1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
24 #include <linux/sched/mm.h>
25 #include <linux/sched/topology.h>
27 #include <linux/latencytop.h>
28 #include <linux/cpumask.h>
29 #include <linux/cpuidle.h>
30 #include <linux/slab.h>
31 #include <linux/profile.h>
32 #include <linux/interrupt.h>
33 #include <linux/mempolicy.h>
34 #include <linux/migrate.h>
35 #include <linux/task_work.h>
36 #include <linux/sched/isolation.h>
38 #include <trace/events/sched.h>
43 * Targeted preemption latency for CPU-bound tasks:
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
55 unsigned int sysctl_sched_latency = 6000000ULL;
56 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
59 * The initial- and re-scaling of tunables is configurable
63 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
64 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
65 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
67 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
69 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
72 * Minimal preemption granularity for CPU-bound tasks:
74 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 unsigned int sysctl_sched_min_granularity = 750000ULL;
77 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
80 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
82 static unsigned int sched_nr_latency = 8;
85 * After fork, child runs first. If set to 0 (default) then
86 * parent will (try to) run first.
88 unsigned int sysctl_sched_child_runs_first __read_mostly;
91 * SCHED_OTHER wake-up granularity.
93 * This option delays the preemption effects of decoupled workloads
94 * and reduces their over-scheduling. Synchronous workloads will still
95 * have immediate wakeup/sleep latencies.
97 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
99 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
100 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
102 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
106 * For asym packing, by default the lower numbered cpu has higher priority.
108 int __weak arch_asym_cpu_priority(int cpu)
114 #ifdef CONFIG_CFS_BANDWIDTH
116 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
117 * each time a cfs_rq requests quota.
119 * Note: in the case that the slice exceeds the runtime remaining (either due
120 * to consumption or the quota being specified to be smaller than the slice)
121 * we will always only issue the remaining available time.
123 * (default: 5 msec, units: microseconds)
125 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
129 * The margin used when comparing utilization with CPU capacity:
130 * util * margin < capacity * 1024
134 unsigned int capacity_margin = 1280;
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
172 case SCHED_TUNABLESCALING_LINEAR:
175 case SCHED_TUNABLESCALING_LOG:
177 factor = 1 + ilog2(cpus);
184 static void update_sysctl(void)
186 unsigned int factor = get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity);
191 SET_SYSCTL(sched_latency);
192 SET_SYSCTL(sched_wakeup_granularity);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
208 if (likely(lw->inv_weight))
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
235 u64 fact = scale_load_down(weight);
236 int shift = WMULT_SHIFT;
238 __update_inv_weight(lw);
240 if (unlikely(fact >> 32)) {
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
255 return mul_u64_u32_shr(delta_exec, fact, shift);
259 const struct sched_class fair_sched_class;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct *task_of(struct sched_entity *se)
278 SCHED_WARN_ON(!entity_is_task(se));
279 return container_of(se, struct task_struct, se);
282 /* Walk up scheduling entities hierarchy */
283 #define for_each_sched_entity(se) \
284 for (; se; se = se->parent)
286 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
291 /* runqueue on which this entity is (to be) queued */
292 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
297 /* runqueue "owned" by this group */
298 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
303 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
305 if (!cfs_rq->on_list) {
306 struct rq *rq = rq_of(cfs_rq);
307 int cpu = cpu_of(rq);
309 * Ensure we either appear before our parent (if already
310 * enqueued) or force our parent to appear after us when it is
311 * enqueued. The fact that we always enqueue bottom-up
312 * reduces this to two cases and a special case for the root
313 * cfs_rq. Furthermore, it also means that we will always reset
314 * tmp_alone_branch either when the branch is connected
315 * to a tree or when we reach the beg of the tree
317 if (cfs_rq->tg->parent &&
318 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
320 * If parent is already on the list, we add the child
321 * just before. Thanks to circular linked property of
322 * the list, this means to put the child at the tail
323 * of the list that starts by parent.
325 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
326 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
328 * The branch is now connected to its tree so we can
329 * reset tmp_alone_branch to the beginning of the
332 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
333 } else if (!cfs_rq->tg->parent) {
335 * cfs rq without parent should be put
336 * at the tail of the list.
338 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
339 &rq->leaf_cfs_rq_list);
341 * We have reach the beg of a tree so we can reset
342 * tmp_alone_branch to the beginning of the list.
344 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
347 * The parent has not already been added so we want to
348 * make sure that it will be put after us.
349 * tmp_alone_branch points to the beg of the branch
350 * where we will add parent.
352 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
353 rq->tmp_alone_branch);
355 * update tmp_alone_branch to points to the new beg
358 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
365 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
367 if (cfs_rq->on_list) {
368 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
373 /* Iterate thr' all leaf cfs_rq's on a runqueue */
374 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
375 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
378 /* Do the two (enqueued) entities belong to the same group ? */
379 static inline struct cfs_rq *
380 is_same_group(struct sched_entity *se, struct sched_entity *pse)
382 if (se->cfs_rq == pse->cfs_rq)
388 static inline struct sched_entity *parent_entity(struct sched_entity *se)
394 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
396 int se_depth, pse_depth;
399 * preemption test can be made between sibling entities who are in the
400 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
401 * both tasks until we find their ancestors who are siblings of common
405 /* First walk up until both entities are at same depth */
406 se_depth = (*se)->depth;
407 pse_depth = (*pse)->depth;
409 while (se_depth > pse_depth) {
411 *se = parent_entity(*se);
414 while (pse_depth > se_depth) {
416 *pse = parent_entity(*pse);
419 while (!is_same_group(*se, *pse)) {
420 *se = parent_entity(*se);
421 *pse = parent_entity(*pse);
425 #else /* !CONFIG_FAIR_GROUP_SCHED */
427 static inline struct task_struct *task_of(struct sched_entity *se)
429 return container_of(se, struct task_struct, se);
432 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
434 return container_of(cfs_rq, struct rq, cfs);
437 #define entity_is_task(se) 1
439 #define for_each_sched_entity(se) \
440 for (; se; se = NULL)
442 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
444 return &task_rq(p)->cfs;
447 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
449 struct task_struct *p = task_of(se);
450 struct rq *rq = task_rq(p);
455 /* runqueue "owned" by this group */
456 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
461 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
465 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
469 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
470 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
472 static inline struct sched_entity *parent_entity(struct sched_entity *se)
478 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
482 #endif /* CONFIG_FAIR_GROUP_SCHED */
484 static __always_inline
485 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
487 /**************************************************************
488 * Scheduling class tree data structure manipulation methods:
491 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
493 s64 delta = (s64)(vruntime - max_vruntime);
495 max_vruntime = vruntime;
500 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
502 s64 delta = (s64)(vruntime - min_vruntime);
504 min_vruntime = vruntime;
509 static inline int entity_before(struct sched_entity *a,
510 struct sched_entity *b)
512 return (s64)(a->vruntime - b->vruntime) < 0;
515 static void update_min_vruntime(struct cfs_rq *cfs_rq)
517 struct sched_entity *curr = cfs_rq->curr;
518 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
520 u64 vruntime = cfs_rq->min_vruntime;
524 vruntime = curr->vruntime;
529 if (leftmost) { /* non-empty tree */
530 struct sched_entity *se;
531 se = rb_entry(leftmost, struct sched_entity, run_node);
534 vruntime = se->vruntime;
536 vruntime = min_vruntime(vruntime, se->vruntime);
539 /* ensure we never gain time by being placed backwards. */
540 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
543 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
548 * Enqueue an entity into the rb-tree:
550 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
553 struct rb_node *parent = NULL;
554 struct sched_entity *entry;
555 bool leftmost = true;
558 * Find the right place in the rbtree:
562 entry = rb_entry(parent, struct sched_entity, run_node);
564 * We dont care about collisions. Nodes with
565 * the same key stay together.
567 if (entity_before(se, entry)) {
568 link = &parent->rb_left;
570 link = &parent->rb_right;
575 rb_link_node(&se->run_node, parent, link);
576 rb_insert_color_cached(&se->run_node,
577 &cfs_rq->tasks_timeline, leftmost);
580 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
582 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
585 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
587 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
592 return rb_entry(left, struct sched_entity, run_node);
595 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
597 struct rb_node *next = rb_next(&se->run_node);
602 return rb_entry(next, struct sched_entity, run_node);
605 #ifdef CONFIG_SCHED_DEBUG
606 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
608 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
613 return rb_entry(last, struct sched_entity, run_node);
616 /**************************************************************
617 * Scheduling class statistics methods:
620 int sched_proc_update_handler(struct ctl_table *table, int write,
621 void __user *buffer, size_t *lenp,
624 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
625 unsigned int factor = get_update_sysctl_factor();
630 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
631 sysctl_sched_min_granularity);
633 #define WRT_SYSCTL(name) \
634 (normalized_sysctl_##name = sysctl_##name / (factor))
635 WRT_SYSCTL(sched_min_granularity);
636 WRT_SYSCTL(sched_latency);
637 WRT_SYSCTL(sched_wakeup_granularity);
647 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
649 if (unlikely(se->load.weight != NICE_0_LOAD))
650 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
656 * The idea is to set a period in which each task runs once.
658 * When there are too many tasks (sched_nr_latency) we have to stretch
659 * this period because otherwise the slices get too small.
661 * p = (nr <= nl) ? l : l*nr/nl
663 static u64 __sched_period(unsigned long nr_running)
665 if (unlikely(nr_running > sched_nr_latency))
666 return nr_running * sysctl_sched_min_granularity;
668 return sysctl_sched_latency;
672 * We calculate the wall-time slice from the period by taking a part
673 * proportional to the weight.
677 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
679 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
681 for_each_sched_entity(se) {
682 struct load_weight *load;
683 struct load_weight lw;
685 cfs_rq = cfs_rq_of(se);
686 load = &cfs_rq->load;
688 if (unlikely(!se->on_rq)) {
691 update_load_add(&lw, se->load.weight);
694 slice = __calc_delta(slice, se->load.weight, load);
700 * We calculate the vruntime slice of a to-be-inserted task.
704 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
706 return calc_delta_fair(sched_slice(cfs_rq, se), se);
711 #include "sched-pelt.h"
713 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
714 static unsigned long task_h_load(struct task_struct *p);
716 /* Give new sched_entity start runnable values to heavy its load in infant time */
717 void init_entity_runnable_average(struct sched_entity *se)
719 struct sched_avg *sa = &se->avg;
721 memset(sa, 0, sizeof(*sa));
724 * Tasks are intialized with full load to be seen as heavy tasks until
725 * they get a chance to stabilize to their real load level.
726 * Group entities are intialized with zero load to reflect the fact that
727 * nothing has been attached to the task group yet.
729 if (entity_is_task(se))
730 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
732 se->runnable_weight = se->load.weight;
734 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
737 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
738 static void attach_entity_cfs_rq(struct sched_entity *se);
741 * With new tasks being created, their initial util_avgs are extrapolated
742 * based on the cfs_rq's current util_avg:
744 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
746 * However, in many cases, the above util_avg does not give a desired
747 * value. Moreover, the sum of the util_avgs may be divergent, such
748 * as when the series is a harmonic series.
750 * To solve this problem, we also cap the util_avg of successive tasks to
751 * only 1/2 of the left utilization budget:
753 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
755 * where n denotes the nth task.
757 * For example, a simplest series from the beginning would be like:
759 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
760 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
762 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
763 * if util_avg > util_avg_cap.
765 void post_init_entity_util_avg(struct sched_entity *se)
767 struct cfs_rq *cfs_rq = cfs_rq_of(se);
768 struct sched_avg *sa = &se->avg;
769 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
772 if (cfs_rq->avg.util_avg != 0) {
773 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
774 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
776 if (sa->util_avg > cap)
783 if (entity_is_task(se)) {
784 struct task_struct *p = task_of(se);
785 if (p->sched_class != &fair_sched_class) {
787 * For !fair tasks do:
789 update_cfs_rq_load_avg(now, cfs_rq);
790 attach_entity_load_avg(cfs_rq, se);
791 switched_from_fair(rq, p);
793 * such that the next switched_to_fair() has the
796 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
801 attach_entity_cfs_rq(se);
804 #else /* !CONFIG_SMP */
805 void init_entity_runnable_average(struct sched_entity *se)
808 void post_init_entity_util_avg(struct sched_entity *se)
811 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
814 #endif /* CONFIG_SMP */
817 * Update the current task's runtime statistics.
819 static void update_curr(struct cfs_rq *cfs_rq)
821 struct sched_entity *curr = cfs_rq->curr;
822 u64 now = rq_clock_task(rq_of(cfs_rq));
828 delta_exec = now - curr->exec_start;
829 if (unlikely((s64)delta_exec <= 0))
832 curr->exec_start = now;
834 schedstat_set(curr->statistics.exec_max,
835 max(delta_exec, curr->statistics.exec_max));
837 curr->sum_exec_runtime += delta_exec;
838 schedstat_add(cfs_rq->exec_clock, delta_exec);
840 curr->vruntime += calc_delta_fair(delta_exec, curr);
841 update_min_vruntime(cfs_rq);
843 if (entity_is_task(curr)) {
844 struct task_struct *curtask = task_of(curr);
846 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
847 cgroup_account_cputime(curtask, delta_exec);
848 account_group_exec_runtime(curtask, delta_exec);
851 account_cfs_rq_runtime(cfs_rq, delta_exec);
854 static void update_curr_fair(struct rq *rq)
856 update_curr(cfs_rq_of(&rq->curr->se));
860 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
862 u64 wait_start, prev_wait_start;
864 if (!schedstat_enabled())
867 wait_start = rq_clock(rq_of(cfs_rq));
868 prev_wait_start = schedstat_val(se->statistics.wait_start);
870 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
871 likely(wait_start > prev_wait_start))
872 wait_start -= prev_wait_start;
874 __schedstat_set(se->statistics.wait_start, wait_start);
878 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
880 struct task_struct *p;
883 if (!schedstat_enabled())
886 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
888 if (entity_is_task(se)) {
890 if (task_on_rq_migrating(p)) {
892 * Preserve migrating task's wait time so wait_start
893 * time stamp can be adjusted to accumulate wait time
894 * prior to migration.
896 __schedstat_set(se->statistics.wait_start, delta);
899 trace_sched_stat_wait(p, delta);
902 __schedstat_set(se->statistics.wait_max,
903 max(schedstat_val(se->statistics.wait_max), delta));
904 __schedstat_inc(se->statistics.wait_count);
905 __schedstat_add(se->statistics.wait_sum, delta);
906 __schedstat_set(se->statistics.wait_start, 0);
910 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
912 struct task_struct *tsk = NULL;
913 u64 sleep_start, block_start;
915 if (!schedstat_enabled())
918 sleep_start = schedstat_val(se->statistics.sleep_start);
919 block_start = schedstat_val(se->statistics.block_start);
921 if (entity_is_task(se))
925 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
930 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
931 __schedstat_set(se->statistics.sleep_max, delta);
933 __schedstat_set(se->statistics.sleep_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 account_scheduler_latency(tsk, delta >> 10, 1);
938 trace_sched_stat_sleep(tsk, delta);
942 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
947 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
948 __schedstat_set(se->statistics.block_max, delta);
950 __schedstat_set(se->statistics.block_start, 0);
951 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
954 if (tsk->in_iowait) {
955 __schedstat_add(se->statistics.iowait_sum, delta);
956 __schedstat_inc(se->statistics.iowait_count);
957 trace_sched_stat_iowait(tsk, delta);
960 trace_sched_stat_blocked(tsk, delta);
963 * Blocking time is in units of nanosecs, so shift by
964 * 20 to get a milliseconds-range estimation of the
965 * amount of time that the task spent sleeping:
967 if (unlikely(prof_on == SLEEP_PROFILING)) {
968 profile_hits(SLEEP_PROFILING,
969 (void *)get_wchan(tsk),
972 account_scheduler_latency(tsk, delta >> 10, 0);
978 * Task is being enqueued - update stats:
981 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
983 if (!schedstat_enabled())
987 * Are we enqueueing a waiting task? (for current tasks
988 * a dequeue/enqueue event is a NOP)
990 if (se != cfs_rq->curr)
991 update_stats_wait_start(cfs_rq, se);
993 if (flags & ENQUEUE_WAKEUP)
994 update_stats_enqueue_sleeper(cfs_rq, se);
998 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1001 if (!schedstat_enabled())
1005 * Mark the end of the wait period if dequeueing a
1008 if (se != cfs_rq->curr)
1009 update_stats_wait_end(cfs_rq, se);
1011 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1012 struct task_struct *tsk = task_of(se);
1014 if (tsk->state & TASK_INTERRUPTIBLE)
1015 __schedstat_set(se->statistics.sleep_start,
1016 rq_clock(rq_of(cfs_rq)));
1017 if (tsk->state & TASK_UNINTERRUPTIBLE)
1018 __schedstat_set(se->statistics.block_start,
1019 rq_clock(rq_of(cfs_rq)));
1024 * We are picking a new current task - update its stats:
1027 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1030 * We are starting a new run period:
1032 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1035 /**************************************************
1036 * Scheduling class queueing methods:
1039 #ifdef CONFIG_NUMA_BALANCING
1041 * Approximate time to scan a full NUMA task in ms. The task scan period is
1042 * calculated based on the tasks virtual memory size and
1043 * numa_balancing_scan_size.
1045 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1046 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1048 /* Portion of address space to scan in MB */
1049 unsigned int sysctl_numa_balancing_scan_size = 256;
1051 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1052 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1057 spinlock_t lock; /* nr_tasks, tasks */
1062 struct rcu_head rcu;
1063 unsigned long total_faults;
1064 unsigned long max_faults_cpu;
1066 * Faults_cpu is used to decide whether memory should move
1067 * towards the CPU. As a consequence, these stats are weighted
1068 * more by CPU use than by memory faults.
1070 unsigned long *faults_cpu;
1071 unsigned long faults[0];
1074 static inline unsigned long group_faults_priv(struct numa_group *ng);
1075 static inline unsigned long group_faults_shared(struct numa_group *ng);
1077 static unsigned int task_nr_scan_windows(struct task_struct *p)
1079 unsigned long rss = 0;
1080 unsigned long nr_scan_pages;
1083 * Calculations based on RSS as non-present and empty pages are skipped
1084 * by the PTE scanner and NUMA hinting faults should be trapped based
1087 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1088 rss = get_mm_rss(p->mm);
1090 rss = nr_scan_pages;
1092 rss = round_up(rss, nr_scan_pages);
1093 return rss / nr_scan_pages;
1096 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1097 #define MAX_SCAN_WINDOW 2560
1099 static unsigned int task_scan_min(struct task_struct *p)
1101 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1102 unsigned int scan, floor;
1103 unsigned int windows = 1;
1105 if (scan_size < MAX_SCAN_WINDOW)
1106 windows = MAX_SCAN_WINDOW / scan_size;
1107 floor = 1000 / windows;
1109 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1110 return max_t(unsigned int, floor, scan);
1113 static unsigned int task_scan_start(struct task_struct *p)
1115 unsigned long smin = task_scan_min(p);
1116 unsigned long period = smin;
1118 /* Scale the maximum scan period with the amount of shared memory. */
1119 if (p->numa_group) {
1120 struct numa_group *ng = p->numa_group;
1121 unsigned long shared = group_faults_shared(ng);
1122 unsigned long private = group_faults_priv(ng);
1124 period *= atomic_read(&ng->refcount);
1125 period *= shared + 1;
1126 period /= private + shared + 1;
1129 return max(smin, period);
1132 static unsigned int task_scan_max(struct task_struct *p)
1134 unsigned long smin = task_scan_min(p);
1137 /* Watch for min being lower than max due to floor calculations */
1138 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1140 /* Scale the maximum scan period with the amount of shared memory. */
1141 if (p->numa_group) {
1142 struct numa_group *ng = p->numa_group;
1143 unsigned long shared = group_faults_shared(ng);
1144 unsigned long private = group_faults_priv(ng);
1145 unsigned long period = smax;
1147 period *= atomic_read(&ng->refcount);
1148 period *= shared + 1;
1149 period /= private + shared + 1;
1151 smax = max(smax, period);
1154 return max(smin, smax);
1157 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1159 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1160 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1163 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1165 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1166 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1169 /* Shared or private faults. */
1170 #define NR_NUMA_HINT_FAULT_TYPES 2
1172 /* Memory and CPU locality */
1173 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1175 /* Averaged statistics, and temporary buffers. */
1176 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1178 pid_t task_numa_group_id(struct task_struct *p)
1180 return p->numa_group ? p->numa_group->gid : 0;
1184 * The averaged statistics, shared & private, memory & cpu,
1185 * occupy the first half of the array. The second half of the
1186 * array is for current counters, which are averaged into the
1187 * first set by task_numa_placement.
1189 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1191 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1194 static inline unsigned long task_faults(struct task_struct *p, int nid)
1196 if (!p->numa_faults)
1199 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1200 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1203 static inline unsigned long group_faults(struct task_struct *p, int nid)
1208 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1209 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1212 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1214 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1215 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1218 static inline unsigned long group_faults_priv(struct numa_group *ng)
1220 unsigned long faults = 0;
1223 for_each_online_node(node) {
1224 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1230 static inline unsigned long group_faults_shared(struct numa_group *ng)
1232 unsigned long faults = 0;
1235 for_each_online_node(node) {
1236 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1243 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1244 * considered part of a numa group's pseudo-interleaving set. Migrations
1245 * between these nodes are slowed down, to allow things to settle down.
1247 #define ACTIVE_NODE_FRACTION 3
1249 static bool numa_is_active_node(int nid, struct numa_group *ng)
1251 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1254 /* Handle placement on systems where not all nodes are directly connected. */
1255 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1256 int maxdist, bool task)
1258 unsigned long score = 0;
1262 * All nodes are directly connected, and the same distance
1263 * from each other. No need for fancy placement algorithms.
1265 if (sched_numa_topology_type == NUMA_DIRECT)
1269 * This code is called for each node, introducing N^2 complexity,
1270 * which should be ok given the number of nodes rarely exceeds 8.
1272 for_each_online_node(node) {
1273 unsigned long faults;
1274 int dist = node_distance(nid, node);
1277 * The furthest away nodes in the system are not interesting
1278 * for placement; nid was already counted.
1280 if (dist == sched_max_numa_distance || node == nid)
1284 * On systems with a backplane NUMA topology, compare groups
1285 * of nodes, and move tasks towards the group with the most
1286 * memory accesses. When comparing two nodes at distance
1287 * "hoplimit", only nodes closer by than "hoplimit" are part
1288 * of each group. Skip other nodes.
1290 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1294 /* Add up the faults from nearby nodes. */
1296 faults = task_faults(p, node);
1298 faults = group_faults(p, node);
1301 * On systems with a glueless mesh NUMA topology, there are
1302 * no fixed "groups of nodes". Instead, nodes that are not
1303 * directly connected bounce traffic through intermediate
1304 * nodes; a numa_group can occupy any set of nodes.
1305 * The further away a node is, the less the faults count.
1306 * This seems to result in good task placement.
1308 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1309 faults *= (sched_max_numa_distance - dist);
1310 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1320 * These return the fraction of accesses done by a particular task, or
1321 * task group, on a particular numa node. The group weight is given a
1322 * larger multiplier, in order to group tasks together that are almost
1323 * evenly spread out between numa nodes.
1325 static inline unsigned long task_weight(struct task_struct *p, int nid,
1328 unsigned long faults, total_faults;
1330 if (!p->numa_faults)
1333 total_faults = p->total_numa_faults;
1338 faults = task_faults(p, nid);
1339 faults += score_nearby_nodes(p, nid, dist, true);
1341 return 1000 * faults / total_faults;
1344 static inline unsigned long group_weight(struct task_struct *p, int nid,
1347 unsigned long faults, total_faults;
1352 total_faults = p->numa_group->total_faults;
1357 faults = group_faults(p, nid);
1358 faults += score_nearby_nodes(p, nid, dist, false);
1360 return 1000 * faults / total_faults;
1363 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1364 int src_nid, int dst_cpu)
1366 struct numa_group *ng = p->numa_group;
1367 int dst_nid = cpu_to_node(dst_cpu);
1368 int last_cpupid, this_cpupid;
1370 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1373 * Multi-stage node selection is used in conjunction with a periodic
1374 * migration fault to build a temporal task<->page relation. By using
1375 * a two-stage filter we remove short/unlikely relations.
1377 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1378 * a task's usage of a particular page (n_p) per total usage of this
1379 * page (n_t) (in a given time-span) to a probability.
1381 * Our periodic faults will sample this probability and getting the
1382 * same result twice in a row, given these samples are fully
1383 * independent, is then given by P(n)^2, provided our sample period
1384 * is sufficiently short compared to the usage pattern.
1386 * This quadric squishes small probabilities, making it less likely we
1387 * act on an unlikely task<->page relation.
1389 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1390 if (!cpupid_pid_unset(last_cpupid) &&
1391 cpupid_to_nid(last_cpupid) != dst_nid)
1394 /* Always allow migrate on private faults */
1395 if (cpupid_match_pid(p, last_cpupid))
1398 /* A shared fault, but p->numa_group has not been set up yet. */
1403 * Destination node is much more heavily used than the source
1404 * node? Allow migration.
1406 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1407 ACTIVE_NODE_FRACTION)
1411 * Distribute memory according to CPU & memory use on each node,
1412 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1414 * faults_cpu(dst) 3 faults_cpu(src)
1415 * --------------- * - > ---------------
1416 * faults_mem(dst) 4 faults_mem(src)
1418 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1419 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1422 static unsigned long weighted_cpuload(struct rq *rq);
1423 static unsigned long source_load(int cpu, int type);
1424 static unsigned long target_load(int cpu, int type);
1425 static unsigned long capacity_of(int cpu);
1427 /* Cached statistics for all CPUs within a node */
1429 unsigned long nr_running;
1432 /* Total compute capacity of CPUs on a node */
1433 unsigned long compute_capacity;
1435 /* Approximate capacity in terms of runnable tasks on a node */
1436 unsigned long task_capacity;
1437 int has_free_capacity;
1441 * XXX borrowed from update_sg_lb_stats
1443 static void update_numa_stats(struct numa_stats *ns, int nid)
1445 int smt, cpu, cpus = 0;
1446 unsigned long capacity;
1448 memset(ns, 0, sizeof(*ns));
1449 for_each_cpu(cpu, cpumask_of_node(nid)) {
1450 struct rq *rq = cpu_rq(cpu);
1452 ns->nr_running += rq->nr_running;
1453 ns->load += weighted_cpuload(rq);
1454 ns->compute_capacity += capacity_of(cpu);
1460 * If we raced with hotplug and there are no CPUs left in our mask
1461 * the @ns structure is NULL'ed and task_numa_compare() will
1462 * not find this node attractive.
1464 * We'll either bail at !has_free_capacity, or we'll detect a huge
1465 * imbalance and bail there.
1470 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1471 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1472 capacity = cpus / smt; /* cores */
1474 ns->task_capacity = min_t(unsigned, capacity,
1475 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1476 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1479 struct task_numa_env {
1480 struct task_struct *p;
1482 int src_cpu, src_nid;
1483 int dst_cpu, dst_nid;
1485 struct numa_stats src_stats, dst_stats;
1490 struct task_struct *best_task;
1495 static void task_numa_assign(struct task_numa_env *env,
1496 struct task_struct *p, long imp)
1499 put_task_struct(env->best_task);
1504 env->best_imp = imp;
1505 env->best_cpu = env->dst_cpu;
1508 static bool load_too_imbalanced(long src_load, long dst_load,
1509 struct task_numa_env *env)
1512 long orig_src_load, orig_dst_load;
1513 long src_capacity, dst_capacity;
1516 * The load is corrected for the CPU capacity available on each node.
1519 * ------------ vs ---------
1520 * src_capacity dst_capacity
1522 src_capacity = env->src_stats.compute_capacity;
1523 dst_capacity = env->dst_stats.compute_capacity;
1525 /* We care about the slope of the imbalance, not the direction. */
1526 if (dst_load < src_load)
1527 swap(dst_load, src_load);
1529 /* Is the difference below the threshold? */
1530 imb = dst_load * src_capacity * 100 -
1531 src_load * dst_capacity * env->imbalance_pct;
1536 * The imbalance is above the allowed threshold.
1537 * Compare it with the old imbalance.
1539 orig_src_load = env->src_stats.load;
1540 orig_dst_load = env->dst_stats.load;
1542 if (orig_dst_load < orig_src_load)
1543 swap(orig_dst_load, orig_src_load);
1545 old_imb = orig_dst_load * src_capacity * 100 -
1546 orig_src_load * dst_capacity * env->imbalance_pct;
1548 /* Would this change make things worse? */
1549 return (imb > old_imb);
1553 * This checks if the overall compute and NUMA accesses of the system would
1554 * be improved if the source tasks was migrated to the target dst_cpu taking
1555 * into account that it might be best if task running on the dst_cpu should
1556 * be exchanged with the source task
1558 static void task_numa_compare(struct task_numa_env *env,
1559 long taskimp, long groupimp)
1561 struct rq *src_rq = cpu_rq(env->src_cpu);
1562 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1563 struct task_struct *cur;
1564 long src_load, dst_load;
1566 long imp = env->p->numa_group ? groupimp : taskimp;
1568 int dist = env->dist;
1571 cur = task_rcu_dereference(&dst_rq->curr);
1572 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1576 * Because we have preemption enabled we can get migrated around and
1577 * end try selecting ourselves (current == env->p) as a swap candidate.
1583 * "imp" is the fault differential for the source task between the
1584 * source and destination node. Calculate the total differential for
1585 * the source task and potential destination task. The more negative
1586 * the value is, the more rmeote accesses that would be expected to
1587 * be incurred if the tasks were swapped.
1590 /* Skip this swap candidate if cannot move to the source cpu */
1591 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1595 * If dst and source tasks are in the same NUMA group, or not
1596 * in any group then look only at task weights.
1598 if (cur->numa_group == env->p->numa_group) {
1599 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1600 task_weight(cur, env->dst_nid, dist);
1602 * Add some hysteresis to prevent swapping the
1603 * tasks within a group over tiny differences.
1605 if (cur->numa_group)
1609 * Compare the group weights. If a task is all by
1610 * itself (not part of a group), use the task weight
1613 if (cur->numa_group)
1614 imp += group_weight(cur, env->src_nid, dist) -
1615 group_weight(cur, env->dst_nid, dist);
1617 imp += task_weight(cur, env->src_nid, dist) -
1618 task_weight(cur, env->dst_nid, dist);
1622 if (imp <= env->best_imp && moveimp <= env->best_imp)
1626 /* Is there capacity at our destination? */
1627 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1628 !env->dst_stats.has_free_capacity)
1634 /* Balance doesn't matter much if we're running a task per cpu */
1635 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1636 dst_rq->nr_running == 1)
1640 * In the overloaded case, try and keep the load balanced.
1643 load = task_h_load(env->p);
1644 dst_load = env->dst_stats.load + load;
1645 src_load = env->src_stats.load - load;
1647 if (moveimp > imp && moveimp > env->best_imp) {
1649 * If the improvement from just moving env->p direction is
1650 * better than swapping tasks around, check if a move is
1651 * possible. Store a slightly smaller score than moveimp,
1652 * so an actually idle CPU will win.
1654 if (!load_too_imbalanced(src_load, dst_load, env)) {
1661 if (imp <= env->best_imp)
1665 load = task_h_load(cur);
1670 if (load_too_imbalanced(src_load, dst_load, env))
1674 * One idle CPU per node is evaluated for a task numa move.
1675 * Call select_idle_sibling to maybe find a better one.
1679 * select_idle_siblings() uses an per-cpu cpumask that
1680 * can be used from IRQ context.
1682 local_irq_disable();
1683 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1689 task_numa_assign(env, cur, imp);
1694 static void task_numa_find_cpu(struct task_numa_env *env,
1695 long taskimp, long groupimp)
1699 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1700 /* Skip this CPU if the source task cannot migrate */
1701 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1705 task_numa_compare(env, taskimp, groupimp);
1709 /* Only move tasks to a NUMA node less busy than the current node. */
1710 static bool numa_has_capacity(struct task_numa_env *env)
1712 struct numa_stats *src = &env->src_stats;
1713 struct numa_stats *dst = &env->dst_stats;
1715 if (src->has_free_capacity && !dst->has_free_capacity)
1719 * Only consider a task move if the source has a higher load
1720 * than the destination, corrected for CPU capacity on each node.
1722 * src->load dst->load
1723 * --------------------- vs ---------------------
1724 * src->compute_capacity dst->compute_capacity
1726 if (src->load * dst->compute_capacity * env->imbalance_pct >
1728 dst->load * src->compute_capacity * 100)
1734 static int task_numa_migrate(struct task_struct *p)
1736 struct task_numa_env env = {
1739 .src_cpu = task_cpu(p),
1740 .src_nid = task_node(p),
1742 .imbalance_pct = 112,
1748 struct sched_domain *sd;
1749 unsigned long taskweight, groupweight;
1751 long taskimp, groupimp;
1754 * Pick the lowest SD_NUMA domain, as that would have the smallest
1755 * imbalance and would be the first to start moving tasks about.
1757 * And we want to avoid any moving of tasks about, as that would create
1758 * random movement of tasks -- counter the numa conditions we're trying
1762 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1764 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1768 * Cpusets can break the scheduler domain tree into smaller
1769 * balance domains, some of which do not cross NUMA boundaries.
1770 * Tasks that are "trapped" in such domains cannot be migrated
1771 * elsewhere, so there is no point in (re)trying.
1773 if (unlikely(!sd)) {
1774 p->numa_preferred_nid = task_node(p);
1778 env.dst_nid = p->numa_preferred_nid;
1779 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1780 taskweight = task_weight(p, env.src_nid, dist);
1781 groupweight = group_weight(p, env.src_nid, dist);
1782 update_numa_stats(&env.src_stats, env.src_nid);
1783 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1784 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1785 update_numa_stats(&env.dst_stats, env.dst_nid);
1787 /* Try to find a spot on the preferred nid. */
1788 if (numa_has_capacity(&env))
1789 task_numa_find_cpu(&env, taskimp, groupimp);
1792 * Look at other nodes in these cases:
1793 * - there is no space available on the preferred_nid
1794 * - the task is part of a numa_group that is interleaved across
1795 * multiple NUMA nodes; in order to better consolidate the group,
1796 * we need to check other locations.
1798 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1799 for_each_online_node(nid) {
1800 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1803 dist = node_distance(env.src_nid, env.dst_nid);
1804 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1806 taskweight = task_weight(p, env.src_nid, dist);
1807 groupweight = group_weight(p, env.src_nid, dist);
1810 /* Only consider nodes where both task and groups benefit */
1811 taskimp = task_weight(p, nid, dist) - taskweight;
1812 groupimp = group_weight(p, nid, dist) - groupweight;
1813 if (taskimp < 0 && groupimp < 0)
1818 update_numa_stats(&env.dst_stats, env.dst_nid);
1819 if (numa_has_capacity(&env))
1820 task_numa_find_cpu(&env, taskimp, groupimp);
1825 * If the task is part of a workload that spans multiple NUMA nodes,
1826 * and is migrating into one of the workload's active nodes, remember
1827 * this node as the task's preferred numa node, so the workload can
1829 * A task that migrated to a second choice node will be better off
1830 * trying for a better one later. Do not set the preferred node here.
1832 if (p->numa_group) {
1833 struct numa_group *ng = p->numa_group;
1835 if (env.best_cpu == -1)
1840 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1841 sched_setnuma(p, env.dst_nid);
1844 /* No better CPU than the current one was found. */
1845 if (env.best_cpu == -1)
1849 * Reset the scan period if the task is being rescheduled on an
1850 * alternative node to recheck if the tasks is now properly placed.
1852 p->numa_scan_period = task_scan_start(p);
1854 if (env.best_task == NULL) {
1855 ret = migrate_task_to(p, env.best_cpu);
1857 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1861 ret = migrate_swap(p, env.best_task);
1863 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864 put_task_struct(env.best_task);
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
1871 unsigned long interval = HZ;
1873 /* This task has no NUMA fault statistics yet */
1874 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1877 /* Periodically retry migrating the task to the preferred node */
1878 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879 p->numa_migrate_retry = jiffies + interval;
1881 /* Success if task is already running on preferred CPU */
1882 if (task_node(p) == p->numa_preferred_nid)
1885 /* Otherwise, try migrate to a CPU on the preferred node */
1886 task_numa_migrate(p);
1890 * Find out how many nodes on the workload is actively running on. Do this by
1891 * tracking the nodes from which NUMA hinting faults are triggered. This can
1892 * be different from the set of nodes where the workload's memory is currently
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1897 unsigned long faults, max_faults = 0;
1898 int nid, active_nodes = 0;
1900 for_each_online_node(nid) {
1901 faults = group_faults_cpu(numa_group, nid);
1902 if (faults > max_faults)
1903 max_faults = faults;
1906 for_each_online_node(nid) {
1907 faults = group_faults_cpu(numa_group, nid);
1908 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
1917 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918 * increments. The more local the fault statistics are, the higher the scan
1919 * period will be for the next scan window. If local/(local+remote) ratio is
1920 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921 * the scan period will decrease. Aim for 70% local accesses.
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
1927 * Increase the scan period (slow down scanning) if the majority of
1928 * our memory is already on our local node, or if the majority of
1929 * the page accesses are shared with other processes.
1930 * Otherwise, decrease the scan period.
1932 static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
1935 unsigned int period_slot;
1936 int lr_ratio, ps_ratio;
1939 unsigned long remote = p->numa_faults_locality[0];
1940 unsigned long local = p->numa_faults_locality[1];
1943 * If there were no record hinting faults then either the task is
1944 * completely idle or all activity is areas that are not of interest
1945 * to automatic numa balancing. Related to that, if there were failed
1946 * migration then it implies we are migrating too quickly or the local
1947 * node is overloaded. In either case, scan slower
1949 if (local + shared == 0 || p->numa_faults_locality[2]) {
1950 p->numa_scan_period = min(p->numa_scan_period_max,
1951 p->numa_scan_period << 1);
1953 p->mm->numa_next_scan = jiffies +
1954 msecs_to_jiffies(p->numa_scan_period);
1960 * Prepare to scale scan period relative to the current period.
1961 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1962 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1965 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1969 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
1974 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1977 diff = slot * period_slot;
1978 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1980 * Most memory accesses are shared with other tasks.
1981 * There is no point in continuing fast NUMA scanning,
1982 * since other tasks may just move the memory elsewhere.
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1987 diff = slot * period_slot;
1990 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991 * yet they are not on the local NUMA node. Speed up
1992 * NUMA scanning to get the memory moved over.
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1998 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999 task_scan_min(p), task_scan_max(p));
2000 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2004 * Get the fraction of time the task has been running since the last
2005 * NUMA placement cycle. The scheduler keeps similar statistics, but
2006 * decays those on a 32ms period, which is orders of magnitude off
2007 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008 * stats only if the task is so new there are no NUMA statistics yet.
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2012 u64 runtime, delta, now;
2013 /* Use the start of this time slice to avoid calculations. */
2014 now = p->se.exec_start;
2015 runtime = p->se.sum_exec_runtime;
2017 if (p->last_task_numa_placement) {
2018 delta = runtime - p->last_sum_exec_runtime;
2019 *period = now - p->last_task_numa_placement;
2021 delta = p->se.avg.load_sum;
2022 *period = LOAD_AVG_MAX;
2025 p->last_sum_exec_runtime = runtime;
2026 p->last_task_numa_placement = now;
2032 * Determine the preferred nid for a task in a numa_group. This needs to
2033 * be done in a way that produces consistent results with group_weight,
2034 * otherwise workloads might not converge.
2036 static int preferred_group_nid(struct task_struct *p, int nid)
2041 /* Direct connections between all NUMA nodes. */
2042 if (sched_numa_topology_type == NUMA_DIRECT)
2046 * On a system with glueless mesh NUMA topology, group_weight
2047 * scores nodes according to the number of NUMA hinting faults on
2048 * both the node itself, and on nearby nodes.
2050 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2051 unsigned long score, max_score = 0;
2052 int node, max_node = nid;
2054 dist = sched_max_numa_distance;
2056 for_each_online_node(node) {
2057 score = group_weight(p, node, dist);
2058 if (score > max_score) {
2067 * Finding the preferred nid in a system with NUMA backplane
2068 * interconnect topology is more involved. The goal is to locate
2069 * tasks from numa_groups near each other in the system, and
2070 * untangle workloads from different sides of the system. This requires
2071 * searching down the hierarchy of node groups, recursively searching
2072 * inside the highest scoring group of nodes. The nodemask tricks
2073 * keep the complexity of the search down.
2075 nodes = node_online_map;
2076 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2077 unsigned long max_faults = 0;
2078 nodemask_t max_group = NODE_MASK_NONE;
2081 /* Are there nodes at this distance from each other? */
2082 if (!find_numa_distance(dist))
2085 for_each_node_mask(a, nodes) {
2086 unsigned long faults = 0;
2087 nodemask_t this_group;
2088 nodes_clear(this_group);
2090 /* Sum group's NUMA faults; includes a==b case. */
2091 for_each_node_mask(b, nodes) {
2092 if (node_distance(a, b) < dist) {
2093 faults += group_faults(p, b);
2094 node_set(b, this_group);
2095 node_clear(b, nodes);
2099 /* Remember the top group. */
2100 if (faults > max_faults) {
2101 max_faults = faults;
2102 max_group = this_group;
2104 * subtle: at the smallest distance there is
2105 * just one node left in each "group", the
2106 * winner is the preferred nid.
2111 /* Next round, evaluate the nodes within max_group. */
2119 static void task_numa_placement(struct task_struct *p)
2121 int seq, nid, max_nid = -1, max_group_nid = -1;
2122 unsigned long max_faults = 0, max_group_faults = 0;
2123 unsigned long fault_types[2] = { 0, 0 };
2124 unsigned long total_faults;
2125 u64 runtime, period;
2126 spinlock_t *group_lock = NULL;
2129 * The p->mm->numa_scan_seq field gets updated without
2130 * exclusive access. Use READ_ONCE() here to ensure
2131 * that the field is read in a single access:
2133 seq = READ_ONCE(p->mm->numa_scan_seq);
2134 if (p->numa_scan_seq == seq)
2136 p->numa_scan_seq = seq;
2137 p->numa_scan_period_max = task_scan_max(p);
2139 total_faults = p->numa_faults_locality[0] +
2140 p->numa_faults_locality[1];
2141 runtime = numa_get_avg_runtime(p, &period);
2143 /* If the task is part of a group prevent parallel updates to group stats */
2144 if (p->numa_group) {
2145 group_lock = &p->numa_group->lock;
2146 spin_lock_irq(group_lock);
2149 /* Find the node with the highest number of faults */
2150 for_each_online_node(nid) {
2151 /* Keep track of the offsets in numa_faults array */
2152 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2153 unsigned long faults = 0, group_faults = 0;
2156 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2157 long diff, f_diff, f_weight;
2159 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2160 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2161 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2162 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2164 /* Decay existing window, copy faults since last scan */
2165 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2166 fault_types[priv] += p->numa_faults[membuf_idx];
2167 p->numa_faults[membuf_idx] = 0;
2170 * Normalize the faults_from, so all tasks in a group
2171 * count according to CPU use, instead of by the raw
2172 * number of faults. Tasks with little runtime have
2173 * little over-all impact on throughput, and thus their
2174 * faults are less important.
2176 f_weight = div64_u64(runtime << 16, period + 1);
2177 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2179 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2180 p->numa_faults[cpubuf_idx] = 0;
2182 p->numa_faults[mem_idx] += diff;
2183 p->numa_faults[cpu_idx] += f_diff;
2184 faults += p->numa_faults[mem_idx];
2185 p->total_numa_faults += diff;
2186 if (p->numa_group) {
2188 * safe because we can only change our own group
2190 * mem_idx represents the offset for a given
2191 * nid and priv in a specific region because it
2192 * is at the beginning of the numa_faults array.
2194 p->numa_group->faults[mem_idx] += diff;
2195 p->numa_group->faults_cpu[mem_idx] += f_diff;
2196 p->numa_group->total_faults += diff;
2197 group_faults += p->numa_group->faults[mem_idx];
2201 if (faults > max_faults) {
2202 max_faults = faults;
2206 if (group_faults > max_group_faults) {
2207 max_group_faults = group_faults;
2208 max_group_nid = nid;
2212 update_task_scan_period(p, fault_types[0], fault_types[1]);
2214 if (p->numa_group) {
2215 numa_group_count_active_nodes(p->numa_group);
2216 spin_unlock_irq(group_lock);
2217 max_nid = preferred_group_nid(p, max_group_nid);
2221 /* Set the new preferred node */
2222 if (max_nid != p->numa_preferred_nid)
2223 sched_setnuma(p, max_nid);
2225 if (task_node(p) != p->numa_preferred_nid)
2226 numa_migrate_preferred(p);
2230 static inline int get_numa_group(struct numa_group *grp)
2232 return atomic_inc_not_zero(&grp->refcount);
2235 static inline void put_numa_group(struct numa_group *grp)
2237 if (atomic_dec_and_test(&grp->refcount))
2238 kfree_rcu(grp, rcu);
2241 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2244 struct numa_group *grp, *my_grp;
2245 struct task_struct *tsk;
2247 int cpu = cpupid_to_cpu(cpupid);
2250 if (unlikely(!p->numa_group)) {
2251 unsigned int size = sizeof(struct numa_group) +
2252 4*nr_node_ids*sizeof(unsigned long);
2254 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2258 atomic_set(&grp->refcount, 1);
2259 grp->active_nodes = 1;
2260 grp->max_faults_cpu = 0;
2261 spin_lock_init(&grp->lock);
2263 /* Second half of the array tracks nids where faults happen */
2264 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2267 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2268 grp->faults[i] = p->numa_faults[i];
2270 grp->total_faults = p->total_numa_faults;
2273 rcu_assign_pointer(p->numa_group, grp);
2277 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2279 if (!cpupid_match_pid(tsk, cpupid))
2282 grp = rcu_dereference(tsk->numa_group);
2286 my_grp = p->numa_group;
2291 * Only join the other group if its bigger; if we're the bigger group,
2292 * the other task will join us.
2294 if (my_grp->nr_tasks > grp->nr_tasks)
2298 * Tie-break on the grp address.
2300 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2303 /* Always join threads in the same process. */
2304 if (tsk->mm == current->mm)
2307 /* Simple filter to avoid false positives due to PID collisions */
2308 if (flags & TNF_SHARED)
2311 /* Update priv based on whether false sharing was detected */
2314 if (join && !get_numa_group(grp))
2322 BUG_ON(irqs_disabled());
2323 double_lock_irq(&my_grp->lock, &grp->lock);
2325 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2326 my_grp->faults[i] -= p->numa_faults[i];
2327 grp->faults[i] += p->numa_faults[i];
2329 my_grp->total_faults -= p->total_numa_faults;
2330 grp->total_faults += p->total_numa_faults;
2335 spin_unlock(&my_grp->lock);
2336 spin_unlock_irq(&grp->lock);
2338 rcu_assign_pointer(p->numa_group, grp);
2340 put_numa_group(my_grp);
2348 void task_numa_free(struct task_struct *p)
2350 struct numa_group *grp = p->numa_group;
2351 void *numa_faults = p->numa_faults;
2352 unsigned long flags;
2356 spin_lock_irqsave(&grp->lock, flags);
2357 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2358 grp->faults[i] -= p->numa_faults[i];
2359 grp->total_faults -= p->total_numa_faults;
2362 spin_unlock_irqrestore(&grp->lock, flags);
2363 RCU_INIT_POINTER(p->numa_group, NULL);
2364 put_numa_group(grp);
2367 p->numa_faults = NULL;
2372 * Got a PROT_NONE fault for a page on @node.
2374 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2376 struct task_struct *p = current;
2377 bool migrated = flags & TNF_MIGRATED;
2378 int cpu_node = task_node(current);
2379 int local = !!(flags & TNF_FAULT_LOCAL);
2380 struct numa_group *ng;
2383 if (!static_branch_likely(&sched_numa_balancing))
2386 /* for example, ksmd faulting in a user's mm */
2390 /* Allocate buffer to track faults on a per-node basis */
2391 if (unlikely(!p->numa_faults)) {
2392 int size = sizeof(*p->numa_faults) *
2393 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2395 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2396 if (!p->numa_faults)
2399 p->total_numa_faults = 0;
2400 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2404 * First accesses are treated as private, otherwise consider accesses
2405 * to be private if the accessing pid has not changed
2407 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2410 priv = cpupid_match_pid(p, last_cpupid);
2411 if (!priv && !(flags & TNF_NO_GROUP))
2412 task_numa_group(p, last_cpupid, flags, &priv);
2416 * If a workload spans multiple NUMA nodes, a shared fault that
2417 * occurs wholly within the set of nodes that the workload is
2418 * actively using should be counted as local. This allows the
2419 * scan rate to slow down when a workload has settled down.
2422 if (!priv && !local && ng && ng->active_nodes > 1 &&
2423 numa_is_active_node(cpu_node, ng) &&
2424 numa_is_active_node(mem_node, ng))
2427 task_numa_placement(p);
2430 * Retry task to preferred node migration periodically, in case it
2431 * case it previously failed, or the scheduler moved us.
2433 if (time_after(jiffies, p->numa_migrate_retry))
2434 numa_migrate_preferred(p);
2437 p->numa_pages_migrated += pages;
2438 if (flags & TNF_MIGRATE_FAIL)
2439 p->numa_faults_locality[2] += pages;
2441 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2442 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2443 p->numa_faults_locality[local] += pages;
2446 static void reset_ptenuma_scan(struct task_struct *p)
2449 * We only did a read acquisition of the mmap sem, so
2450 * p->mm->numa_scan_seq is written to without exclusive access
2451 * and the update is not guaranteed to be atomic. That's not
2452 * much of an issue though, since this is just used for
2453 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2454 * expensive, to avoid any form of compiler optimizations:
2456 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2457 p->mm->numa_scan_offset = 0;
2461 * The expensive part of numa migration is done from task_work context.
2462 * Triggered from task_tick_numa().
2464 void task_numa_work(struct callback_head *work)
2466 unsigned long migrate, next_scan, now = jiffies;
2467 struct task_struct *p = current;
2468 struct mm_struct *mm = p->mm;
2469 u64 runtime = p->se.sum_exec_runtime;
2470 struct vm_area_struct *vma;
2471 unsigned long start, end;
2472 unsigned long nr_pte_updates = 0;
2473 long pages, virtpages;
2475 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2477 work->next = work; /* protect against double add */
2479 * Who cares about NUMA placement when they're dying.
2481 * NOTE: make sure not to dereference p->mm before this check,
2482 * exit_task_work() happens _after_ exit_mm() so we could be called
2483 * without p->mm even though we still had it when we enqueued this
2486 if (p->flags & PF_EXITING)
2489 if (!mm->numa_next_scan) {
2490 mm->numa_next_scan = now +
2491 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2495 * Enforce maximal scan/migration frequency..
2497 migrate = mm->numa_next_scan;
2498 if (time_before(now, migrate))
2501 if (p->numa_scan_period == 0) {
2502 p->numa_scan_period_max = task_scan_max(p);
2503 p->numa_scan_period = task_scan_start(p);
2506 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2507 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2511 * Delay this task enough that another task of this mm will likely win
2512 * the next time around.
2514 p->node_stamp += 2 * TICK_NSEC;
2516 start = mm->numa_scan_offset;
2517 pages = sysctl_numa_balancing_scan_size;
2518 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2519 virtpages = pages * 8; /* Scan up to this much virtual space */
2524 if (!down_read_trylock(&mm->mmap_sem))
2526 vma = find_vma(mm, start);
2528 reset_ptenuma_scan(p);
2532 for (; vma; vma = vma->vm_next) {
2533 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2534 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2539 * Shared library pages mapped by multiple processes are not
2540 * migrated as it is expected they are cache replicated. Avoid
2541 * hinting faults in read-only file-backed mappings or the vdso
2542 * as migrating the pages will be of marginal benefit.
2545 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2549 * Skip inaccessible VMAs to avoid any confusion between
2550 * PROT_NONE and NUMA hinting ptes
2552 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2556 start = max(start, vma->vm_start);
2557 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2558 end = min(end, vma->vm_end);
2559 nr_pte_updates = change_prot_numa(vma, start, end);
2562 * Try to scan sysctl_numa_balancing_size worth of
2563 * hpages that have at least one present PTE that
2564 * is not already pte-numa. If the VMA contains
2565 * areas that are unused or already full of prot_numa
2566 * PTEs, scan up to virtpages, to skip through those
2570 pages -= (end - start) >> PAGE_SHIFT;
2571 virtpages -= (end - start) >> PAGE_SHIFT;
2574 if (pages <= 0 || virtpages <= 0)
2578 } while (end != vma->vm_end);
2583 * It is possible to reach the end of the VMA list but the last few
2584 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2585 * would find the !migratable VMA on the next scan but not reset the
2586 * scanner to the start so check it now.
2589 mm->numa_scan_offset = start;
2591 reset_ptenuma_scan(p);
2592 up_read(&mm->mmap_sem);
2595 * Make sure tasks use at least 32x as much time to run other code
2596 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2597 * Usually update_task_scan_period slows down scanning enough; on an
2598 * overloaded system we need to limit overhead on a per task basis.
2600 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2601 u64 diff = p->se.sum_exec_runtime - runtime;
2602 p->node_stamp += 32 * diff;
2607 * Drive the periodic memory faults..
2609 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2611 struct callback_head *work = &curr->numa_work;
2615 * We don't care about NUMA placement if we don't have memory.
2617 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2621 * Using runtime rather than walltime has the dual advantage that
2622 * we (mostly) drive the selection from busy threads and that the
2623 * task needs to have done some actual work before we bother with
2626 now = curr->se.sum_exec_runtime;
2627 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2629 if (now > curr->node_stamp + period) {
2630 if (!curr->node_stamp)
2631 curr->numa_scan_period = task_scan_start(curr);
2632 curr->node_stamp += period;
2634 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2635 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2636 task_work_add(curr, work, true);
2642 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2646 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2650 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2654 #endif /* CONFIG_NUMA_BALANCING */
2657 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2659 update_load_add(&cfs_rq->load, se->load.weight);
2660 if (!parent_entity(se))
2661 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2663 if (entity_is_task(se)) {
2664 struct rq *rq = rq_of(cfs_rq);
2666 account_numa_enqueue(rq, task_of(se));
2667 list_add(&se->group_node, &rq->cfs_tasks);
2670 cfs_rq->nr_running++;
2674 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2676 update_load_sub(&cfs_rq->load, se->load.weight);
2677 if (!parent_entity(se))
2678 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2680 if (entity_is_task(se)) {
2681 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2682 list_del_init(&se->group_node);
2685 cfs_rq->nr_running--;
2689 * Signed add and clamp on underflow.
2691 * Explicitly do a load-store to ensure the intermediate value never hits
2692 * memory. This allows lockless observations without ever seeing the negative
2695 #define add_positive(_ptr, _val) do { \
2696 typeof(_ptr) ptr = (_ptr); \
2697 typeof(_val) val = (_val); \
2698 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2702 if (val < 0 && res > var) \
2705 WRITE_ONCE(*ptr, res); \
2709 * Unsigned subtract and clamp on underflow.
2711 * Explicitly do a load-store to ensure the intermediate value never hits
2712 * memory. This allows lockless observations without ever seeing the negative
2715 #define sub_positive(_ptr, _val) do { \
2716 typeof(_ptr) ptr = (_ptr); \
2717 typeof(*ptr) val = (_val); \
2718 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2722 WRITE_ONCE(*ptr, res); \
2727 * XXX we want to get rid of these helpers and use the full load resolution.
2729 static inline long se_weight(struct sched_entity *se)
2731 return scale_load_down(se->load.weight);
2734 static inline long se_runnable(struct sched_entity *se)
2736 return scale_load_down(se->runnable_weight);
2740 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2742 cfs_rq->runnable_weight += se->runnable_weight;
2744 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2745 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2749 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2751 cfs_rq->runnable_weight -= se->runnable_weight;
2753 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2754 sub_positive(&cfs_rq->avg.runnable_load_sum,
2755 se_runnable(se) * se->avg.runnable_load_sum);
2759 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2761 cfs_rq->avg.load_avg += se->avg.load_avg;
2762 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2766 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2768 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2769 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2773 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2775 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2777 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2779 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2782 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2783 unsigned long weight, unsigned long runnable)
2786 /* commit outstanding execution time */
2787 if (cfs_rq->curr == se)
2788 update_curr(cfs_rq);
2789 account_entity_dequeue(cfs_rq, se);
2790 dequeue_runnable_load_avg(cfs_rq, se);
2792 dequeue_load_avg(cfs_rq, se);
2794 se->runnable_weight = runnable;
2795 update_load_set(&se->load, weight);
2799 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2801 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2802 se->avg.runnable_load_avg =
2803 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2807 enqueue_load_avg(cfs_rq, se);
2809 account_entity_enqueue(cfs_rq, se);
2810 enqueue_runnable_load_avg(cfs_rq, se);
2814 void reweight_task(struct task_struct *p, int prio)
2816 struct sched_entity *se = &p->se;
2817 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2818 struct load_weight *load = &se->load;
2819 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2821 reweight_entity(cfs_rq, se, weight, weight);
2822 load->inv_weight = sched_prio_to_wmult[prio];
2825 #ifdef CONFIG_FAIR_GROUP_SCHED
2828 * All this does is approximate the hierarchical proportion which includes that
2829 * global sum we all love to hate.
2831 * That is, the weight of a group entity, is the proportional share of the
2832 * group weight based on the group runqueue weights. That is:
2834 * tg->weight * grq->load.weight
2835 * ge->load.weight = ----------------------------- (1)
2836 * \Sum grq->load.weight
2838 * Now, because computing that sum is prohibitively expensive to compute (been
2839 * there, done that) we approximate it with this average stuff. The average
2840 * moves slower and therefore the approximation is cheaper and more stable.
2842 * So instead of the above, we substitute:
2844 * grq->load.weight -> grq->avg.load_avg (2)
2846 * which yields the following:
2848 * tg->weight * grq->avg.load_avg
2849 * ge->load.weight = ------------------------------ (3)
2852 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2854 * That is shares_avg, and it is right (given the approximation (2)).
2856 * The problem with it is that because the average is slow -- it was designed
2857 * to be exactly that of course -- this leads to transients in boundary
2858 * conditions. In specific, the case where the group was idle and we start the
2859 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2860 * yielding bad latency etc..
2862 * Now, in that special case (1) reduces to:
2864 * tg->weight * grq->load.weight
2865 * ge->load.weight = ----------------------------- = tg->weight (4)
2868 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2870 * So what we do is modify our approximation (3) to approach (4) in the (near)
2875 * tg->weight * grq->load.weight
2876 * --------------------------------------------------- (5)
2877 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2879 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2880 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2883 * tg->weight * grq->load.weight
2884 * ge->load.weight = ----------------------------- (6)
2889 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2890 * max(grq->load.weight, grq->avg.load_avg)
2892 * And that is shares_weight and is icky. In the (near) UP case it approaches
2893 * (4) while in the normal case it approaches (3). It consistently
2894 * overestimates the ge->load.weight and therefore:
2896 * \Sum ge->load.weight >= tg->weight
2900 static long calc_group_shares(struct cfs_rq *cfs_rq)
2902 long tg_weight, tg_shares, load, shares;
2903 struct task_group *tg = cfs_rq->tg;
2905 tg_shares = READ_ONCE(tg->shares);
2907 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2909 tg_weight = atomic_long_read(&tg->load_avg);
2911 /* Ensure tg_weight >= load */
2912 tg_weight -= cfs_rq->tg_load_avg_contrib;
2915 shares = (tg_shares * load);
2917 shares /= tg_weight;
2920 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2921 * of a group with small tg->shares value. It is a floor value which is
2922 * assigned as a minimum load.weight to the sched_entity representing
2923 * the group on a CPU.
2925 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2926 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2927 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2928 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2931 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2935 * This calculates the effective runnable weight for a group entity based on
2936 * the group entity weight calculated above.
2938 * Because of the above approximation (2), our group entity weight is
2939 * an load_avg based ratio (3). This means that it includes blocked load and
2940 * does not represent the runnable weight.
2942 * Approximate the group entity's runnable weight per ratio from the group
2945 * grq->avg.runnable_load_avg
2946 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2949 * However, analogous to above, since the avg numbers are slow, this leads to
2950 * transients in the from-idle case. Instead we use:
2952 * ge->runnable_weight = ge->load.weight *
2954 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2955 * ----------------------------------------------------- (8)
2956 * max(grq->avg.load_avg, grq->load.weight)
2958 * Where these max() serve both to use the 'instant' values to fix the slow
2959 * from-idle and avoid the /0 on to-idle, similar to (6).
2961 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2963 long runnable, load_avg;
2965 load_avg = max(cfs_rq->avg.load_avg,
2966 scale_load_down(cfs_rq->load.weight));
2968 runnable = max(cfs_rq->avg.runnable_load_avg,
2969 scale_load_down(cfs_rq->runnable_weight));
2973 runnable /= load_avg;
2975 return clamp_t(long, runnable, MIN_SHARES, shares);
2977 #endif /* CONFIG_SMP */
2979 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2982 * Recomputes the group entity based on the current state of its group
2985 static void update_cfs_group(struct sched_entity *se)
2987 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2988 long shares, runnable;
2993 if (throttled_hierarchy(gcfs_rq))
2997 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2999 if (likely(se->load.weight == shares))
3002 shares = calc_group_shares(gcfs_rq);
3003 runnable = calc_group_runnable(gcfs_rq, shares);
3006 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3009 #else /* CONFIG_FAIR_GROUP_SCHED */
3010 static inline void update_cfs_group(struct sched_entity *se)
3013 #endif /* CONFIG_FAIR_GROUP_SCHED */
3015 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3017 struct rq *rq = rq_of(cfs_rq);
3019 if (&rq->cfs == cfs_rq) {
3021 * There are a few boundary cases this might miss but it should
3022 * get called often enough that that should (hopefully) not be
3025 * It will not get called when we go idle, because the idle
3026 * thread is a different class (!fair), nor will the utilization
3027 * number include things like RT tasks.
3029 * As is, the util number is not freq-invariant (we'd have to
3030 * implement arch_scale_freq_capacity() for that).
3034 cpufreq_update_util(rq, 0);
3041 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3043 static u64 decay_load(u64 val, u64 n)
3045 unsigned int local_n;
3047 if (unlikely(n > LOAD_AVG_PERIOD * 63))
3050 /* after bounds checking we can collapse to 32-bit */
3054 * As y^PERIOD = 1/2, we can combine
3055 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3056 * With a look-up table which covers y^n (n<PERIOD)
3058 * To achieve constant time decay_load.
3060 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3061 val >>= local_n / LOAD_AVG_PERIOD;
3062 local_n %= LOAD_AVG_PERIOD;
3065 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3069 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3071 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3076 c1 = decay_load((u64)d1, periods);
3080 * c2 = 1024 \Sum y^n
3084 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3087 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3089 return c1 + c2 + c3;
3093 * Accumulate the three separate parts of the sum; d1 the remainder
3094 * of the last (incomplete) period, d2 the span of full periods and d3
3095 * the remainder of the (incomplete) current period.
3100 * |<->|<----------------->|<--->|
3101 * ... |---x---|------| ... |------|-----x (now)
3104 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3107 * = u y^p + (Step 1)
3110 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3113 static __always_inline u32
3114 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3115 unsigned long load, unsigned long runnable, int running)
3117 unsigned long scale_freq, scale_cpu;
3118 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3121 scale_freq = arch_scale_freq_capacity(cpu);
3122 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3124 delta += sa->period_contrib;
3125 periods = delta / 1024; /* A period is 1024us (~1ms) */
3128 * Step 1: decay old *_sum if we crossed period boundaries.
3131 sa->load_sum = decay_load(sa->load_sum, periods);
3132 sa->runnable_load_sum =
3133 decay_load(sa->runnable_load_sum, periods);
3134 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3140 contrib = __accumulate_pelt_segments(periods,
3141 1024 - sa->period_contrib, delta);
3143 sa->period_contrib = delta;
3145 contrib = cap_scale(contrib, scale_freq);
3147 sa->load_sum += load * contrib;
3149 sa->runnable_load_sum += runnable * contrib;
3151 sa->util_sum += contrib * scale_cpu;
3157 * We can represent the historical contribution to runnable average as the
3158 * coefficients of a geometric series. To do this we sub-divide our runnable
3159 * history into segments of approximately 1ms (1024us); label the segment that
3160 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3162 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3164 * (now) (~1ms ago) (~2ms ago)
3166 * Let u_i denote the fraction of p_i that the entity was runnable.
3168 * We then designate the fractions u_i as our co-efficients, yielding the
3169 * following representation of historical load:
3170 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3172 * We choose y based on the with of a reasonably scheduling period, fixing:
3175 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3176 * approximately half as much as the contribution to load within the last ms
3179 * When a period "rolls over" and we have new u_0`, multiplying the previous
3180 * sum again by y is sufficient to update:
3181 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3182 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3184 static __always_inline int
3185 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3186 unsigned long load, unsigned long runnable, int running)
3190 delta = now - sa->last_update_time;
3192 * This should only happen when time goes backwards, which it
3193 * unfortunately does during sched clock init when we swap over to TSC.
3195 if ((s64)delta < 0) {
3196 sa->last_update_time = now;
3201 * Use 1024ns as the unit of measurement since it's a reasonable
3202 * approximation of 1us and fast to compute.
3208 sa->last_update_time += delta << 10;
3211 * running is a subset of runnable (weight) so running can't be set if
3212 * runnable is clear. But there are some corner cases where the current
3213 * se has been already dequeued but cfs_rq->curr still points to it.
3214 * This means that weight will be 0 but not running for a sched_entity
3215 * but also for a cfs_rq if the latter becomes idle. As an example,
3216 * this happens during idle_balance() which calls
3217 * update_blocked_averages()
3220 runnable = running = 0;
3223 * Now we know we crossed measurement unit boundaries. The *_avg
3224 * accrues by two steps:
3226 * Step 1: accumulate *_sum since last_update_time. If we haven't
3227 * crossed period boundaries, finish.
3229 if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3235 static __always_inline void
3236 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3238 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3241 * Step 2: update *_avg.
3243 sa->load_avg = div_u64(load * sa->load_sum, divider);
3244 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3245 sa->util_avg = sa->util_sum / divider;
3252 * se_runnable() == se_weight()
3254 * group: [ see update_cfs_group() ]
3255 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3256 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3258 * load_sum := runnable_sum
3259 * load_avg = se_weight(se) * runnable_avg
3261 * runnable_load_sum := runnable_sum
3262 * runnable_load_avg = se_runnable(se) * runnable_avg
3264 * XXX collapse load_sum and runnable_load_sum
3268 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3269 * load_avg = \Sum se->avg.load_avg
3271 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3272 * runnable_load_avg = \Sum se->avg.runable_load_avg
3276 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3278 if (entity_is_task(se))
3279 se->runnable_weight = se->load.weight;
3281 if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3282 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3290 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3292 if (entity_is_task(se))
3293 se->runnable_weight = se->load.weight;
3295 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3296 cfs_rq->curr == se)) {
3298 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3306 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3308 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3309 scale_load_down(cfs_rq->load.weight),
3310 scale_load_down(cfs_rq->runnable_weight),
3311 cfs_rq->curr != NULL)) {
3313 ___update_load_avg(&cfs_rq->avg, 1, 1);
3320 #ifdef CONFIG_FAIR_GROUP_SCHED
3322 * update_tg_load_avg - update the tg's load avg
3323 * @cfs_rq: the cfs_rq whose avg changed
3324 * @force: update regardless of how small the difference
3326 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3327 * However, because tg->load_avg is a global value there are performance
3330 * In order to avoid having to look at the other cfs_rq's, we use a
3331 * differential update where we store the last value we propagated. This in
3332 * turn allows skipping updates if the differential is 'small'.
3334 * Updating tg's load_avg is necessary before update_cfs_share().
3336 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3338 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3341 * No need to update load_avg for root_task_group as it is not used.
3343 if (cfs_rq->tg == &root_task_group)
3346 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3347 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3348 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3353 * Called within set_task_rq() right before setting a task's cpu. The
3354 * caller only guarantees p->pi_lock is held; no other assumptions,
3355 * including the state of rq->lock, should be made.
3357 void set_task_rq_fair(struct sched_entity *se,
3358 struct cfs_rq *prev, struct cfs_rq *next)
3360 u64 p_last_update_time;
3361 u64 n_last_update_time;
3363 if (!sched_feat(ATTACH_AGE_LOAD))
3367 * We are supposed to update the task to "current" time, then its up to
3368 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3369 * getting what current time is, so simply throw away the out-of-date
3370 * time. This will result in the wakee task is less decayed, but giving
3371 * the wakee more load sounds not bad.
3373 if (!(se->avg.last_update_time && prev))
3376 #ifndef CONFIG_64BIT
3378 u64 p_last_update_time_copy;
3379 u64 n_last_update_time_copy;
3382 p_last_update_time_copy = prev->load_last_update_time_copy;
3383 n_last_update_time_copy = next->load_last_update_time_copy;
3387 p_last_update_time = prev->avg.last_update_time;
3388 n_last_update_time = next->avg.last_update_time;
3390 } while (p_last_update_time != p_last_update_time_copy ||
3391 n_last_update_time != n_last_update_time_copy);
3394 p_last_update_time = prev->avg.last_update_time;
3395 n_last_update_time = next->avg.last_update_time;
3397 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3398 se->avg.last_update_time = n_last_update_time;
3403 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3404 * propagate its contribution. The key to this propagation is the invariant
3405 * that for each group:
3407 * ge->avg == grq->avg (1)
3409 * _IFF_ we look at the pure running and runnable sums. Because they
3410 * represent the very same entity, just at different points in the hierarchy.
3412 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3413 * sum over (but still wrong, because the group entity and group rq do not have
3414 * their PELT windows aligned).
3416 * However, update_tg_cfs_runnable() is more complex. So we have:
3418 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3420 * And since, like util, the runnable part should be directly transferable,
3421 * the following would _appear_ to be the straight forward approach:
3423 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3425 * And per (1) we have:
3427 * ge->avg.runnable_avg == grq->avg.runnable_avg
3431 * ge->load.weight * grq->avg.load_avg
3432 * ge->avg.load_avg = ----------------------------------- (4)
3435 * Except that is wrong!
3437 * Because while for entities historical weight is not important and we
3438 * really only care about our future and therefore can consider a pure
3439 * runnable sum, runqueues can NOT do this.
3441 * We specifically want runqueues to have a load_avg that includes
3442 * historical weights. Those represent the blocked load, the load we expect
3443 * to (shortly) return to us. This only works by keeping the weights as
3444 * integral part of the sum. We therefore cannot decompose as per (3).
3446 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3447 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3448 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3449 * runnable section of these tasks overlap (or not). If they were to perfectly
3450 * align the rq as a whole would be runnable 2/3 of the time. If however we
3451 * always have at least 1 runnable task, the rq as a whole is always runnable.
3453 * So we'll have to approximate.. :/
3455 * Given the constraint:
3457 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3459 * We can construct a rule that adds runnable to a rq by assuming minimal
3462 * On removal, we'll assume each task is equally runnable; which yields:
3464 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3466 * XXX: only do this for the part of runnable > running ?
3471 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3473 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3475 /* Nothing to update */
3480 * The relation between sum and avg is:
3482 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3484 * however, the PELT windows are not aligned between grq and gse.
3487 /* Set new sched_entity's utilization */
3488 se->avg.util_avg = gcfs_rq->avg.util_avg;
3489 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3491 /* Update parent cfs_rq utilization */
3492 add_positive(&cfs_rq->avg.util_avg, delta);
3493 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3497 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3499 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3500 unsigned long runnable_load_avg, load_avg;
3501 u64 runnable_load_sum, load_sum = 0;
3507 gcfs_rq->prop_runnable_sum = 0;
3509 if (runnable_sum >= 0) {
3511 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3512 * the CPU is saturated running == runnable.
3514 runnable_sum += se->avg.load_sum;
3515 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3518 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3519 * assuming all tasks are equally runnable.
3521 if (scale_load_down(gcfs_rq->load.weight)) {
3522 load_sum = div_s64(gcfs_rq->avg.load_sum,
3523 scale_load_down(gcfs_rq->load.weight));
3526 /* But make sure to not inflate se's runnable */
3527 runnable_sum = min(se->avg.load_sum, load_sum);
3531 * runnable_sum can't be lower than running_sum
3532 * As running sum is scale with cpu capacity wehreas the runnable sum
3533 * is not we rescale running_sum 1st
3535 running_sum = se->avg.util_sum /
3536 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3537 runnable_sum = max(runnable_sum, running_sum);
3539 load_sum = (s64)se_weight(se) * runnable_sum;
3540 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3542 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3543 delta_avg = load_avg - se->avg.load_avg;
3545 se->avg.load_sum = runnable_sum;
3546 se->avg.load_avg = load_avg;
3547 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3548 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3550 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3551 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3552 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3553 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3555 se->avg.runnable_load_sum = runnable_sum;
3556 se->avg.runnable_load_avg = runnable_load_avg;
3559 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3560 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3564 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3566 cfs_rq->propagate = 1;
3567 cfs_rq->prop_runnable_sum += runnable_sum;
3570 /* Update task and its cfs_rq load average */
3571 static inline int propagate_entity_load_avg(struct sched_entity *se)
3573 struct cfs_rq *cfs_rq, *gcfs_rq;
3575 if (entity_is_task(se))
3578 gcfs_rq = group_cfs_rq(se);
3579 if (!gcfs_rq->propagate)
3582 gcfs_rq->propagate = 0;
3584 cfs_rq = cfs_rq_of(se);
3586 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3588 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3589 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3595 * Check if we need to update the load and the utilization of a blocked
3598 static inline bool skip_blocked_update(struct sched_entity *se)
3600 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3603 * If sched_entity still have not zero load or utilization, we have to
3606 if (se->avg.load_avg || se->avg.util_avg)
3610 * If there is a pending propagation, we have to update the load and
3611 * the utilization of the sched_entity:
3613 if (gcfs_rq->propagate)
3617 * Otherwise, the load and the utilization of the sched_entity is
3618 * already zero and there is no pending propagation, so it will be a
3619 * waste of time to try to decay it:
3624 #else /* CONFIG_FAIR_GROUP_SCHED */
3626 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3628 static inline int propagate_entity_load_avg(struct sched_entity *se)
3633 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3635 #endif /* CONFIG_FAIR_GROUP_SCHED */
3638 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3639 * @now: current time, as per cfs_rq_clock_task()
3640 * @cfs_rq: cfs_rq to update
3642 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3643 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3644 * post_init_entity_util_avg().
3646 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3648 * Returns true if the load decayed or we removed load.
3650 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3651 * call update_tg_load_avg() when this function returns true.
3654 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3656 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3657 struct sched_avg *sa = &cfs_rq->avg;
3660 if (cfs_rq->removed.nr) {
3662 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3664 raw_spin_lock(&cfs_rq->removed.lock);
3665 swap(cfs_rq->removed.util_avg, removed_util);
3666 swap(cfs_rq->removed.load_avg, removed_load);
3667 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3668 cfs_rq->removed.nr = 0;
3669 raw_spin_unlock(&cfs_rq->removed.lock);
3672 sub_positive(&sa->load_avg, r);
3673 sub_positive(&sa->load_sum, r * divider);
3676 sub_positive(&sa->util_avg, r);
3677 sub_positive(&sa->util_sum, r * divider);
3679 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3684 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3686 #ifndef CONFIG_64BIT
3688 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3692 cfs_rq_util_change(cfs_rq);
3698 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3699 * @cfs_rq: cfs_rq to attach to
3700 * @se: sched_entity to attach
3702 * Must call update_cfs_rq_load_avg() before this, since we rely on
3703 * cfs_rq->avg.last_update_time being current.
3705 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3707 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3710 * When we attach the @se to the @cfs_rq, we must align the decay
3711 * window because without that, really weird and wonderful things can
3716 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3717 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3720 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3721 * period_contrib. This isn't strictly correct, but since we're
3722 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3725 se->avg.util_sum = se->avg.util_avg * divider;
3727 se->avg.load_sum = divider;
3728 if (se_weight(se)) {
3730 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3733 se->avg.runnable_load_sum = se->avg.load_sum;
3735 enqueue_load_avg(cfs_rq, se);
3736 cfs_rq->avg.util_avg += se->avg.util_avg;
3737 cfs_rq->avg.util_sum += se->avg.util_sum;
3739 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3741 cfs_rq_util_change(cfs_rq);
3745 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3746 * @cfs_rq: cfs_rq to detach from
3747 * @se: sched_entity to detach
3749 * Must call update_cfs_rq_load_avg() before this, since we rely on
3750 * cfs_rq->avg.last_update_time being current.
3752 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3754 dequeue_load_avg(cfs_rq, se);
3755 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3756 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3758 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3760 cfs_rq_util_change(cfs_rq);
3764 * Optional action to be done while updating the load average
3766 #define UPDATE_TG 0x1
3767 #define SKIP_AGE_LOAD 0x2
3768 #define DO_ATTACH 0x4
3770 /* Update task and its cfs_rq load average */
3771 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3773 u64 now = cfs_rq_clock_task(cfs_rq);
3774 struct rq *rq = rq_of(cfs_rq);
3775 int cpu = cpu_of(rq);
3779 * Track task load average for carrying it to new CPU after migrated, and
3780 * track group sched_entity load average for task_h_load calc in migration
3782 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3783 __update_load_avg_se(now, cpu, cfs_rq, se);
3785 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3786 decayed |= propagate_entity_load_avg(se);
3788 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3790 attach_entity_load_avg(cfs_rq, se);
3791 update_tg_load_avg(cfs_rq, 0);
3793 } else if (decayed && (flags & UPDATE_TG))
3794 update_tg_load_avg(cfs_rq, 0);
3797 #ifndef CONFIG_64BIT
3798 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3800 u64 last_update_time_copy;
3801 u64 last_update_time;
3804 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3806 last_update_time = cfs_rq->avg.last_update_time;
3807 } while (last_update_time != last_update_time_copy);
3809 return last_update_time;
3812 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3814 return cfs_rq->avg.last_update_time;
3819 * Synchronize entity load avg of dequeued entity without locking
3822 void sync_entity_load_avg(struct sched_entity *se)
3824 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3825 u64 last_update_time;
3827 last_update_time = cfs_rq_last_update_time(cfs_rq);
3828 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3832 * Task first catches up with cfs_rq, and then subtract
3833 * itself from the cfs_rq (task must be off the queue now).
3835 void remove_entity_load_avg(struct sched_entity *se)
3837 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3838 unsigned long flags;
3841 * tasks cannot exit without having gone through wake_up_new_task() ->
3842 * post_init_entity_util_avg() which will have added things to the
3843 * cfs_rq, so we can remove unconditionally.
3845 * Similarly for groups, they will have passed through
3846 * post_init_entity_util_avg() before unregister_sched_fair_group()
3850 sync_entity_load_avg(se);
3852 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3853 ++cfs_rq->removed.nr;
3854 cfs_rq->removed.util_avg += se->avg.util_avg;
3855 cfs_rq->removed.load_avg += se->avg.load_avg;
3856 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3857 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3860 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3862 return cfs_rq->avg.runnable_load_avg;
3865 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3867 return cfs_rq->avg.load_avg;
3870 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3872 #else /* CONFIG_SMP */
3875 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3880 #define UPDATE_TG 0x0
3881 #define SKIP_AGE_LOAD 0x0
3882 #define DO_ATTACH 0x0
3884 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3886 cfs_rq_util_change(cfs_rq);
3889 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3892 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3894 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3896 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3901 #endif /* CONFIG_SMP */
3903 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3905 #ifdef CONFIG_SCHED_DEBUG
3906 s64 d = se->vruntime - cfs_rq->min_vruntime;
3911 if (d > 3*sysctl_sched_latency)
3912 schedstat_inc(cfs_rq->nr_spread_over);
3917 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3919 u64 vruntime = cfs_rq->min_vruntime;
3922 * The 'current' period is already promised to the current tasks,
3923 * however the extra weight of the new task will slow them down a
3924 * little, place the new task so that it fits in the slot that
3925 * stays open at the end.
3927 if (initial && sched_feat(START_DEBIT))
3928 vruntime += sched_vslice(cfs_rq, se);
3930 /* sleeps up to a single latency don't count. */
3932 unsigned long thresh = sysctl_sched_latency;
3935 * Halve their sleep time's effect, to allow
3936 * for a gentler effect of sleepers:
3938 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3944 /* ensure we never gain time by being placed backwards. */
3945 se->vruntime = max_vruntime(se->vruntime, vruntime);
3948 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3950 static inline void check_schedstat_required(void)
3952 #ifdef CONFIG_SCHEDSTATS
3953 if (schedstat_enabled())
3956 /* Force schedstat enabled if a dependent tracepoint is active */
3957 if (trace_sched_stat_wait_enabled() ||
3958 trace_sched_stat_sleep_enabled() ||
3959 trace_sched_stat_iowait_enabled() ||
3960 trace_sched_stat_blocked_enabled() ||
3961 trace_sched_stat_runtime_enabled()) {
3962 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3963 "stat_blocked and stat_runtime require the "
3964 "kernel parameter schedstats=enable or "
3965 "kernel.sched_schedstats=1\n");
3976 * update_min_vruntime()
3977 * vruntime -= min_vruntime
3981 * update_min_vruntime()
3982 * vruntime += min_vruntime
3984 * this way the vruntime transition between RQs is done when both
3985 * min_vruntime are up-to-date.
3989 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3990 * vruntime -= min_vruntime
3994 * update_min_vruntime()
3995 * vruntime += min_vruntime
3997 * this way we don't have the most up-to-date min_vruntime on the originating
3998 * CPU and an up-to-date min_vruntime on the destination CPU.
4002 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4004 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4005 bool curr = cfs_rq->curr == se;
4008 * If we're the current task, we must renormalise before calling
4012 se->vruntime += cfs_rq->min_vruntime;
4014 update_curr(cfs_rq);
4017 * Otherwise, renormalise after, such that we're placed at the current
4018 * moment in time, instead of some random moment in the past. Being
4019 * placed in the past could significantly boost this task to the
4020 * fairness detriment of existing tasks.
4022 if (renorm && !curr)
4023 se->vruntime += cfs_rq->min_vruntime;
4026 * When enqueuing a sched_entity, we must:
4027 * - Update loads to have both entity and cfs_rq synced with now.
4028 * - Add its load to cfs_rq->runnable_avg
4029 * - For group_entity, update its weight to reflect the new share of
4031 * - Add its new weight to cfs_rq->load.weight
4033 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4034 update_cfs_group(se);
4035 enqueue_runnable_load_avg(cfs_rq, se);
4036 account_entity_enqueue(cfs_rq, se);
4038 if (flags & ENQUEUE_WAKEUP)
4039 place_entity(cfs_rq, se, 0);
4041 check_schedstat_required();
4042 update_stats_enqueue(cfs_rq, se, flags);
4043 check_spread(cfs_rq, se);
4045 __enqueue_entity(cfs_rq, se);
4048 if (cfs_rq->nr_running == 1) {
4049 list_add_leaf_cfs_rq(cfs_rq);
4050 check_enqueue_throttle(cfs_rq);
4054 static void __clear_buddies_last(struct sched_entity *se)
4056 for_each_sched_entity(se) {
4057 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4058 if (cfs_rq->last != se)
4061 cfs_rq->last = NULL;
4065 static void __clear_buddies_next(struct sched_entity *se)
4067 for_each_sched_entity(se) {
4068 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4069 if (cfs_rq->next != se)
4072 cfs_rq->next = NULL;
4076 static void __clear_buddies_skip(struct sched_entity *se)
4078 for_each_sched_entity(se) {
4079 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4080 if (cfs_rq->skip != se)
4083 cfs_rq->skip = NULL;
4087 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4089 if (cfs_rq->last == se)
4090 __clear_buddies_last(se);
4092 if (cfs_rq->next == se)
4093 __clear_buddies_next(se);
4095 if (cfs_rq->skip == se)
4096 __clear_buddies_skip(se);
4099 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4102 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4105 * Update run-time statistics of the 'current'.
4107 update_curr(cfs_rq);
4110 * When dequeuing a sched_entity, we must:
4111 * - Update loads to have both entity and cfs_rq synced with now.
4112 * - Substract its load from the cfs_rq->runnable_avg.
4113 * - Substract its previous weight from cfs_rq->load.weight.
4114 * - For group entity, update its weight to reflect the new share
4115 * of its group cfs_rq.
4117 update_load_avg(cfs_rq, se, UPDATE_TG);
4118 dequeue_runnable_load_avg(cfs_rq, se);
4120 update_stats_dequeue(cfs_rq, se, flags);
4122 clear_buddies(cfs_rq, se);
4124 if (se != cfs_rq->curr)
4125 __dequeue_entity(cfs_rq, se);
4127 account_entity_dequeue(cfs_rq, se);
4130 * Normalize after update_curr(); which will also have moved
4131 * min_vruntime if @se is the one holding it back. But before doing
4132 * update_min_vruntime() again, which will discount @se's position and
4133 * can move min_vruntime forward still more.
4135 if (!(flags & DEQUEUE_SLEEP))
4136 se->vruntime -= cfs_rq->min_vruntime;
4138 /* return excess runtime on last dequeue */
4139 return_cfs_rq_runtime(cfs_rq);
4141 update_cfs_group(se);
4144 * Now advance min_vruntime if @se was the entity holding it back,
4145 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4146 * put back on, and if we advance min_vruntime, we'll be placed back
4147 * further than we started -- ie. we'll be penalized.
4149 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4150 update_min_vruntime(cfs_rq);
4154 * Preempt the current task with a newly woken task if needed:
4157 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4159 unsigned long ideal_runtime, delta_exec;
4160 struct sched_entity *se;
4163 ideal_runtime = sched_slice(cfs_rq, curr);
4164 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4165 if (delta_exec > ideal_runtime) {
4166 resched_curr(rq_of(cfs_rq));
4168 * The current task ran long enough, ensure it doesn't get
4169 * re-elected due to buddy favours.
4171 clear_buddies(cfs_rq, curr);
4176 * Ensure that a task that missed wakeup preemption by a
4177 * narrow margin doesn't have to wait for a full slice.
4178 * This also mitigates buddy induced latencies under load.
4180 if (delta_exec < sysctl_sched_min_granularity)
4183 se = __pick_first_entity(cfs_rq);
4184 delta = curr->vruntime - se->vruntime;
4189 if (delta > ideal_runtime)
4190 resched_curr(rq_of(cfs_rq));
4194 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4196 /* 'current' is not kept within the tree. */
4199 * Any task has to be enqueued before it get to execute on
4200 * a CPU. So account for the time it spent waiting on the
4203 update_stats_wait_end(cfs_rq, se);
4204 __dequeue_entity(cfs_rq, se);
4205 update_load_avg(cfs_rq, se, UPDATE_TG);
4208 update_stats_curr_start(cfs_rq, se);
4212 * Track our maximum slice length, if the CPU's load is at
4213 * least twice that of our own weight (i.e. dont track it
4214 * when there are only lesser-weight tasks around):
4216 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4217 schedstat_set(se->statistics.slice_max,
4218 max((u64)schedstat_val(se->statistics.slice_max),
4219 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4222 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4226 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4229 * Pick the next process, keeping these things in mind, in this order:
4230 * 1) keep things fair between processes/task groups
4231 * 2) pick the "next" process, since someone really wants that to run
4232 * 3) pick the "last" process, for cache locality
4233 * 4) do not run the "skip" process, if something else is available
4235 static struct sched_entity *
4236 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4238 struct sched_entity *left = __pick_first_entity(cfs_rq);
4239 struct sched_entity *se;
4242 * If curr is set we have to see if its left of the leftmost entity
4243 * still in the tree, provided there was anything in the tree at all.
4245 if (!left || (curr && entity_before(curr, left)))
4248 se = left; /* ideally we run the leftmost entity */
4251 * Avoid running the skip buddy, if running something else can
4252 * be done without getting too unfair.
4254 if (cfs_rq->skip == se) {
4255 struct sched_entity *second;
4258 second = __pick_first_entity(cfs_rq);
4260 second = __pick_next_entity(se);
4261 if (!second || (curr && entity_before(curr, second)))
4265 if (second && wakeup_preempt_entity(second, left) < 1)
4270 * Prefer last buddy, try to return the CPU to a preempted task.
4272 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4276 * Someone really wants this to run. If it's not unfair, run it.
4278 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4281 clear_buddies(cfs_rq, se);
4286 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4288 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4291 * If still on the runqueue then deactivate_task()
4292 * was not called and update_curr() has to be done:
4295 update_curr(cfs_rq);
4297 /* throttle cfs_rqs exceeding runtime */
4298 check_cfs_rq_runtime(cfs_rq);
4300 check_spread(cfs_rq, prev);
4303 update_stats_wait_start(cfs_rq, prev);
4304 /* Put 'current' back into the tree. */
4305 __enqueue_entity(cfs_rq, prev);
4306 /* in !on_rq case, update occurred at dequeue */
4307 update_load_avg(cfs_rq, prev, 0);
4309 cfs_rq->curr = NULL;
4313 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4316 * Update run-time statistics of the 'current'.
4318 update_curr(cfs_rq);
4321 * Ensure that runnable average is periodically updated.
4323 update_load_avg(cfs_rq, curr, UPDATE_TG);
4324 update_cfs_group(curr);
4326 #ifdef CONFIG_SCHED_HRTICK
4328 * queued ticks are scheduled to match the slice, so don't bother
4329 * validating it and just reschedule.
4332 resched_curr(rq_of(cfs_rq));
4336 * don't let the period tick interfere with the hrtick preemption
4338 if (!sched_feat(DOUBLE_TICK) &&
4339 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4343 if (cfs_rq->nr_running > 1)
4344 check_preempt_tick(cfs_rq, curr);
4348 /**************************************************
4349 * CFS bandwidth control machinery
4352 #ifdef CONFIG_CFS_BANDWIDTH
4354 #ifdef HAVE_JUMP_LABEL
4355 static struct static_key __cfs_bandwidth_used;
4357 static inline bool cfs_bandwidth_used(void)
4359 return static_key_false(&__cfs_bandwidth_used);
4362 void cfs_bandwidth_usage_inc(void)
4364 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4367 void cfs_bandwidth_usage_dec(void)
4369 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4371 #else /* HAVE_JUMP_LABEL */
4372 static bool cfs_bandwidth_used(void)
4377 void cfs_bandwidth_usage_inc(void) {}
4378 void cfs_bandwidth_usage_dec(void) {}
4379 #endif /* HAVE_JUMP_LABEL */
4382 * default period for cfs group bandwidth.
4383 * default: 0.1s, units: nanoseconds
4385 static inline u64 default_cfs_period(void)
4387 return 100000000ULL;
4390 static inline u64 sched_cfs_bandwidth_slice(void)
4392 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4396 * Replenish runtime according to assigned quota and update expiration time.
4397 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4398 * additional synchronization around rq->lock.
4400 * requires cfs_b->lock
4402 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4406 if (cfs_b->quota == RUNTIME_INF)
4409 now = sched_clock_cpu(smp_processor_id());
4410 cfs_b->runtime = cfs_b->quota;
4411 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4414 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4416 return &tg->cfs_bandwidth;
4419 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4420 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4422 if (unlikely(cfs_rq->throttle_count))
4423 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4425 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4428 /* returns 0 on failure to allocate runtime */
4429 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4431 struct task_group *tg = cfs_rq->tg;
4432 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4433 u64 amount = 0, min_amount, expires;
4435 /* note: this is a positive sum as runtime_remaining <= 0 */
4436 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4438 raw_spin_lock(&cfs_b->lock);
4439 if (cfs_b->quota == RUNTIME_INF)
4440 amount = min_amount;
4442 start_cfs_bandwidth(cfs_b);
4444 if (cfs_b->runtime > 0) {
4445 amount = min(cfs_b->runtime, min_amount);
4446 cfs_b->runtime -= amount;
4450 expires = cfs_b->runtime_expires;
4451 raw_spin_unlock(&cfs_b->lock);
4453 cfs_rq->runtime_remaining += amount;
4455 * we may have advanced our local expiration to account for allowed
4456 * spread between our sched_clock and the one on which runtime was
4459 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4460 cfs_rq->runtime_expires = expires;
4462 return cfs_rq->runtime_remaining > 0;
4466 * Note: This depends on the synchronization provided by sched_clock and the
4467 * fact that rq->clock snapshots this value.
4469 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4471 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4473 /* if the deadline is ahead of our clock, nothing to do */
4474 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4477 if (cfs_rq->runtime_remaining < 0)
4481 * If the local deadline has passed we have to consider the
4482 * possibility that our sched_clock is 'fast' and the global deadline
4483 * has not truly expired.
4485 * Fortunately we can check determine whether this the case by checking
4486 * whether the global deadline has advanced. It is valid to compare
4487 * cfs_b->runtime_expires without any locks since we only care about
4488 * exact equality, so a partial write will still work.
4491 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4492 /* extend local deadline, drift is bounded above by 2 ticks */
4493 cfs_rq->runtime_expires += TICK_NSEC;
4495 /* global deadline is ahead, expiration has passed */
4496 cfs_rq->runtime_remaining = 0;
4500 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4502 /* dock delta_exec before expiring quota (as it could span periods) */
4503 cfs_rq->runtime_remaining -= delta_exec;
4504 expire_cfs_rq_runtime(cfs_rq);
4506 if (likely(cfs_rq->runtime_remaining > 0))
4510 * if we're unable to extend our runtime we resched so that the active
4511 * hierarchy can be throttled
4513 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4514 resched_curr(rq_of(cfs_rq));
4517 static __always_inline
4518 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4520 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4523 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4526 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4528 return cfs_bandwidth_used() && cfs_rq->throttled;
4531 /* check whether cfs_rq, or any parent, is throttled */
4532 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4534 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4538 * Ensure that neither of the group entities corresponding to src_cpu or
4539 * dest_cpu are members of a throttled hierarchy when performing group
4540 * load-balance operations.
4542 static inline int throttled_lb_pair(struct task_group *tg,
4543 int src_cpu, int dest_cpu)
4545 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4547 src_cfs_rq = tg->cfs_rq[src_cpu];
4548 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4550 return throttled_hierarchy(src_cfs_rq) ||
4551 throttled_hierarchy(dest_cfs_rq);
4554 /* updated child weight may affect parent so we have to do this bottom up */
4555 static int tg_unthrottle_up(struct task_group *tg, void *data)
4557 struct rq *rq = data;
4558 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4560 cfs_rq->throttle_count--;
4561 if (!cfs_rq->throttle_count) {
4562 /* adjust cfs_rq_clock_task() */
4563 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4564 cfs_rq->throttled_clock_task;
4570 static int tg_throttle_down(struct task_group *tg, void *data)
4572 struct rq *rq = data;
4573 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4575 /* group is entering throttled state, stop time */
4576 if (!cfs_rq->throttle_count)
4577 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4578 cfs_rq->throttle_count++;
4583 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4585 struct rq *rq = rq_of(cfs_rq);
4586 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4587 struct sched_entity *se;
4588 long task_delta, dequeue = 1;
4591 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4593 /* freeze hierarchy runnable averages while throttled */
4595 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4598 task_delta = cfs_rq->h_nr_running;
4599 for_each_sched_entity(se) {
4600 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4601 /* throttled entity or throttle-on-deactivate */
4606 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4607 qcfs_rq->h_nr_running -= task_delta;
4609 if (qcfs_rq->load.weight)
4614 sub_nr_running(rq, task_delta);
4616 cfs_rq->throttled = 1;
4617 cfs_rq->throttled_clock = rq_clock(rq);
4618 raw_spin_lock(&cfs_b->lock);
4619 empty = list_empty(&cfs_b->throttled_cfs_rq);
4622 * Add to the _head_ of the list, so that an already-started
4623 * distribute_cfs_runtime will not see us
4625 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4628 * If we're the first throttled task, make sure the bandwidth
4632 start_cfs_bandwidth(cfs_b);
4634 raw_spin_unlock(&cfs_b->lock);
4637 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4639 struct rq *rq = rq_of(cfs_rq);
4640 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4641 struct sched_entity *se;
4645 se = cfs_rq->tg->se[cpu_of(rq)];
4647 cfs_rq->throttled = 0;
4649 update_rq_clock(rq);
4651 raw_spin_lock(&cfs_b->lock);
4652 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4653 list_del_rcu(&cfs_rq->throttled_list);
4654 raw_spin_unlock(&cfs_b->lock);
4656 /* update hierarchical throttle state */
4657 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4659 if (!cfs_rq->load.weight)
4662 task_delta = cfs_rq->h_nr_running;
4663 for_each_sched_entity(se) {
4667 cfs_rq = cfs_rq_of(se);
4669 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4670 cfs_rq->h_nr_running += task_delta;
4672 if (cfs_rq_throttled(cfs_rq))
4677 add_nr_running(rq, task_delta);
4679 /* determine whether we need to wake up potentially idle cpu */
4680 if (rq->curr == rq->idle && rq->cfs.nr_running)
4684 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4685 u64 remaining, u64 expires)
4687 struct cfs_rq *cfs_rq;
4689 u64 starting_runtime = remaining;
4692 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4694 struct rq *rq = rq_of(cfs_rq);
4698 if (!cfs_rq_throttled(cfs_rq))
4701 runtime = -cfs_rq->runtime_remaining + 1;
4702 if (runtime > remaining)
4703 runtime = remaining;
4704 remaining -= runtime;
4706 cfs_rq->runtime_remaining += runtime;
4707 cfs_rq->runtime_expires = expires;
4709 /* we check whether we're throttled above */
4710 if (cfs_rq->runtime_remaining > 0)
4711 unthrottle_cfs_rq(cfs_rq);
4721 return starting_runtime - remaining;
4725 * Responsible for refilling a task_group's bandwidth and unthrottling its
4726 * cfs_rqs as appropriate. If there has been no activity within the last
4727 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4728 * used to track this state.
4730 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4732 u64 runtime, runtime_expires;
4735 /* no need to continue the timer with no bandwidth constraint */
4736 if (cfs_b->quota == RUNTIME_INF)
4737 goto out_deactivate;
4739 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4740 cfs_b->nr_periods += overrun;
4743 * idle depends on !throttled (for the case of a large deficit), and if
4744 * we're going inactive then everything else can be deferred
4746 if (cfs_b->idle && !throttled)
4747 goto out_deactivate;
4749 __refill_cfs_bandwidth_runtime(cfs_b);
4752 /* mark as potentially idle for the upcoming period */
4757 /* account preceding periods in which throttling occurred */
4758 cfs_b->nr_throttled += overrun;
4760 runtime_expires = cfs_b->runtime_expires;
4763 * This check is repeated as we are holding onto the new bandwidth while
4764 * we unthrottle. This can potentially race with an unthrottled group
4765 * trying to acquire new bandwidth from the global pool. This can result
4766 * in us over-using our runtime if it is all used during this loop, but
4767 * only by limited amounts in that extreme case.
4769 while (throttled && cfs_b->runtime > 0) {
4770 runtime = cfs_b->runtime;
4771 raw_spin_unlock(&cfs_b->lock);
4772 /* we can't nest cfs_b->lock while distributing bandwidth */
4773 runtime = distribute_cfs_runtime(cfs_b, runtime,
4775 raw_spin_lock(&cfs_b->lock);
4777 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4779 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4783 * While we are ensured activity in the period following an
4784 * unthrottle, this also covers the case in which the new bandwidth is
4785 * insufficient to cover the existing bandwidth deficit. (Forcing the
4786 * timer to remain active while there are any throttled entities.)
4796 /* a cfs_rq won't donate quota below this amount */
4797 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4798 /* minimum remaining period time to redistribute slack quota */
4799 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4800 /* how long we wait to gather additional slack before distributing */
4801 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4804 * Are we near the end of the current quota period?
4806 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4807 * hrtimer base being cleared by hrtimer_start. In the case of
4808 * migrate_hrtimers, base is never cleared, so we are fine.
4810 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4812 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4815 /* if the call-back is running a quota refresh is already occurring */
4816 if (hrtimer_callback_running(refresh_timer))
4819 /* is a quota refresh about to occur? */
4820 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4821 if (remaining < min_expire)
4827 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4829 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4831 /* if there's a quota refresh soon don't bother with slack */
4832 if (runtime_refresh_within(cfs_b, min_left))
4835 hrtimer_start(&cfs_b->slack_timer,
4836 ns_to_ktime(cfs_bandwidth_slack_period),
4840 /* we know any runtime found here is valid as update_curr() precedes return */
4841 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4843 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4844 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4846 if (slack_runtime <= 0)
4849 raw_spin_lock(&cfs_b->lock);
4850 if (cfs_b->quota != RUNTIME_INF &&
4851 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4852 cfs_b->runtime += slack_runtime;
4854 /* we are under rq->lock, defer unthrottling using a timer */
4855 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4856 !list_empty(&cfs_b->throttled_cfs_rq))
4857 start_cfs_slack_bandwidth(cfs_b);
4859 raw_spin_unlock(&cfs_b->lock);
4861 /* even if it's not valid for return we don't want to try again */
4862 cfs_rq->runtime_remaining -= slack_runtime;
4865 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4867 if (!cfs_bandwidth_used())
4870 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4873 __return_cfs_rq_runtime(cfs_rq);
4877 * This is done with a timer (instead of inline with bandwidth return) since
4878 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4880 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4882 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4885 /* confirm we're still not at a refresh boundary */
4886 raw_spin_lock(&cfs_b->lock);
4887 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4888 raw_spin_unlock(&cfs_b->lock);
4892 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4893 runtime = cfs_b->runtime;
4895 expires = cfs_b->runtime_expires;
4896 raw_spin_unlock(&cfs_b->lock);
4901 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4903 raw_spin_lock(&cfs_b->lock);
4904 if (expires == cfs_b->runtime_expires)
4905 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4906 raw_spin_unlock(&cfs_b->lock);
4910 * When a group wakes up we want to make sure that its quota is not already
4911 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4912 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4914 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4916 if (!cfs_bandwidth_used())
4919 /* an active group must be handled by the update_curr()->put() path */
4920 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4923 /* ensure the group is not already throttled */
4924 if (cfs_rq_throttled(cfs_rq))
4927 /* update runtime allocation */
4928 account_cfs_rq_runtime(cfs_rq, 0);
4929 if (cfs_rq->runtime_remaining <= 0)
4930 throttle_cfs_rq(cfs_rq);
4933 static void sync_throttle(struct task_group *tg, int cpu)
4935 struct cfs_rq *pcfs_rq, *cfs_rq;
4937 if (!cfs_bandwidth_used())
4943 cfs_rq = tg->cfs_rq[cpu];
4944 pcfs_rq = tg->parent->cfs_rq[cpu];
4946 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4947 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4950 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4951 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4953 if (!cfs_bandwidth_used())
4956 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4960 * it's possible for a throttled entity to be forced into a running
4961 * state (e.g. set_curr_task), in this case we're finished.
4963 if (cfs_rq_throttled(cfs_rq))
4966 throttle_cfs_rq(cfs_rq);
4970 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4972 struct cfs_bandwidth *cfs_b =
4973 container_of(timer, struct cfs_bandwidth, slack_timer);
4975 do_sched_cfs_slack_timer(cfs_b);
4977 return HRTIMER_NORESTART;
4980 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4982 struct cfs_bandwidth *cfs_b =
4983 container_of(timer, struct cfs_bandwidth, period_timer);
4987 raw_spin_lock(&cfs_b->lock);
4989 overrun = hrtimer_forward_now(timer, cfs_b->period);
4993 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4996 cfs_b->period_active = 0;
4997 raw_spin_unlock(&cfs_b->lock);
4999 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5002 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5004 raw_spin_lock_init(&cfs_b->lock);
5006 cfs_b->quota = RUNTIME_INF;
5007 cfs_b->period = ns_to_ktime(default_cfs_period());
5009 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5010 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5011 cfs_b->period_timer.function = sched_cfs_period_timer;
5012 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5013 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5016 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5018 cfs_rq->runtime_enabled = 0;
5019 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5022 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5024 lockdep_assert_held(&cfs_b->lock);
5026 if (!cfs_b->period_active) {
5027 cfs_b->period_active = 1;
5028 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5029 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5033 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5035 /* init_cfs_bandwidth() was not called */
5036 if (!cfs_b->throttled_cfs_rq.next)
5039 hrtimer_cancel(&cfs_b->period_timer);
5040 hrtimer_cancel(&cfs_b->slack_timer);
5044 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
5046 * The race is harmless, since modifying bandwidth settings of unhooked group
5047 * bits doesn't do much.
5050 /* cpu online calback */
5051 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5053 struct task_group *tg;
5055 lockdep_assert_held(&rq->lock);
5058 list_for_each_entry_rcu(tg, &task_groups, list) {
5059 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5060 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5062 raw_spin_lock(&cfs_b->lock);
5063 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5064 raw_spin_unlock(&cfs_b->lock);
5069 /* cpu offline callback */
5070 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5072 struct task_group *tg;
5074 lockdep_assert_held(&rq->lock);
5077 list_for_each_entry_rcu(tg, &task_groups, list) {
5078 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5080 if (!cfs_rq->runtime_enabled)
5084 * clock_task is not advancing so we just need to make sure
5085 * there's some valid quota amount
5087 cfs_rq->runtime_remaining = 1;
5089 * Offline rq is schedulable till cpu is completely disabled
5090 * in take_cpu_down(), so we prevent new cfs throttling here.
5092 cfs_rq->runtime_enabled = 0;
5094 if (cfs_rq_throttled(cfs_rq))
5095 unthrottle_cfs_rq(cfs_rq);
5100 #else /* CONFIG_CFS_BANDWIDTH */
5101 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5103 return rq_clock_task(rq_of(cfs_rq));
5106 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5107 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5108 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5109 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5110 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5112 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5117 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5122 static inline int throttled_lb_pair(struct task_group *tg,
5123 int src_cpu, int dest_cpu)
5128 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5130 #ifdef CONFIG_FAIR_GROUP_SCHED
5131 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5134 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5138 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5139 static inline void update_runtime_enabled(struct rq *rq) {}
5140 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5142 #endif /* CONFIG_CFS_BANDWIDTH */
5144 /**************************************************
5145 * CFS operations on tasks:
5148 #ifdef CONFIG_SCHED_HRTICK
5149 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5151 struct sched_entity *se = &p->se;
5152 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5154 SCHED_WARN_ON(task_rq(p) != rq);
5156 if (rq->cfs.h_nr_running > 1) {
5157 u64 slice = sched_slice(cfs_rq, se);
5158 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5159 s64 delta = slice - ran;
5166 hrtick_start(rq, delta);
5171 * called from enqueue/dequeue and updates the hrtick when the
5172 * current task is from our class and nr_running is low enough
5175 static void hrtick_update(struct rq *rq)
5177 struct task_struct *curr = rq->curr;
5179 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5182 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5183 hrtick_start_fair(rq, curr);
5185 #else /* !CONFIG_SCHED_HRTICK */
5187 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5191 static inline void hrtick_update(struct rq *rq)
5197 * The enqueue_task method is called before nr_running is
5198 * increased. Here we update the fair scheduling stats and
5199 * then put the task into the rbtree:
5202 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5204 struct cfs_rq *cfs_rq;
5205 struct sched_entity *se = &p->se;
5208 * If in_iowait is set, the code below may not trigger any cpufreq
5209 * utilization updates, so do it here explicitly with the IOWAIT flag
5213 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5215 for_each_sched_entity(se) {
5218 cfs_rq = cfs_rq_of(se);
5219 enqueue_entity(cfs_rq, se, flags);
5222 * end evaluation on encountering a throttled cfs_rq
5224 * note: in the case of encountering a throttled cfs_rq we will
5225 * post the final h_nr_running increment below.
5227 if (cfs_rq_throttled(cfs_rq))
5229 cfs_rq->h_nr_running++;
5231 flags = ENQUEUE_WAKEUP;
5234 for_each_sched_entity(se) {
5235 cfs_rq = cfs_rq_of(se);
5236 cfs_rq->h_nr_running++;
5238 if (cfs_rq_throttled(cfs_rq))
5241 update_load_avg(cfs_rq, se, UPDATE_TG);
5242 update_cfs_group(se);
5246 add_nr_running(rq, 1);
5251 static void set_next_buddy(struct sched_entity *se);
5254 * The dequeue_task method is called before nr_running is
5255 * decreased. We remove the task from the rbtree and
5256 * update the fair scheduling stats:
5258 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5260 struct cfs_rq *cfs_rq;
5261 struct sched_entity *se = &p->se;
5262 int task_sleep = flags & DEQUEUE_SLEEP;
5264 for_each_sched_entity(se) {
5265 cfs_rq = cfs_rq_of(se);
5266 dequeue_entity(cfs_rq, se, flags);
5269 * end evaluation on encountering a throttled cfs_rq
5271 * note: in the case of encountering a throttled cfs_rq we will
5272 * post the final h_nr_running decrement below.
5274 if (cfs_rq_throttled(cfs_rq))
5276 cfs_rq->h_nr_running--;
5278 /* Don't dequeue parent if it has other entities besides us */
5279 if (cfs_rq->load.weight) {
5280 /* Avoid re-evaluating load for this entity: */
5281 se = parent_entity(se);
5283 * Bias pick_next to pick a task from this cfs_rq, as
5284 * p is sleeping when it is within its sched_slice.
5286 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5290 flags |= DEQUEUE_SLEEP;
5293 for_each_sched_entity(se) {
5294 cfs_rq = cfs_rq_of(se);
5295 cfs_rq->h_nr_running--;
5297 if (cfs_rq_throttled(cfs_rq))
5300 update_load_avg(cfs_rq, se, UPDATE_TG);
5301 update_cfs_group(se);
5305 sub_nr_running(rq, 1);
5312 /* Working cpumask for: load_balance, load_balance_newidle. */
5313 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5314 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5316 #ifdef CONFIG_NO_HZ_COMMON
5318 * per rq 'load' arrray crap; XXX kill this.
5322 * The exact cpuload calculated at every tick would be:
5324 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5326 * If a cpu misses updates for n ticks (as it was idle) and update gets
5327 * called on the n+1-th tick when cpu may be busy, then we have:
5329 * load_n = (1 - 1/2^i)^n * load_0
5330 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5332 * decay_load_missed() below does efficient calculation of
5334 * load' = (1 - 1/2^i)^n * load
5336 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5337 * This allows us to precompute the above in said factors, thereby allowing the
5338 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5339 * fixed_power_int())
5341 * The calculation is approximated on a 128 point scale.
5343 #define DEGRADE_SHIFT 7
5345 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5346 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5347 { 0, 0, 0, 0, 0, 0, 0, 0 },
5348 { 64, 32, 8, 0, 0, 0, 0, 0 },
5349 { 96, 72, 40, 12, 1, 0, 0, 0 },
5350 { 112, 98, 75, 43, 15, 1, 0, 0 },
5351 { 120, 112, 98, 76, 45, 16, 2, 0 }
5355 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5356 * would be when CPU is idle and so we just decay the old load without
5357 * adding any new load.
5359 static unsigned long
5360 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5364 if (!missed_updates)
5367 if (missed_updates >= degrade_zero_ticks[idx])
5371 return load >> missed_updates;
5373 while (missed_updates) {
5374 if (missed_updates % 2)
5375 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5377 missed_updates >>= 1;
5382 #endif /* CONFIG_NO_HZ_COMMON */
5385 * __cpu_load_update - update the rq->cpu_load[] statistics
5386 * @this_rq: The rq to update statistics for
5387 * @this_load: The current load
5388 * @pending_updates: The number of missed updates
5390 * Update rq->cpu_load[] statistics. This function is usually called every
5391 * scheduler tick (TICK_NSEC).
5393 * This function computes a decaying average:
5395 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5397 * Because of NOHZ it might not get called on every tick which gives need for
5398 * the @pending_updates argument.
5400 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5401 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5402 * = A * (A * load[i]_n-2 + B) + B
5403 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5404 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5405 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5406 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5407 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5409 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5410 * any change in load would have resulted in the tick being turned back on.
5412 * For regular NOHZ, this reduces to:
5414 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5416 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5419 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5420 unsigned long pending_updates)
5422 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5425 this_rq->nr_load_updates++;
5427 /* Update our load: */
5428 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5429 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5430 unsigned long old_load, new_load;
5432 /* scale is effectively 1 << i now, and >> i divides by scale */
5434 old_load = this_rq->cpu_load[i];
5435 #ifdef CONFIG_NO_HZ_COMMON
5436 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5437 if (tickless_load) {
5438 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5440 * old_load can never be a negative value because a
5441 * decayed tickless_load cannot be greater than the
5442 * original tickless_load.
5444 old_load += tickless_load;
5447 new_load = this_load;
5449 * Round up the averaging division if load is increasing. This
5450 * prevents us from getting stuck on 9 if the load is 10, for
5453 if (new_load > old_load)
5454 new_load += scale - 1;
5456 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5459 sched_avg_update(this_rq);
5462 /* Used instead of source_load when we know the type == 0 */
5463 static unsigned long weighted_cpuload(struct rq *rq)
5465 return cfs_rq_runnable_load_avg(&rq->cfs);
5468 #ifdef CONFIG_NO_HZ_COMMON
5470 * There is no sane way to deal with nohz on smp when using jiffies because the
5471 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5472 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5474 * Therefore we need to avoid the delta approach from the regular tick when
5475 * possible since that would seriously skew the load calculation. This is why we
5476 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5477 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5478 * loop exit, nohz_idle_balance, nohz full exit...)
5480 * This means we might still be one tick off for nohz periods.
5483 static void cpu_load_update_nohz(struct rq *this_rq,
5484 unsigned long curr_jiffies,
5487 unsigned long pending_updates;
5489 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5490 if (pending_updates) {
5491 this_rq->last_load_update_tick = curr_jiffies;
5493 * In the regular NOHZ case, we were idle, this means load 0.
5494 * In the NOHZ_FULL case, we were non-idle, we should consider
5495 * its weighted load.
5497 cpu_load_update(this_rq, load, pending_updates);
5502 * Called from nohz_idle_balance() to update the load ratings before doing the
5505 static void cpu_load_update_idle(struct rq *this_rq)
5508 * bail if there's load or we're actually up-to-date.
5510 if (weighted_cpuload(this_rq))
5513 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5517 * Record CPU load on nohz entry so we know the tickless load to account
5518 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5519 * than other cpu_load[idx] but it should be fine as cpu_load readers
5520 * shouldn't rely into synchronized cpu_load[*] updates.
5522 void cpu_load_update_nohz_start(void)
5524 struct rq *this_rq = this_rq();
5527 * This is all lockless but should be fine. If weighted_cpuload changes
5528 * concurrently we'll exit nohz. And cpu_load write can race with
5529 * cpu_load_update_idle() but both updater would be writing the same.
5531 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5535 * Account the tickless load in the end of a nohz frame.
5537 void cpu_load_update_nohz_stop(void)
5539 unsigned long curr_jiffies = READ_ONCE(jiffies);
5540 struct rq *this_rq = this_rq();
5544 if (curr_jiffies == this_rq->last_load_update_tick)
5547 load = weighted_cpuload(this_rq);
5548 rq_lock(this_rq, &rf);
5549 update_rq_clock(this_rq);
5550 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5551 rq_unlock(this_rq, &rf);
5553 #else /* !CONFIG_NO_HZ_COMMON */
5554 static inline void cpu_load_update_nohz(struct rq *this_rq,
5555 unsigned long curr_jiffies,
5556 unsigned long load) { }
5557 #endif /* CONFIG_NO_HZ_COMMON */
5559 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5561 #ifdef CONFIG_NO_HZ_COMMON
5562 /* See the mess around cpu_load_update_nohz(). */
5563 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5565 cpu_load_update(this_rq, load, 1);
5569 * Called from scheduler_tick()
5571 void cpu_load_update_active(struct rq *this_rq)
5573 unsigned long load = weighted_cpuload(this_rq);
5575 if (tick_nohz_tick_stopped())
5576 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5578 cpu_load_update_periodic(this_rq, load);
5582 * Return a low guess at the load of a migration-source cpu weighted
5583 * according to the scheduling class and "nice" value.
5585 * We want to under-estimate the load of migration sources, to
5586 * balance conservatively.
5588 static unsigned long source_load(int cpu, int type)
5590 struct rq *rq = cpu_rq(cpu);
5591 unsigned long total = weighted_cpuload(rq);
5593 if (type == 0 || !sched_feat(LB_BIAS))
5596 return min(rq->cpu_load[type-1], total);
5600 * Return a high guess at the load of a migration-target cpu weighted
5601 * according to the scheduling class and "nice" value.
5603 static unsigned long target_load(int cpu, int type)
5605 struct rq *rq = cpu_rq(cpu);
5606 unsigned long total = weighted_cpuload(rq);
5608 if (type == 0 || !sched_feat(LB_BIAS))
5611 return max(rq->cpu_load[type-1], total);
5614 static unsigned long capacity_of(int cpu)
5616 return cpu_rq(cpu)->cpu_capacity;
5619 static unsigned long capacity_orig_of(int cpu)
5621 return cpu_rq(cpu)->cpu_capacity_orig;
5624 static unsigned long cpu_avg_load_per_task(int cpu)
5626 struct rq *rq = cpu_rq(cpu);
5627 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5628 unsigned long load_avg = weighted_cpuload(rq);
5631 return load_avg / nr_running;
5636 static void record_wakee(struct task_struct *p)
5639 * Only decay a single time; tasks that have less then 1 wakeup per
5640 * jiffy will not have built up many flips.
5642 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5643 current->wakee_flips >>= 1;
5644 current->wakee_flip_decay_ts = jiffies;
5647 if (current->last_wakee != p) {
5648 current->last_wakee = p;
5649 current->wakee_flips++;
5654 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5656 * A waker of many should wake a different task than the one last awakened
5657 * at a frequency roughly N times higher than one of its wakees.
5659 * In order to determine whether we should let the load spread vs consolidating
5660 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5661 * partner, and a factor of lls_size higher frequency in the other.
5663 * With both conditions met, we can be relatively sure that the relationship is
5664 * non-monogamous, with partner count exceeding socket size.
5666 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5667 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5670 static int wake_wide(struct task_struct *p)
5672 unsigned int master = current->wakee_flips;
5673 unsigned int slave = p->wakee_flips;
5674 int factor = this_cpu_read(sd_llc_size);
5677 swap(master, slave);
5678 if (slave < factor || master < slave * factor)
5684 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5685 * soonest. For the purpose of speed we only consider the waking and previous
5688 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5689 * cache-affine and is (or will be) idle.
5691 * wake_affine_weight() - considers the weight to reflect the average
5692 * scheduling latency of the CPUs. This seems to work
5693 * for the overloaded case.
5696 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5699 * If this_cpu is idle, it implies the wakeup is from interrupt
5700 * context. Only allow the move if cache is shared. Otherwise an
5701 * interrupt intensive workload could force all tasks onto one
5702 * node depending on the IO topology or IRQ affinity settings.
5704 * If the prev_cpu is idle and cache affine then avoid a migration.
5705 * There is no guarantee that the cache hot data from an interrupt
5706 * is more important than cache hot data on the prev_cpu and from
5707 * a cpufreq perspective, it's better to have higher utilisation
5710 if (idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5711 return idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5713 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5716 return nr_cpumask_bits;
5720 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5721 int this_cpu, int prev_cpu, int sync)
5723 s64 this_eff_load, prev_eff_load;
5724 unsigned long task_load;
5726 this_eff_load = target_load(this_cpu, sd->wake_idx);
5729 unsigned long current_load = task_h_load(current);
5731 if (current_load > this_eff_load)
5734 this_eff_load -= current_load;
5737 task_load = task_h_load(p);
5739 this_eff_load += task_load;
5740 if (sched_feat(WA_BIAS))
5741 this_eff_load *= 100;
5742 this_eff_load *= capacity_of(prev_cpu);
5744 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5745 prev_eff_load -= task_load;
5746 if (sched_feat(WA_BIAS))
5747 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5748 prev_eff_load *= capacity_of(this_cpu);
5751 * If sync, adjust the weight of prev_eff_load such that if
5752 * prev_eff == this_eff that select_idle_sibling() will consider
5753 * stacking the wakee on top of the waker if no other CPU is
5759 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5762 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5763 int this_cpu, int prev_cpu, int sync)
5765 int target = nr_cpumask_bits;
5767 if (sched_feat(WA_IDLE))
5768 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5770 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5771 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5773 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5774 if (target == nr_cpumask_bits)
5777 schedstat_inc(sd->ttwu_move_affine);
5778 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5782 static inline unsigned long task_util(struct task_struct *p);
5783 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5785 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5787 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5791 * find_idlest_group finds and returns the least busy CPU group within the
5794 * Assumes p is allowed on at least one CPU in sd.
5796 static struct sched_group *
5797 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5798 int this_cpu, int sd_flag)
5800 struct sched_group *idlest = NULL, *group = sd->groups;
5801 struct sched_group *most_spare_sg = NULL;
5802 unsigned long min_runnable_load = ULONG_MAX;
5803 unsigned long this_runnable_load = ULONG_MAX;
5804 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5805 unsigned long most_spare = 0, this_spare = 0;
5806 int load_idx = sd->forkexec_idx;
5807 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5808 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5809 (sd->imbalance_pct-100) / 100;
5811 if (sd_flag & SD_BALANCE_WAKE)
5812 load_idx = sd->wake_idx;
5815 unsigned long load, avg_load, runnable_load;
5816 unsigned long spare_cap, max_spare_cap;
5820 /* Skip over this group if it has no CPUs allowed */
5821 if (!cpumask_intersects(sched_group_span(group),
5825 local_group = cpumask_test_cpu(this_cpu,
5826 sched_group_span(group));
5829 * Tally up the load of all CPUs in the group and find
5830 * the group containing the CPU with most spare capacity.
5836 for_each_cpu(i, sched_group_span(group)) {
5837 /* Bias balancing toward cpus of our domain */
5839 load = source_load(i, load_idx);
5841 load = target_load(i, load_idx);
5843 runnable_load += load;
5845 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5847 spare_cap = capacity_spare_wake(i, p);
5849 if (spare_cap > max_spare_cap)
5850 max_spare_cap = spare_cap;
5853 /* Adjust by relative CPU capacity of the group */
5854 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5855 group->sgc->capacity;
5856 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5857 group->sgc->capacity;
5860 this_runnable_load = runnable_load;
5861 this_avg_load = avg_load;
5862 this_spare = max_spare_cap;
5864 if (min_runnable_load > (runnable_load + imbalance)) {
5866 * The runnable load is significantly smaller
5867 * so we can pick this new cpu
5869 min_runnable_load = runnable_load;
5870 min_avg_load = avg_load;
5872 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5873 (100*min_avg_load > imbalance_scale*avg_load)) {
5875 * The runnable loads are close so take the
5876 * blocked load into account through avg_load.
5878 min_avg_load = avg_load;
5882 if (most_spare < max_spare_cap) {
5883 most_spare = max_spare_cap;
5884 most_spare_sg = group;
5887 } while (group = group->next, group != sd->groups);
5890 * The cross-over point between using spare capacity or least load
5891 * is too conservative for high utilization tasks on partially
5892 * utilized systems if we require spare_capacity > task_util(p),
5893 * so we allow for some task stuffing by using
5894 * spare_capacity > task_util(p)/2.
5896 * Spare capacity can't be used for fork because the utilization has
5897 * not been set yet, we must first select a rq to compute the initial
5900 if (sd_flag & SD_BALANCE_FORK)
5903 if (this_spare > task_util(p) / 2 &&
5904 imbalance_scale*this_spare > 100*most_spare)
5907 if (most_spare > task_util(p) / 2)
5908 return most_spare_sg;
5914 if (min_runnable_load > (this_runnable_load + imbalance))
5917 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5918 (100*this_avg_load < imbalance_scale*min_avg_load))
5925 * find_idlest_group_cpu - find the idlest cpu among the cpus in group.
5928 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5930 unsigned long load, min_load = ULONG_MAX;
5931 unsigned int min_exit_latency = UINT_MAX;
5932 u64 latest_idle_timestamp = 0;
5933 int least_loaded_cpu = this_cpu;
5934 int shallowest_idle_cpu = -1;
5937 /* Check if we have any choice: */
5938 if (group->group_weight == 1)
5939 return cpumask_first(sched_group_span(group));
5941 /* Traverse only the allowed CPUs */
5942 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5944 struct rq *rq = cpu_rq(i);
5945 struct cpuidle_state *idle = idle_get_state(rq);
5946 if (idle && idle->exit_latency < min_exit_latency) {
5948 * We give priority to a CPU whose idle state
5949 * has the smallest exit latency irrespective
5950 * of any idle timestamp.
5952 min_exit_latency = idle->exit_latency;
5953 latest_idle_timestamp = rq->idle_stamp;
5954 shallowest_idle_cpu = i;
5955 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5956 rq->idle_stamp > latest_idle_timestamp) {
5958 * If equal or no active idle state, then
5959 * the most recently idled CPU might have
5962 latest_idle_timestamp = rq->idle_stamp;
5963 shallowest_idle_cpu = i;
5965 } else if (shallowest_idle_cpu == -1) {
5966 load = weighted_cpuload(cpu_rq(i));
5967 if (load < min_load) {
5969 least_loaded_cpu = i;
5974 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5977 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5978 int cpu, int prev_cpu, int sd_flag)
5982 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5986 struct sched_group *group;
5987 struct sched_domain *tmp;
5990 if (!(sd->flags & sd_flag)) {
5995 group = find_idlest_group(sd, p, cpu, sd_flag);
6001 new_cpu = find_idlest_group_cpu(group, p, cpu);
6002 if (new_cpu == cpu) {
6003 /* Now try balancing at a lower domain level of cpu */
6008 /* Now try balancing at a lower domain level of new_cpu */
6010 weight = sd->span_weight;
6012 for_each_domain(cpu, tmp) {
6013 if (weight <= tmp->span_weight)
6015 if (tmp->flags & sd_flag)
6018 /* while loop will break here if sd == NULL */
6024 #ifdef CONFIG_SCHED_SMT
6026 static inline void set_idle_cores(int cpu, int val)
6028 struct sched_domain_shared *sds;
6030 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6032 WRITE_ONCE(sds->has_idle_cores, val);
6035 static inline bool test_idle_cores(int cpu, bool def)
6037 struct sched_domain_shared *sds;
6039 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6041 return READ_ONCE(sds->has_idle_cores);
6047 * Scans the local SMT mask to see if the entire core is idle, and records this
6048 * information in sd_llc_shared->has_idle_cores.
6050 * Since SMT siblings share all cache levels, inspecting this limited remote
6051 * state should be fairly cheap.
6053 void __update_idle_core(struct rq *rq)
6055 int core = cpu_of(rq);
6059 if (test_idle_cores(core, true))
6062 for_each_cpu(cpu, cpu_smt_mask(core)) {
6070 set_idle_cores(core, 1);
6076 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6077 * there are no idle cores left in the system; tracked through
6078 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6080 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6082 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6085 if (!static_branch_likely(&sched_smt_present))
6088 if (!test_idle_cores(target, false))
6091 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6093 for_each_cpu_wrap(core, cpus, target) {
6096 for_each_cpu(cpu, cpu_smt_mask(core)) {
6097 cpumask_clear_cpu(cpu, cpus);
6107 * Failed to find an idle core; stop looking for one.
6109 set_idle_cores(target, 0);
6115 * Scan the local SMT mask for idle CPUs.
6117 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6121 if (!static_branch_likely(&sched_smt_present))
6124 for_each_cpu(cpu, cpu_smt_mask(target)) {
6125 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6134 #else /* CONFIG_SCHED_SMT */
6136 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6141 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6146 #endif /* CONFIG_SCHED_SMT */
6149 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6150 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6151 * average idle time for this rq (as found in rq->avg_idle).
6153 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6155 struct sched_domain *this_sd;
6156 u64 avg_cost, avg_idle;
6159 int cpu, nr = INT_MAX;
6161 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6166 * Due to large variance we need a large fuzz factor; hackbench in
6167 * particularly is sensitive here.
6169 avg_idle = this_rq()->avg_idle / 512;
6170 avg_cost = this_sd->avg_scan_cost + 1;
6172 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6175 if (sched_feat(SIS_PROP)) {
6176 u64 span_avg = sd->span_weight * avg_idle;
6177 if (span_avg > 4*avg_cost)
6178 nr = div_u64(span_avg, avg_cost);
6183 time = local_clock();
6185 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6188 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6194 time = local_clock() - time;
6195 cost = this_sd->avg_scan_cost;
6196 delta = (s64)(time - cost) / 8;
6197 this_sd->avg_scan_cost += delta;
6203 * Try and locate an idle core/thread in the LLC cache domain.
6205 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6207 struct sched_domain *sd;
6208 int i, recent_used_cpu;
6210 if (idle_cpu(target))
6214 * If the previous cpu is cache affine and idle, don't be stupid.
6216 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6219 /* Check a recently used CPU as a potential idle candidate */
6220 recent_used_cpu = p->recent_used_cpu;
6221 if (recent_used_cpu != prev &&
6222 recent_used_cpu != target &&
6223 cpus_share_cache(recent_used_cpu, target) &&
6224 idle_cpu(recent_used_cpu) &&
6225 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6227 * Replace recent_used_cpu with prev as it is a potential
6228 * candidate for the next wake.
6230 p->recent_used_cpu = prev;
6231 return recent_used_cpu;
6234 sd = rcu_dereference(per_cpu(sd_llc, target));
6238 i = select_idle_core(p, sd, target);
6239 if ((unsigned)i < nr_cpumask_bits)
6242 i = select_idle_cpu(p, sd, target);
6243 if ((unsigned)i < nr_cpumask_bits)
6246 i = select_idle_smt(p, sd, target);
6247 if ((unsigned)i < nr_cpumask_bits)
6254 * cpu_util returns the amount of capacity of a CPU that is used by CFS
6255 * tasks. The unit of the return value must be the one of capacity so we can
6256 * compare the utilization with the capacity of the CPU that is available for
6257 * CFS task (ie cpu_capacity).
6259 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6260 * recent utilization of currently non-runnable tasks on a CPU. It represents
6261 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6262 * capacity_orig is the cpu_capacity available at the highest frequency
6263 * (arch_scale_freq_capacity()).
6264 * The utilization of a CPU converges towards a sum equal to or less than the
6265 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6266 * the running time on this CPU scaled by capacity_curr.
6268 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6269 * higher than capacity_orig because of unfortunate rounding in
6270 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6271 * the average stabilizes with the new running time. We need to check that the
6272 * utilization stays within the range of [0..capacity_orig] and cap it if
6273 * necessary. Without utilization capping, a group could be seen as overloaded
6274 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6275 * available capacity. We allow utilization to overshoot capacity_curr (but not
6276 * capacity_orig) as it useful for predicting the capacity required after task
6277 * migrations (scheduler-driven DVFS).
6279 static unsigned long cpu_util(int cpu)
6281 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
6282 unsigned long capacity = capacity_orig_of(cpu);
6284 return (util >= capacity) ? capacity : util;
6287 static inline unsigned long task_util(struct task_struct *p)
6289 return p->se.avg.util_avg;
6293 * cpu_util_wake: Compute cpu utilization with any contributions from
6294 * the waking task p removed.
6296 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6298 unsigned long util, capacity;
6300 /* Task has no contribution or is new */
6301 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
6302 return cpu_util(cpu);
6304 capacity = capacity_orig_of(cpu);
6305 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
6307 return (util >= capacity) ? capacity : util;
6311 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6312 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6314 * In that case WAKE_AFFINE doesn't make sense and we'll let
6315 * BALANCE_WAKE sort things out.
6317 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6319 long min_cap, max_cap;
6321 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6322 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6324 /* Minimum capacity is close to max, no need to abort wake_affine */
6325 if (max_cap - min_cap < max_cap >> 3)
6328 /* Bring task utilization in sync with prev_cpu */
6329 sync_entity_load_avg(&p->se);
6331 return min_cap * 1024 < task_util(p) * capacity_margin;
6335 * select_task_rq_fair: Select target runqueue for the waking task in domains
6336 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6337 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6339 * Balances load by selecting the idlest cpu in the idlest group, or under
6340 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
6342 * Returns the target cpu number.
6344 * preempt must be disabled.
6347 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6349 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6350 int cpu = smp_processor_id();
6351 int new_cpu = prev_cpu;
6352 int want_affine = 0;
6353 int sync = wake_flags & WF_SYNC;
6355 if (sd_flag & SD_BALANCE_WAKE) {
6357 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6358 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6362 for_each_domain(cpu, tmp) {
6363 if (!(tmp->flags & SD_LOAD_BALANCE))
6367 * If both cpu and prev_cpu are part of this domain,
6368 * cpu is a valid SD_WAKE_AFFINE target.
6370 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6371 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6376 if (tmp->flags & sd_flag)
6378 else if (!want_affine)
6383 sd = NULL; /* Prefer wake_affine over balance flags */
6384 if (cpu == prev_cpu)
6387 new_cpu = wake_affine(affine_sd, p, cpu, prev_cpu, sync);
6390 if (sd && !(sd_flag & SD_BALANCE_FORK)) {
6392 * We're going to need the task's util for capacity_spare_wake
6393 * in find_idlest_group. Sync it up to prev_cpu's
6396 sync_entity_load_avg(&p->se);
6401 if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6402 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6405 current->recent_used_cpu = cpu;
6408 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6415 static void detach_entity_cfs_rq(struct sched_entity *se);
6418 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6419 * cfs_rq_of(p) references at time of call are still valid and identify the
6420 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6422 static void migrate_task_rq_fair(struct task_struct *p)
6425 * As blocked tasks retain absolute vruntime the migration needs to
6426 * deal with this by subtracting the old and adding the new
6427 * min_vruntime -- the latter is done by enqueue_entity() when placing
6428 * the task on the new runqueue.
6430 if (p->state == TASK_WAKING) {
6431 struct sched_entity *se = &p->se;
6432 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6435 #ifndef CONFIG_64BIT
6436 u64 min_vruntime_copy;
6439 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6441 min_vruntime = cfs_rq->min_vruntime;
6442 } while (min_vruntime != min_vruntime_copy);
6444 min_vruntime = cfs_rq->min_vruntime;
6447 se->vruntime -= min_vruntime;
6450 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6452 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6453 * rq->lock and can modify state directly.
6455 lockdep_assert_held(&task_rq(p)->lock);
6456 detach_entity_cfs_rq(&p->se);
6460 * We are supposed to update the task to "current" time, then
6461 * its up to date and ready to go to new CPU/cfs_rq. But we
6462 * have difficulty in getting what current time is, so simply
6463 * throw away the out-of-date time. This will result in the
6464 * wakee task is less decayed, but giving the wakee more load
6467 remove_entity_load_avg(&p->se);
6470 /* Tell new CPU we are migrated */
6471 p->se.avg.last_update_time = 0;
6473 /* We have migrated, no longer consider this task hot */
6474 p->se.exec_start = 0;
6477 static void task_dead_fair(struct task_struct *p)
6479 remove_entity_load_avg(&p->se);
6481 #endif /* CONFIG_SMP */
6483 static unsigned long wakeup_gran(struct sched_entity *se)
6485 unsigned long gran = sysctl_sched_wakeup_granularity;
6488 * Since its curr running now, convert the gran from real-time
6489 * to virtual-time in his units.
6491 * By using 'se' instead of 'curr' we penalize light tasks, so
6492 * they get preempted easier. That is, if 'se' < 'curr' then
6493 * the resulting gran will be larger, therefore penalizing the
6494 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6495 * be smaller, again penalizing the lighter task.
6497 * This is especially important for buddies when the leftmost
6498 * task is higher priority than the buddy.
6500 return calc_delta_fair(gran, se);
6504 * Should 'se' preempt 'curr'.
6518 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6520 s64 gran, vdiff = curr->vruntime - se->vruntime;
6525 gran = wakeup_gran(se);
6532 static void set_last_buddy(struct sched_entity *se)
6534 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6537 for_each_sched_entity(se) {
6538 if (SCHED_WARN_ON(!se->on_rq))
6540 cfs_rq_of(se)->last = se;
6544 static void set_next_buddy(struct sched_entity *se)
6546 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6549 for_each_sched_entity(se) {
6550 if (SCHED_WARN_ON(!se->on_rq))
6552 cfs_rq_of(se)->next = se;
6556 static void set_skip_buddy(struct sched_entity *se)
6558 for_each_sched_entity(se)
6559 cfs_rq_of(se)->skip = se;
6563 * Preempt the current task with a newly woken task if needed:
6565 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6567 struct task_struct *curr = rq->curr;
6568 struct sched_entity *se = &curr->se, *pse = &p->se;
6569 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6570 int scale = cfs_rq->nr_running >= sched_nr_latency;
6571 int next_buddy_marked = 0;
6573 if (unlikely(se == pse))
6577 * This is possible from callers such as attach_tasks(), in which we
6578 * unconditionally check_prempt_curr() after an enqueue (which may have
6579 * lead to a throttle). This both saves work and prevents false
6580 * next-buddy nomination below.
6582 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6585 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6586 set_next_buddy(pse);
6587 next_buddy_marked = 1;
6591 * We can come here with TIF_NEED_RESCHED already set from new task
6594 * Note: this also catches the edge-case of curr being in a throttled
6595 * group (e.g. via set_curr_task), since update_curr() (in the
6596 * enqueue of curr) will have resulted in resched being set. This
6597 * prevents us from potentially nominating it as a false LAST_BUDDY
6600 if (test_tsk_need_resched(curr))
6603 /* Idle tasks are by definition preempted by non-idle tasks. */
6604 if (unlikely(curr->policy == SCHED_IDLE) &&
6605 likely(p->policy != SCHED_IDLE))
6609 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6610 * is driven by the tick):
6612 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6615 find_matching_se(&se, &pse);
6616 update_curr(cfs_rq_of(se));
6618 if (wakeup_preempt_entity(se, pse) == 1) {
6620 * Bias pick_next to pick the sched entity that is
6621 * triggering this preemption.
6623 if (!next_buddy_marked)
6624 set_next_buddy(pse);
6633 * Only set the backward buddy when the current task is still
6634 * on the rq. This can happen when a wakeup gets interleaved
6635 * with schedule on the ->pre_schedule() or idle_balance()
6636 * point, either of which can * drop the rq lock.
6638 * Also, during early boot the idle thread is in the fair class,
6639 * for obvious reasons its a bad idea to schedule back to it.
6641 if (unlikely(!se->on_rq || curr == rq->idle))
6644 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6648 static struct task_struct *
6649 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6651 struct cfs_rq *cfs_rq = &rq->cfs;
6652 struct sched_entity *se;
6653 struct task_struct *p;
6657 if (!cfs_rq->nr_running)
6660 #ifdef CONFIG_FAIR_GROUP_SCHED
6661 if (prev->sched_class != &fair_sched_class)
6665 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6666 * likely that a next task is from the same cgroup as the current.
6668 * Therefore attempt to avoid putting and setting the entire cgroup
6669 * hierarchy, only change the part that actually changes.
6673 struct sched_entity *curr = cfs_rq->curr;
6676 * Since we got here without doing put_prev_entity() we also
6677 * have to consider cfs_rq->curr. If it is still a runnable
6678 * entity, update_curr() will update its vruntime, otherwise
6679 * forget we've ever seen it.
6683 update_curr(cfs_rq);
6688 * This call to check_cfs_rq_runtime() will do the
6689 * throttle and dequeue its entity in the parent(s).
6690 * Therefore the nr_running test will indeed
6693 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6696 if (!cfs_rq->nr_running)
6703 se = pick_next_entity(cfs_rq, curr);
6704 cfs_rq = group_cfs_rq(se);
6710 * Since we haven't yet done put_prev_entity and if the selected task
6711 * is a different task than we started out with, try and touch the
6712 * least amount of cfs_rqs.
6715 struct sched_entity *pse = &prev->se;
6717 while (!(cfs_rq = is_same_group(se, pse))) {
6718 int se_depth = se->depth;
6719 int pse_depth = pse->depth;
6721 if (se_depth <= pse_depth) {
6722 put_prev_entity(cfs_rq_of(pse), pse);
6723 pse = parent_entity(pse);
6725 if (se_depth >= pse_depth) {
6726 set_next_entity(cfs_rq_of(se), se);
6727 se = parent_entity(se);
6731 put_prev_entity(cfs_rq, pse);
6732 set_next_entity(cfs_rq, se);
6739 put_prev_task(rq, prev);
6742 se = pick_next_entity(cfs_rq, NULL);
6743 set_next_entity(cfs_rq, se);
6744 cfs_rq = group_cfs_rq(se);
6749 done: __maybe_unused
6752 * Move the next running task to the front of
6753 * the list, so our cfs_tasks list becomes MRU
6756 list_move(&p->se.group_node, &rq->cfs_tasks);
6759 if (hrtick_enabled(rq))
6760 hrtick_start_fair(rq, p);
6765 new_tasks = idle_balance(rq, rf);
6768 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6769 * possible for any higher priority task to appear. In that case we
6770 * must re-start the pick_next_entity() loop.
6782 * Account for a descheduled task:
6784 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6786 struct sched_entity *se = &prev->se;
6787 struct cfs_rq *cfs_rq;
6789 for_each_sched_entity(se) {
6790 cfs_rq = cfs_rq_of(se);
6791 put_prev_entity(cfs_rq, se);
6796 * sched_yield() is very simple
6798 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6800 static void yield_task_fair(struct rq *rq)
6802 struct task_struct *curr = rq->curr;
6803 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6804 struct sched_entity *se = &curr->se;
6807 * Are we the only task in the tree?
6809 if (unlikely(rq->nr_running == 1))
6812 clear_buddies(cfs_rq, se);
6814 if (curr->policy != SCHED_BATCH) {
6815 update_rq_clock(rq);
6817 * Update run-time statistics of the 'current'.
6819 update_curr(cfs_rq);
6821 * Tell update_rq_clock() that we've just updated,
6822 * so we don't do microscopic update in schedule()
6823 * and double the fastpath cost.
6825 rq_clock_skip_update(rq, true);
6831 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6833 struct sched_entity *se = &p->se;
6835 /* throttled hierarchies are not runnable */
6836 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6839 /* Tell the scheduler that we'd really like pse to run next. */
6842 yield_task_fair(rq);
6848 /**************************************************
6849 * Fair scheduling class load-balancing methods.
6853 * The purpose of load-balancing is to achieve the same basic fairness the
6854 * per-cpu scheduler provides, namely provide a proportional amount of compute
6855 * time to each task. This is expressed in the following equation:
6857 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6859 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6860 * W_i,0 is defined as:
6862 * W_i,0 = \Sum_j w_i,j (2)
6864 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6865 * is derived from the nice value as per sched_prio_to_weight[].
6867 * The weight average is an exponential decay average of the instantaneous
6870 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6872 * C_i is the compute capacity of cpu i, typically it is the
6873 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6874 * can also include other factors [XXX].
6876 * To achieve this balance we define a measure of imbalance which follows
6877 * directly from (1):
6879 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6881 * We them move tasks around to minimize the imbalance. In the continuous
6882 * function space it is obvious this converges, in the discrete case we get
6883 * a few fun cases generally called infeasible weight scenarios.
6886 * - infeasible weights;
6887 * - local vs global optima in the discrete case. ]
6892 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6893 * for all i,j solution, we create a tree of cpus that follows the hardware
6894 * topology where each level pairs two lower groups (or better). This results
6895 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6896 * tree to only the first of the previous level and we decrease the frequency
6897 * of load-balance at each level inv. proportional to the number of cpus in
6903 * \Sum { --- * --- * 2^i } = O(n) (5)
6905 * `- size of each group
6906 * | | `- number of cpus doing load-balance
6908 * `- sum over all levels
6910 * Coupled with a limit on how many tasks we can migrate every balance pass,
6911 * this makes (5) the runtime complexity of the balancer.
6913 * An important property here is that each CPU is still (indirectly) connected
6914 * to every other cpu in at most O(log n) steps:
6916 * The adjacency matrix of the resulting graph is given by:
6919 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6922 * And you'll find that:
6924 * A^(log_2 n)_i,j != 0 for all i,j (7)
6926 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6927 * The task movement gives a factor of O(m), giving a convergence complexity
6930 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6935 * In order to avoid CPUs going idle while there's still work to do, new idle
6936 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6937 * tree itself instead of relying on other CPUs to bring it work.
6939 * This adds some complexity to both (5) and (8) but it reduces the total idle
6947 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6950 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6955 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6957 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6959 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6962 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6963 * rewrite all of this once again.]
6966 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6968 enum fbq_type { regular, remote, all };
6970 #define LBF_ALL_PINNED 0x01
6971 #define LBF_NEED_BREAK 0x02
6972 #define LBF_DST_PINNED 0x04
6973 #define LBF_SOME_PINNED 0x08
6976 struct sched_domain *sd;
6984 struct cpumask *dst_grpmask;
6986 enum cpu_idle_type idle;
6988 /* The set of CPUs under consideration for load-balancing */
6989 struct cpumask *cpus;
6994 unsigned int loop_break;
6995 unsigned int loop_max;
6997 enum fbq_type fbq_type;
6998 struct list_head tasks;
7002 * Is this task likely cache-hot:
7004 static int task_hot(struct task_struct *p, struct lb_env *env)
7008 lockdep_assert_held(&env->src_rq->lock);
7010 if (p->sched_class != &fair_sched_class)
7013 if (unlikely(p->policy == SCHED_IDLE))
7017 * Buddy candidates are cache hot:
7019 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7020 (&p->se == cfs_rq_of(&p->se)->next ||
7021 &p->se == cfs_rq_of(&p->se)->last))
7024 if (sysctl_sched_migration_cost == -1)
7026 if (sysctl_sched_migration_cost == 0)
7029 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7031 return delta < (s64)sysctl_sched_migration_cost;
7034 #ifdef CONFIG_NUMA_BALANCING
7036 * Returns 1, if task migration degrades locality
7037 * Returns 0, if task migration improves locality i.e migration preferred.
7038 * Returns -1, if task migration is not affected by locality.
7040 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7042 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7043 unsigned long src_faults, dst_faults;
7044 int src_nid, dst_nid;
7046 if (!static_branch_likely(&sched_numa_balancing))
7049 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7052 src_nid = cpu_to_node(env->src_cpu);
7053 dst_nid = cpu_to_node(env->dst_cpu);
7055 if (src_nid == dst_nid)
7058 /* Migrating away from the preferred node is always bad. */
7059 if (src_nid == p->numa_preferred_nid) {
7060 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7066 /* Encourage migration to the preferred node. */
7067 if (dst_nid == p->numa_preferred_nid)
7070 /* Leaving a core idle is often worse than degrading locality. */
7071 if (env->idle != CPU_NOT_IDLE)
7075 src_faults = group_faults(p, src_nid);
7076 dst_faults = group_faults(p, dst_nid);
7078 src_faults = task_faults(p, src_nid);
7079 dst_faults = task_faults(p, dst_nid);
7082 return dst_faults < src_faults;
7086 static inline int migrate_degrades_locality(struct task_struct *p,
7094 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7097 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7101 lockdep_assert_held(&env->src_rq->lock);
7104 * We do not migrate tasks that are:
7105 * 1) throttled_lb_pair, or
7106 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7107 * 3) running (obviously), or
7108 * 4) are cache-hot on their current CPU.
7110 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7113 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7116 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7118 env->flags |= LBF_SOME_PINNED;
7121 * Remember if this task can be migrated to any other cpu in
7122 * our sched_group. We may want to revisit it if we couldn't
7123 * meet load balance goals by pulling other tasks on src_cpu.
7125 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7126 * already computed one in current iteration.
7128 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7131 /* Prevent to re-select dst_cpu via env's cpus */
7132 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7133 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7134 env->flags |= LBF_DST_PINNED;
7135 env->new_dst_cpu = cpu;
7143 /* Record that we found atleast one task that could run on dst_cpu */
7144 env->flags &= ~LBF_ALL_PINNED;
7146 if (task_running(env->src_rq, p)) {
7147 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7152 * Aggressive migration if:
7153 * 1) destination numa is preferred
7154 * 2) task is cache cold, or
7155 * 3) too many balance attempts have failed.
7157 tsk_cache_hot = migrate_degrades_locality(p, env);
7158 if (tsk_cache_hot == -1)
7159 tsk_cache_hot = task_hot(p, env);
7161 if (tsk_cache_hot <= 0 ||
7162 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7163 if (tsk_cache_hot == 1) {
7164 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7165 schedstat_inc(p->se.statistics.nr_forced_migrations);
7170 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7175 * detach_task() -- detach the task for the migration specified in env
7177 static void detach_task(struct task_struct *p, struct lb_env *env)
7179 lockdep_assert_held(&env->src_rq->lock);
7181 p->on_rq = TASK_ON_RQ_MIGRATING;
7182 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7183 set_task_cpu(p, env->dst_cpu);
7187 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7188 * part of active balancing operations within "domain".
7190 * Returns a task if successful and NULL otherwise.
7192 static struct task_struct *detach_one_task(struct lb_env *env)
7194 struct task_struct *p;
7196 lockdep_assert_held(&env->src_rq->lock);
7198 list_for_each_entry_reverse(p,
7199 &env->src_rq->cfs_tasks, se.group_node) {
7200 if (!can_migrate_task(p, env))
7203 detach_task(p, env);
7206 * Right now, this is only the second place where
7207 * lb_gained[env->idle] is updated (other is detach_tasks)
7208 * so we can safely collect stats here rather than
7209 * inside detach_tasks().
7211 schedstat_inc(env->sd->lb_gained[env->idle]);
7217 static const unsigned int sched_nr_migrate_break = 32;
7220 * detach_tasks() -- tries to detach up to imbalance weighted load from
7221 * busiest_rq, as part of a balancing operation within domain "sd".
7223 * Returns number of detached tasks if successful and 0 otherwise.
7225 static int detach_tasks(struct lb_env *env)
7227 struct list_head *tasks = &env->src_rq->cfs_tasks;
7228 struct task_struct *p;
7232 lockdep_assert_held(&env->src_rq->lock);
7234 if (env->imbalance <= 0)
7237 while (!list_empty(tasks)) {
7239 * We don't want to steal all, otherwise we may be treated likewise,
7240 * which could at worst lead to a livelock crash.
7242 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7245 p = list_last_entry(tasks, struct task_struct, se.group_node);
7248 /* We've more or less seen every task there is, call it quits */
7249 if (env->loop > env->loop_max)
7252 /* take a breather every nr_migrate tasks */
7253 if (env->loop > env->loop_break) {
7254 env->loop_break += sched_nr_migrate_break;
7255 env->flags |= LBF_NEED_BREAK;
7259 if (!can_migrate_task(p, env))
7262 load = task_h_load(p);
7264 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7267 if ((load / 2) > env->imbalance)
7270 detach_task(p, env);
7271 list_add(&p->se.group_node, &env->tasks);
7274 env->imbalance -= load;
7276 #ifdef CONFIG_PREEMPT
7278 * NEWIDLE balancing is a source of latency, so preemptible
7279 * kernels will stop after the first task is detached to minimize
7280 * the critical section.
7282 if (env->idle == CPU_NEWLY_IDLE)
7287 * We only want to steal up to the prescribed amount of
7290 if (env->imbalance <= 0)
7295 list_move(&p->se.group_node, tasks);
7299 * Right now, this is one of only two places we collect this stat
7300 * so we can safely collect detach_one_task() stats here rather
7301 * than inside detach_one_task().
7303 schedstat_add(env->sd->lb_gained[env->idle], detached);
7309 * attach_task() -- attach the task detached by detach_task() to its new rq.
7311 static void attach_task(struct rq *rq, struct task_struct *p)
7313 lockdep_assert_held(&rq->lock);
7315 BUG_ON(task_rq(p) != rq);
7316 activate_task(rq, p, ENQUEUE_NOCLOCK);
7317 p->on_rq = TASK_ON_RQ_QUEUED;
7318 check_preempt_curr(rq, p, 0);
7322 * attach_one_task() -- attaches the task returned from detach_one_task() to
7325 static void attach_one_task(struct rq *rq, struct task_struct *p)
7330 update_rq_clock(rq);
7336 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7339 static void attach_tasks(struct lb_env *env)
7341 struct list_head *tasks = &env->tasks;
7342 struct task_struct *p;
7345 rq_lock(env->dst_rq, &rf);
7346 update_rq_clock(env->dst_rq);
7348 while (!list_empty(tasks)) {
7349 p = list_first_entry(tasks, struct task_struct, se.group_node);
7350 list_del_init(&p->se.group_node);
7352 attach_task(env->dst_rq, p);
7355 rq_unlock(env->dst_rq, &rf);
7358 #ifdef CONFIG_FAIR_GROUP_SCHED
7360 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7362 if (cfs_rq->load.weight)
7365 if (cfs_rq->avg.load_sum)
7368 if (cfs_rq->avg.util_sum)
7371 if (cfs_rq->avg.runnable_load_sum)
7377 static void update_blocked_averages(int cpu)
7379 struct rq *rq = cpu_rq(cpu);
7380 struct cfs_rq *cfs_rq, *pos;
7383 rq_lock_irqsave(rq, &rf);
7384 update_rq_clock(rq);
7387 * Iterates the task_group tree in a bottom up fashion, see
7388 * list_add_leaf_cfs_rq() for details.
7390 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7391 struct sched_entity *se;
7393 /* throttled entities do not contribute to load */
7394 if (throttled_hierarchy(cfs_rq))
7397 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7398 update_tg_load_avg(cfs_rq, 0);
7400 /* Propagate pending load changes to the parent, if any: */
7401 se = cfs_rq->tg->se[cpu];
7402 if (se && !skip_blocked_update(se))
7403 update_load_avg(cfs_rq_of(se), se, 0);
7406 * There can be a lot of idle CPU cgroups. Don't let fully
7407 * decayed cfs_rqs linger on the list.
7409 if (cfs_rq_is_decayed(cfs_rq))
7410 list_del_leaf_cfs_rq(cfs_rq);
7412 rq_unlock_irqrestore(rq, &rf);
7416 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7417 * This needs to be done in a top-down fashion because the load of a child
7418 * group is a fraction of its parents load.
7420 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7422 struct rq *rq = rq_of(cfs_rq);
7423 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7424 unsigned long now = jiffies;
7427 if (cfs_rq->last_h_load_update == now)
7430 cfs_rq->h_load_next = NULL;
7431 for_each_sched_entity(se) {
7432 cfs_rq = cfs_rq_of(se);
7433 cfs_rq->h_load_next = se;
7434 if (cfs_rq->last_h_load_update == now)
7439 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7440 cfs_rq->last_h_load_update = now;
7443 while ((se = cfs_rq->h_load_next) != NULL) {
7444 load = cfs_rq->h_load;
7445 load = div64_ul(load * se->avg.load_avg,
7446 cfs_rq_load_avg(cfs_rq) + 1);
7447 cfs_rq = group_cfs_rq(se);
7448 cfs_rq->h_load = load;
7449 cfs_rq->last_h_load_update = now;
7453 static unsigned long task_h_load(struct task_struct *p)
7455 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7457 update_cfs_rq_h_load(cfs_rq);
7458 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7459 cfs_rq_load_avg(cfs_rq) + 1);
7462 static inline void update_blocked_averages(int cpu)
7464 struct rq *rq = cpu_rq(cpu);
7465 struct cfs_rq *cfs_rq = &rq->cfs;
7468 rq_lock_irqsave(rq, &rf);
7469 update_rq_clock(rq);
7470 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7471 rq_unlock_irqrestore(rq, &rf);
7474 static unsigned long task_h_load(struct task_struct *p)
7476 return p->se.avg.load_avg;
7480 /********** Helpers for find_busiest_group ************************/
7489 * sg_lb_stats - stats of a sched_group required for load_balancing
7491 struct sg_lb_stats {
7492 unsigned long avg_load; /*Avg load across the CPUs of the group */
7493 unsigned long group_load; /* Total load over the CPUs of the group */
7494 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7495 unsigned long load_per_task;
7496 unsigned long group_capacity;
7497 unsigned long group_util; /* Total utilization of the group */
7498 unsigned int sum_nr_running; /* Nr tasks running in the group */
7499 unsigned int idle_cpus;
7500 unsigned int group_weight;
7501 enum group_type group_type;
7502 int group_no_capacity;
7503 #ifdef CONFIG_NUMA_BALANCING
7504 unsigned int nr_numa_running;
7505 unsigned int nr_preferred_running;
7510 * sd_lb_stats - Structure to store the statistics of a sched_domain
7511 * during load balancing.
7513 struct sd_lb_stats {
7514 struct sched_group *busiest; /* Busiest group in this sd */
7515 struct sched_group *local; /* Local group in this sd */
7516 unsigned long total_running;
7517 unsigned long total_load; /* Total load of all groups in sd */
7518 unsigned long total_capacity; /* Total capacity of all groups in sd */
7519 unsigned long avg_load; /* Average load across all groups in sd */
7521 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7522 struct sg_lb_stats local_stat; /* Statistics of the local group */
7525 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7528 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7529 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7530 * We must however clear busiest_stat::avg_load because
7531 * update_sd_pick_busiest() reads this before assignment.
7533 *sds = (struct sd_lb_stats){
7536 .total_running = 0UL,
7538 .total_capacity = 0UL,
7541 .sum_nr_running = 0,
7542 .group_type = group_other,
7548 * get_sd_load_idx - Obtain the load index for a given sched domain.
7549 * @sd: The sched_domain whose load_idx is to be obtained.
7550 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7552 * Return: The load index.
7554 static inline int get_sd_load_idx(struct sched_domain *sd,
7555 enum cpu_idle_type idle)
7561 load_idx = sd->busy_idx;
7564 case CPU_NEWLY_IDLE:
7565 load_idx = sd->newidle_idx;
7568 load_idx = sd->idle_idx;
7575 static unsigned long scale_rt_capacity(int cpu)
7577 struct rq *rq = cpu_rq(cpu);
7578 u64 total, used, age_stamp, avg;
7582 * Since we're reading these variables without serialization make sure
7583 * we read them once before doing sanity checks on them.
7585 age_stamp = READ_ONCE(rq->age_stamp);
7586 avg = READ_ONCE(rq->rt_avg);
7587 delta = __rq_clock_broken(rq) - age_stamp;
7589 if (unlikely(delta < 0))
7592 total = sched_avg_period() + delta;
7594 used = div_u64(avg, total);
7596 if (likely(used < SCHED_CAPACITY_SCALE))
7597 return SCHED_CAPACITY_SCALE - used;
7602 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7604 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7605 struct sched_group *sdg = sd->groups;
7607 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7609 capacity *= scale_rt_capacity(cpu);
7610 capacity >>= SCHED_CAPACITY_SHIFT;
7615 cpu_rq(cpu)->cpu_capacity = capacity;
7616 sdg->sgc->capacity = capacity;
7617 sdg->sgc->min_capacity = capacity;
7620 void update_group_capacity(struct sched_domain *sd, int cpu)
7622 struct sched_domain *child = sd->child;
7623 struct sched_group *group, *sdg = sd->groups;
7624 unsigned long capacity, min_capacity;
7625 unsigned long interval;
7627 interval = msecs_to_jiffies(sd->balance_interval);
7628 interval = clamp(interval, 1UL, max_load_balance_interval);
7629 sdg->sgc->next_update = jiffies + interval;
7632 update_cpu_capacity(sd, cpu);
7637 min_capacity = ULONG_MAX;
7639 if (child->flags & SD_OVERLAP) {
7641 * SD_OVERLAP domains cannot assume that child groups
7642 * span the current group.
7645 for_each_cpu(cpu, sched_group_span(sdg)) {
7646 struct sched_group_capacity *sgc;
7647 struct rq *rq = cpu_rq(cpu);
7650 * build_sched_domains() -> init_sched_groups_capacity()
7651 * gets here before we've attached the domains to the
7654 * Use capacity_of(), which is set irrespective of domains
7655 * in update_cpu_capacity().
7657 * This avoids capacity from being 0 and
7658 * causing divide-by-zero issues on boot.
7660 if (unlikely(!rq->sd)) {
7661 capacity += capacity_of(cpu);
7663 sgc = rq->sd->groups->sgc;
7664 capacity += sgc->capacity;
7667 min_capacity = min(capacity, min_capacity);
7671 * !SD_OVERLAP domains can assume that child groups
7672 * span the current group.
7675 group = child->groups;
7677 struct sched_group_capacity *sgc = group->sgc;
7679 capacity += sgc->capacity;
7680 min_capacity = min(sgc->min_capacity, min_capacity);
7681 group = group->next;
7682 } while (group != child->groups);
7685 sdg->sgc->capacity = capacity;
7686 sdg->sgc->min_capacity = min_capacity;
7690 * Check whether the capacity of the rq has been noticeably reduced by side
7691 * activity. The imbalance_pct is used for the threshold.
7692 * Return true is the capacity is reduced
7695 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7697 return ((rq->cpu_capacity * sd->imbalance_pct) <
7698 (rq->cpu_capacity_orig * 100));
7702 * Group imbalance indicates (and tries to solve) the problem where balancing
7703 * groups is inadequate due to ->cpus_allowed constraints.
7705 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7706 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7709 * { 0 1 2 3 } { 4 5 6 7 }
7712 * If we were to balance group-wise we'd place two tasks in the first group and
7713 * two tasks in the second group. Clearly this is undesired as it will overload
7714 * cpu 3 and leave one of the cpus in the second group unused.
7716 * The current solution to this issue is detecting the skew in the first group
7717 * by noticing the lower domain failed to reach balance and had difficulty
7718 * moving tasks due to affinity constraints.
7720 * When this is so detected; this group becomes a candidate for busiest; see
7721 * update_sd_pick_busiest(). And calculate_imbalance() and
7722 * find_busiest_group() avoid some of the usual balance conditions to allow it
7723 * to create an effective group imbalance.
7725 * This is a somewhat tricky proposition since the next run might not find the
7726 * group imbalance and decide the groups need to be balanced again. A most
7727 * subtle and fragile situation.
7730 static inline int sg_imbalanced(struct sched_group *group)
7732 return group->sgc->imbalance;
7736 * group_has_capacity returns true if the group has spare capacity that could
7737 * be used by some tasks.
7738 * We consider that a group has spare capacity if the * number of task is
7739 * smaller than the number of CPUs or if the utilization is lower than the
7740 * available capacity for CFS tasks.
7741 * For the latter, we use a threshold to stabilize the state, to take into
7742 * account the variance of the tasks' load and to return true if the available
7743 * capacity in meaningful for the load balancer.
7744 * As an example, an available capacity of 1% can appear but it doesn't make
7745 * any benefit for the load balance.
7748 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7750 if (sgs->sum_nr_running < sgs->group_weight)
7753 if ((sgs->group_capacity * 100) >
7754 (sgs->group_util * env->sd->imbalance_pct))
7761 * group_is_overloaded returns true if the group has more tasks than it can
7763 * group_is_overloaded is not equals to !group_has_capacity because a group
7764 * with the exact right number of tasks, has no more spare capacity but is not
7765 * overloaded so both group_has_capacity and group_is_overloaded return
7769 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7771 if (sgs->sum_nr_running <= sgs->group_weight)
7774 if ((sgs->group_capacity * 100) <
7775 (sgs->group_util * env->sd->imbalance_pct))
7782 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7783 * per-CPU capacity than sched_group ref.
7786 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7788 return sg->sgc->min_capacity * capacity_margin <
7789 ref->sgc->min_capacity * 1024;
7793 group_type group_classify(struct sched_group *group,
7794 struct sg_lb_stats *sgs)
7796 if (sgs->group_no_capacity)
7797 return group_overloaded;
7799 if (sg_imbalanced(group))
7800 return group_imbalanced;
7806 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7807 * @env: The load balancing environment.
7808 * @group: sched_group whose statistics are to be updated.
7809 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7810 * @local_group: Does group contain this_cpu.
7811 * @sgs: variable to hold the statistics for this group.
7812 * @overload: Indicate more than one runnable task for any CPU.
7814 static inline void update_sg_lb_stats(struct lb_env *env,
7815 struct sched_group *group, int load_idx,
7816 int local_group, struct sg_lb_stats *sgs,
7822 memset(sgs, 0, sizeof(*sgs));
7824 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7825 struct rq *rq = cpu_rq(i);
7827 /* Bias balancing toward cpus of our domain */
7829 load = target_load(i, load_idx);
7831 load = source_load(i, load_idx);
7833 sgs->group_load += load;
7834 sgs->group_util += cpu_util(i);
7835 sgs->sum_nr_running += rq->cfs.h_nr_running;
7837 nr_running = rq->nr_running;
7841 #ifdef CONFIG_NUMA_BALANCING
7842 sgs->nr_numa_running += rq->nr_numa_running;
7843 sgs->nr_preferred_running += rq->nr_preferred_running;
7845 sgs->sum_weighted_load += weighted_cpuload(rq);
7847 * No need to call idle_cpu() if nr_running is not 0
7849 if (!nr_running && idle_cpu(i))
7853 /* Adjust by relative CPU capacity of the group */
7854 sgs->group_capacity = group->sgc->capacity;
7855 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7857 if (sgs->sum_nr_running)
7858 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7860 sgs->group_weight = group->group_weight;
7862 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7863 sgs->group_type = group_classify(group, sgs);
7867 * update_sd_pick_busiest - return 1 on busiest group
7868 * @env: The load balancing environment.
7869 * @sds: sched_domain statistics
7870 * @sg: sched_group candidate to be checked for being the busiest
7871 * @sgs: sched_group statistics
7873 * Determine if @sg is a busier group than the previously selected
7876 * Return: %true if @sg is a busier group than the previously selected
7877 * busiest group. %false otherwise.
7879 static bool update_sd_pick_busiest(struct lb_env *env,
7880 struct sd_lb_stats *sds,
7881 struct sched_group *sg,
7882 struct sg_lb_stats *sgs)
7884 struct sg_lb_stats *busiest = &sds->busiest_stat;
7886 if (sgs->group_type > busiest->group_type)
7889 if (sgs->group_type < busiest->group_type)
7892 if (sgs->avg_load <= busiest->avg_load)
7895 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7899 * Candidate sg has no more than one task per CPU and
7900 * has higher per-CPU capacity. Migrating tasks to less
7901 * capable CPUs may harm throughput. Maximize throughput,
7902 * power/energy consequences are not considered.
7904 if (sgs->sum_nr_running <= sgs->group_weight &&
7905 group_smaller_cpu_capacity(sds->local, sg))
7909 /* This is the busiest node in its class. */
7910 if (!(env->sd->flags & SD_ASYM_PACKING))
7913 /* No ASYM_PACKING if target cpu is already busy */
7914 if (env->idle == CPU_NOT_IDLE)
7917 * ASYM_PACKING needs to move all the work to the highest
7918 * prority CPUs in the group, therefore mark all groups
7919 * of lower priority than ourself as busy.
7921 if (sgs->sum_nr_running &&
7922 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7926 /* Prefer to move from lowest priority cpu's work */
7927 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7928 sg->asym_prefer_cpu))
7935 #ifdef CONFIG_NUMA_BALANCING
7936 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7938 if (sgs->sum_nr_running > sgs->nr_numa_running)
7940 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7945 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7947 if (rq->nr_running > rq->nr_numa_running)
7949 if (rq->nr_running > rq->nr_preferred_running)
7954 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7959 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7963 #endif /* CONFIG_NUMA_BALANCING */
7966 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7967 * @env: The load balancing environment.
7968 * @sds: variable to hold the statistics for this sched_domain.
7970 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7972 struct sched_domain *child = env->sd->child;
7973 struct sched_group *sg = env->sd->groups;
7974 struct sg_lb_stats *local = &sds->local_stat;
7975 struct sg_lb_stats tmp_sgs;
7976 int load_idx, prefer_sibling = 0;
7977 bool overload = false;
7979 if (child && child->flags & SD_PREFER_SIBLING)
7982 load_idx = get_sd_load_idx(env->sd, env->idle);
7985 struct sg_lb_stats *sgs = &tmp_sgs;
7988 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7993 if (env->idle != CPU_NEWLY_IDLE ||
7994 time_after_eq(jiffies, sg->sgc->next_update))
7995 update_group_capacity(env->sd, env->dst_cpu);
7998 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8005 * In case the child domain prefers tasks go to siblings
8006 * first, lower the sg capacity so that we'll try
8007 * and move all the excess tasks away. We lower the capacity
8008 * of a group only if the local group has the capacity to fit
8009 * these excess tasks. The extra check prevents the case where
8010 * you always pull from the heaviest group when it is already
8011 * under-utilized (possible with a large weight task outweighs
8012 * the tasks on the system).
8014 if (prefer_sibling && sds->local &&
8015 group_has_capacity(env, local) &&
8016 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8017 sgs->group_no_capacity = 1;
8018 sgs->group_type = group_classify(sg, sgs);
8021 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8023 sds->busiest_stat = *sgs;
8027 /* Now, start updating sd_lb_stats */
8028 sds->total_running += sgs->sum_nr_running;
8029 sds->total_load += sgs->group_load;
8030 sds->total_capacity += sgs->group_capacity;
8033 } while (sg != env->sd->groups);
8035 if (env->sd->flags & SD_NUMA)
8036 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8038 if (!env->sd->parent) {
8039 /* update overload indicator if we are at root domain */
8040 if (env->dst_rq->rd->overload != overload)
8041 env->dst_rq->rd->overload = overload;
8046 * check_asym_packing - Check to see if the group is packed into the
8049 * This is primarily intended to used at the sibling level. Some
8050 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8051 * case of POWER7, it can move to lower SMT modes only when higher
8052 * threads are idle. When in lower SMT modes, the threads will
8053 * perform better since they share less core resources. Hence when we
8054 * have idle threads, we want them to be the higher ones.
8056 * This packing function is run on idle threads. It checks to see if
8057 * the busiest CPU in this domain (core in the P7 case) has a higher
8058 * CPU number than the packing function is being run on. Here we are
8059 * assuming lower CPU number will be equivalent to lower a SMT thread
8062 * Return: 1 when packing is required and a task should be moved to
8063 * this CPU. The amount of the imbalance is returned in env->imbalance.
8065 * @env: The load balancing environment.
8066 * @sds: Statistics of the sched_domain which is to be packed
8068 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8072 if (!(env->sd->flags & SD_ASYM_PACKING))
8075 if (env->idle == CPU_NOT_IDLE)
8081 busiest_cpu = sds->busiest->asym_prefer_cpu;
8082 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8085 env->imbalance = DIV_ROUND_CLOSEST(
8086 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8087 SCHED_CAPACITY_SCALE);
8093 * fix_small_imbalance - Calculate the minor imbalance that exists
8094 * amongst the groups of a sched_domain, during
8096 * @env: The load balancing environment.
8097 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8100 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8102 unsigned long tmp, capa_now = 0, capa_move = 0;
8103 unsigned int imbn = 2;
8104 unsigned long scaled_busy_load_per_task;
8105 struct sg_lb_stats *local, *busiest;
8107 local = &sds->local_stat;
8108 busiest = &sds->busiest_stat;
8110 if (!local->sum_nr_running)
8111 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8112 else if (busiest->load_per_task > local->load_per_task)
8115 scaled_busy_load_per_task =
8116 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8117 busiest->group_capacity;
8119 if (busiest->avg_load + scaled_busy_load_per_task >=
8120 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8121 env->imbalance = busiest->load_per_task;
8126 * OK, we don't have enough imbalance to justify moving tasks,
8127 * however we may be able to increase total CPU capacity used by
8131 capa_now += busiest->group_capacity *
8132 min(busiest->load_per_task, busiest->avg_load);
8133 capa_now += local->group_capacity *
8134 min(local->load_per_task, local->avg_load);
8135 capa_now /= SCHED_CAPACITY_SCALE;
8137 /* Amount of load we'd subtract */
8138 if (busiest->avg_load > scaled_busy_load_per_task) {
8139 capa_move += busiest->group_capacity *
8140 min(busiest->load_per_task,
8141 busiest->avg_load - scaled_busy_load_per_task);
8144 /* Amount of load we'd add */
8145 if (busiest->avg_load * busiest->group_capacity <
8146 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8147 tmp = (busiest->avg_load * busiest->group_capacity) /
8148 local->group_capacity;
8150 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8151 local->group_capacity;
8153 capa_move += local->group_capacity *
8154 min(local->load_per_task, local->avg_load + tmp);
8155 capa_move /= SCHED_CAPACITY_SCALE;
8157 /* Move if we gain throughput */
8158 if (capa_move > capa_now)
8159 env->imbalance = busiest->load_per_task;
8163 * calculate_imbalance - Calculate the amount of imbalance present within the
8164 * groups of a given sched_domain during load balance.
8165 * @env: load balance environment
8166 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8168 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8170 unsigned long max_pull, load_above_capacity = ~0UL;
8171 struct sg_lb_stats *local, *busiest;
8173 local = &sds->local_stat;
8174 busiest = &sds->busiest_stat;
8176 if (busiest->group_type == group_imbalanced) {
8178 * In the group_imb case we cannot rely on group-wide averages
8179 * to ensure cpu-load equilibrium, look at wider averages. XXX
8181 busiest->load_per_task =
8182 min(busiest->load_per_task, sds->avg_load);
8186 * Avg load of busiest sg can be less and avg load of local sg can
8187 * be greater than avg load across all sgs of sd because avg load
8188 * factors in sg capacity and sgs with smaller group_type are
8189 * skipped when updating the busiest sg:
8191 if (busiest->avg_load <= sds->avg_load ||
8192 local->avg_load >= sds->avg_load) {
8194 return fix_small_imbalance(env, sds);
8198 * If there aren't any idle cpus, avoid creating some.
8200 if (busiest->group_type == group_overloaded &&
8201 local->group_type == group_overloaded) {
8202 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8203 if (load_above_capacity > busiest->group_capacity) {
8204 load_above_capacity -= busiest->group_capacity;
8205 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8206 load_above_capacity /= busiest->group_capacity;
8208 load_above_capacity = ~0UL;
8212 * We're trying to get all the cpus to the average_load, so we don't
8213 * want to push ourselves above the average load, nor do we wish to
8214 * reduce the max loaded cpu below the average load. At the same time,
8215 * we also don't want to reduce the group load below the group
8216 * capacity. Thus we look for the minimum possible imbalance.
8218 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8220 /* How much load to actually move to equalise the imbalance */
8221 env->imbalance = min(
8222 max_pull * busiest->group_capacity,
8223 (sds->avg_load - local->avg_load) * local->group_capacity
8224 ) / SCHED_CAPACITY_SCALE;
8227 * if *imbalance is less than the average load per runnable task
8228 * there is no guarantee that any tasks will be moved so we'll have
8229 * a think about bumping its value to force at least one task to be
8232 if (env->imbalance < busiest->load_per_task)
8233 return fix_small_imbalance(env, sds);
8236 /******* find_busiest_group() helpers end here *********************/
8239 * find_busiest_group - Returns the busiest group within the sched_domain
8240 * if there is an imbalance.
8242 * Also calculates the amount of weighted load which should be moved
8243 * to restore balance.
8245 * @env: The load balancing environment.
8247 * Return: - The busiest group if imbalance exists.
8249 static struct sched_group *find_busiest_group(struct lb_env *env)
8251 struct sg_lb_stats *local, *busiest;
8252 struct sd_lb_stats sds;
8254 init_sd_lb_stats(&sds);
8257 * Compute the various statistics relavent for load balancing at
8260 update_sd_lb_stats(env, &sds);
8261 local = &sds.local_stat;
8262 busiest = &sds.busiest_stat;
8264 /* ASYM feature bypasses nice load balance check */
8265 if (check_asym_packing(env, &sds))
8268 /* There is no busy sibling group to pull tasks from */
8269 if (!sds.busiest || busiest->sum_nr_running == 0)
8272 /* XXX broken for overlapping NUMA groups */
8273 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8274 / sds.total_capacity;
8277 * If the busiest group is imbalanced the below checks don't
8278 * work because they assume all things are equal, which typically
8279 * isn't true due to cpus_allowed constraints and the like.
8281 if (busiest->group_type == group_imbalanced)
8285 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8286 * capacities from resulting in underutilization due to avg_load.
8288 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8289 busiest->group_no_capacity)
8293 * If the local group is busier than the selected busiest group
8294 * don't try and pull any tasks.
8296 if (local->avg_load >= busiest->avg_load)
8300 * Don't pull any tasks if this group is already above the domain
8303 if (local->avg_load >= sds.avg_load)
8306 if (env->idle == CPU_IDLE) {
8308 * This cpu is idle. If the busiest group is not overloaded
8309 * and there is no imbalance between this and busiest group
8310 * wrt idle cpus, it is balanced. The imbalance becomes
8311 * significant if the diff is greater than 1 otherwise we
8312 * might end up to just move the imbalance on another group
8314 if ((busiest->group_type != group_overloaded) &&
8315 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8319 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8320 * imbalance_pct to be conservative.
8322 if (100 * busiest->avg_load <=
8323 env->sd->imbalance_pct * local->avg_load)
8328 /* Looks like there is an imbalance. Compute it */
8329 calculate_imbalance(env, &sds);
8338 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8340 static struct rq *find_busiest_queue(struct lb_env *env,
8341 struct sched_group *group)
8343 struct rq *busiest = NULL, *rq;
8344 unsigned long busiest_load = 0, busiest_capacity = 1;
8347 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8348 unsigned long capacity, wl;
8352 rt = fbq_classify_rq(rq);
8355 * We classify groups/runqueues into three groups:
8356 * - regular: there are !numa tasks
8357 * - remote: there are numa tasks that run on the 'wrong' node
8358 * - all: there is no distinction
8360 * In order to avoid migrating ideally placed numa tasks,
8361 * ignore those when there's better options.
8363 * If we ignore the actual busiest queue to migrate another
8364 * task, the next balance pass can still reduce the busiest
8365 * queue by moving tasks around inside the node.
8367 * If we cannot move enough load due to this classification
8368 * the next pass will adjust the group classification and
8369 * allow migration of more tasks.
8371 * Both cases only affect the total convergence complexity.
8373 if (rt > env->fbq_type)
8376 capacity = capacity_of(i);
8378 wl = weighted_cpuload(rq);
8381 * When comparing with imbalance, use weighted_cpuload()
8382 * which is not scaled with the cpu capacity.
8385 if (rq->nr_running == 1 && wl > env->imbalance &&
8386 !check_cpu_capacity(rq, env->sd))
8390 * For the load comparisons with the other cpu's, consider
8391 * the weighted_cpuload() scaled with the cpu capacity, so
8392 * that the load can be moved away from the cpu that is
8393 * potentially running at a lower capacity.
8395 * Thus we're looking for max(wl_i / capacity_i), crosswise
8396 * multiplication to rid ourselves of the division works out
8397 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8398 * our previous maximum.
8400 if (wl * busiest_capacity > busiest_load * capacity) {
8402 busiest_capacity = capacity;
8411 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8412 * so long as it is large enough.
8414 #define MAX_PINNED_INTERVAL 512
8416 static int need_active_balance(struct lb_env *env)
8418 struct sched_domain *sd = env->sd;
8420 if (env->idle == CPU_NEWLY_IDLE) {
8423 * ASYM_PACKING needs to force migrate tasks from busy but
8424 * lower priority CPUs in order to pack all tasks in the
8425 * highest priority CPUs.
8427 if ((sd->flags & SD_ASYM_PACKING) &&
8428 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8433 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8434 * It's worth migrating the task if the src_cpu's capacity is reduced
8435 * because of other sched_class or IRQs if more capacity stays
8436 * available on dst_cpu.
8438 if ((env->idle != CPU_NOT_IDLE) &&
8439 (env->src_rq->cfs.h_nr_running == 1)) {
8440 if ((check_cpu_capacity(env->src_rq, sd)) &&
8441 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8445 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8448 static int active_load_balance_cpu_stop(void *data);
8450 static int should_we_balance(struct lb_env *env)
8452 struct sched_group *sg = env->sd->groups;
8453 int cpu, balance_cpu = -1;
8456 * Ensure the balancing environment is consistent; can happen
8457 * when the softirq triggers 'during' hotplug.
8459 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8463 * In the newly idle case, we will allow all the cpu's
8464 * to do the newly idle load balance.
8466 if (env->idle == CPU_NEWLY_IDLE)
8469 /* Try to find first idle cpu */
8470 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8478 if (balance_cpu == -1)
8479 balance_cpu = group_balance_cpu(sg);
8482 * First idle cpu or the first cpu(busiest) in this sched group
8483 * is eligible for doing load balancing at this and above domains.
8485 return balance_cpu == env->dst_cpu;
8489 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8490 * tasks if there is an imbalance.
8492 static int load_balance(int this_cpu, struct rq *this_rq,
8493 struct sched_domain *sd, enum cpu_idle_type idle,
8494 int *continue_balancing)
8496 int ld_moved, cur_ld_moved, active_balance = 0;
8497 struct sched_domain *sd_parent = sd->parent;
8498 struct sched_group *group;
8501 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8503 struct lb_env env = {
8505 .dst_cpu = this_cpu,
8507 .dst_grpmask = sched_group_span(sd->groups),
8509 .loop_break = sched_nr_migrate_break,
8512 .tasks = LIST_HEAD_INIT(env.tasks),
8515 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8517 schedstat_inc(sd->lb_count[idle]);
8520 if (!should_we_balance(&env)) {
8521 *continue_balancing = 0;
8525 group = find_busiest_group(&env);
8527 schedstat_inc(sd->lb_nobusyg[idle]);
8531 busiest = find_busiest_queue(&env, group);
8533 schedstat_inc(sd->lb_nobusyq[idle]);
8537 BUG_ON(busiest == env.dst_rq);
8539 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8541 env.src_cpu = busiest->cpu;
8542 env.src_rq = busiest;
8545 if (busiest->nr_running > 1) {
8547 * Attempt to move tasks. If find_busiest_group has found
8548 * an imbalance but busiest->nr_running <= 1, the group is
8549 * still unbalanced. ld_moved simply stays zero, so it is
8550 * correctly treated as an imbalance.
8552 env.flags |= LBF_ALL_PINNED;
8553 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8556 rq_lock_irqsave(busiest, &rf);
8557 update_rq_clock(busiest);
8560 * cur_ld_moved - load moved in current iteration
8561 * ld_moved - cumulative load moved across iterations
8563 cur_ld_moved = detach_tasks(&env);
8566 * We've detached some tasks from busiest_rq. Every
8567 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8568 * unlock busiest->lock, and we are able to be sure
8569 * that nobody can manipulate the tasks in parallel.
8570 * See task_rq_lock() family for the details.
8573 rq_unlock(busiest, &rf);
8577 ld_moved += cur_ld_moved;
8580 local_irq_restore(rf.flags);
8582 if (env.flags & LBF_NEED_BREAK) {
8583 env.flags &= ~LBF_NEED_BREAK;
8588 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8589 * us and move them to an alternate dst_cpu in our sched_group
8590 * where they can run. The upper limit on how many times we
8591 * iterate on same src_cpu is dependent on number of cpus in our
8594 * This changes load balance semantics a bit on who can move
8595 * load to a given_cpu. In addition to the given_cpu itself
8596 * (or a ilb_cpu acting on its behalf where given_cpu is
8597 * nohz-idle), we now have balance_cpu in a position to move
8598 * load to given_cpu. In rare situations, this may cause
8599 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8600 * _independently_ and at _same_ time to move some load to
8601 * given_cpu) causing exceess load to be moved to given_cpu.
8602 * This however should not happen so much in practice and
8603 * moreover subsequent load balance cycles should correct the
8604 * excess load moved.
8606 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8608 /* Prevent to re-select dst_cpu via env's cpus */
8609 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8611 env.dst_rq = cpu_rq(env.new_dst_cpu);
8612 env.dst_cpu = env.new_dst_cpu;
8613 env.flags &= ~LBF_DST_PINNED;
8615 env.loop_break = sched_nr_migrate_break;
8618 * Go back to "more_balance" rather than "redo" since we
8619 * need to continue with same src_cpu.
8625 * We failed to reach balance because of affinity.
8628 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8630 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8631 *group_imbalance = 1;
8634 /* All tasks on this runqueue were pinned by CPU affinity */
8635 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8636 cpumask_clear_cpu(cpu_of(busiest), cpus);
8638 * Attempting to continue load balancing at the current
8639 * sched_domain level only makes sense if there are
8640 * active CPUs remaining as possible busiest CPUs to
8641 * pull load from which are not contained within the
8642 * destination group that is receiving any migrated
8645 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8647 env.loop_break = sched_nr_migrate_break;
8650 goto out_all_pinned;
8655 schedstat_inc(sd->lb_failed[idle]);
8657 * Increment the failure counter only on periodic balance.
8658 * We do not want newidle balance, which can be very
8659 * frequent, pollute the failure counter causing
8660 * excessive cache_hot migrations and active balances.
8662 if (idle != CPU_NEWLY_IDLE)
8663 sd->nr_balance_failed++;
8665 if (need_active_balance(&env)) {
8666 unsigned long flags;
8668 raw_spin_lock_irqsave(&busiest->lock, flags);
8670 /* don't kick the active_load_balance_cpu_stop,
8671 * if the curr task on busiest cpu can't be
8674 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8675 raw_spin_unlock_irqrestore(&busiest->lock,
8677 env.flags |= LBF_ALL_PINNED;
8678 goto out_one_pinned;
8682 * ->active_balance synchronizes accesses to
8683 * ->active_balance_work. Once set, it's cleared
8684 * only after active load balance is finished.
8686 if (!busiest->active_balance) {
8687 busiest->active_balance = 1;
8688 busiest->push_cpu = this_cpu;
8691 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8693 if (active_balance) {
8694 stop_one_cpu_nowait(cpu_of(busiest),
8695 active_load_balance_cpu_stop, busiest,
8696 &busiest->active_balance_work);
8699 /* We've kicked active balancing, force task migration. */
8700 sd->nr_balance_failed = sd->cache_nice_tries+1;
8703 sd->nr_balance_failed = 0;
8705 if (likely(!active_balance)) {
8706 /* We were unbalanced, so reset the balancing interval */
8707 sd->balance_interval = sd->min_interval;
8710 * If we've begun active balancing, start to back off. This
8711 * case may not be covered by the all_pinned logic if there
8712 * is only 1 task on the busy runqueue (because we don't call
8715 if (sd->balance_interval < sd->max_interval)
8716 sd->balance_interval *= 2;
8723 * We reach balance although we may have faced some affinity
8724 * constraints. Clear the imbalance flag if it was set.
8727 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8729 if (*group_imbalance)
8730 *group_imbalance = 0;
8735 * We reach balance because all tasks are pinned at this level so
8736 * we can't migrate them. Let the imbalance flag set so parent level
8737 * can try to migrate them.
8739 schedstat_inc(sd->lb_balanced[idle]);
8741 sd->nr_balance_failed = 0;
8744 /* tune up the balancing interval */
8745 if (((env.flags & LBF_ALL_PINNED) &&
8746 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8747 (sd->balance_interval < sd->max_interval))
8748 sd->balance_interval *= 2;
8755 static inline unsigned long
8756 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8758 unsigned long interval = sd->balance_interval;
8761 interval *= sd->busy_factor;
8763 /* scale ms to jiffies */
8764 interval = msecs_to_jiffies(interval);
8765 interval = clamp(interval, 1UL, max_load_balance_interval);
8771 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8773 unsigned long interval, next;
8775 /* used by idle balance, so cpu_busy = 0 */
8776 interval = get_sd_balance_interval(sd, 0);
8777 next = sd->last_balance + interval;
8779 if (time_after(*next_balance, next))
8780 *next_balance = next;
8784 * idle_balance is called by schedule() if this_cpu is about to become
8785 * idle. Attempts to pull tasks from other CPUs.
8787 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
8789 unsigned long next_balance = jiffies + HZ;
8790 int this_cpu = this_rq->cpu;
8791 struct sched_domain *sd;
8792 int pulled_task = 0;
8796 * We must set idle_stamp _before_ calling idle_balance(), such that we
8797 * measure the duration of idle_balance() as idle time.
8799 this_rq->idle_stamp = rq_clock(this_rq);
8802 * Do not pull tasks towards !active CPUs...
8804 if (!cpu_active(this_cpu))
8808 * This is OK, because current is on_cpu, which avoids it being picked
8809 * for load-balance and preemption/IRQs are still disabled avoiding
8810 * further scheduler activity on it and we're being very careful to
8811 * re-start the picking loop.
8813 rq_unpin_lock(this_rq, rf);
8815 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8816 !this_rq->rd->overload) {
8818 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8820 update_next_balance(sd, &next_balance);
8826 raw_spin_unlock(&this_rq->lock);
8828 update_blocked_averages(this_cpu);
8830 for_each_domain(this_cpu, sd) {
8831 int continue_balancing = 1;
8832 u64 t0, domain_cost;
8834 if (!(sd->flags & SD_LOAD_BALANCE))
8837 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8838 update_next_balance(sd, &next_balance);
8842 if (sd->flags & SD_BALANCE_NEWIDLE) {
8843 t0 = sched_clock_cpu(this_cpu);
8845 pulled_task = load_balance(this_cpu, this_rq,
8847 &continue_balancing);
8849 domain_cost = sched_clock_cpu(this_cpu) - t0;
8850 if (domain_cost > sd->max_newidle_lb_cost)
8851 sd->max_newidle_lb_cost = domain_cost;
8853 curr_cost += domain_cost;
8856 update_next_balance(sd, &next_balance);
8859 * Stop searching for tasks to pull if there are
8860 * now runnable tasks on this rq.
8862 if (pulled_task || this_rq->nr_running > 0)
8867 raw_spin_lock(&this_rq->lock);
8869 if (curr_cost > this_rq->max_idle_balance_cost)
8870 this_rq->max_idle_balance_cost = curr_cost;
8873 * While browsing the domains, we released the rq lock, a task could
8874 * have been enqueued in the meantime. Since we're not going idle,
8875 * pretend we pulled a task.
8877 if (this_rq->cfs.h_nr_running && !pulled_task)
8881 /* Move the next balance forward */
8882 if (time_after(this_rq->next_balance, next_balance))
8883 this_rq->next_balance = next_balance;
8885 /* Is there a task of a high priority class? */
8886 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8890 this_rq->idle_stamp = 0;
8892 rq_repin_lock(this_rq, rf);
8898 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8899 * running tasks off the busiest CPU onto idle CPUs. It requires at
8900 * least 1 task to be running on each physical CPU where possible, and
8901 * avoids physical / logical imbalances.
8903 static int active_load_balance_cpu_stop(void *data)
8905 struct rq *busiest_rq = data;
8906 int busiest_cpu = cpu_of(busiest_rq);
8907 int target_cpu = busiest_rq->push_cpu;
8908 struct rq *target_rq = cpu_rq(target_cpu);
8909 struct sched_domain *sd;
8910 struct task_struct *p = NULL;
8913 rq_lock_irq(busiest_rq, &rf);
8915 * Between queueing the stop-work and running it is a hole in which
8916 * CPUs can become inactive. We should not move tasks from or to
8919 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8922 /* make sure the requested cpu hasn't gone down in the meantime */
8923 if (unlikely(busiest_cpu != smp_processor_id() ||
8924 !busiest_rq->active_balance))
8927 /* Is there any task to move? */
8928 if (busiest_rq->nr_running <= 1)
8932 * This condition is "impossible", if it occurs
8933 * we need to fix it. Originally reported by
8934 * Bjorn Helgaas on a 128-cpu setup.
8936 BUG_ON(busiest_rq == target_rq);
8938 /* Search for an sd spanning us and the target CPU. */
8940 for_each_domain(target_cpu, sd) {
8941 if ((sd->flags & SD_LOAD_BALANCE) &&
8942 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8947 struct lb_env env = {
8949 .dst_cpu = target_cpu,
8950 .dst_rq = target_rq,
8951 .src_cpu = busiest_rq->cpu,
8952 .src_rq = busiest_rq,
8955 * can_migrate_task() doesn't need to compute new_dst_cpu
8956 * for active balancing. Since we have CPU_IDLE, but no
8957 * @dst_grpmask we need to make that test go away with lying
8960 .flags = LBF_DST_PINNED,
8963 schedstat_inc(sd->alb_count);
8964 update_rq_clock(busiest_rq);
8966 p = detach_one_task(&env);
8968 schedstat_inc(sd->alb_pushed);
8969 /* Active balancing done, reset the failure counter. */
8970 sd->nr_balance_failed = 0;
8972 schedstat_inc(sd->alb_failed);
8977 busiest_rq->active_balance = 0;
8978 rq_unlock(busiest_rq, &rf);
8981 attach_one_task(target_rq, p);
8988 static inline int on_null_domain(struct rq *rq)
8990 return unlikely(!rcu_dereference_sched(rq->sd));
8993 #ifdef CONFIG_NO_HZ_COMMON
8995 * idle load balancing details
8996 * - When one of the busy CPUs notice that there may be an idle rebalancing
8997 * needed, they will kick the idle load balancer, which then does idle
8998 * load balancing for all the idle CPUs.
9001 cpumask_var_t idle_cpus_mask;
9003 unsigned long next_balance; /* in jiffy units */
9004 } nohz ____cacheline_aligned;
9006 static inline int find_new_ilb(void)
9008 int ilb = cpumask_first(nohz.idle_cpus_mask);
9010 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9017 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9018 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9019 * CPU (if there is one).
9021 static void nohz_balancer_kick(void)
9025 nohz.next_balance++;
9027 ilb_cpu = find_new_ilb();
9029 if (ilb_cpu >= nr_cpu_ids)
9032 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
9035 * Use smp_send_reschedule() instead of resched_cpu().
9036 * This way we generate a sched IPI on the target cpu which
9037 * is idle. And the softirq performing nohz idle load balance
9038 * will be run before returning from the IPI.
9040 smp_send_reschedule(ilb_cpu);
9044 void nohz_balance_exit_idle(unsigned int cpu)
9046 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
9048 * Completely isolated CPUs don't ever set, so we must test.
9050 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
9051 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
9052 atomic_dec(&nohz.nr_cpus);
9054 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9058 static inline void set_cpu_sd_state_busy(void)
9060 struct sched_domain *sd;
9061 int cpu = smp_processor_id();
9064 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9066 if (!sd || !sd->nohz_idle)
9070 atomic_inc(&sd->shared->nr_busy_cpus);
9075 void set_cpu_sd_state_idle(void)
9077 struct sched_domain *sd;
9078 int cpu = smp_processor_id();
9081 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9083 if (!sd || sd->nohz_idle)
9087 atomic_dec(&sd->shared->nr_busy_cpus);
9093 * This routine will record that the cpu is going idle with tick stopped.
9094 * This info will be used in performing idle load balancing in the future.
9096 void nohz_balance_enter_idle(int cpu)
9099 * If this cpu is going down, then nothing needs to be done.
9101 if (!cpu_active(cpu))
9104 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9105 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9108 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
9112 * If we're a completely isolated CPU, we don't play.
9114 if (on_null_domain(cpu_rq(cpu)))
9117 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9118 atomic_inc(&nohz.nr_cpus);
9119 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
9123 static DEFINE_SPINLOCK(balancing);
9126 * Scale the max load_balance interval with the number of CPUs in the system.
9127 * This trades load-balance latency on larger machines for less cross talk.
9129 void update_max_interval(void)
9131 max_load_balance_interval = HZ*num_online_cpus()/10;
9135 * It checks each scheduling domain to see if it is due to be balanced,
9136 * and initiates a balancing operation if so.
9138 * Balancing parameters are set up in init_sched_domains.
9140 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9142 int continue_balancing = 1;
9144 unsigned long interval;
9145 struct sched_domain *sd;
9146 /* Earliest time when we have to do rebalance again */
9147 unsigned long next_balance = jiffies + 60*HZ;
9148 int update_next_balance = 0;
9149 int need_serialize, need_decay = 0;
9152 update_blocked_averages(cpu);
9155 for_each_domain(cpu, sd) {
9157 * Decay the newidle max times here because this is a regular
9158 * visit to all the domains. Decay ~1% per second.
9160 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9161 sd->max_newidle_lb_cost =
9162 (sd->max_newidle_lb_cost * 253) / 256;
9163 sd->next_decay_max_lb_cost = jiffies + HZ;
9166 max_cost += sd->max_newidle_lb_cost;
9168 if (!(sd->flags & SD_LOAD_BALANCE))
9172 * Stop the load balance at this level. There is another
9173 * CPU in our sched group which is doing load balancing more
9176 if (!continue_balancing) {
9182 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9184 need_serialize = sd->flags & SD_SERIALIZE;
9185 if (need_serialize) {
9186 if (!spin_trylock(&balancing))
9190 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9191 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9193 * The LBF_DST_PINNED logic could have changed
9194 * env->dst_cpu, so we can't know our idle
9195 * state even if we migrated tasks. Update it.
9197 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9199 sd->last_balance = jiffies;
9200 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9203 spin_unlock(&balancing);
9205 if (time_after(next_balance, sd->last_balance + interval)) {
9206 next_balance = sd->last_balance + interval;
9207 update_next_balance = 1;
9212 * Ensure the rq-wide value also decays but keep it at a
9213 * reasonable floor to avoid funnies with rq->avg_idle.
9215 rq->max_idle_balance_cost =
9216 max((u64)sysctl_sched_migration_cost, max_cost);
9221 * next_balance will be updated only when there is a need.
9222 * When the cpu is attached to null domain for ex, it will not be
9225 if (likely(update_next_balance)) {
9226 rq->next_balance = next_balance;
9228 #ifdef CONFIG_NO_HZ_COMMON
9230 * If this CPU has been elected to perform the nohz idle
9231 * balance. Other idle CPUs have already rebalanced with
9232 * nohz_idle_balance() and nohz.next_balance has been
9233 * updated accordingly. This CPU is now running the idle load
9234 * balance for itself and we need to update the
9235 * nohz.next_balance accordingly.
9237 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9238 nohz.next_balance = rq->next_balance;
9243 #ifdef CONFIG_NO_HZ_COMMON
9245 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9246 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9248 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9250 int this_cpu = this_rq->cpu;
9253 /* Earliest time when we have to do rebalance again */
9254 unsigned long next_balance = jiffies + 60*HZ;
9255 int update_next_balance = 0;
9257 if (idle != CPU_IDLE ||
9258 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
9261 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9262 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9266 * If this cpu gets work to do, stop the load balancing
9267 * work being done for other cpus. Next load
9268 * balancing owner will pick it up.
9273 rq = cpu_rq(balance_cpu);
9276 * If time for next balance is due,
9279 if (time_after_eq(jiffies, rq->next_balance)) {
9282 rq_lock_irq(rq, &rf);
9283 update_rq_clock(rq);
9284 cpu_load_update_idle(rq);
9285 rq_unlock_irq(rq, &rf);
9287 rebalance_domains(rq, CPU_IDLE);
9290 if (time_after(next_balance, rq->next_balance)) {
9291 next_balance = rq->next_balance;
9292 update_next_balance = 1;
9297 * next_balance will be updated only when there is a need.
9298 * When the CPU is attached to null domain for ex, it will not be
9301 if (likely(update_next_balance))
9302 nohz.next_balance = next_balance;
9304 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
9308 * Current heuristic for kicking the idle load balancer in the presence
9309 * of an idle cpu in the system.
9310 * - This rq has more than one task.
9311 * - This rq has at least one CFS task and the capacity of the CPU is
9312 * significantly reduced because of RT tasks or IRQs.
9313 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9314 * multiple busy cpu.
9315 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9316 * domain span are idle.
9318 static inline bool nohz_kick_needed(struct rq *rq)
9320 unsigned long now = jiffies;
9321 struct sched_domain_shared *sds;
9322 struct sched_domain *sd;
9323 int nr_busy, i, cpu = rq->cpu;
9326 if (unlikely(rq->idle_balance))
9330 * We may be recently in ticked or tickless idle mode. At the first
9331 * busy tick after returning from idle, we will update the busy stats.
9333 set_cpu_sd_state_busy();
9334 nohz_balance_exit_idle(cpu);
9337 * None are in tickless mode and hence no need for NOHZ idle load
9340 if (likely(!atomic_read(&nohz.nr_cpus)))
9343 if (time_before(now, nohz.next_balance))
9346 if (rq->nr_running >= 2)
9350 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9353 * XXX: write a coherent comment on why we do this.
9354 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9356 nr_busy = atomic_read(&sds->nr_busy_cpus);
9364 sd = rcu_dereference(rq->sd);
9366 if ((rq->cfs.h_nr_running >= 1) &&
9367 check_cpu_capacity(rq, sd)) {
9373 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9375 for_each_cpu(i, sched_domain_span(sd)) {
9377 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9380 if (sched_asym_prefer(i, cpu)) {
9391 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9395 * run_rebalance_domains is triggered when needed from the scheduler tick.
9396 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9398 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9400 struct rq *this_rq = this_rq();
9401 enum cpu_idle_type idle = this_rq->idle_balance ?
9402 CPU_IDLE : CPU_NOT_IDLE;
9405 * If this cpu has a pending nohz_balance_kick, then do the
9406 * balancing on behalf of the other idle cpus whose ticks are
9407 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9408 * give the idle cpus a chance to load balance. Else we may
9409 * load balance only within the local sched_domain hierarchy
9410 * and abort nohz_idle_balance altogether if we pull some load.
9412 nohz_idle_balance(this_rq, idle);
9413 rebalance_domains(this_rq, idle);
9417 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9419 void trigger_load_balance(struct rq *rq)
9421 /* Don't need to rebalance while attached to NULL domain */
9422 if (unlikely(on_null_domain(rq)))
9425 if (time_after_eq(jiffies, rq->next_balance))
9426 raise_softirq(SCHED_SOFTIRQ);
9427 #ifdef CONFIG_NO_HZ_COMMON
9428 if (nohz_kick_needed(rq))
9429 nohz_balancer_kick();
9433 static void rq_online_fair(struct rq *rq)
9437 update_runtime_enabled(rq);
9440 static void rq_offline_fair(struct rq *rq)
9444 /* Ensure any throttled groups are reachable by pick_next_task */
9445 unthrottle_offline_cfs_rqs(rq);
9448 #endif /* CONFIG_SMP */
9451 * scheduler tick hitting a task of our scheduling class:
9453 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9455 struct cfs_rq *cfs_rq;
9456 struct sched_entity *se = &curr->se;
9458 for_each_sched_entity(se) {
9459 cfs_rq = cfs_rq_of(se);
9460 entity_tick(cfs_rq, se, queued);
9463 if (static_branch_unlikely(&sched_numa_balancing))
9464 task_tick_numa(rq, curr);
9468 * called on fork with the child task as argument from the parent's context
9469 * - child not yet on the tasklist
9470 * - preemption disabled
9472 static void task_fork_fair(struct task_struct *p)
9474 struct cfs_rq *cfs_rq;
9475 struct sched_entity *se = &p->se, *curr;
9476 struct rq *rq = this_rq();
9480 update_rq_clock(rq);
9482 cfs_rq = task_cfs_rq(current);
9483 curr = cfs_rq->curr;
9485 update_curr(cfs_rq);
9486 se->vruntime = curr->vruntime;
9488 place_entity(cfs_rq, se, 1);
9490 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9492 * Upon rescheduling, sched_class::put_prev_task() will place
9493 * 'current' within the tree based on its new key value.
9495 swap(curr->vruntime, se->vruntime);
9499 se->vruntime -= cfs_rq->min_vruntime;
9504 * Priority of the task has changed. Check to see if we preempt
9508 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9510 if (!task_on_rq_queued(p))
9514 * Reschedule if we are currently running on this runqueue and
9515 * our priority decreased, or if we are not currently running on
9516 * this runqueue and our priority is higher than the current's
9518 if (rq->curr == p) {
9519 if (p->prio > oldprio)
9522 check_preempt_curr(rq, p, 0);
9525 static inline bool vruntime_normalized(struct task_struct *p)
9527 struct sched_entity *se = &p->se;
9530 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9531 * the dequeue_entity(.flags=0) will already have normalized the
9538 * When !on_rq, vruntime of the task has usually NOT been normalized.
9539 * But there are some cases where it has already been normalized:
9541 * - A forked child which is waiting for being woken up by
9542 * wake_up_new_task().
9543 * - A task which has been woken up by try_to_wake_up() and
9544 * waiting for actually being woken up by sched_ttwu_pending().
9546 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9552 #ifdef CONFIG_FAIR_GROUP_SCHED
9554 * Propagate the changes of the sched_entity across the tg tree to make it
9555 * visible to the root
9557 static void propagate_entity_cfs_rq(struct sched_entity *se)
9559 struct cfs_rq *cfs_rq;
9561 /* Start to propagate at parent */
9564 for_each_sched_entity(se) {
9565 cfs_rq = cfs_rq_of(se);
9567 if (cfs_rq_throttled(cfs_rq))
9570 update_load_avg(cfs_rq, se, UPDATE_TG);
9574 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9577 static void detach_entity_cfs_rq(struct sched_entity *se)
9579 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9581 /* Catch up with the cfs_rq and remove our load when we leave */
9582 update_load_avg(cfs_rq, se, 0);
9583 detach_entity_load_avg(cfs_rq, se);
9584 update_tg_load_avg(cfs_rq, false);
9585 propagate_entity_cfs_rq(se);
9588 static void attach_entity_cfs_rq(struct sched_entity *se)
9590 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9592 #ifdef CONFIG_FAIR_GROUP_SCHED
9594 * Since the real-depth could have been changed (only FAIR
9595 * class maintain depth value), reset depth properly.
9597 se->depth = se->parent ? se->parent->depth + 1 : 0;
9600 /* Synchronize entity with its cfs_rq */
9601 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9602 attach_entity_load_avg(cfs_rq, se);
9603 update_tg_load_avg(cfs_rq, false);
9604 propagate_entity_cfs_rq(se);
9607 static void detach_task_cfs_rq(struct task_struct *p)
9609 struct sched_entity *se = &p->se;
9610 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9612 if (!vruntime_normalized(p)) {
9614 * Fix up our vruntime so that the current sleep doesn't
9615 * cause 'unlimited' sleep bonus.
9617 place_entity(cfs_rq, se, 0);
9618 se->vruntime -= cfs_rq->min_vruntime;
9621 detach_entity_cfs_rq(se);
9624 static void attach_task_cfs_rq(struct task_struct *p)
9626 struct sched_entity *se = &p->se;
9627 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9629 attach_entity_cfs_rq(se);
9631 if (!vruntime_normalized(p))
9632 se->vruntime += cfs_rq->min_vruntime;
9635 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9637 detach_task_cfs_rq(p);
9640 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9642 attach_task_cfs_rq(p);
9644 if (task_on_rq_queued(p)) {
9646 * We were most likely switched from sched_rt, so
9647 * kick off the schedule if running, otherwise just see
9648 * if we can still preempt the current task.
9653 check_preempt_curr(rq, p, 0);
9657 /* Account for a task changing its policy or group.
9659 * This routine is mostly called to set cfs_rq->curr field when a task
9660 * migrates between groups/classes.
9662 static void set_curr_task_fair(struct rq *rq)
9664 struct sched_entity *se = &rq->curr->se;
9666 for_each_sched_entity(se) {
9667 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9669 set_next_entity(cfs_rq, se);
9670 /* ensure bandwidth has been allocated on our new cfs_rq */
9671 account_cfs_rq_runtime(cfs_rq, 0);
9675 void init_cfs_rq(struct cfs_rq *cfs_rq)
9677 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9678 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9679 #ifndef CONFIG_64BIT
9680 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9683 raw_spin_lock_init(&cfs_rq->removed.lock);
9687 #ifdef CONFIG_FAIR_GROUP_SCHED
9688 static void task_set_group_fair(struct task_struct *p)
9690 struct sched_entity *se = &p->se;
9692 set_task_rq(p, task_cpu(p));
9693 se->depth = se->parent ? se->parent->depth + 1 : 0;
9696 static void task_move_group_fair(struct task_struct *p)
9698 detach_task_cfs_rq(p);
9699 set_task_rq(p, task_cpu(p));
9702 /* Tell se's cfs_rq has been changed -- migrated */
9703 p->se.avg.last_update_time = 0;
9705 attach_task_cfs_rq(p);
9708 static void task_change_group_fair(struct task_struct *p, int type)
9711 case TASK_SET_GROUP:
9712 task_set_group_fair(p);
9715 case TASK_MOVE_GROUP:
9716 task_move_group_fair(p);
9721 void free_fair_sched_group(struct task_group *tg)
9725 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9727 for_each_possible_cpu(i) {
9729 kfree(tg->cfs_rq[i]);
9738 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9740 struct sched_entity *se;
9741 struct cfs_rq *cfs_rq;
9744 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9747 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9751 tg->shares = NICE_0_LOAD;
9753 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9755 for_each_possible_cpu(i) {
9756 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9757 GFP_KERNEL, cpu_to_node(i));
9761 se = kzalloc_node(sizeof(struct sched_entity),
9762 GFP_KERNEL, cpu_to_node(i));
9766 init_cfs_rq(cfs_rq);
9767 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9768 init_entity_runnable_average(se);
9779 void online_fair_sched_group(struct task_group *tg)
9781 struct sched_entity *se;
9785 for_each_possible_cpu(i) {
9789 raw_spin_lock_irq(&rq->lock);
9790 update_rq_clock(rq);
9791 attach_entity_cfs_rq(se);
9792 sync_throttle(tg, i);
9793 raw_spin_unlock_irq(&rq->lock);
9797 void unregister_fair_sched_group(struct task_group *tg)
9799 unsigned long flags;
9803 for_each_possible_cpu(cpu) {
9805 remove_entity_load_avg(tg->se[cpu]);
9808 * Only empty task groups can be destroyed; so we can speculatively
9809 * check on_list without danger of it being re-added.
9811 if (!tg->cfs_rq[cpu]->on_list)
9816 raw_spin_lock_irqsave(&rq->lock, flags);
9817 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9818 raw_spin_unlock_irqrestore(&rq->lock, flags);
9822 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9823 struct sched_entity *se, int cpu,
9824 struct sched_entity *parent)
9826 struct rq *rq = cpu_rq(cpu);
9830 init_cfs_rq_runtime(cfs_rq);
9832 tg->cfs_rq[cpu] = cfs_rq;
9835 /* se could be NULL for root_task_group */
9840 se->cfs_rq = &rq->cfs;
9843 se->cfs_rq = parent->my_q;
9844 se->depth = parent->depth + 1;
9848 /* guarantee group entities always have weight */
9849 update_load_set(&se->load, NICE_0_LOAD);
9850 se->parent = parent;
9853 static DEFINE_MUTEX(shares_mutex);
9855 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9860 * We can't change the weight of the root cgroup.
9865 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9867 mutex_lock(&shares_mutex);
9868 if (tg->shares == shares)
9871 tg->shares = shares;
9872 for_each_possible_cpu(i) {
9873 struct rq *rq = cpu_rq(i);
9874 struct sched_entity *se = tg->se[i];
9877 /* Propagate contribution to hierarchy */
9878 rq_lock_irqsave(rq, &rf);
9879 update_rq_clock(rq);
9880 for_each_sched_entity(se) {
9881 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9882 update_cfs_group(se);
9884 rq_unlock_irqrestore(rq, &rf);
9888 mutex_unlock(&shares_mutex);
9891 #else /* CONFIG_FAIR_GROUP_SCHED */
9893 void free_fair_sched_group(struct task_group *tg) { }
9895 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9900 void online_fair_sched_group(struct task_group *tg) { }
9902 void unregister_fair_sched_group(struct task_group *tg) { }
9904 #endif /* CONFIG_FAIR_GROUP_SCHED */
9907 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9909 struct sched_entity *se = &task->se;
9910 unsigned int rr_interval = 0;
9913 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9916 if (rq->cfs.load.weight)
9917 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9923 * All the scheduling class methods:
9925 const struct sched_class fair_sched_class = {
9926 .next = &idle_sched_class,
9927 .enqueue_task = enqueue_task_fair,
9928 .dequeue_task = dequeue_task_fair,
9929 .yield_task = yield_task_fair,
9930 .yield_to_task = yield_to_task_fair,
9932 .check_preempt_curr = check_preempt_wakeup,
9934 .pick_next_task = pick_next_task_fair,
9935 .put_prev_task = put_prev_task_fair,
9938 .select_task_rq = select_task_rq_fair,
9939 .migrate_task_rq = migrate_task_rq_fair,
9941 .rq_online = rq_online_fair,
9942 .rq_offline = rq_offline_fair,
9944 .task_dead = task_dead_fair,
9945 .set_cpus_allowed = set_cpus_allowed_common,
9948 .set_curr_task = set_curr_task_fair,
9949 .task_tick = task_tick_fair,
9950 .task_fork = task_fork_fair,
9952 .prio_changed = prio_changed_fair,
9953 .switched_from = switched_from_fair,
9954 .switched_to = switched_to_fair,
9956 .get_rr_interval = get_rr_interval_fair,
9958 .update_curr = update_curr_fair,
9960 #ifdef CONFIG_FAIR_GROUP_SCHED
9961 .task_change_group = task_change_group_fair,
9965 #ifdef CONFIG_SCHED_DEBUG
9966 void print_cfs_stats(struct seq_file *m, int cpu)
9968 struct cfs_rq *cfs_rq, *pos;
9971 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
9972 print_cfs_rq(m, cpu, cfs_rq);
9976 #ifdef CONFIG_NUMA_BALANCING
9977 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9980 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9982 for_each_online_node(node) {
9983 if (p->numa_faults) {
9984 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9985 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9987 if (p->numa_group) {
9988 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9989 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9991 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9994 #endif /* CONFIG_NUMA_BALANCING */
9995 #endif /* CONFIG_SCHED_DEBUG */
9997 __init void init_sched_fair_class(void)
10000 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10002 #ifdef CONFIG_NO_HZ_COMMON
10003 nohz.next_balance = jiffies;
10004 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);