled(cfs_rq))
4894 cfs_rq->h_nr_running++;
4896 flags = ENQUEUE_WAKEUP;
4899 for_each_sched_entity(se) {
4900 cfs_rq = cfs_rq_of(se);
4901 cfs_rq->h_nr_running++;
4903 if (cfs_rq_throttled(cfs_rq))
4906 update_load_avg(se, UPDATE_TG);
4907 update_cfs_shares(se);
4911 add_nr_running(rq, 1);
4916 static void set_next_buddy(struct sched_entity *se);
4919 * The dequeue_task method is called before nr_running is
4920 * decreased. We remove the task from the rbtree and
4921 * update the fair scheduling stats:
4923 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4925 struct cfs_rq *cfs_rq;
4926 struct sched_entity *se = &p->se;
4927 int task_sleep = flags & DEQUEUE_SLEEP;
4929 for_each_sched_entity(se) {
4930 cfs_rq = cfs_rq_of(se);
4931 dequeue_entity(cfs_rq, se, flags);
4934 * end evaluation on encountering a throttled cfs_rq
4936 * note: in the case of encountering a throttled cfs_rq we will
4937 * post the final h_nr_running decrement below.
4939 if (cfs_rq_throttled(cfs_rq))
4941 cfs_rq->h_nr_running--;
4943 /* Don't dequeue parent if it has other entities besides us */
4944 if (cfs_rq->load.weight) {
4945 /* Avoid re-evaluating load for this entity: */
4946 se = parent_entity(se);
4948 * Bias pick_next to pick a task from this cfs_rq, as
4949 * p is sleeping when it is within its sched_slice.
4951 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4955 flags |= DEQUEUE_SLEEP;
4958 for_each_sched_entity(se) {
4959 cfs_rq = cfs_rq_of(se);
4960 cfs_rq->h_nr_running--;
4962 if (cfs_rq_throttled(cfs_rq))
4965 update_load_avg(se, UPDATE_TG);
4966 update_cfs_shares(se);
4970 sub_nr_running(rq, 1);
4977 /* Working cpumask for: load_balance, load_balance_newidle. */
4978 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4979 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4981 #ifdef CONFIG_NO_HZ_COMMON
4983 * per rq 'load' arrray crap; XXX kill this.
4987 * The exact cpuload calculated at every tick would be:
4989 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4991 * If a cpu misses updates for n ticks (as it was idle) and update gets
4992 * called on the n+1-th tick when cpu may be busy, then we have:
4994 * load_n = (1 - 1/2^i)^n * load_0
4995 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4997 * decay_load_missed() below does efficient calculation of
4999 * load' = (1 - 1/2^i)^n * load
5001 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5002 * This allows us to precompute the above in said factors, thereby allowing the
5003 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5004 * fixed_power_int())
5006 * The calculation is approximated on a 128 point scale.
5008 #define DEGRADE_SHIFT 7
5010 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5011 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5012 { 0, 0, 0, 0, 0, 0, 0, 0 },
5013 { 64, 32, 8, 0, 0, 0, 0, 0 },
5014 { 96, 72, 40, 12, 1, 0, 0, 0 },
5015 { 112, 98, 75, 43, 15, 1, 0, 0 },
5016 { 120, 112, 98, 76, 45, 16, 2, 0 }
5020 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5021 * would be when CPU is idle and so we just decay the old load without
5022 * adding any new load.
5024 static unsigned long
5025 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5029 if (!missed_updates)
5032 if (missed_updates >= degrade_zero_ticks[idx])
5036 return load >> missed_updates;
5038 while (missed_updates) {
5039 if (missed_updates % 2)
5040 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5042 missed_updates >>= 1;
5047 #endif /* CONFIG_NO_HZ_COMMON */
5050 * __cpu_load_update - update the rq->cpu_load[] statistics
5051 * @this_rq: The rq to update statistics for
5052 * @this_load: The current load
5053 * @pending_updates: The number of missed updates
5055 * Update rq->cpu_load[] statistics. This function is usually called every
5056 * scheduler tick (TICK_NSEC).
5058 * This function computes a decaying average:
5060 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5062 * Because of NOHZ it might not get called on every tick which gives need for
5063 * the @pending_updates argument.
5065 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5066 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5067 * = A * (A * load[i]_n-2 + B) + B
5068 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5069 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5070 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5071 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5072 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5074 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5075 * any change in load would have resulted in the tick being turned back on.
5077 * For regular NOHZ, this reduces to:
5079 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5081 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5084 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5085 unsigned long pending_updates)
5087 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5090 this_rq->nr_load_updates++;
5092 /* Update our load: */
5093 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5094 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5095 unsigned long old_load, new_load;
5097 /* scale is effectively 1 << i now, and >> i divides by scale */
5099 old_load = this_rq->cpu_load[i];
5100 #ifdef CONFIG_NO_HZ_COMMON
5101 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5102 if (tickless_load) {
5103 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5105 * old_load can never be a negative value because a
5106 * decayed tickless_load cannot be greater than the
5107 * original tickless_load.
5109 old_load += tickless_load;
5112 new_load = this_load;
5114 * Round up the averaging division if load is increasing. This
5115 * prevents us from getting stuck on 9 if the load is 10, for
5118 if (new_load > old_load)
5119 new_load += scale - 1;
5121 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5124 sched_avg_update(this_rq);
5127 /* Used instead of source_load when we know the type == 0 */
5128 static unsigned long weighted_cpuload(const int cpu)
5130 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
5133 #ifdef CONFIG_NO_HZ_COMMON
5135 * There is no sane way to deal with nohz on smp when using jiffies because the
5136 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5137 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5139 * Therefore we need to avoid the delta approach from the regular tick when
5140 * possible since that would seriously skew the load calculation. This is why we
5141 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5142 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5143 * loop exit, nohz_idle_balance, nohz full exit...)
5145 * This means we might still be one tick off for nohz periods.
5148 static void cpu_load_update_nohz(struct rq *this_rq,
5149 unsigned long curr_jiffies,
5152 unsigned long pending_updates;
5154 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5155 if (pending_updates) {
5156 this_rq->last_load_update_tick = curr_jiffies;
5158 * In the regular NOHZ case, we were idle, this means load 0.
5159 * In the NOHZ_FULL case, we were non-idle, we should consider
5160 * its weighted load.
5162 cpu_load_update(this_rq, load, pending_updates);
5167 * Called from nohz_idle_balance() to update the load ratings before doing the
5170 static void cpu_load_update_idle(struct rq *this_rq)
5173 * bail if there's load or we're actually up-to-date.
5175 if (weighted_cpuload(cpu_of(this_rq)))
5178 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5182 * Record CPU load on nohz entry so we know the tickless load to account
5183 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5184 * than other cpu_load[idx] but it should be fine as cpu_load readers
5185 * shouldn't rely into synchronized cpu_load[*] updates.
5187 void cpu_load_update_nohz_start(void)
5189 struct rq *this_rq = this_rq();
5192 * This is all lockless but should be fine. If weighted_cpuload changes
5193 * concurrently we'll exit nohz. And cpu_load write can race with
5194 * cpu_load_update_idle() but both updater would be writing the same.
5196 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
5200 * Account the tickless load in the end of a nohz frame.
5202 void cpu_load_update_nohz_stop(void)
5204 unsigned long curr_jiffies = READ_ONCE(jiffies);
5205 struct rq *this_rq = this_rq();
5209 if (curr_jiffies == this_rq->last_load_update_tick)
5212 load = weighted_cpuload(cpu_of(this_rq));
5213 rq_lock(this_rq, &rf);
5214 update_rq_clock(this_rq);
5215 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5216 rq_unlock(this_rq, &rf);
5218 #else /* !CONFIG_NO_HZ_COMMON */
5219 static inline void cpu_load_update_nohz(struct rq *this_rq,
5220 unsigned long curr_jiffies,
5221 unsigned long load) { }
5222 #endif /* CONFIG_NO_HZ_COMMON */
5224 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5226 #ifdef CONFIG_NO_HZ_COMMON
5227 /* See the mess around cpu_load_update_nohz(). */
5228 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5230 cpu_load_update(this_rq, load, 1);
5234 * Called from scheduler_tick()
5236 void cpu_load_update_active(struct rq *this_rq)
5238 unsigned long load = weighted_cpuload(cpu_of(this_rq));
5240 if (tick_nohz_tick_stopped())
5241 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5243 cpu_load_update_periodic(this_rq, load);
5247 * Return a low guess at the load of a migration-source cpu weighted
5248 * according to the scheduling class and "nice" value.
5250 * We want to under-estimate the load of migration sources, to
5251 * balance conservatively.
5253 static unsigned long source_load(int cpu, int type)
5255 struct rq *rq = cpu_rq(cpu);
5256 unsigned long total = weighted_cpuload(cpu);
5258 if (type == 0 || !sched_feat(LB_BIAS))
5261 return min(rq->cpu_load[type-1], total);
5265 * Return a high guess at the load of a migration-target cpu weighted
5266 * according to the scheduling class and "nice" value.
5268 static unsigned long target_load(int cpu, int type)
5270 struct rq *rq = cpu_rq(cpu);
5271 unsigned long total = weighted_cpuload(cpu);
5273 if (type == 0 || !sched_feat(LB_BIAS))
5276 return max(rq->cpu_load[type-1], total);
5279 static unsigned long capacity_of(int cpu)
5281 return cpu_rq(cpu)->cpu_capacity;
5284 static unsigned long capacity_orig_of(int cpu)
5286 return cpu_rq(cpu)->cpu_capacity_orig;
5289 static unsigned long cpu_avg_load_per_task(int cpu)
5291 struct rq *rq = cpu_rq(cpu);
5292 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5293 unsigned long load_avg = weighted_cpuload(cpu);
5296 return load_avg / nr_running;
5301 #ifdef CONFIG_FAIR_GROUP_SCHED
5303 * effective_load() calculates the load change as seen from the root_task_group
5305 * Adding load to a group doesn't make a group heavier, but can cause movement
5306 * of group shares between cpus. Assuming the shares were perfectly aligned one
5307 * can calculate the shift in shares.
5309 * Calculate the effective load difference if @wl is added (subtracted) to @tg
5310 * on this @cpu and results in a total addition (subtraction) of @wg to the
5311 * total group weight.
5313 * Given a runqueue weight distribution (rw_i) we can compute a shares
5314 * distribution (s_i) using:
5316 * s_i = rw_i / \Sum rw_j (1)
5318 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5319 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5320 * shares distribution (s_i):
5322 * rw_i = { 2, 4, 1, 0 }
5323 * s_i = { 2/7, 4/7, 1/7, 0 }
5325 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5326 * task used to run on and the CPU the waker is running on), we need to
5327 * compute the effect of waking a task on either CPU and, in case of a sync
5328 * wakeup, compute the effect of the current task going to sleep.
5330 * So for a change of @wl to the local @cpu with an overall group weight change
5331 * of @wl we can compute the new shares distribution (s'_i) using:
5333 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
5335 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5336 * differences in waking a task to CPU 0. The additional task changes the
5337 * weight and shares distributions like:
5339 * rw'_i = { 3, 4, 1, 0 }
5340 * s'_i = { 3/8, 4/8, 1/8, 0 }
5342 * We can then compute the difference in effective weight by using:
5344 * dw_i = S * (s'_i - s_i) (3)
5346 * Where 'S' is the group weight as seen by its parent.
5348 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5349 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5350 * 4/7) times the weight of the group.
5352 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5354 struct sched_entity *se = tg->se[cpu];
5356 if (!tg->parent) /* the trivial, non-cgroup case */
5359 for_each_sched_entity(se) {
5360 struct cfs_rq *cfs_rq = se->my_q;
5361 long W, w = cfs_rq_load_avg(cfs_rq);
5366 * W = @wg + \Sum rw_j
5368 W = wg + atomic_long_read(&tg->load_avg);
5370 /* Ensure \Sum rw_j >= rw_i */
5371 W -= cfs_rq->tg_load_avg_contrib;
5380 * wl = S * s'_i; see (2)
5383 wl = (w * (long)scale_load_down(tg->shares)) / W;
5385 wl = scale_load_down(tg->shares);
5388 * Per the above, wl is the new se->load.weight value; since
5389 * those are clipped to [MIN_SHARES, ...) do so now. See
5390 * calc_cfs_shares().
5392 if (wl < MIN_SHARES)
5396 * wl = dw_i = S * (s'_i - s_i); see (3)
5398 wl -= se->avg.load_avg;
5401 * Recursively apply this logic to all parent groups to compute
5402 * the final effective load change on the root group. Since
5403 * only the @tg group gets extra weight, all parent groups can
5404 * only redistribute existing shares. @wl is the shift in shares
5405 * resulting from this level per the above.
5414 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5421 static void record_wakee(struct task_struct *p)
5424 * Only decay a single time; tasks that have less then 1 wakeup per
5425 * jiffy will not have built up many flips.
5427 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5428 current->wakee_flips >>= 1;
5429 current->wakee_flip_decay_ts = jiffies;
5432 if (current->last_wakee != p) {
5433 current->last_wakee = p;
5434 current->wakee_flips++;
5439 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5441 * A waker of many should wake a different task than the one last awakened
5442 * at a frequency roughly N times higher than one of its wakees.
5444 * In order to determine whether we should let the load spread vs consolidating
5445 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5446 * partner, and a factor of lls_size higher frequency in the other.
5448 * With both conditions met, we can be relatively sure that the relationship is
5449 * non-monogamous, with partner count exceeding socket size.
5451 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5452 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5455 static int wake_wide(struct task_struct *p)
5457 unsigned int master = current->wakee_flips;
5458 unsigned int slave = p->wakee_flips;
5459 int factor = this_cpu_read(sd_llc_size);
5462 swap(master, slave);
5463 if (slave < factor || master < slave * factor)
5468 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5469 int prev_cpu, int sync)
5471 int this_cpu = smp_processor_id();
5472 bool affine = false;
5475 * Common case: CPUs are in the same socket, and select_idle_sibling()
5476 * will do its thing regardless of what we return:
5478 if (cpus_share_cache(prev_cpu, this_cpu))
5481 affine = numa_wake_affine(sd, p, this_cpu, prev_cpu, sync);
5483 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5485 schedstat_inc(sd->ttwu_move_affine);
5486 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5492 static inline int task_util(struct task_struct *p);
5493 static int cpu_util_wake(int cpu, struct task_struct *p);
5495 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5497 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5501 * find_idlest_group finds and returns the least busy CPU group within the
5504 static struct sched_group *
5505 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5506 int this_cpu, int sd_flag)
5508 struct sched_group *idlest = NULL, *group = sd->groups;
5509 struct sched_group *most_spare_sg = NULL;
5510 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5511 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5512 unsigned long most_spare = 0, this_spare = 0;
5513 int load_idx = sd->forkexec_idx;
5514 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5515 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5516 (sd->imbalance_pct-100) / 100;
5518 if (sd_flag & SD_BALANCE_WAKE)
5519 load_idx = sd->wake_idx;
5522 unsigned long load, avg_load, runnable_load;
5523 unsigned long spare_cap, max_spare_cap;
5527 /* Skip over this group if it has no CPUs allowed */
5528 if (!cpumask_intersects(sched_group_span(group),
5532 local_group = cpumask_test_cpu(this_cpu,
5533 sched_group_span(group));
5536 * Tally up the load of all CPUs in the group and find
5537 * the group containing the CPU with most spare capacity.
5543 for_each_cpu(i, sched_group_span(group)) {
5544 /* Bias balancing toward cpus of our domain */
5546 load = source_load(i, load_idx);
5548 load = target_load(i, load_idx);
5550 runnable_load += load;
5552 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5554 spare_cap = capacity_spare_wake(i, p);
5556 if (spare_cap > max_spare_cap)
5557 max_spare_cap = spare_cap;
5560 /* Adjust by relative CPU capacity of the group */
5561 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5562 group->sgc->capacity;
5563 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5564 group->sgc->capacity;
5567 this_runnable_load = runnable_load;
5568 this_avg_load = avg_load;
5569 this_spare = max_spare_cap;
5571 if (min_runnable_load > (runnable_load + imbalance)) {
5573 * The runnable load is significantly smaller
5574 * so we can pick this new cpu
5576 min_runnable_load = runnable_load;
5577 min_avg_load = avg_load;
5579 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5580 (100*min_avg_load > imbalance_scale*avg_load)) {
5582 * The runnable loads are close so take the
5583 * blocked load into account through avg_load.
5585 min_avg_load = avg_load;
5589 if (most_spare < max_spare_cap) {
5590 most_spare = max_spare_cap;
5591 most_spare_sg = group;
5594 } while (group = group->next, group != sd->groups);
5597 * The cross-over point between using spare capacity or least load
5598 * is too conservative for high utilization tasks on partially
5599 * utilized systems if we require spare_capacity > task_util(p),
5600 * so we allow for some task stuffing by using
5601 * spare_capacity > task_util(p)/2.
5603 * Spare capacity can't be used for fork because the utilization has
5604 * not been set yet, we must first select a rq to compute the initial
5607 if (sd_flag & SD_BALANCE_FORK)
5610 if (this_spare > task_util(p) / 2 &&
5611 imbalance_scale*this_spare > 100*most_spare)
5614 if (most_spare > task_util(p) / 2)
5615 return most_spare_sg;
5621 if (min_runnable_load > (this_runnable_load + imbalance))
5624 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5625 (100*this_avg_load < imbalance_scale*min_avg_load))
5632 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5635 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5637 unsigned long load, min_load = ULONG_MAX;
5638 unsigned int min_exit_latency = UINT_MAX;
5639 u64 latest_idle_timestamp = 0;
5640 int least_loaded_cpu = this_cpu;
5641 int shallowest_idle_cpu = -1;
5644 /* Check if we have any choice: */
5645 if (group->group_weight == 1)
5646 return cpumask_first(sched_group_span(group));
5648 /* Traverse only the allowed CPUs */
5649 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5651 struct rq *rq = cpu_rq(i);
5652 struct cpuidle_state *idle = idle_get_state(rq);
5653 if (idle && idle->exit_latency < min_exit_latency) {
5655 * We give priority to a CPU whose idle state
5656 * has the smallest exit latency irrespective
5657 * of any idle timestamp.
5659 min_exit_latency = idle->exit_latency;
5660 latest_idle_timestamp = rq->idle_stamp;
5661 shallowest_idle_cpu = i;
5662 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5663 rq->idle_stamp > latest_idle_timestamp) {
5665 * If equal or no active idle state, then
5666 * the most recently idled CPU might have
5669 latest_idle_timestamp = rq->idle_stamp;
5670 shallowest_idle_cpu = i;
5672 } else if (shallowest_idle_cpu == -1) {
5673 load = weighted_cpuload(i);
5674 if (load < min_load || (load == min_load && i == this_cpu)) {
5676 least_loaded_cpu = i;
5681 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5684 #ifdef CONFIG_SCHED_SMT
5686 static inline void set_idle_cores(int cpu, int val)
5688 struct sched_domain_shared *sds;
5690 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5692 WRITE_ONCE(sds->has_idle_cores, val);
5695 static inline bool test_idle_cores(int cpu, bool def)
5697 struct sched_domain_shared *sds;
5699 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5701 return READ_ONCE(sds->has_idle_cores);
5707 * Scans the local SMT mask to see if the entire core is idle, and records this
5708 * information in sd_llc_shared->has_idle_cores.
5710 * Since SMT siblings share all cache levels, inspecting this limited remote
5711 * state should be fairly cheap.
5713 void __update_idle_core(struct rq *rq)
5715 int core = cpu_of(rq);
5719 if (test_idle_cores(core, true))
5722 for_each_cpu(cpu, cpu_smt_mask(core)) {
5730 set_idle_cores(core, 1);
5736 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5737 * there are no idle cores left in the system; tracked through
5738 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5740 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5742 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5745 if (!static_branch_likely(&sched_smt_present))
5748 if (!test_idle_cores(target, false))
5751 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5753 for_each_cpu_wrap(core, cpus, target) {
5756 for_each_cpu(cpu, cpu_smt_mask(core)) {
5757 cpumask_clear_cpu(cpu, cpus);
5767 * Failed to find an idle core; stop looking for one.
5769 set_idle_cores(target, 0);
5775 * Scan the local SMT mask for idle CPUs.
5777 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5781 if (!static_branch_likely(&sched_smt_present))
5784 for_each_cpu(cpu, cpu_smt_mask(target)) {
5785 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5794 #else /* CONFIG_SCHED_SMT */
5796 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5801 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5806 #endif /* CONFIG_SCHED_SMT */
5809 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5810 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5811 * average idle time for this rq (as found in rq->avg_idle).
5813 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5815 struct sched_domain *this_sd;
5816 u64 avg_cost, avg_idle;
5819 int cpu, nr = INT_MAX;
5821 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5826 * Due to large variance we need a large fuzz factor; hackbench in
5827 * particularly is sensitive here.
5829 avg_idle = this_rq()->avg_idle / 512;
5830 avg_cost = this_sd->avg_scan_cost + 1;
5832 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5835 if (sched_feat(SIS_PROP)) {
5836 u64 span_avg = sd->span_weight * avg_idle;
5837 if (span_avg > 4*avg_cost)
5838 nr = div_u64(span_avg, avg_cost);
5843 time = local_clock();
5845 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5848 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5854 time = local_clock() - time;
5855 cost = this_sd->avg_scan_cost;
5856 delta = (s64)(time - cost) / 8;
5857 this_sd->avg_scan_cost += delta;
5863 * Try and locate an idle core/thread in the LLC cache domain.
5865 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5867 struct sched_domain *sd;
5870 if (idle_cpu(target))
5874 * If the previous cpu is cache affine and idle, don't be stupid.
5876 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5879 sd = rcu_dereference(per_cpu(sd_llc, target));
5883 i = select_idle_core(p, sd, target);
5884 if ((unsigned)i < nr_cpumask_bits)
5887 i = select_idle_cpu(p, sd, target);
5888 if ((unsigned)i < nr_cpumask_bits)
5891 i = select_idle_smt(p, sd, target);
5892 if ((unsigned)i < nr_cpumask_bits)
5899 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5900 * tasks. The unit of the return value must be the one of capacity so we can
5901 * compare the utilization with the capacity of the CPU that is available for
5902 * CFS task (ie cpu_capacity).
5904 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5905 * recent utilization of currently non-runnable tasks on a CPU. It represents
5906 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5907 * capacity_orig is the cpu_capacity available at the highest frequency
5908 * (arch_scale_freq_capacity()).
5909 * The utilization of a CPU converges towards a sum equal to or less than the
5910 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5911 * the running time on this CPU scaled by capacity_curr.
5913 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5914 * higher than capacity_orig because of unfortunate rounding in
5915 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5916 * the average stabilizes with the new running time. We need to check that the
5917 * utilization stays within the range of [0..capacity_orig] and cap it if
5918 * necessary. Without utilization capping, a group could be seen as overloaded
5919 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5920 * available capacity. We allow utilization to overshoot capacity_curr (but not
5921 * capacity_orig) as it useful for predicting the capacity required after task
5922 * migrations (scheduler-driven DVFS).
5924 static int cpu_util(int cpu)
5926 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5927 unsigned long capacity = capacity_orig_of(cpu);
5929 return (util >= capacity) ? capacity : util;
5932 static inline int task_util(struct task_struct *p)
5934 return p->se.avg.util_avg;
5938 * cpu_util_wake: Compute cpu utilization with any contributions from
5939 * the waking task p removed.
5941 static int cpu_util_wake(int cpu, struct task_struct *p)
5943 unsigned long util, capacity;
5945 /* Task has no contribution or is new */
5946 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5947 return cpu_util(cpu);
5949 capacity = capacity_orig_of(cpu);
5950 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5952 return (util >= capacity) ? capacity : util;
5956 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5957 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5959 * In that case WAKE_AFFINE doesn't make sense and we'll let
5960 * BALANCE_WAKE sort things out.
5962 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5964 long min_cap, max_cap;
5966 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5967 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5969 /* Minimum capacity is close to max, no need to abort wake_affine */
5970 if (max_cap - min_cap < max_cap >> 3)
5973 /* Bring task utilization in sync with prev_cpu */
5974 sync_entity_load_avg(&p->se);
5976 return min_cap * 1024 < task_util(p) * capacity_margin;
5980 * select_task_rq_fair: Select target runqueue for the waking task in domains
5981 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5982 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5984 * Balances load by selecting the idlest cpu in the idlest group, or under
5985 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5987 * Returns the target cpu number.
5989 * preempt must be disabled.
5992 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5994 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5995 int cpu = smp_processor_id();
5996 int new_cpu = prev_cpu;
5997 int want_affine = 0;
5998 int sync = wake_flags & WF_SYNC;
6000 if (sd_flag & SD_BALANCE_WAKE) {
6002 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6003 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6007 for_each_domain(cpu, tmp) {
6008 if (!(tmp->flags & SD_LOAD_BALANCE))
6012 * If both cpu and prev_cpu are part of this domain,
6013 * cpu is a valid SD_WAKE_AFFINE target.
6015 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6016 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6021 if (tmp->flags & sd_flag)
6023 else if (!want_affine)
6028 sd = NULL; /* Prefer wake_affine over balance flags */
6029 if (cpu == prev_cpu)
6032 if (wake_affine(affine_sd, p, prev_cpu, sync))
6038 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6039 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6042 struct sched_group *group;
6045 if (!(sd->flags & sd_flag)) {
6050 group = find_idlest_group(sd, p, cpu, sd_flag);
6056 new_cpu = find_idlest_cpu(group, p, cpu);
6057 if (new_cpu == -1 || new_cpu == cpu) {
6058 /* Now try balancing at a lower domain level of cpu */
6063 /* Now try balancing at a lower domain level of new_cpu */
6065 weight = sd->span_weight;
6067 for_each_domain(cpu, tmp) {
6068 if (weight <= tmp->span_weight)
6070 if (tmp->flags & sd_flag)
6073 /* while loop will break here if sd == NULL */
6081 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6082 * cfs_rq_of(p) references at time of call are still valid and identify the
6083 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6085 static void migrate_task_rq_fair(struct task_struct *p)
6088 * As blocked tasks retain absolute vruntime the migration needs to
6089 * deal with this by subtracting the old and adding the new
6090 * min_vruntime -- the latter is done by enqueue_entity() when placing
6091 * the task on the new runqueue.
6093 if (p->state == TASK_WAKING) {
6094 struct sched_entity *se = &p->se;
6095 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6098 #ifndef CONFIG_64BIT
6099 u64 min_vruntime_copy;
6102 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6104 min_vruntime = cfs_rq->min_vruntime;
6105 } while (min_vruntime != min_vruntime_copy);
6107 min_vruntime = cfs_rq->min_vruntime;
6110 se->vruntime -= min_vruntime;
6114 * We are supposed to update the task to "current" time, then its up to date
6115 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6116 * what current time is, so simply throw away the out-of-date time. This
6117 * will result in the wakee task is less decayed, but giving the wakee more
6118 * load sounds not bad.
6120 remove_entity_load_avg(&p->se);
6122 /* Tell new CPU we are migrated */
6123 p->se.avg.last_update_time = 0;
6125 /* We have migrated, no longer consider this task hot */
6126 p->se.exec_start = 0;
6129 static void task_dead_fair(struct task_struct *p)
6131 remove_entity_load_avg(&p->se);
6133 #endif /* CONFIG_SMP */
6135 static unsigned long
6136 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6138 unsigned long gran = sysctl_sched_wakeup_granularity;
6141 * Since its curr running now, convert the gran from real-time
6142 * to virtual-time in his units.
6144 * By using 'se' instead of 'curr' we penalize light tasks, so
6145 * they get preempted easier. That is, if 'se' < 'curr' then
6146 * the resulting gran will be larger, therefore penalizing the
6147 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6148 * be smaller, again penalizing the lighter task.
6150 * This is especially important for buddies when the leftmost
6151 * task is higher priority than the buddy.
6153 return calc_delta_fair(gran, se);
6157 * Should 'se' preempt 'curr'.
6171 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6173 s64 gran, vdiff = curr->vruntime - se->vruntime;
6178 gran = wakeup_gran(curr, se);
6185 static void set_last_buddy(struct sched_entity *se)
6187 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6190 for_each_sched_entity(se) {
6191 if (SCHED_WARN_ON(!se->on_rq))
6193 cfs_rq_of(se)->last = se;
6197 static void set_next_buddy(struct sched_entity *se)
6199 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6202 for_each_sched_entity(se) {
6203 if (SCHED_WARN_ON(!se->on_rq))
6205 cfs_rq_of(se)->next = se;
6209 static void set_skip_buddy(struct sched_entity *se)
6211 for_each_sched_entity(se)
6212 cfs_rq_of(se)->skip = se;
6216 * Preempt the current task with a newly woken task if needed:
6218 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6220 struct task_struct *curr = rq->curr;
6221 struct sched_entity *se = &curr->se, *pse = &p->se;
6222 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6223 int scale = cfs_rq->nr_running >= sched_nr_latency;
6224 int next_buddy_marked = 0;
6226 if (unlikely(se == pse))
6230 * This is possible from callers such as attach_tasks(), in which we
6231 * unconditionally check_prempt_curr() after an enqueue (which may have
6232 * lead to a throttle). This both saves work and prevents false
6233 * next-buddy nomination below.
6235 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6238 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6239 set_next_buddy(pse);
6240 next_buddy_marked = 1;
6244 * We can come here with TIF_NEED_RESCHED already set from new task
6247 * Note: this also catches the edge-case of curr being in a throttled
6248 * group (e.g. via set_curr_task), since update_curr() (in the
6249 * enqueue of curr) will have resulted in resched being set. This
6250 * prevents us from potentially nominating it as a false LAST_BUDDY
6253 if (test_tsk_need_resched(curr))
6256 /* Idle tasks are by definition preempted by non-idle tasks. */
6257 if (unlikely(curr->policy == SCHED_IDLE) &&
6258 likely(p->policy != SCHED_IDLE))
6262 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6263 * is driven by the tick):
6265 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6268 find_matching_se(&se, &pse);
6269 update_curr(cfs_rq_of(se));
6271 if (wakeup_preempt_entity(se, pse) == 1) {
6273 * Bias pick_next to pick the sched entity that is
6274 * triggering this preemption.
6276 if (!next_buddy_marked)
6277 set_next_buddy(pse);
6286 * Only set the backward buddy when the current task is still
6287 * on the rq. This can happen when a wakeup gets interleaved
6288 * with schedule on the ->pre_schedule() or idle_balance()
6289 * point, either of which can * drop the rq lock.
6291 * Also, during early boot the idle thread is in the fair class,
6292 * for obvious reasons its a bad idea to schedule back to it.
6294 if (unlikely(!se->on_rq || curr == rq->idle))
6297 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6301 static struct task_struct *
6302 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6304 struct cfs_rq *cfs_rq = &rq->cfs;
6305 struct sched_entity *se;
6306 struct task_struct *p;
6310 #ifdef CONFIG_FAIR_GROUP_SCHED
6311 if (!cfs_rq->nr_running)
6314 if (prev->sched_class != &fair_sched_class)
6318 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6319 * likely that a next task is from the same cgroup as the current.
6321 * Therefore attempt to avoid putting and setting the entire cgroup
6322 * hierarchy, only change the part that actually changes.
6326 struct sched_entity *curr = cfs_rq->curr;
6329 * Since we got here without doing put_prev_entity() we also
6330 * have to consider cfs_rq->curr. If it is still a runnable
6331 * entity, update_curr() will update its vruntime, otherwise
6332 * forget we've ever seen it.
6336 update_curr(cfs_rq);
6341 * This call to check_cfs_rq_runtime() will do the
6342 * throttle and dequeue its entity in the parent(s).
6343 * Therefore the 'simple' nr_running test will indeed
6346 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6350 se = pick_next_entity(cfs_rq, curr);
6351 cfs_rq = group_cfs_rq(se);
6357 * Since we haven't yet done put_prev_entity and if the selected task
6358 * is a different task than we started out with, try and touch the
6359 * least amount of cfs_rqs.
6362 struct sched_entity *pse = &prev->se;
6364 while (!(cfs_rq = is_same_group(se, pse))) {
6365 int se_depth = se->depth;
6366 int pse_depth = pse->depth;
6368 if (se_depth <= pse_depth) {
6369 put_prev_entity(cfs_rq_of(pse), pse);
6370 pse = parent_entity(pse);
6372 if (se_depth >= pse_depth) {
6373 set_next_entity(cfs_rq_of(se), se);
6374 se = parent_entity(se);
6378 put_prev_entity(cfs_rq, pse);
6379 set_next_entity(cfs_rq, se);
6382 if (hrtick_enabled(rq))
6383 hrtick_start_fair(rq, p);
6390 if (!cfs_rq->nr_running)
6393 put_prev_task(rq, prev);
6396 se = pick_next_entity(cfs_rq, NULL);
6397 set_next_entity(cfs_rq, se);
6398 cfs_rq = group_cfs_rq(se);
6403 if (hrtick_enabled(rq))
6404 hrtick_start_fair(rq, p);
6409 new_tasks = idle_balance(rq, rf);
6412 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6413 * possible for any higher priority task to appear. In that case we
6414 * must re-start the pick_next_entity() loop.
6426 * Account for a descheduled task:
6428 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6430 struct sched_entity *se = &prev->se;
6431 struct cfs_rq *cfs_rq;
6433 for_each_sched_entity(se) {
6434 cfs_rq = cfs_rq_of(se);
6435 put_prev_entity(cfs_rq, se);
6440 * sched_yield() is very simple
6442 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6444 static void yield_task_fair(struct rq *rq)
6446 struct task_struct *curr = rq->curr;
6447 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6448 struct sched_entity *se = &curr->se;
6451 * Are we the only task in the tree?
6453 if (unlikely(rq->nr_running == 1))
6456 clear_buddies(cfs_rq, se);
6458 if (curr->policy != SCHED_BATCH) {
6459 update_rq_clock(rq);
6461 * Update run-time statistics of the 'current'.
6463 update_curr(cfs_rq);
6465 * Tell update_rq_clock() that we've just updated,
6466 * so we don't do microscopic update in schedule()
6467 * and double the fastpath cost.
6469 rq_clock_skip_update(rq, true);
6475 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6477 struct sched_entity *se = &p->se;
6479 /* throttled hierarchies are not runnable */
6480 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6483 /* Tell the scheduler that we'd really like pse to run next. */
6486 yield_task_fair(rq);
6492 /**************************************************
6493 * Fair scheduling class load-balancing methods.
6497 * The purpose of load-balancing is to achieve the same basic fairness the
6498 * per-cpu scheduler provides, namely provide a proportional amount of compute
6499 * time to each task. This is expressed in the following equation:
6501 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6503 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6504 * W_i,0 is defined as:
6506 * W_i,0 = \Sum_j w_i,j (2)
6508 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6509 * is derived from the nice value as per sched_prio_to_weight[].
6511 * The weight average is an exponential decay average of the instantaneous
6514 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6516 * C_i is the compute capacity of cpu i, typically it is the
6517 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6518 * can also include other factors [XXX].
6520 * To achieve this balance we define a measure of imbalance which follows
6521 * directly from (1):
6523 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6525 * We them move tasks around to minimize the imbalance. In the continuous
6526 * function space it is obvious this converges, in the discrete case we get
6527 * a few fun cases generally called infeasible weight scenarios.
6530 * - infeasible weights;
6531 * - local vs global optima in the discrete case. ]
6536 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6537 * for all i,j solution, we create a tree of cpus that follows the hardware
6538 * topology where each level pairs two lower groups (or better). This results
6539 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6540 * tree to only the first of the previous level and we decrease the frequency
6541 * of load-balance at each level inv. proportional to the number of cpus in
6547 * \Sum { --- * --- * 2^i } = O(n) (5)
6549 * `- size of each group
6550 * | | `- number of cpus doing load-balance
6552 * `- sum over all levels
6554 * Coupled with a limit on how many tasks we can migrate every balance pass,
6555 * this makes (5) the runtime complexity of the balancer.
6557 * An important property here is that each CPU is still (indirectly) connected
6558 * to every other cpu in at most O(log n) steps:
6560 * The adjacency matrix of the resulting graph is given by:
6563 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6566 * And you'll find that:
6568 * A^(log_2 n)_i,j != 0 for all i,j (7)
6570 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6571 * The task movement gives a factor of O(m), giving a convergence complexity
6574 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6579 * In order to avoid CPUs going idle while there's still work to do, new idle
6580 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6581 * tree itself instead of relying on other CPUs to bring it work.
6583 * This adds some complexity to both (5) and (8) but it reduces the total idle
6591 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6594 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6599 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6601 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6603 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6606 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6607 * rewrite all of this once again.]
6610 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6612 enum fbq_type { regular, remote, all };
6614 #define LBF_ALL_PINNED 0x01
6615 #define LBF_NEED_BREAK 0x02
6616 #define LBF_DST_PINNED 0x04
6617 #define LBF_SOME_PINNED 0x08
6620 struct sched_domain *sd;
6628 struct cpumask *dst_grpmask;
6630 enum cpu_idle_type idle;
6632 /* The set of CPUs under consideration for load-balancing */
6633 struct cpumask *cpus;
6638 unsigned int loop_break;
6639 unsigned int loop_max;
6641 enum fbq_type fbq_type;
6642 struct list_head tasks;
6646 * Is this task likely cache-hot:
6648 static int task_hot(struct task_struct *p, struct lb_env *env)
6652 lockdep_assert_held(&env->src_rq->lock);
6654 if (p->sched_class != &fair_sched_class)
6657 if (unlikely(p->policy == SCHED_IDLE))
6661 * Buddy candidates are cache hot:
6663 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6664 (&p->se == cfs_rq_of(&p->se)->next ||
6665 &p->se == cfs_rq_of(&p->se)->last))
6668 if (sysctl_sched_migration_cost == -1)
6670 if (sysctl_sched_migration_cost == 0)
6673 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6675 return delta < (s64)sysctl_sched_migration_cost;
6678 #ifdef CONFIG_NUMA_BALANCING
6680 * Returns 1, if task migration degrades locality
6681 * Returns 0, if task migration improves locality i.e migration preferred.
6682 * Returns -1, if task migration is not affected by locality.
6684 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6686 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6687 unsigned long src_faults, dst_faults;
6688 int src_nid, dst_nid;
6690 if (!static_branch_likely(&sched_numa_balancing))
6693 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6696 src_nid = cpu_to_node(env->src_cpu);
6697 dst_nid = cpu_to_node(env->dst_cpu);
6699 if (src_nid == dst_nid)
6702 /* Migrating away from the preferred node is always bad. */
6703 if (src_nid == p->numa_preferred_nid) {
6704 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6710 /* Encourage migration to the preferred node. */
6711 if (dst_nid == p->numa_preferred_nid)
6714 /* Leaving a core idle is often worse than degrading locality. */
6715 if (env->idle != CPU_NOT_IDLE)
6719 src_faults = group_faults(p, src_nid);
6720 dst_faults = group_faults(p, dst_nid);
6722 src_faults = task_faults(p, src_nid);
6723 dst_faults = task_faults(p, dst_nid);
6726 return dst_faults < src_faults;
6730 static inline int migrate_degrades_locality(struct task_struct *p,
6738 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6741 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6745 lockdep_assert_held(&env->src_rq->lock);
6748 * We do not migrate tasks that are:
6749 * 1) throttled_lb_pair, or
6750 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6751 * 3) running (obviously), or
6752 * 4) are cache-hot on their current CPU.
6754 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6757 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6760 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6762 env->flags |= LBF_SOME_PINNED;
6765 * Remember if this task can be migrated to any other cpu in
6766 * our sched_group. We may want to revisit it if we couldn't
6767 * meet load balance goals by pulling other tasks on src_cpu.
6769 * Also avoid computing new_dst_cpu if we have already computed
6770 * one in current iteration.
6772 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6775 /* Prevent to re-select dst_cpu via env's cpus */
6776 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6777 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
6778 env->flags |= LBF_DST_PINNED;
6779 env->new_dst_cpu = cpu;
6787 /* Record that we found atleast one task that could run on dst_cpu */
6788 env->flags &= ~LBF_ALL_PINNED;
6790 if (task_running(env->src_rq, p)) {
6791 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6796 * Aggressive migration if:
6797 * 1) destination numa is preferred
6798 * 2) task is cache cold, or
6799 * 3) too many balance attempts have failed.
6801 tsk_cache_hot = migrate_degrades_locality(p, env);
6802 if (tsk_cache_hot == -1)
6803 tsk_cache_hot = task_hot(p, env);
6805 if (tsk_cache_hot <= 0 ||
6806 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6807 if (tsk_cache_hot == 1) {
6808 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6809 schedstat_inc(p->se.statistics.nr_forced_migrations);
6814 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6819 * detach_task() -- detach the task for the migration specified in env
6821 static void detach_task(struct task_struct *p, struct lb_env *env)
6823 lockdep_assert_held(&env->src_rq->lock);
6825 p->on_rq = TASK_ON_RQ_MIGRATING;
6826 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
6827 set_task_cpu(p, env->dst_cpu);
6831 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6832 * part of active balancing operations within "domain".
6834 * Returns a task if successful and NULL otherwise.
6836 static struct task_struct *detach_one_task(struct lb_env *env)
6838 struct task_struct *p, *n;
6840 lockdep_assert_held(&env->src_rq->lock);
6842 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6843 if (!can_migrate_task(p, env))
6846 detach_task(p, env);
6849 * Right now, this is only the second place where
6850 * lb_gained[env->idle] is updated (other is detach_tasks)
6851 * so we can safely collect stats here rather than
6852 * inside detach_tasks().
6854 schedstat_inc(env->sd->lb_gained[env->idle]);
6860 static const unsigned int sched_nr_migrate_break = 32;
6863 * detach_tasks() -- tries to detach up to imbalance weighted load from
6864 * busiest_rq, as part of a balancing operation within domain "sd".
6866 * Returns number of detached tasks if successful and 0 otherwise.
6868 static int detach_tasks(struct lb_env *env)
6870 struct list_head *tasks = &env->src_rq->cfs_tasks;
6871 struct task_struct *p;
6875 lockdep_assert_held(&env->src_rq->lock);
6877 if (env->imbalance <= 0)
6880 while (!list_empty(tasks)) {
6882 * We don't want to steal all, otherwise we may be treated likewise,
6883 * which could at worst lead to a livelock crash.
6885 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6888 p = list_first_entry(tasks, struct task_struct, se.group_node);
6891 /* We've more or less seen every task there is, call it quits */
6892 if (env->loop > env->loop_max)
6895 /* take a breather every nr_migrate tasks */
6896 if (env->loop > env->loop_break) {
6897 env->loop_break += sched_nr_migrate_break;
6898 env->flags |= LBF_NEED_BREAK;
6902 if (!can_migrate_task(p, env))
6905 load = task_h_load(p);
6907 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6910 if ((load / 2) > env->imbalance)
6913 detach_task(p, env);
6914 list_add(&p->se.group_node, &env->tasks);
6917 env->imbalance -= load;
6919 #ifdef CONFIG_PREEMPT
6921 * NEWIDLE balancing is a source of latency, so preemptible
6922 * kernels will stop after the first task is detached to minimize
6923 * the critical section.
6925 if (env->idle == CPU_NEWLY_IDLE)
6930 * We only want to steal up to the prescribed amount of
6933 if (env->imbalance <= 0)
6938 list_move_tail(&p->se.group_node, tasks);
6942 * Right now, this is one of only two places we collect this stat
6943 * so we can safely collect detach_one_task() stats here rather
6944 * than inside detach_one_task().
6946 schedstat_add(env->sd->lb_gained[env->idle], detached);
6952 * attach_task() -- attach the task detached by detach_task() to its new rq.
6954 static void attach_task(struct rq *rq, struct task_struct *p)
6956 lockdep_assert_held(&rq->lock);
6958 BUG_ON(task_rq(p) != rq);
6959 activate_task(rq, p, ENQUEUE_NOCLOCK);
6960 p->on_rq = TASK_ON_RQ_QUEUED;
6961 check_preempt_curr(rq, p, 0);
6965 * attach_one_task() -- attaches the task returned from detach_one_task() to
6968 static void attach_one_task(struct rq *rq, struct task_struct *p)
6973 update_rq_clock(rq);
6979 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6982 static void attach_tasks(struct lb_env *env)
6984 struct list_head *tasks = &env->tasks;
6985 struct task_struct *p;
6988 rq_lock(env->dst_rq, &rf);
6989 update_rq_clock(env->dst_rq);
6991 while (!list_empty(tasks)) {
6992 p = list_first_entry(tasks, struct task_struct, se.group_node);
6993 list_del_init(&p->se.group_node);
6995 attach_task(env->dst_rq, p);
6998 rq_unlock(env->dst_rq, &rf);
7001 #ifdef CONFIG_FAIR_GROUP_SCHED
7003 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7005 if (cfs_rq->load.weight)
7008 if (cfs_rq->avg.load_sum)
7011 if (cfs_rq->avg.util_sum)
7014 if (cfs_rq->runnable_load_sum)
7020 static void update_blocked_averages(int cpu)
7022 struct rq *rq = cpu_rq(cpu);
7023 struct cfs_rq *cfs_rq, *pos;
7026 rq_lock_irqsave(rq, &rf);
7027 update_rq_clock(rq);
7030 * Iterates the task_group tree in a bottom up fashion, see
7031 * list_add_leaf_cfs_rq() for details.
7033 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7034 struct sched_entity *se;
7036 /* throttled entities do not contribute to load */
7037 if (throttled_hierarchy(cfs_rq))
7040 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
7041 update_tg_load_avg(cfs_rq, 0);
7043 /* Propagate pending load changes to the parent, if any: */
7044 se = cfs_rq->tg->se[cpu];
7045 if (se && !skip_blocked_update(se))
7046 update_load_avg(se, 0);
7049 * There can be a lot of idle CPU cgroups. Don't let fully
7050 * decayed cfs_rqs linger on the list.
7052 if (cfs_rq_is_decayed(cfs_rq))
7053 list_del_leaf_cfs_rq(cfs_rq);
7055 rq_unlock_irqrestore(rq, &rf);
7059 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7060 * This needs to be done in a top-down fashion because the load of a child
7061 * group is a fraction of its parents load.
7063 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7065 struct rq *rq = rq_of(cfs_rq);
7066 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7067 unsigned long now = jiffies;
7070 if (cfs_rq->last_h_load_update == now)
7073 cfs_rq->h_load_next = NULL;
7074 for_each_sched_entity(se) {
7075 cfs_rq = cfs_rq_of(se);
7076 cfs_rq->h_load_next = se;
7077 if (cfs_rq->last_h_load_update == now)
7082 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7083 cfs_rq->last_h_load_update = now;
7086 while ((se = cfs_rq->h_load_next) != NULL) {
7087 load = cfs_rq->h_load;
7088 load = div64_ul(load * se->avg.load_avg,
7089 cfs_rq_load_avg(cfs_rq) + 1);
7090 cfs_rq = group_cfs_rq(se);
7091 cfs_rq->h_load = load;
7092 cfs_rq->last_h_load_update = now;
7096 static unsigned long task_h_load(struct task_struct *p)
7098 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7100 update_cfs_rq_h_load(cfs_rq);
7101 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7102 cfs_rq_load_avg(cfs_rq) + 1);
7105 static inline void update_blocked_averages(int cpu)
7107 struct rq *rq = cpu_rq(cpu);
7108 struct cfs_rq *cfs_rq = &rq->cfs;
7111 rq_lock_irqsave(rq, &rf);
7112 update_rq_clock(rq);
7113 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7114 rq_unlock_irqrestore(rq, &rf);
7117 static unsigned long task_h_load(struct task_struct *p)
7119 return p->se.avg.load_avg;
7123 /********** Helpers for find_busiest_group ************************/
7132 * sg_lb_stats - stats of a sched_group required for load_balancing
7134 struct sg_lb_stats {
7135 unsigned long avg_load; /*Avg load across the CPUs of the group */
7136 unsigned long group_load; /* Total load over the CPUs of the group */
7137 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7138 unsigned long load_per_task;
7139 unsigned long group_capacity;
7140 unsigned long group_util; /* Total utilization of the group */
7141 unsigned int sum_nr_running; /* Nr tasks running in the group */
7142 unsigned int idle_cpus;
7143 unsigned int group_weight;
7144 enum group_type group_type;
7145 int group_no_capacity;
7146 #ifdef CONFIG_NUMA_BALANCING
7147 unsigned int nr_numa_running;
7148 unsigned int nr_preferred_running;
7153 * sd_lb_stats - Structure to store the statistics of a sched_domain
7154 * during load balancing.
7156 struct sd_lb_stats {
7157 struct sched_group *busiest; /* Busiest group in this sd */
7158 struct sched_group *local; /* Local group in this sd */
7159 unsigned long total_load; /* Total load of all groups in sd */
7160 unsigned long total_capacity; /* Total capacity of all groups in sd */
7161 unsigned long avg_load; /* Average load across all groups in sd */
7163 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7164 struct sg_lb_stats local_stat; /* Statistics of the local group */
7167 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7170 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7171 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7172 * We must however clear busiest_stat::avg_load because
7173 * update_sd_pick_busiest() reads this before assignment.
7175 *sds = (struct sd_lb_stats){
7179 .total_capacity = 0UL,
7182 .sum_nr_running = 0,
7183 .group_type = group_other,
7189 * get_sd_load_idx - Obtain the load index for a given sched domain.
7190 * @sd: The sched_domain whose load_idx is to be obtained.
7191 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7193 * Return: The load index.
7195 static inline int get_sd_load_idx(struct sched_domain *sd,
7196 enum cpu_idle_type idle)
7202 load_idx = sd->busy_idx;
7205 case CPU_NEWLY_IDLE:
7206 load_idx = sd->newidle_idx;
7209 load_idx = sd->idle_idx;
7216 static unsigned long scale_rt_capacity(int cpu)
7218 struct rq *rq = cpu_rq(cpu);
7219 u64 total, used, age_stamp, avg;
7223 * Since we're reading these variables without serialization make sure
7224 * we read them once before doing sanity checks on them.
7226 age_stamp = READ_ONCE(rq->age_stamp);
7227 avg = READ_ONCE(rq->rt_avg);
7228 delta = __rq_clock_broken(rq) - age_stamp;
7230 if (unlikely(delta < 0))
7233 total = sched_avg_period() + delta;
7235 used = div_u64(avg, total);
7237 if (likely(used < SCHED_CAPACITY_SCALE))
7238 return SCHED_CAPACITY_SCALE - used;
7243 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7245 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7246 struct sched_group *sdg = sd->groups;
7248 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7250 capacity *= scale_rt_capacity(cpu);
7251 capacity >>= SCHED_CAPACITY_SHIFT;
7256 cpu_rq(cpu)->cpu_capacity = capacity;
7257 sdg->sgc->capacity = capacity;
7258 sdg->sgc->min_capacity = capacity;
7261 void update_group_capacity(struct sched_domain *sd, int cpu)
7263 struct sched_domain *child = sd->child;
7264 struct sched_group *group, *sdg = sd->groups;
7265 unsigned long capacity, min_capacity;
7266 unsigned long interval;
7268 interval = msecs_to_jiffies(sd->balance_interval);
7269 interval = clamp(interval, 1UL, max_load_balance_interval);
7270 sdg->sgc->next_update = jiffies + interval;
7273 update_cpu_capacity(sd, cpu);
7278 min_capacity = ULONG_MAX;
7280 if (child->flags & SD_OVERLAP) {
7282 * SD_OVERLAP domains cannot assume that child groups
7283 * span the current group.
7286 for_each_cpu(cpu, sched_group_span(sdg)) {
7287 struct sched_group_capacity *sgc;
7288 struct rq *rq = cpu_rq(cpu);
7291 * build_sched_domains() -> init_sched_groups_capacity()
7292 * gets here before we've attached the domains to the
7295 * Use capacity_of(), which is set irrespective of domains
7296 * in update_cpu_capacity().
7298 * This avoids capacity from being 0 and
7299 * causing divide-by-zero issues on boot.
7301 if (unlikely(!rq->sd)) {
7302 capacity += capacity_of(cpu);
7304 sgc = rq->sd->groups->sgc;
7305 capacity += sgc->capacity;
7308 min_capacity = min(capacity, min_capacity);
7312 * !SD_OVERLAP domains can assume that child groups
7313 * span the current group.
7316 group = child->groups;
7318 struct sched_group_capacity *sgc = group->sgc;
7320 capacity += sgc->capacity;
7321 min_capacity = min(sgc->min_capacity, min_capacity);
7322 group = group->next;
7323 } while (group != child->groups);
7326 sdg->sgc->capacity = capacity;
7327 sdg->sgc->min_capacity = min_capacity;
7331 * Check whether the capacity of the rq has been noticeably reduced by side
7332 * activity. The imbalance_pct is used for the threshold.
7333 * Return true is the capacity is reduced
7336 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7338 return ((rq->cpu_capacity * sd->imbalance_pct) <
7339 (rq->cpu_capacity_orig * 100));
7343 * Group imbalance indicates (and tries to solve) the problem where balancing
7344 * groups is inadequate due to ->cpus_allowed constraints.
7346 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7347 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7350 * { 0 1 2 3 } { 4 5 6 7 }
7353 * If we were to balance group-wise we'd place two tasks in the first group and
7354 * two tasks in the second group. Clearly this is undesired as it will overload
7355 * cpu 3 and leave one of the cpus in the second group unused.
7357 * The current solution to this issue is detecting the skew in the first group
7358 * by noticing the lower domain failed to reach balance and had difficulty
7359 * moving tasks due to affinity constraints.
7361 * When this is so detected; this group becomes a candidate for busiest; see
7362 * update_sd_pick_busiest(). And calculate_imbalance() and
7363 * find_busiest_group() avoid some of the usual balance conditions to allow it
7364 * to create an effective group imbalance.
7366 * This is a somewhat tricky proposition since the next run might not find the
7367 * group imbalance and decide the groups need to be balanced again. A most
7368 * subtle and fragile situation.
7371 static inline int sg_imbalanced(struct sched_group *group)
7373 return group->sgc->imbalance;
7377 * group_has_capacity returns true if the group has spare capacity that could
7378 * be used by some tasks.
7379 * We consider that a group has spare capacity if the * number of task is
7380 * smaller than the number of CPUs or if the utilization is lower than the
7381 * available capacity for CFS tasks.
7382 * For the latter, we use a threshold to stabilize the state, to take into
7383 * account the variance of the tasks' load and to return true if the available
7384 * capacity in meaningful for the load balancer.
7385 * As an example, an available capacity of 1% can appear but it doesn't make
7386 * any benefit for the load balance.
7389 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7391 if (sgs->sum_nr_running < sgs->group_weight)
7394 if ((sgs->group_capacity * 100) >
7395 (sgs->group_util * env->sd->imbalance_pct))
7402 * group_is_overloaded returns true if the group has more tasks than it can
7404 * group_is_overloaded is not equals to !group_has_capacity because a group
7405 * with the exact right number of tasks, has no more spare capacity but is not
7406 * overloaded so both group_has_capacity and group_is_overloaded return
7410 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7412 if (sgs->sum_nr_running <= sgs->group_weight)
7415 if ((sgs->group_capacity * 100) <
7416 (sgs->group_util * env->sd->imbalance_pct))
7423 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7424 * per-CPU capacity than sched_group ref.
7427 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7429 return sg->sgc->min_capacity * capacity_margin <
7430 ref->sgc->min_capacity * 1024;
7434 group_type group_classify(struct sched_group *group,
7435 struct sg_lb_stats *sgs)
7437 if (sgs->group_no_capacity)
7438 return group_overloaded;
7440 if (sg_imbalanced(group))
7441 return group_imbalanced;
7447 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7448 * @env: The load balancing environment.
7449 * @group: sched_group whose statistics are to be updated.
7450 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7451 * @local_group: Does group contain this_cpu.
7452 * @sgs: variable to hold the statistics for this group.
7453 * @overload: Indicate more than one runnable task for any CPU.
7455 static inline void update_sg_lb_stats(struct lb_env *env,
7456 struct sched_group *group, int load_idx,
7457 int local_group, struct sg_lb_stats *sgs,
7463 memset(sgs, 0, sizeof(*sgs));
7465 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7466 struct rq *rq = cpu_rq(i);
7468 /* Bias balancing toward cpus of our domain */
7470 load = target_load(i, load_idx);
7472 load = source_load(i, load_idx);
7474 sgs->group_load += load;
7475 sgs->group_util += cpu_util(i);
7476 sgs->sum_nr_running += rq->cfs.h_nr_running;
7478 nr_running = rq->nr_running;
7482 #ifdef CONFIG_NUMA_BALANCING
7483 sgs->nr_numa_running += rq->nr_numa_running;
7484 sgs->nr_preferred_running += rq->nr_preferred_running;
7486 sgs->sum_weighted_load += weighted_cpuload(i);
7488 * No need to call idle_cpu() if nr_running is not 0
7490 if (!nr_running && idle_cpu(i))
7494 /* Adjust by relative CPU capacity of the group */
7495 sgs->group_capacity = group->sgc->capacity;
7496 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7498 if (sgs->sum_nr_running)
7499 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7501 sgs->group_weight = group->group_weight;
7503 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7504 sgs->group_type = group_classify(group, sgs);
7508 * update_sd_pick_busiest - return 1 on busiest group
7509 * @env: The load balancing environment.
7510 * @sds: sched_domain statistics
7511 * @sg: sched_group candidate to be checked for being the busiest
7512 * @sgs: sched_group statistics
7514 * Determine if @sg is a busier group than the previously selected
7517 * Return: %true if @sg is a busier group than the previously selected
7518 * busiest group. %false otherwise.
7520 static bool update_sd_pick_busiest(struct lb_env *env,
7521 struct sd_lb_stats *sds,
7522 struct sched_group *sg,
7523 struct sg_lb_stats *sgs)
7525 struct sg_lb_stats *busiest = &sds->busiest_stat;
7527 if (sgs->group_type > busiest->group_type)
7530 if (sgs->group_type < busiest->group_type)
7533 if (sgs->avg_load <= busiest->avg_load)
7536 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7540 * Candidate sg has no more than one task per CPU and
7541 * has higher per-CPU capacity. Migrating tasks to less
7542 * capable CPUs may harm throughput. Maximize throughput,
7543 * power/energy consequences are not considered.
7545 if (sgs->sum_nr_running <= sgs->group_weight &&
7546 group_smaller_cpu_capacity(sds->local, sg))
7550 /* This is the busiest node in its class. */
7551 if (!(env->sd->flags & SD_ASYM_PACKING))
7554 /* No ASYM_PACKING if target cpu is already busy */
7555 if (env->idle == CPU_NOT_IDLE)
7558 * ASYM_PACKING needs to move all the work to the highest
7559 * prority CPUs in the group, therefore mark all groups
7560 * of lower priority than ourself as busy.
7562 if (sgs->sum_nr_running &&
7563 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7567 /* Prefer to move from lowest priority cpu's work */
7568 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7569 sg->asym_prefer_cpu))
7576 #ifdef CONFIG_NUMA_BALANCING
7577 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7579 if (sgs->sum_nr_running > sgs->nr_numa_running)
7581 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7586 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7588 if (rq->nr_running > rq->nr_numa_running)
7590 if (rq->nr_running > rq->nr_preferred_running)
7595 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7600 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7604 #endif /* CONFIG_NUMA_BALANCING */
7607 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7608 * @env: The load balancing environment.
7609 * @sds: variable to hold the statistics for this sched_domain.
7611 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7613 struct sched_domain *child = env->sd->child;
7614 struct sched_group *sg = env->sd->groups;
7615 struct sg_lb_stats *local = &sds->local_stat;
7616 struct sg_lb_stats tmp_sgs;
7617 int load_idx, prefer_sibling = 0;
7618 bool overload = false;
7620 if (child && child->flags & SD_PREFER_SIBLING)
7623 load_idx = get_sd_load_idx(env->sd, env->idle);
7626 struct sg_lb_stats *sgs = &tmp_sgs;
7629 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7634 if (env->idle != CPU_NEWLY_IDLE ||
7635 time_after_eq(jiffies, sg->sgc->next_update))
7636 update_group_capacity(env->sd, env->dst_cpu);
7639 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7646 * In case the child domain prefers tasks go to siblings
7647 * first, lower the sg capacity so that we'll try
7648 * and move all the excess tasks away. We lower the capacity
7649 * of a group only if the local group has the capacity to fit
7650 * these excess tasks. The extra check prevents the case where
7651 * you always pull from the heaviest group when it is already
7652 * under-utilized (possible with a large weight task outweighs
7653 * the tasks on the system).
7655 if (prefer_sibling && sds->local &&
7656 group_has_capacity(env, local) &&
7657 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7658 sgs->group_no_capacity = 1;
7659 sgs->group_type = group_classify(sg, sgs);
7662 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7664 sds->busiest_stat = *sgs;
7668 /* Now, start updating sd_lb_stats */
7669 sds->total_load += sgs->group_load;
7670 sds->total_capacity += sgs->group_capacity;
7673 } while (sg != env->sd->groups);
7675 if (env->sd->flags & SD_NUMA)
7676 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7678 if (!env->sd->parent) {
7679 /* update overload indicator if we are at root domain */
7680 if (env->dst_rq->rd->overload != overload)
7681 env->dst_rq->rd->overload = overload;
7687 * check_asym_packing - Check to see if the group is packed into the
7690 * This is primarily intended to used at the sibling level. Some
7691 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7692 * case of POWER7, it can move to lower SMT modes only when higher
7693 * threads are idle. When in lower SMT modes, the threads will
7694 * perform better since they share less core resources. Hence when we
7695 * have idle threads, we want them to be the higher ones.
7697 * This packing function is run on idle threads. It checks to see if
7698 * the busiest CPU in this domain (core in the P7 case) has a higher
7699 * CPU number than the packing function is being run on. Here we are
7700 * assuming lower CPU number will be equivalent to lower a SMT thread
7703 * Return: 1 when packing is required and a task should be moved to
7704 * this CPU. The amount of the imbalance is returned in *imbalance.
7706 * @env: The load balancing environment.
7707 * @sds: Statistics of the sched_domain which is to be packed
7709 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7713 if (!(env->sd->flags & SD_ASYM_PACKING))
7716 if (env->idle == CPU_NOT_IDLE)
7722 busiest_cpu = sds->busiest->asym_prefer_cpu;
7723 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7726 env->imbalance = DIV_ROUND_CLOSEST(
7727 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7728 SCHED_CAPACITY_SCALE);
7734 * fix_small_imbalance - Calculate the minor imbalance that exists
7735 * amongst the groups of a sched_domain, during
7737 * @env: The load balancing environment.
7738 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7741 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7743 unsigned long tmp, capa_now = 0, capa_move = 0;
7744 unsigned int imbn = 2;
7745 unsigned long scaled_busy_load_per_task;
7746 struct sg_lb_stats *local, *busiest;
7748 local = &sds->local_stat;
7749 busiest = &sds->busiest_stat;
7751 if (!local->sum_nr_running)
7752 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7753 else if (busiest->load_per_task > local->load_per_task)
7756 scaled_busy_load_per_task =
7757 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7758 busiest->group_capacity;
7760 if (busiest->avg_load + scaled_busy_load_per_task >=
7761 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7762 env->imbalance = busiest->load_per_task;
7767 * OK, we don't have enough imbalance to justify moving tasks,
7768 * however we may be able to increase total CPU capacity used by
7772 capa_now += busiest->group_capacity *
7773 min(busiest->load_per_task, busiest->avg_load);
7774 capa_now += local->group_capacity *
7775 min(local->load_per_task, local->avg_load);
7776 capa_now /= SCHED_CAPACITY_SCALE;
7778 /* Amount of load we'd subtract */
7779 if (busiest->avg_load > scaled_busy_load_per_task) {
7780 capa_move += busiest->group_capacity *
7781 min(busiest->load_per_task,
7782 busiest->avg_load - scaled_busy_load_per_task);
7785 /* Amount of load we'd add */
7786 if (busiest->avg_load * busiest->group_capacity <
7787 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7788 tmp = (busiest->avg_load * busiest->group_capacity) /
7789 local->group_capacity;
7791 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7792 local->group_capacity;
7794 capa_move += local->group_capacity *
7795 min(local->load_per_task, local->avg_load + tmp);
7796 capa_move /= SCHED_CAPACITY_SCALE;
7798 /* Move if we gain throughput */
7799 if (capa_move > capa_now)
7800 env->imbalance = busiest->load_per_task;
7804 * calculate_imbalance - Calculate the amount of imbalance present within the
7805 * groups of a given sched_domain during load balance.
7806 * @env: load balance environment
7807 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7809 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7811 unsigned long max_pull, load_above_capacity = ~0UL;
7812 struct sg_lb_stats *local, *busiest;
7814 local = &sds->local_stat;
7815 busiest = &sds->busiest_stat;
7817 if (busiest->group_type == group_imbalanced) {
7819 * In the group_imb case we cannot rely on group-wide averages
7820 * to ensure cpu-load equilibrium, look at wider averages. XXX
7822 busiest->load_per_task =
7823 min(busiest->load_per_task, sds->avg_load);
7827 * Avg load of busiest sg can be less and avg load of local sg can
7828 * be greater than avg load across all sgs of sd because avg load
7829 * factors in sg capacity and sgs with smaller group_type are
7830 * skipped when updating the busiest sg:
7832 if (busiest->avg_load <= sds->avg_load ||
7833 local->avg_load >= sds->avg_load) {
7835 return fix_small_imbalance(env, sds);
7839 * If there aren't any idle cpus, avoid creating some.
7841 if (busiest->group_type == group_overloaded &&
7842 local->group_type == group_overloaded) {
7843 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7844 if (load_above_capacity > busiest->group_capacity) {
7845 load_above_capacity -= busiest->group_capacity;
7846 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7847 load_above_capacity /= busiest->group_capacity;
7849 load_above_capacity = ~0UL;
7853 * We're trying to get all the cpus to the average_load, so we don't
7854 * want to push ourselves above the average load, nor do we wish to
7855 * reduce the max loaded cpu below the average load. At the same time,
7856 * we also don't want to reduce the group load below the group
7857 * capacity. Thus we look for the minimum possible imbalance.
7859 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7861 /* How much load to actually move to equalise the imbalance */
7862 env->imbalance = min(
7863 max_pull * busiest->group_capacity,
7864 (sds->avg_load - local->avg_load) * local->group_capacity
7865 ) / SCHED_CAPACITY_SCALE;
7868 * if *imbalance is less than the average load per runnable task
7869 * there is no guarantee that any tasks will be moved so we'll have
7870 * a think about bumping its value to force at least one task to be
7873 if (env->imbalance < busiest->load_per_task)
7874 return fix_small_imbalance(env, sds);
7877 /******* find_busiest_group() helpers end here *********************/
7880 * find_busiest_group - Returns the busiest group within the sched_domain
7881 * if there is an imbalance.
7883 * Also calculates the amount of weighted load which should be moved
7884 * to restore balance.
7886 * @env: The load balancing environment.
7888 * Return: - The busiest group if imbalance exists.
7890 static struct sched_group *find_busiest_group(struct lb_env *env)
7892 struct sg_lb_stats *local, *busiest;
7893 struct sd_lb_stats sds;
7895 init_sd_lb_stats(&sds);
7898 * Compute the various statistics relavent for load balancing at
7901 update_sd_lb_stats(env, &sds);
7902 local = &sds.local_stat;
7903 busiest = &sds.busiest_stat;
7905 /* ASYM feature bypasses nice load balance check */
7906 if (check_asym_packing(env, &sds))
7909 /* There is no busy sibling group to pull tasks from */
7910 if (!sds.busiest || busiest->sum_nr_running == 0)
7913 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7914 / sds.total_capacity;
7917 * If the busiest group is imbalanced the below checks don't
7918 * work because they assume all things are equal, which typically
7919 * isn't true due to cpus_allowed constraints and the like.
7921 if (busiest->group_type == group_imbalanced)
7924 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7925 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7926 busiest->group_no_capacity)
7930 * If the local group is busier than the selected busiest group
7931 * don't try and pull any tasks.
7933 if (local->avg_load >= busiest->avg_load)
7937 * Don't pull any tasks if this group is already above the domain
7940 if (local->avg_load >= sds.avg_load)
7943 if (env->idle == CPU_IDLE) {
7945 * This cpu is idle. If the busiest group is not overloaded
7946 * and there is no imbalance between this and busiest group
7947 * wrt idle cpus, it is balanced. The imbalance becomes
7948 * significant if the diff is greater than 1 otherwise we
7949 * might end up to just move the imbalance on another group
7951 if ((busiest->group_type != group_overloaded) &&
7952 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7956 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7957 * imbalance_pct to be conservative.
7959 if (100 * busiest->avg_load <=
7960 env->sd->imbalance_pct * local->avg_load)
7965 /* Looks like there is an imbalance. Compute it */
7966 calculate_imbalance(env, &sds);
7975 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7977 static struct rq *find_busiest_queue(struct lb_env *env,
7978 struct sched_group *group)
7980 struct rq *busiest = NULL, *rq;
7981 unsigned long busiest_load = 0, busiest_capacity = 1;
7984 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7985 unsigned long capacity, wl;
7989 rt = fbq_classify_rq(rq);
7992 * We classify groups/runqueues into three groups:
7993 * - regular: there are !numa tasks
7994 * - remote: there are numa tasks that run on the 'wrong' node
7995 * - all: there is no distinction
7997 * In order to avoid migrating ideally placed numa tasks,
7998 * ignore those when there's better options.
8000 * If we ignore the actual busiest queue to migrate another
8001 * task, the next balance pass can still reduce the busiest
8002 * queue by moving tasks around inside the node.
8004 * If we cannot move enough load due to this classification
8005 * the next pass will adjust the group classification and
8006 * allow migration of more tasks.
8008 * Both cases only affect the total convergence complexity.
8010 if (rt > env->fbq_type)
8013 capacity = capacity_of(i);
8015 wl = weighted_cpuload(i);
8018 * When comparing with imbalance, use weighted_cpuload()
8019 * which is not scaled with the cpu capacity.
8022 if (rq->nr_running == 1 && wl > env->imbalance &&
8023 !check_cpu_capacity(rq, env->sd))
8027 * For the load comparisons with the other cpu's, consider
8028 * the weighted_cpuload() scaled with the cpu capacity, so
8029 * that the load can be moved away from the cpu that is
8030 * potentially running at a lower capacity.
8032 * Thus we're looking for max(wl_i / capacity_i), crosswise
8033 * multiplication to rid ourselves of the division works out
8034 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8035 * our previous maximum.
8037 if (wl * busiest_capacity > busiest_load * capacity) {
8039 busiest_capacity = capacity;
8048 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8049 * so long as it is large enough.
8051 #define MAX_PINNED_INTERVAL 512
8053 static int need_active_balance(struct lb_env *env)
8055 struct sched_domain *sd = env->sd;
8057 if (env->idle == CPU_NEWLY_IDLE) {
8060 * ASYM_PACKING needs to force migrate tasks from busy but
8061 * lower priority CPUs in order to pack all tasks in the
8062 * highest priority CPUs.
8064 if ((sd->flags & SD_ASYM_PACKING) &&
8065 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8070 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8071 * It's worth migrating the task if the src_cpu's capacity is reduced
8072 * because of other sched_class or IRQs if more capacity stays
8073 * available on dst_cpu.
8075 if ((env->idle != CPU_NOT_IDLE) &&
8076 (env->src_rq->cfs.h_nr_running == 1)) {
8077 if ((check_cpu_capacity(env->src_rq, sd)) &&
8078 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8082 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8085 static int active_load_balance_cpu_stop(void *data);
8087 static int should_we_balance(struct lb_env *env)
8089 struct sched_group *sg = env->sd->groups;
8090 int cpu, balance_cpu = -1;
8093 * In the newly idle case, we will allow all the cpu's
8094 * to do the newly idle load balance.
8096 if (env->idle == CPU_NEWLY_IDLE)
8099 /* Try to find first idle cpu */
8100 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8108 if (balance_cpu == -1)
8109 balance_cpu = group_balance_cpu(sg);
8112 * First idle cpu or the first cpu(busiest) in this sched group
8113 * is eligible for doing load balancing at this and above domains.
8115 return balance_cpu == env->dst_cpu;
8119 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8120 * tasks if there is an imbalance.
8122 static int load_balance(int this_cpu, struct rq *this_rq,
8123 struct sched_domain *sd, enum cpu_idle_type idle,
8124 int *continue_balancing)
8126 int ld_moved, cur_ld_moved, active_balance = 0;
8127 struct sched_domain *sd_parent = sd->parent;
8128 struct sched_group *group;
8131 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8133 struct lb_env env = {
8135 .dst_cpu = this_cpu,
8137 .dst_grpmask = sched_group_span(sd->groups),
8139 .loop_break = sched_nr_migrate_break,
8142 .tasks = LIST_HEAD_INIT(env.tasks),
8146 * For NEWLY_IDLE load_balancing, we don't need to consider
8147 * other cpus in our group
8149 if (idle == CPU_NEWLY_IDLE)
8150 env.dst_grpmask = NULL;
8152 cpumask_copy(cpus, cpu_active_mask);
8154 schedstat_inc(sd->lb_count[idle]);
8157 if (!should_we_balance(&env)) {
8158 *continue_balancing = 0;
8162 group = find_busiest_group(&env);
8164 schedstat_inc(sd->lb_nobusyg[idle]);
8168 busiest = find_busiest_queue(&env, group);
8170 schedstat_inc(sd->lb_nobusyq[idle]);
8174 BUG_ON(busiest == env.dst_rq);
8176 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8178 env.src_cpu = busiest->cpu;
8179 env.src_rq = busiest;
8182 if (busiest->nr_running > 1) {
8184 * Attempt to move tasks. If find_busiest_group has found
8185 * an imbalance but busiest->nr_running <= 1, the group is
8186 * still unbalanced. ld_moved simply stays zero, so it is
8187 * correctly treated as an imbalance.
8189 env.flags |= LBF_ALL_PINNED;
8190 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8193 rq_lock_irqsave(busiest, &rf);
8194 update_rq_clock(busiest);
8197 * cur_ld_moved - load moved in current iteration
8198 * ld_moved - cumulative load moved across iterations
8200 cur_ld_moved = detach_tasks(&env);
8203 * We've detached some tasks from busiest_rq. Every
8204 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8205 * unlock busiest->lock, and we are able to be sure
8206 * that nobody can manipulate the tasks in parallel.
8207 * See task_rq_lock() family for the details.
8210 rq_unlock(busiest, &rf);
8214 ld_moved += cur_ld_moved;
8217 local_irq_restore(rf.flags);
8219 if (env.flags & LBF_NEED_BREAK) {
8220 env.flags &= ~LBF_NEED_BREAK;
8225 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8226 * us and move them to an alternate dst_cpu in our sched_group
8227 * where they can run. The upper limit on how many times we
8228 * iterate on same src_cpu is dependent on number of cpus in our
8231 * This changes load balance semantics a bit on who can move
8232 * load to a given_cpu. In addition to the given_cpu itself
8233 * (or a ilb_cpu acting on its behalf where given_cpu is
8234 * nohz-idle), we now have balance_cpu in a position to move
8235 * load to given_cpu. In rare situations, this may cause
8236 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8237 * _independently_ and at _same_ time to move some load to
8238 * given_cpu) causing exceess load to be moved to given_cpu.
8239 * This however should not happen so much in practice and
8240 * moreover subsequent load balance cycles should correct the
8241 * excess load moved.
8243 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8245 /* Prevent to re-select dst_cpu via env's cpus */
8246 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8248 env.dst_rq = cpu_rq(env.new_dst_cpu);
8249 env.dst_cpu = env.new_dst_cpu;
8250 env.flags &= ~LBF_DST_PINNED;
8252 env.loop_break = sched_nr_migrate_break;
8255 * Go back to "more_balance" rather than "redo" since we
8256 * need to continue with same src_cpu.
8262 * We failed to reach balance because of affinity.
8265 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8267 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8268 *group_imbalance = 1;
8271 /* All tasks on this runqueue were pinned by CPU affinity */
8272 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8273 cpumask_clear_cpu(cpu_of(busiest), cpus);
8274 if (!cpumask_empty(cpus)) {
8276 env.loop_break = sched_nr_migrate_break;
8279 goto out_all_pinned;
8284 schedstat_inc(sd->lb_failed[idle]);
8286 * Increment the failure counter only on periodic balance.
8287 * We do not want newidle balance, which can be very
8288 * frequent, pollute the failure counter causing
8289 * excessive cache_hot migrations and active balances.
8291 if (idle != CPU_NEWLY_IDLE)
8292 sd->nr_balance_failed++;
8294 if (need_active_balance(&env)) {
8295 unsigned long flags;
8297 raw_spin_lock_irqsave(&busiest->lock, flags);
8299 /* don't kick the active_load_balance_cpu_stop,
8300 * if the curr task on busiest cpu can't be
8303 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8304 raw_spin_unlock_irqrestore(&busiest->lock,
8306 env.flags |= LBF_ALL_PINNED;
8307 goto out_one_pinned;
8311 * ->active_balance synchronizes accesses to
8312 * ->active_balance_work. Once set, it's cleared
8313 * only after active load balance is finished.
8315 if (!busiest->active_balance) {
8316 busiest->active_balance = 1;
8317 busiest->push_cpu = this_cpu;
8320 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8322 if (active_balance) {
8323 stop_one_cpu_nowait(cpu_of(busiest),
8324 active_load_balance_cpu_stop, busiest,
8325 &busiest->active_balance_work);
8328 /* We've kicked active balancing, force task migration. */
8329 sd->nr_balance_failed = sd->cache_nice_tries+1;
8332 sd->nr_balance_failed = 0;
8334 if (likely(!active_balance)) {
8335 /* We were unbalanced, so reset the balancing interval */
8336 sd->balance_interval = sd->min_interval;
8339 * If we've begun active balancing, start to back off. This
8340 * case may not be covered by the all_pinned logic if there
8341 * is only 1 task on the busy runqueue (because we don't call
8344 if (sd->balance_interval < sd->max_interval)
8345 sd->balance_interval *= 2;
8352 * We reach balance although we may have faced some affinity
8353 * constraints. Clear the imbalance flag if it was set.
8356 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8358 if (*group_imbalance)
8359 *group_imbalance = 0;
8364 * We reach balance because all tasks are pinned at this level so
8365 * we can't migrate them. Let the imbalance flag set so parent level
8366 * can try to migrate them.
8368 schedstat_inc(sd->lb_balanced[idle]);
8370 sd->nr_balance_failed = 0;
8373 /* tune up the balancing interval */
8374 if (((env.flags & LBF_ALL_PINNED) &&
8375 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8376 (sd->balance_interval < sd->max_interval))
8377 sd->balance_interval *= 2;
8384 static inline unsigned long
8385 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8387 unsigned long interval = sd->balance_interval;
8390 interval *= sd->busy_factor;
8392 /* scale ms to jiffies */
8393 interval = msecs_to_jiffies(interval);
8394 interval = clamp(interval, 1UL, max_load_balance_interval);
8400 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8402 unsigned long interval, next;
8404 /* used by idle balance, so cpu_busy = 0 */
8405 interval = get_sd_balance_interval(sd, 0);
8406 next = sd->last_balance + interval;
8408 if (time_after(*next_balance, next))
8409 *next_balance = next;
8413 * idle_balance is called by schedule() if this_cpu is about to become
8414 * idle. Attempts to pull tasks from other CPUs.
8416 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8418 unsigned long next_balance = jiffies + HZ;
8419 int this_cpu = this_rq->cpu;
8420 struct sched_domain *sd;
8421 int pulled_task = 0;
8425 * We must set idle_stamp _before_ calling idle_balance(), such that we
8426 * measure the duration of idle_balance() as idle time.
8428 this_rq->idle_stamp = rq_clock(this_rq);
8431 * This is OK, because current is on_cpu, which avoids it being picked
8432 * for load-balance and preemption/IRQs are still disabled avoiding
8433 * further scheduler activity on it and we're being very careful to
8434 * re-start the picking loop.
8436 rq_unpin_lock(this_rq, rf);
8438 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8439 !this_rq->rd->overload) {
8441 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8443 update_next_balance(sd, &next_balance);
8449 raw_spin_unlock(&this_rq->lock);
8451 update_blocked_averages(this_cpu);
8453 for_each_domain(this_cpu, sd) {
8454 int continue_balancing = 1;
8455 u64 t0, domain_cost;
8457 if (!(sd->flags & SD_LOAD_BALANCE))
8460 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8461 update_next_balance(sd, &next_balance);
8465 if (sd->flags & SD_BALANCE_NEWIDLE) {
8466 t0 = sched_clock_cpu(this_cpu);
8468 pulled_task = load_balance(this_cpu, this_rq,
8470 &continue_balancing);
8472 domain_cost = sched_clock_cpu(this_cpu) - t0;
8473 if (domain_cost > sd->max_newidle_lb_cost)
8474 sd->max_newidle_lb_cost = domain_cost;
8476 curr_cost += domain_cost;
8479 update_next_balance(sd, &next_balance);
8482 * Stop searching for tasks to pull if there are
8483 * now runnable tasks on this rq.
8485 if (pulled_task || this_rq->nr_running > 0)
8490 raw_spin_lock(&this_rq->lock);
8492 if (curr_cost > this_rq->max_idle_balance_cost)
8493 this_rq->max_idle_balance_cost = curr_cost;
8496 * While browsing the domains, we released the rq lock, a task could
8497 * have been enqueued in the meantime. Since we're not going idle,
8498 * pretend we pulled a task.
8500 if (this_rq->cfs.h_nr_running && !pulled_task)
8504 /* Move the next balance forward */
8505 if (time_after(this_rq->next_balance, next_balance))
8506 this_rq->next_balance = next_balance;
8508 /* Is there a task of a high priority class? */
8509 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8513 this_rq->idle_stamp = 0;
8515 rq_repin_lock(this_rq, rf);
8521 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8522 * running tasks off the busiest CPU onto idle CPUs. It requires at
8523 * least 1 task to be running on each physical CPU where possible, and
8524 * avoids physical / logical imbalances.
8526 static int active_load_balance_cpu_stop(void *data)
8528 struct rq *busiest_rq = data;
8529 int busiest_cpu = cpu_of(busiest_rq);
8530 int target_cpu = busiest_rq->push_cpu;
8531 struct rq *target_rq = cpu_rq(target_cpu);
8532 struct sched_domain *sd;
8533 struct task_struct *p = NULL;
8536 rq_lock_irq(busiest_rq, &rf);
8538 /* make sure the requested cpu hasn't gone down in the meantime */
8539 if (unlikely(busiest_cpu != smp_processor_id() ||
8540 !busiest_rq->active_balance))
8543 /* Is there any task to move? */
8544 if (busiest_rq->nr_running <= 1)
8548 * This condition is "impossible", if it occurs
8549 * we need to fix it. Originally reported by
8550 * Bjorn Helgaas on a 128-cpu setup.
8552 BUG_ON(busiest_rq == target_rq);
8554 /* Search for an sd spanning us and the target CPU. */
8556 for_each_domain(target_cpu, sd) {
8557 if ((sd->flags & SD_LOAD_BALANCE) &&
8558 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8563 struct lb_env env = {
8565 .dst_cpu = target_cpu,
8566 .dst_rq = target_rq,
8567 .src_cpu = busiest_rq->cpu,
8568 .src_rq = busiest_rq,
8572 schedstat_inc(sd->alb_count);
8573 update_rq_clock(busiest_rq);
8575 p = detach_one_task(&env);
8577 schedstat_inc(sd->alb_pushed);
8578 /* Active balancing done, reset the failure counter. */
8579 sd->nr_balance_failed = 0;
8581 schedstat_inc(sd->alb_failed);
8586 busiest_rq->active_balance = 0;
8587 rq_unlock(busiest_rq, &rf);
8590 attach_one_task(target_rq, p);
8597 static inline int on_null_domain(struct rq *rq)
8599 return unlikely(!rcu_dereference_sched(rq->sd));
8602 #ifdef CONFIG_NO_HZ_COMMON
8604 * idle load balancing details
8605 * - When one of the busy CPUs notice that there may be an idle rebalancing
8606 * needed, they will kick the idle load balancer, which then does idle
8607 * load balancing for all the idle CPUs.
8610 cpumask_var_t idle_cpus_mask;
8612 unsigned long next_balance; /* in jiffy units */
8613 } nohz ____cacheline_aligned;
8615 static inline int find_new_ilb(void)
8617 int ilb = cpumask_first(nohz.idle_cpus_mask);
8619 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8626 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8627 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8628 * CPU (if there is one).
8630 static void nohz_balancer_kick(void)
8634 nohz.next_balance++;
8636 ilb_cpu = find_new_ilb();
8638 if (ilb_cpu >= nr_cpu_ids)
8641 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8644 * Use smp_send_reschedule() instead of resched_cpu().
8645 * This way we generate a sched IPI on the target cpu which
8646 * is idle. And the softirq performing nohz idle load balance
8647 * will be run before returning from the IPI.
8649 smp_send_reschedule(ilb_cpu);
8653 void nohz_balance_exit_idle(unsigned int cpu)
8655 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8657 * Completely isolated CPUs don't ever set, so we must test.
8659 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8660 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8661 atomic_dec(&nohz.nr_cpus);
8663 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8667 static inline void set_cpu_sd_state_busy(void)
8669 struct sched_domain *sd;
8670 int cpu = smp_processor_id();
8673 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8675 if (!sd || !sd->nohz_idle)
8679 atomic_inc(&sd->shared->nr_busy_cpus);
8684 void set_cpu_sd_state_idle(void)
8686 struct sched_domain *sd;
8687 int cpu = smp_processor_id();
8690 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8692 if (!sd || sd->nohz_idle)
8696 atomic_dec(&sd->shared->nr_busy_cpus);
8702 * This routine will record that the cpu is going idle with tick stopped.
8703 * This info will be used in performing idle load balancing in the future.
8705 void nohz_balance_enter_idle(int cpu)
8708 * If this cpu is going down, then nothing needs to be done.
8710 if (!cpu_active(cpu))
8713 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8714 if (!is_housekeeping_cpu(cpu))
8717 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8721 * If we're a completely isolated CPU, we don't play.
8723 if (on_null_domain(cpu_rq(cpu)))
8726 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8727 atomic_inc(&nohz.nr_cpus);
8728 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8732 static DEFINE_SPINLOCK(balancing);
8735 * Scale the max load_balance interval with the number of CPUs in the system.
8736 * This trades load-balance latency on larger machines for less cross talk.
8738 void update_max_interval(void)
8740 max_load_balance_interval = HZ*num_online_cpus()/10;
8744 * It checks each scheduling domain to see if it is due to be balanced,
8745 * and initiates a balancing operation if so.
8747 * Balancing parameters are set up in init_sched_domains.
8749 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8751 int continue_balancing = 1;
8753 unsigned long interval;
8754 struct sched_domain *sd;
8755 /* Earliest time when we have to do rebalance again */
8756 unsigned long next_balance = jiffies + 60*HZ;
8757 int update_next_balance = 0;
8758 int need_serialize, need_decay = 0;
8761 update_blocked_averages(cpu);
8764 for_each_domain(cpu, sd) {
8766 * Decay the newidle max times here because this is a regular
8767 * visit to all the domains. Decay ~1% per second.
8769 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8770 sd->max_newidle_lb_cost =
8771 (sd->max_newidle_lb_cost * 253) / 256;
8772 sd->next_decay_max_lb_cost = jiffies + HZ;
8775 max_cost += sd->max_newidle_lb_cost;
8777 if (!(sd->flags & SD_LOAD_BALANCE))
8781 * Stop the load balance at this level. There is another
8782 * CPU in our sched group which is doing load balancing more
8785 if (!continue_balancing) {
8791 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8793 need_serialize = sd->flags & SD_SERIALIZE;
8794 if (need_serialize) {
8795 if (!spin_trylock(&balancing))
8799 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8800 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8802 * The LBF_DST_PINNED logic could have changed
8803 * env->dst_cpu, so we can't know our idle
8804 * state even if we migrated tasks. Update it.
8806 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8808 sd->last_balance = jiffies;
8809 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8812 spin_unlock(&balancing);
8814 if (time_after(next_balance, sd->last_balance + interval)) {
8815 next_balance = sd->last_balance + interval;
8816 update_next_balance = 1;
8821 * Ensure the rq-wide value also decays but keep it at a
8822 * reasonable floor to avoid funnies with rq->avg_idle.
8824 rq->max_idle_balance_cost =
8825 max((u64)sysctl_sched_migration_cost, max_cost);
8830 * next_balance will be updated only when there is a need.
8831 * When the cpu is attached to null domain for ex, it will not be
8834 if (likely(update_next_balance)) {
8835 rq->next_balance = next_balance;
8837 #ifdef CONFIG_NO_HZ_COMMON
8839 * If this CPU has been elected to perform the nohz idle
8840 * balance. Other idle CPUs have already rebalanced with
8841 * nohz_idle_balance() and nohz.next_balance has been
8842 * updated accordingly. This CPU is now running the idle load
8843 * balance for itself and we need to update the
8844 * nohz.next_balance accordingly.
8846 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8847 nohz.next_balance = rq->next_balance;
8852 #ifdef CONFIG_NO_HZ_COMMON
8854 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8855 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8857 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8859 int this_cpu = this_rq->cpu;
8862 /* Earliest time when we have to do rebalance again */
8863 unsigned long next_balance = jiffies + 60*HZ;
8864 int update_next_balance = 0;
8866 if (idle != CPU_IDLE ||
8867 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8870 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8871 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8875 * If this cpu gets work to do, stop the load balancing
8876 * work being done for other cpus. Next load
8877 * balancing owner will pick it up.
8882 rq = cpu_rq(balance_cpu);
8885 * If time for next balance is due,
8888 if (time_after_eq(jiffies, rq->next_balance)) {
8891 rq_lock_irq(rq, &rf);
8892 update_rq_clock(rq);
8893 cpu_load_update_idle(rq);
8894 rq_unlock_irq(rq, &rf);
8896 rebalance_domains(rq, CPU_IDLE);
8899 if (time_after(next_balance, rq->next_balance)) {
8900 next_balance = rq->next_balance;
8901 update_next_balance = 1;
8906 * next_balance will be updated only when there is a need.
8907 * When the CPU is attached to null domain for ex, it will not be
8910 if (likely(update_next_balance))
8911 nohz.next_balance = next_balance;
8913 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8917 * Current heuristic for kicking the idle load balancer in the presence
8918 * of an idle cpu in the system.
8919 * - This rq has more than one task.
8920 * - This rq has at least one CFS task and the capacity of the CPU is
8921 * significantly reduced because of RT tasks or IRQs.
8922 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8923 * multiple busy cpu.
8924 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8925 * domain span are idle.
8927 static inline bool nohz_kick_needed(struct rq *rq)
8929 unsigned long now = jiffies;
8930 struct sched_domain_shared *sds;
8931 struct sched_domain *sd;
8932 int nr_busy, i, cpu = rq->cpu;
8935 if (unlikely(rq->idle_balance))
8939 * We may be recently in ticked or tickless idle mode. At the first
8940 * busy tick after returning from idle, we will update the busy stats.
8942 set_cpu_sd_state_busy();
8943 nohz_balance_exit_idle(cpu);
8946 * None are in tickless mode and hence no need for NOHZ idle load
8949 if (likely(!atomic_read(&nohz.nr_cpus)))
8952 if (time_before(now, nohz.next_balance))
8955 if (rq->nr_running >= 2)
8959 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8962 * XXX: write a coherent comment on why we do this.
8963 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8965 nr_busy = atomic_read(&sds->nr_busy_cpus);
8973 sd = rcu_dereference(rq->sd);
8975 if ((rq->cfs.h_nr_running >= 1) &&
8976 check_cpu_capacity(rq, sd)) {
8982 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8984 for_each_cpu(i, sched_domain_span(sd)) {
8986 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
8989 if (sched_asym_prefer(i, cpu)) {
9000 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9004 * run_rebalance_domains is triggered when needed from the scheduler tick.
9005 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9007 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9009 struct rq *this_rq = this_rq();
9010 enum cpu_idle_type idle = this_rq->idle_balance ?
9011 CPU_IDLE : CPU_NOT_IDLE;
9014 * If this cpu has a pending nohz_balance_kick, then do the
9015 * balancing on behalf of the other idle cpus whose ticks are
9016 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9017 * give the idle cpus a chance to load balance. Else we may
9018 * load balance only within the local sched_domain hierarchy
9019 * and abort nohz_idle_balance altogether if we pull some load.
9021 nohz_idle_balance(this_rq, idle);
9022 rebalance_domains(this_rq, idle);
9026 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9028 void trigger_load_balance(struct rq *rq)
9030 /* Don't need to rebalance while attached to NULL domain */
9031 if (unlikely(on_null_domain(rq)))
9034 if (time_after_eq(jiffies, rq->next_balance))
9035 raise_softirq(SCHED_SOFTIRQ);
9036 #ifdef CONFIG_NO_HZ_COMMON
9037 if (nohz_kick_needed(rq))
9038 nohz_balancer_kick();
9042 static void rq_online_fair(struct rq *rq)
9046 update_runtime_enabled(rq);
9049 static void rq_offline_fair(struct rq *rq)
9053 /* Ensure any throttled groups are reachable by pick_next_task */
9054 unthrottle_offline_cfs_rqs(rq);
9057 #endif /* CONFIG_SMP */
9060 * scheduler tick hitting a task of our scheduling class:
9062 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9064 struct cfs_rq *cfs_rq;
9065 struct sched_entity *se = &curr->se;
9067 for_each_sched_entity(se) {
9068 cfs_rq = cfs_rq_of(se);
9069 entity_tick(cfs_rq, se, queued);
9072 if (static_branch_unlikely(&sched_numa_balancing))
9073 task_tick_numa(rq, curr);
9077 * called on fork with the child task as argument from the parent's context
9078 * - child not yet on the tasklist
9079 * - preemption disabled
9081 static void task_fork_fair(struct task_struct *p)
9083 struct cfs_rq *cfs_rq;
9084 struct sched_entity *se = &p->se, *curr;
9085 struct rq *rq = this_rq();
9089 update_rq_clock(rq);
9091 cfs_rq = task_cfs_rq(current);
9092 curr = cfs_rq->curr;
9094 update_curr(cfs_rq);
9095 se->vruntime = curr->vruntime;
9097 place_entity(cfs_rq, se, 1);
9099 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9101 * Upon rescheduling, sched_class::put_prev_task() will place
9102 * 'current' within the tree based on its new key value.
9104 swap(curr->vruntime, se->vruntime);
9108 se->vruntime -= cfs_rq->min_vruntime;
9113 * Priority of the task has changed. Check to see if we preempt
9117 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9119 if (!task_on_rq_queued(p))
9123 * Reschedule if we are currently running on this runqueue and
9124 * our priority decreased, or if we are not currently running on
9125 * this runqueue and our priority is higher than the current's
9127 if (rq->curr == p) {
9128 if (p->prio > oldprio)
9131 check_preempt_curr(rq, p, 0);
9134 static inline bool vruntime_normalized(struct task_struct *p)
9136 struct sched_entity *se = &p->se;
9139 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9140 * the dequeue_entity(.flags=0) will already have normalized the
9147 * When !on_rq, vruntime of the task has usually NOT been normalized.
9148 * But there are some cases where it has already been normalized:
9150 * - A forked child which is waiting for being woken up by
9151 * wake_up_new_task().
9152 * - A task which has been woken up by try_to_wake_up() and
9153 * waiting for actually being woken up by sched_ttwu_pending().
9155 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9161 #ifdef CONFIG_FAIR_GROUP_SCHED
9163 * Propagate the changes of the sched_entity across the tg tree to make it
9164 * visible to the root
9166 static void propagate_entity_cfs_rq(struct sched_entity *se)
9168 struct cfs_rq *cfs_rq;
9170 /* Start to propagate at parent */
9173 for_each_sched_entity(se) {
9174 cfs_rq = cfs_rq_of(se);
9176 if (cfs_rq_throttled(cfs_rq))
9179 update_load_avg(se, UPDATE_TG);
9183 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9186 static void detach_entity_cfs_rq(struct sched_entity *se)
9188 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9190 /* Catch up with the cfs_rq and remove our load when we leave */
9191 update_load_avg(se, 0);
9192 detach_entity_load_avg(cfs_rq, se);
9193 update_tg_load_avg(cfs_rq, false);
9194 propagate_entity_cfs_rq(se);
9197 static void attach_entity_cfs_rq(struct sched_entity *se)
9199 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9201 #ifdef CONFIG_FAIR_GROUP_SCHED
9203 * Since the real-depth could have been changed (only FAIR
9204 * class maintain depth value), reset depth properly.
9206 se->depth = se->parent ? se->parent->depth + 1 : 0;
9209 /* Synchronize entity with its cfs_rq */
9210 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9211 attach_entity_load_avg(cfs_rq, se);
9212 update_tg_load_avg(cfs_rq, false);
9213 propagate_entity_cfs_rq(se);
9216 static void detach_task_cfs_rq(struct task_struct *p)
9218 struct sched_entity *se = &p->se;
9219 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9221 if (!vruntime_normalized(p)) {
9223 * Fix up our vruntime so that the current sleep doesn't
9224 * cause 'unlimited' sleep bonus.
9226 place_entity(cfs_rq, se, 0);
9227 se->vruntime -= cfs_rq->min_vruntime;
9230 detach_entity_cfs_rq(se);
9233 static void attach_task_cfs_rq(struct task_struct *p)
9235 struct sched_entity *se = &p->se;
9236 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9238 attach_entity_cfs_rq(se);
9240 if (!vruntime_normalized(p))
9241 se->vruntime += cfs_rq->min_vruntime;
9244 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9246 detach_task_cfs_rq(p);
9249 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9251 attach_task_cfs_rq(p);
9253 if (task_on_rq_queued(p)) {
9255 * We were most likely switched from sched_rt, so
9256 * kick off the schedule if running, otherwise just see
9257 * if we can still preempt the current task.
9262 check_preempt_curr(rq, p, 0);
9266 /* Account for a task changing its policy or group.
9268 * This routine is mostly called to set cfs_rq->curr field when a task
9269 * migrates between groups/classes.
9271 static void set_curr_task_fair(struct rq *rq)
9273 struct sched_entity *se = &rq->curr->se;
9275 for_each_sched_entity(se) {
9276 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9278 set_next_entity(cfs_rq, se);
9279 /* ensure bandwidth has been allocated on our new cfs_rq */
9280 account_cfs_rq_runtime(cfs_rq, 0);
9284 void init_cfs_rq(struct cfs_rq *cfs_rq)
9286 cfs_rq->tasks_timeline = RB_ROOT;
9287 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9288 #ifndef CONFIG_64BIT
9289 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9292 #ifdef CONFIG_FAIR_GROUP_SCHED
9293 cfs_rq->propagate_avg = 0;
9295 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9296 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9300 #ifdef CONFIG_FAIR_GROUP_SCHED
9301 static void task_set_group_fair(struct task_struct *p)
9303 struct sched_entity *se = &p->se;
9305 set_task_rq(p, task_cpu(p));
9306 se->depth = se->parent ? se->parent->depth + 1 : 0;
9309 static void task_move_group_fair(struct task_struct *p)
9311 detach_task_cfs_rq(p);
9312 set_task_rq(p, task_cpu(p));
9315 /* Tell se's cfs_rq has been changed -- migrated */
9316 p->se.avg.last_update_time = 0;
9318 attach_task_cfs_rq(p);
9321 static void task_change_group_fair(struct task_struct *p, int type)
9324 case TASK_SET_GROUP:
9325 task_set_group_fair(p);
9328 case TASK_MOVE_GROUP:
9329 task_move_group_fair(p);
9334 void free_fair_sched_group(struct task_group *tg)
9338 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9340 for_each_possible_cpu(i) {
9342 kfree(tg->cfs_rq[i]);
9351 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9353 struct sched_entity *se;
9354 struct cfs_rq *cfs_rq;
9357 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9360 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9364 tg->shares = NICE_0_LOAD;
9366 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9368 for_each_possible_cpu(i) {
9369 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9370 GFP_KERNEL, cpu_to_node(i));
9374 se = kzalloc_node(sizeof(struct sched_entity),
9375 GFP_KERNEL, cpu_to_node(i));
9379 init_cfs_rq(cfs_rq);
9380 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9381 init_entity_runnable_average(se);
9392 void online_fair_sched_group(struct task_group *tg)
9394 struct sched_entity *se;
9398 for_each_possible_cpu(i) {
9402 raw_spin_lock_irq(&rq->lock);
9403 update_rq_clock(rq);
9404 attach_entity_cfs_rq(se);
9405 sync_throttle(tg, i);
9406 raw_spin_unlock_irq(&rq->lock);
9410 void unregister_fair_sched_group(struct task_group *tg)
9412 unsigned long flags;
9416 for_each_possible_cpu(cpu) {
9418 remove_entity_load_avg(tg->se[cpu]);
9421 * Only empty task groups can be destroyed; so we can speculatively
9422 * check on_list without danger of it being re-added.
9424 if (!tg->cfs_rq[cpu]->on_list)
9429 raw_spin_lock_irqsave(&rq->lock, flags);
9430 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9431 raw_spin_unlock_irqrestore(&rq->lock, flags);
9435 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9436 struct sched_entity *se, int cpu,
9437 struct sched_entity *parent)
9439 struct rq *rq = cpu_rq(cpu);
9443 init_cfs_rq_runtime(cfs_rq);
9445 tg->cfs_rq[cpu] = cfs_rq;
9448 /* se could be NULL for root_task_group */
9453 se->cfs_rq = &rq->cfs;
9456 se->cfs_rq = parent->my_q;
9457 se->depth = parent->depth + 1;
9461 /* guarantee group entities always have weight */
9462 update_load_set(&se->load, NICE_0_LOAD);
9463 se->parent = parent;
9466 static DEFINE_MUTEX(shares_mutex);
9468 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9473 * We can't change the weight of the root cgroup.
9478 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9480 mutex_lock(&shares_mutex);
9481 if (tg->shares == shares)
9484 tg->shares = shares;
9485 for_each_possible_cpu(i) {
9486 struct rq *rq = cpu_rq(i);
9487 struct sched_entity *se = tg->se[i];
9490 /* Propagate contribution to hierarchy */
9491 rq_lock_irqsave(rq, &rf);
9492 update_rq_clock(rq);
9493 for_each_sched_entity(se) {
9494 update_load_avg(se, UPDATE_TG);
9495 update_cfs_shares(se);
9497 rq_unlock_irqrestore(rq, &rf);
9501 mutex_unlock(&shares_mutex);
9504 #else /* CONFIG_FAIR_GROUP_SCHED */
9506 void free_fair_sched_group(struct task_group *tg) { }
9508 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9513 void online_fair_sched_group(struct task_group *tg) { }
9515 void unregister_fair_sched_group(struct task_group *tg) { }
9517 #endif /* CONFIG_FAIR_GROUP_SCHED */
9520 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9522 struct sched_entity *se = &task->se;
9523 unsigned int rr_interval = 0;
9526 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9529 if (rq->cfs.load.weight)
9530 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9536 * All the scheduling class methods:
9538 const struct sched_class fair_sched_class = {
9539 .next = &idle_sched_class,
9540 .enqueue_task = enqueue_task_fair,
9541 .dequeue_task = dequeue_task_fair,
9542 .yield_task = yield_task_fair,
9543 .yield_to_task = yield_to_task_fair,
9545 .check_preempt_curr = check_preempt_wakeup,
9547 .pick_next_task = pick_next_task_fair,
9548 .put_prev_task = put_prev_task_fair,
9551 .select_task_rq = select_task_rq_fair,
9552 .migrate_task_rq = migrate_task_rq_fair,
9554 .rq_online = rq_online_fair,
9555 .rq_offline = rq_offline_fair,
9557 .task_dead = task_dead_fair,
9558 .set_cpus_allowed = set_cpus_allowed_common,
9561 .set_curr_task = set_curr_task_fair,
9562 .task_tick = task_tick_fair,
9563 .task_fork = task_fork_fair,
9565 .prio_changed = prio_changed_fair,
9566 .switched_from = switched_from_fair,
9567 .switched_to = switched_to_fair,
9569 .get_rr_interval = get_rr_interval_fair,
9571 .update_curr = update_curr_fair,
9573 #ifdef CONFIG_FAIR_GROUP_SCHED
9574 .task_change_group = task_change_group_fair,
9578 #ifdef CONFIG_SCHED_DEBUG
9579 void print_cfs_stats(struct seq_file *m, int cpu)
9581 struct cfs_rq *cfs_rq, *pos;
9584 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9585 print_cfs_rq(m, cpu, cfs_rq);
9589 #ifdef CONFIG_NUMA_BALANCING
9590 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9593 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9595 for_each_online_node(node) {
9596 if (p->numa_faults) {
9597 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9598 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9600 if (p->numa_group) {
9601 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9602 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9604 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9607 #endif /* CONFIG_NUMA_BALANCING */
9608 #endif /* CONFIG_SCHED_DEBUG */
9610 __init void init_sched_fair_class(void)
9613 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9615 #ifdef CONFIG_NO_HZ_COMMON
9616 nohz.next_balance = jiffies;
9617 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);