2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/sched.h>
24 #include <linux/latencytop.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
118 * The margin used when comparing utilization with CPU capacity:
119 * util * 1024 < capacity * margin
121 unsigned int capacity_margin = 1280; /* ~20% */
123 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
129 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
135 static inline void update_load_set(struct load_weight *lw, unsigned long w)
142 * Increase the granularity value when there are more CPUs,
143 * because with more CPUs the 'effective latency' as visible
144 * to users decreases. But the relationship is not linear,
145 * so pick a second-best guess by going with the log2 of the
148 * This idea comes from the SD scheduler of Con Kolivas:
150 static unsigned int get_update_sysctl_factor(void)
152 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
155 switch (sysctl_sched_tunable_scaling) {
156 case SCHED_TUNABLESCALING_NONE:
159 case SCHED_TUNABLESCALING_LINEAR:
162 case SCHED_TUNABLESCALING_LOG:
164 factor = 1 + ilog2(cpus);
171 static void update_sysctl(void)
173 unsigned int factor = get_update_sysctl_factor();
175 #define SET_SYSCTL(name) \
176 (sysctl_##name = (factor) * normalized_sysctl_##name)
177 SET_SYSCTL(sched_min_granularity);
178 SET_SYSCTL(sched_latency);
179 SET_SYSCTL(sched_wakeup_granularity);
183 void sched_init_granularity(void)
188 #define WMULT_CONST (~0U)
189 #define WMULT_SHIFT 32
191 static void __update_inv_weight(struct load_weight *lw)
195 if (likely(lw->inv_weight))
198 w = scale_load_down(lw->weight);
200 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
202 else if (unlikely(!w))
203 lw->inv_weight = WMULT_CONST;
205 lw->inv_weight = WMULT_CONST / w;
209 * delta_exec * weight / lw.weight
211 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
213 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
214 * we're guaranteed shift stays positive because inv_weight is guaranteed to
215 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
217 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
218 * weight/lw.weight <= 1, and therefore our shift will also be positive.
220 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
222 u64 fact = scale_load_down(weight);
223 int shift = WMULT_SHIFT;
225 __update_inv_weight(lw);
227 if (unlikely(fact >> 32)) {
234 /* hint to use a 32x32->64 mul */
235 fact = (u64)(u32)fact * lw->inv_weight;
242 return mul_u64_u32_shr(delta_exec, fact, shift);
246 const struct sched_class fair_sched_class;
248 /**************************************************************
249 * CFS operations on generic schedulable entities:
252 #ifdef CONFIG_FAIR_GROUP_SCHED
254 /* cpu runqueue to which this cfs_rq is attached */
255 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
260 /* An entity is a task if it doesn't "own" a runqueue */
261 #define entity_is_task(se) (!se->my_q)
263 static inline struct task_struct *task_of(struct sched_entity *se)
265 SCHED_WARN_ON(!entity_is_task(se));
266 return container_of(se, struct task_struct, se);
269 /* Walk up scheduling entities hierarchy */
270 #define for_each_sched_entity(se) \
271 for (; se; se = se->parent)
273 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
278 /* runqueue on which this entity is (to be) queued */
279 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
284 /* runqueue "owned" by this group */
285 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
290 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
292 if (!cfs_rq->on_list) {
294 * Ensure we either appear before our parent (if already
295 * enqueued) or force our parent to appear after us when it is
296 * enqueued. The fact that we always enqueue bottom-up
297 * reduces this to two cases.
299 if (cfs_rq->tg->parent &&
300 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
301 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
302 &rq_of(cfs_rq)->leaf_cfs_rq_list);
304 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
305 &rq_of(cfs_rq)->leaf_cfs_rq_list);
312 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
314 if (cfs_rq->on_list) {
315 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
320 /* Iterate thr' all leaf cfs_rq's on a runqueue */
321 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
322 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
324 /* Do the two (enqueued) entities belong to the same group ? */
325 static inline struct cfs_rq *
326 is_same_group(struct sched_entity *se, struct sched_entity *pse)
328 if (se->cfs_rq == pse->cfs_rq)
334 static inline struct sched_entity *parent_entity(struct sched_entity *se)
340 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
342 int se_depth, pse_depth;
345 * preemption test can be made between sibling entities who are in the
346 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
347 * both tasks until we find their ancestors who are siblings of common
351 /* First walk up until both entities are at same depth */
352 se_depth = (*se)->depth;
353 pse_depth = (*pse)->depth;
355 while (se_depth > pse_depth) {
357 *se = parent_entity(*se);
360 while (pse_depth > se_depth) {
362 *pse = parent_entity(*pse);
365 while (!is_same_group(*se, *pse)) {
366 *se = parent_entity(*se);
367 *pse = parent_entity(*pse);
371 #else /* !CONFIG_FAIR_GROUP_SCHED */
373 static inline struct task_struct *task_of(struct sched_entity *se)
375 return container_of(se, struct task_struct, se);
378 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
380 return container_of(cfs_rq, struct rq, cfs);
383 #define entity_is_task(se) 1
385 #define for_each_sched_entity(se) \
386 for (; se; se = NULL)
388 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
390 return &task_rq(p)->cfs;
393 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
395 struct task_struct *p = task_of(se);
396 struct rq *rq = task_rq(p);
401 /* runqueue "owned" by this group */
402 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
407 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
415 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
416 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
418 static inline struct sched_entity *parent_entity(struct sched_entity *se)
424 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
428 #endif /* CONFIG_FAIR_GROUP_SCHED */
430 static __always_inline
431 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
433 /**************************************************************
434 * Scheduling class tree data structure manipulation methods:
437 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
439 s64 delta = (s64)(vruntime - max_vruntime);
441 max_vruntime = vruntime;
446 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
448 s64 delta = (s64)(vruntime - min_vruntime);
450 min_vruntime = vruntime;
455 static inline int entity_before(struct sched_entity *a,
456 struct sched_entity *b)
458 return (s64)(a->vruntime - b->vruntime) < 0;
461 static void update_min_vruntime(struct cfs_rq *cfs_rq)
463 struct sched_entity *curr = cfs_rq->curr;
465 u64 vruntime = cfs_rq->min_vruntime;
469 vruntime = curr->vruntime;
474 if (cfs_rq->rb_leftmost) {
475 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
480 vruntime = se->vruntime;
482 vruntime = min_vruntime(vruntime, se->vruntime);
485 /* ensure we never gain time by being placed backwards. */
486 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
489 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
494 * Enqueue an entity into the rb-tree:
496 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
498 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
499 struct rb_node *parent = NULL;
500 struct sched_entity *entry;
504 * Find the right place in the rbtree:
508 entry = rb_entry(parent, struct sched_entity, run_node);
510 * We dont care about collisions. Nodes with
511 * the same key stay together.
513 if (entity_before(se, entry)) {
514 link = &parent->rb_left;
516 link = &parent->rb_right;
522 * Maintain a cache of leftmost tree entries (it is frequently
526 cfs_rq->rb_leftmost = &se->run_node;
528 rb_link_node(&se->run_node, parent, link);
529 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
532 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
534 if (cfs_rq->rb_leftmost == &se->run_node) {
535 struct rb_node *next_node;
537 next_node = rb_next(&se->run_node);
538 cfs_rq->rb_leftmost = next_node;
541 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
544 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
546 struct rb_node *left = cfs_rq->rb_leftmost;
551 return rb_entry(left, struct sched_entity, run_node);
554 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
556 struct rb_node *next = rb_next(&se->run_node);
561 return rb_entry(next, struct sched_entity, run_node);
564 #ifdef CONFIG_SCHED_DEBUG
565 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
567 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
572 return rb_entry(last, struct sched_entity, run_node);
575 /**************************************************************
576 * Scheduling class statistics methods:
579 int sched_proc_update_handler(struct ctl_table *table, int write,
580 void __user *buffer, size_t *lenp,
583 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
584 unsigned int factor = get_update_sysctl_factor();
589 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
590 sysctl_sched_min_granularity);
592 #define WRT_SYSCTL(name) \
593 (normalized_sysctl_##name = sysctl_##name / (factor))
594 WRT_SYSCTL(sched_min_granularity);
595 WRT_SYSCTL(sched_latency);
596 WRT_SYSCTL(sched_wakeup_granularity);
606 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
608 if (unlikely(se->load.weight != NICE_0_LOAD))
609 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
615 * The idea is to set a period in which each task runs once.
617 * When there are too many tasks (sched_nr_latency) we have to stretch
618 * this period because otherwise the slices get too small.
620 * p = (nr <= nl) ? l : l*nr/nl
622 static u64 __sched_period(unsigned long nr_running)
624 if (unlikely(nr_running > sched_nr_latency))
625 return nr_running * sysctl_sched_min_granularity;
627 return sysctl_sched_latency;
631 * We calculate the wall-time slice from the period by taking a part
632 * proportional to the weight.
636 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
638 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
640 for_each_sched_entity(se) {
641 struct load_weight *load;
642 struct load_weight lw;
644 cfs_rq = cfs_rq_of(se);
645 load = &cfs_rq->load;
647 if (unlikely(!se->on_rq)) {
650 update_load_add(&lw, se->load.weight);
653 slice = __calc_delta(slice, se->load.weight, load);
659 * We calculate the vruntime slice of a to-be-inserted task.
663 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
665 return calc_delta_fair(sched_slice(cfs_rq, se), se);
669 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
670 static unsigned long task_h_load(struct task_struct *p);
673 * We choose a half-life close to 1 scheduling period.
674 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
675 * dependent on this value.
677 #define LOAD_AVG_PERIOD 32
678 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
679 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
681 /* Give new sched_entity start runnable values to heavy its load in infant time */
682 void init_entity_runnable_average(struct sched_entity *se)
684 struct sched_avg *sa = &se->avg;
686 sa->last_update_time = 0;
688 * sched_avg's period_contrib should be strictly less then 1024, so
689 * we give it 1023 to make sure it is almost a period (1024us), and
690 * will definitely be update (after enqueue).
692 sa->period_contrib = 1023;
693 sa->load_avg = scale_load_down(se->load.weight);
694 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
696 * At this point, util_avg won't be used in select_task_rq_fair anyway
700 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
703 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
704 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
705 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
706 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
709 * With new tasks being created, their initial util_avgs are extrapolated
710 * based on the cfs_rq's current util_avg:
712 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
714 * However, in many cases, the above util_avg does not give a desired
715 * value. Moreover, the sum of the util_avgs may be divergent, such
716 * as when the series is a harmonic series.
718 * To solve this problem, we also cap the util_avg of successive tasks to
719 * only 1/2 of the left utilization budget:
721 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
723 * where n denotes the nth task.
725 * For example, a simplest series from the beginning would be like:
727 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
728 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
730 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
731 * if util_avg > util_avg_cap.
733 void post_init_entity_util_avg(struct sched_entity *se)
735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
736 struct sched_avg *sa = &se->avg;
737 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
738 u64 now = cfs_rq_clock_task(cfs_rq);
741 if (cfs_rq->avg.util_avg != 0) {
742 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
743 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
745 if (sa->util_avg > cap)
750 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
753 if (entity_is_task(se)) {
754 struct task_struct *p = task_of(se);
755 if (p->sched_class != &fair_sched_class) {
757 * For !fair tasks do:
759 update_cfs_rq_load_avg(now, cfs_rq, false);
760 attach_entity_load_avg(cfs_rq, se);
761 switched_from_fair(rq, p);
763 * such that the next switched_to_fair() has the
766 se->avg.last_update_time = now;
771 update_cfs_rq_load_avg(now, cfs_rq, false);
772 attach_entity_load_avg(cfs_rq, se);
773 update_tg_load_avg(cfs_rq, false);
776 #else /* !CONFIG_SMP */
777 void init_entity_runnable_average(struct sched_entity *se)
780 void post_init_entity_util_avg(struct sched_entity *se)
783 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
786 #endif /* CONFIG_SMP */
789 * Update the current task's runtime statistics.
791 static void update_curr(struct cfs_rq *cfs_rq)
793 struct sched_entity *curr = cfs_rq->curr;
794 u64 now = rq_clock_task(rq_of(cfs_rq));
800 delta_exec = now - curr->exec_start;
801 if (unlikely((s64)delta_exec <= 0))
804 curr->exec_start = now;
806 schedstat_set(curr->statistics.exec_max,
807 max(delta_exec, curr->statistics.exec_max));
809 curr->sum_exec_runtime += delta_exec;
810 schedstat_add(cfs_rq->exec_clock, delta_exec);
812 curr->vruntime += calc_delta_fair(delta_exec, curr);
813 update_min_vruntime(cfs_rq);
815 if (entity_is_task(curr)) {
816 struct task_struct *curtask = task_of(curr);
818 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
819 cpuacct_charge(curtask, delta_exec);
820 account_group_exec_runtime(curtask, delta_exec);
823 account_cfs_rq_runtime(cfs_rq, delta_exec);
826 static void update_curr_fair(struct rq *rq)
828 update_curr(cfs_rq_of(&rq->curr->se));
832 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
834 u64 wait_start, prev_wait_start;
836 if (!schedstat_enabled())
839 wait_start = rq_clock(rq_of(cfs_rq));
840 prev_wait_start = schedstat_val(se->statistics.wait_start);
842 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
843 likely(wait_start > prev_wait_start))
844 wait_start -= prev_wait_start;
846 schedstat_set(se->statistics.wait_start, wait_start);
850 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
852 struct task_struct *p;
855 if (!schedstat_enabled())
858 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
860 if (entity_is_task(se)) {
862 if (task_on_rq_migrating(p)) {
864 * Preserve migrating task's wait time so wait_start
865 * time stamp can be adjusted to accumulate wait time
866 * prior to migration.
868 schedstat_set(se->statistics.wait_start, delta);
871 trace_sched_stat_wait(p, delta);
874 schedstat_set(se->statistics.wait_max,
875 max(schedstat_val(se->statistics.wait_max), delta));
876 schedstat_inc(se->statistics.wait_count);
877 schedstat_add(se->statistics.wait_sum, delta);
878 schedstat_set(se->statistics.wait_start, 0);
882 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 struct task_struct *tsk = NULL;
885 u64 sleep_start, block_start;
887 if (!schedstat_enabled())
890 sleep_start = schedstat_val(se->statistics.sleep_start);
891 block_start = schedstat_val(se->statistics.block_start);
893 if (entity_is_task(se))
897 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
902 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
903 schedstat_set(se->statistics.sleep_max, delta);
905 schedstat_set(se->statistics.sleep_start, 0);
906 schedstat_add(se->statistics.sum_sleep_runtime, delta);
909 account_scheduler_latency(tsk, delta >> 10, 1);
910 trace_sched_stat_sleep(tsk, delta);
914 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
919 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
920 schedstat_set(se->statistics.block_max, delta);
922 schedstat_set(se->statistics.block_start, 0);
923 schedstat_add(se->statistics.sum_sleep_runtime, delta);
926 if (tsk->in_iowait) {
927 schedstat_add(se->statistics.iowait_sum, delta);
928 schedstat_inc(se->statistics.iowait_count);
929 trace_sched_stat_iowait(tsk, delta);
932 trace_sched_stat_blocked(tsk, delta);
935 * Blocking time is in units of nanosecs, so shift by
936 * 20 to get a milliseconds-range estimation of the
937 * amount of time that the task spent sleeping:
939 if (unlikely(prof_on == SLEEP_PROFILING)) {
940 profile_hits(SLEEP_PROFILING,
941 (void *)get_wchan(tsk),
944 account_scheduler_latency(tsk, delta >> 10, 0);
950 * Task is being enqueued - update stats:
953 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
955 if (!schedstat_enabled())
959 * Are we enqueueing a waiting task? (for current tasks
960 * a dequeue/enqueue event is a NOP)
962 if (se != cfs_rq->curr)
963 update_stats_wait_start(cfs_rq, se);
965 if (flags & ENQUEUE_WAKEUP)
966 update_stats_enqueue_sleeper(cfs_rq, se);
970 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
973 if (!schedstat_enabled())
977 * Mark the end of the wait period if dequeueing a
980 if (se != cfs_rq->curr)
981 update_stats_wait_end(cfs_rq, se);
983 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
984 struct task_struct *tsk = task_of(se);
986 if (tsk->state & TASK_INTERRUPTIBLE)
987 schedstat_set(se->statistics.sleep_start,
988 rq_clock(rq_of(cfs_rq)));
989 if (tsk->state & TASK_UNINTERRUPTIBLE)
990 schedstat_set(se->statistics.block_start,
991 rq_clock(rq_of(cfs_rq)));
996 * We are picking a new current task - update its stats:
999 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1002 * We are starting a new run period:
1004 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1007 /**************************************************
1008 * Scheduling class queueing methods:
1011 #ifdef CONFIG_NUMA_BALANCING
1013 * Approximate time to scan a full NUMA task in ms. The task scan period is
1014 * calculated based on the tasks virtual memory size and
1015 * numa_balancing_scan_size.
1017 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1018 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1020 /* Portion of address space to scan in MB */
1021 unsigned int sysctl_numa_balancing_scan_size = 256;
1023 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1024 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1026 static unsigned int task_nr_scan_windows(struct task_struct *p)
1028 unsigned long rss = 0;
1029 unsigned long nr_scan_pages;
1032 * Calculations based on RSS as non-present and empty pages are skipped
1033 * by the PTE scanner and NUMA hinting faults should be trapped based
1036 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1037 rss = get_mm_rss(p->mm);
1039 rss = nr_scan_pages;
1041 rss = round_up(rss, nr_scan_pages);
1042 return rss / nr_scan_pages;
1045 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1046 #define MAX_SCAN_WINDOW 2560
1048 static unsigned int task_scan_min(struct task_struct *p)
1050 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1051 unsigned int scan, floor;
1052 unsigned int windows = 1;
1054 if (scan_size < MAX_SCAN_WINDOW)
1055 windows = MAX_SCAN_WINDOW / scan_size;
1056 floor = 1000 / windows;
1058 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1059 return max_t(unsigned int, floor, scan);
1062 static unsigned int task_scan_max(struct task_struct *p)
1064 unsigned int smin = task_scan_min(p);
1067 /* Watch for min being lower than max due to floor calculations */
1068 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1069 return max(smin, smax);
1072 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1074 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1075 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1078 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1080 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1081 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1087 spinlock_t lock; /* nr_tasks, tasks */
1092 struct rcu_head rcu;
1093 unsigned long total_faults;
1094 unsigned long max_faults_cpu;
1096 * Faults_cpu is used to decide whether memory should move
1097 * towards the CPU. As a consequence, these stats are weighted
1098 * more by CPU use than by memory faults.
1100 unsigned long *faults_cpu;
1101 unsigned long faults[0];
1104 /* Shared or private faults. */
1105 #define NR_NUMA_HINT_FAULT_TYPES 2
1107 /* Memory and CPU locality */
1108 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1110 /* Averaged statistics, and temporary buffers. */
1111 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1113 pid_t task_numa_group_id(struct task_struct *p)
1115 return p->numa_group ? p->numa_group->gid : 0;
1119 * The averaged statistics, shared & private, memory & cpu,
1120 * occupy the first half of the array. The second half of the
1121 * array is for current counters, which are averaged into the
1122 * first set by task_numa_placement.
1124 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1126 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1129 static inline unsigned long task_faults(struct task_struct *p, int nid)
1131 if (!p->numa_faults)
1134 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1135 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1138 static inline unsigned long group_faults(struct task_struct *p, int nid)
1143 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1144 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1147 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1149 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1150 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1154 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1155 * considered part of a numa group's pseudo-interleaving set. Migrations
1156 * between these nodes are slowed down, to allow things to settle down.
1158 #define ACTIVE_NODE_FRACTION 3
1160 static bool numa_is_active_node(int nid, struct numa_group *ng)
1162 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1165 /* Handle placement on systems where not all nodes are directly connected. */
1166 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1167 int maxdist, bool task)
1169 unsigned long score = 0;
1173 * All nodes are directly connected, and the same distance
1174 * from each other. No need for fancy placement algorithms.
1176 if (sched_numa_topology_type == NUMA_DIRECT)
1180 * This code is called for each node, introducing N^2 complexity,
1181 * which should be ok given the number of nodes rarely exceeds 8.
1183 for_each_online_node(node) {
1184 unsigned long faults;
1185 int dist = node_distance(nid, node);
1188 * The furthest away nodes in the system are not interesting
1189 * for placement; nid was already counted.
1191 if (dist == sched_max_numa_distance || node == nid)
1195 * On systems with a backplane NUMA topology, compare groups
1196 * of nodes, and move tasks towards the group with the most
1197 * memory accesses. When comparing two nodes at distance
1198 * "hoplimit", only nodes closer by than "hoplimit" are part
1199 * of each group. Skip other nodes.
1201 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1205 /* Add up the faults from nearby nodes. */
1207 faults = task_faults(p, node);
1209 faults = group_faults(p, node);
1212 * On systems with a glueless mesh NUMA topology, there are
1213 * no fixed "groups of nodes". Instead, nodes that are not
1214 * directly connected bounce traffic through intermediate
1215 * nodes; a numa_group can occupy any set of nodes.
1216 * The further away a node is, the less the faults count.
1217 * This seems to result in good task placement.
1219 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1220 faults *= (sched_max_numa_distance - dist);
1221 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1231 * These return the fraction of accesses done by a particular task, or
1232 * task group, on a particular numa node. The group weight is given a
1233 * larger multiplier, in order to group tasks together that are almost
1234 * evenly spread out between numa nodes.
1236 static inline unsigned long task_weight(struct task_struct *p, int nid,
1239 unsigned long faults, total_faults;
1241 if (!p->numa_faults)
1244 total_faults = p->total_numa_faults;
1249 faults = task_faults(p, nid);
1250 faults += score_nearby_nodes(p, nid, dist, true);
1252 return 1000 * faults / total_faults;
1255 static inline unsigned long group_weight(struct task_struct *p, int nid,
1258 unsigned long faults, total_faults;
1263 total_faults = p->numa_group->total_faults;
1268 faults = group_faults(p, nid);
1269 faults += score_nearby_nodes(p, nid, dist, false);
1271 return 1000 * faults / total_faults;
1274 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1275 int src_nid, int dst_cpu)
1277 struct numa_group *ng = p->numa_group;
1278 int dst_nid = cpu_to_node(dst_cpu);
1279 int last_cpupid, this_cpupid;
1281 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1284 * Multi-stage node selection is used in conjunction with a periodic
1285 * migration fault to build a temporal task<->page relation. By using
1286 * a two-stage filter we remove short/unlikely relations.
1288 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1289 * a task's usage of a particular page (n_p) per total usage of this
1290 * page (n_t) (in a given time-span) to a probability.
1292 * Our periodic faults will sample this probability and getting the
1293 * same result twice in a row, given these samples are fully
1294 * independent, is then given by P(n)^2, provided our sample period
1295 * is sufficiently short compared to the usage pattern.
1297 * This quadric squishes small probabilities, making it less likely we
1298 * act on an unlikely task<->page relation.
1300 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1301 if (!cpupid_pid_unset(last_cpupid) &&
1302 cpupid_to_nid(last_cpupid) != dst_nid)
1305 /* Always allow migrate on private faults */
1306 if (cpupid_match_pid(p, last_cpupid))
1309 /* A shared fault, but p->numa_group has not been set up yet. */
1314 * Destination node is much more heavily used than the source
1315 * node? Allow migration.
1317 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1318 ACTIVE_NODE_FRACTION)
1322 * Distribute memory according to CPU & memory use on each node,
1323 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1325 * faults_cpu(dst) 3 faults_cpu(src)
1326 * --------------- * - > ---------------
1327 * faults_mem(dst) 4 faults_mem(src)
1329 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1330 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1333 static unsigned long weighted_cpuload(const int cpu);
1334 static unsigned long source_load(int cpu, int type);
1335 static unsigned long target_load(int cpu, int type);
1336 static unsigned long capacity_of(int cpu);
1337 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1339 /* Cached statistics for all CPUs within a node */
1341 unsigned long nr_running;
1344 /* Total compute capacity of CPUs on a node */
1345 unsigned long compute_capacity;
1347 /* Approximate capacity in terms of runnable tasks on a node */
1348 unsigned long task_capacity;
1349 int has_free_capacity;
1353 * XXX borrowed from update_sg_lb_stats
1355 static void update_numa_stats(struct numa_stats *ns, int nid)
1357 int smt, cpu, cpus = 0;
1358 unsigned long capacity;
1360 memset(ns, 0, sizeof(*ns));
1361 for_each_cpu(cpu, cpumask_of_node(nid)) {
1362 struct rq *rq = cpu_rq(cpu);
1364 ns->nr_running += rq->nr_running;
1365 ns->load += weighted_cpuload(cpu);
1366 ns->compute_capacity += capacity_of(cpu);
1372 * If we raced with hotplug and there are no CPUs left in our mask
1373 * the @ns structure is NULL'ed and task_numa_compare() will
1374 * not find this node attractive.
1376 * We'll either bail at !has_free_capacity, or we'll detect a huge
1377 * imbalance and bail there.
1382 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1383 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1384 capacity = cpus / smt; /* cores */
1386 ns->task_capacity = min_t(unsigned, capacity,
1387 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1388 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1391 struct task_numa_env {
1392 struct task_struct *p;
1394 int src_cpu, src_nid;
1395 int dst_cpu, dst_nid;
1397 struct numa_stats src_stats, dst_stats;
1402 struct task_struct *best_task;
1407 static void task_numa_assign(struct task_numa_env *env,
1408 struct task_struct *p, long imp)
1411 put_task_struct(env->best_task);
1416 env->best_imp = imp;
1417 env->best_cpu = env->dst_cpu;
1420 static bool load_too_imbalanced(long src_load, long dst_load,
1421 struct task_numa_env *env)
1424 long orig_src_load, orig_dst_load;
1425 long src_capacity, dst_capacity;
1428 * The load is corrected for the CPU capacity available on each node.
1431 * ------------ vs ---------
1432 * src_capacity dst_capacity
1434 src_capacity = env->src_stats.compute_capacity;
1435 dst_capacity = env->dst_stats.compute_capacity;
1437 /* We care about the slope of the imbalance, not the direction. */
1438 if (dst_load < src_load)
1439 swap(dst_load, src_load);
1441 /* Is the difference below the threshold? */
1442 imb = dst_load * src_capacity * 100 -
1443 src_load * dst_capacity * env->imbalance_pct;
1448 * The imbalance is above the allowed threshold.
1449 * Compare it with the old imbalance.
1451 orig_src_load = env->src_stats.load;
1452 orig_dst_load = env->dst_stats.load;
1454 if (orig_dst_load < orig_src_load)
1455 swap(orig_dst_load, orig_src_load);
1457 old_imb = orig_dst_load * src_capacity * 100 -
1458 orig_src_load * dst_capacity * env->imbalance_pct;
1460 /* Would this change make things worse? */
1461 return (imb > old_imb);
1465 * This checks if the overall compute and NUMA accesses of the system would
1466 * be improved if the source tasks was migrated to the target dst_cpu taking
1467 * into account that it might be best if task running on the dst_cpu should
1468 * be exchanged with the source task
1470 static void task_numa_compare(struct task_numa_env *env,
1471 long taskimp, long groupimp)
1473 struct rq *src_rq = cpu_rq(env->src_cpu);
1474 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1475 struct task_struct *cur;
1476 long src_load, dst_load;
1478 long imp = env->p->numa_group ? groupimp : taskimp;
1480 int dist = env->dist;
1483 cur = task_rcu_dereference(&dst_rq->curr);
1484 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1488 * Because we have preemption enabled we can get migrated around and
1489 * end try selecting ourselves (current == env->p) as a swap candidate.
1495 * "imp" is the fault differential for the source task between the
1496 * source and destination node. Calculate the total differential for
1497 * the source task and potential destination task. The more negative
1498 * the value is, the more rmeote accesses that would be expected to
1499 * be incurred if the tasks were swapped.
1502 /* Skip this swap candidate if cannot move to the source cpu */
1503 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1507 * If dst and source tasks are in the same NUMA group, or not
1508 * in any group then look only at task weights.
1510 if (cur->numa_group == env->p->numa_group) {
1511 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1512 task_weight(cur, env->dst_nid, dist);
1514 * Add some hysteresis to prevent swapping the
1515 * tasks within a group over tiny differences.
1517 if (cur->numa_group)
1521 * Compare the group weights. If a task is all by
1522 * itself (not part of a group), use the task weight
1525 if (cur->numa_group)
1526 imp += group_weight(cur, env->src_nid, dist) -
1527 group_weight(cur, env->dst_nid, dist);
1529 imp += task_weight(cur, env->src_nid, dist) -
1530 task_weight(cur, env->dst_nid, dist);
1534 if (imp <= env->best_imp && moveimp <= env->best_imp)
1538 /* Is there capacity at our destination? */
1539 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1540 !env->dst_stats.has_free_capacity)
1546 /* Balance doesn't matter much if we're running a task per cpu */
1547 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1548 dst_rq->nr_running == 1)
1552 * In the overloaded case, try and keep the load balanced.
1555 load = task_h_load(env->p);
1556 dst_load = env->dst_stats.load + load;
1557 src_load = env->src_stats.load - load;
1559 if (moveimp > imp && moveimp > env->best_imp) {
1561 * If the improvement from just moving env->p direction is
1562 * better than swapping tasks around, check if a move is
1563 * possible. Store a slightly smaller score than moveimp,
1564 * so an actually idle CPU will win.
1566 if (!load_too_imbalanced(src_load, dst_load, env)) {
1573 if (imp <= env->best_imp)
1577 load = task_h_load(cur);
1582 if (load_too_imbalanced(src_load, dst_load, env))
1586 * One idle CPU per node is evaluated for a task numa move.
1587 * Call select_idle_sibling to maybe find a better one.
1591 * select_idle_siblings() uses an per-cpu cpumask that
1592 * can be used from IRQ context.
1594 local_irq_disable();
1595 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1601 task_numa_assign(env, cur, imp);
1606 static void task_numa_find_cpu(struct task_numa_env *env,
1607 long taskimp, long groupimp)
1611 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1612 /* Skip this CPU if the source task cannot migrate */
1613 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1617 task_numa_compare(env, taskimp, groupimp);
1621 /* Only move tasks to a NUMA node less busy than the current node. */
1622 static bool numa_has_capacity(struct task_numa_env *env)
1624 struct numa_stats *src = &env->src_stats;
1625 struct numa_stats *dst = &env->dst_stats;
1627 if (src->has_free_capacity && !dst->has_free_capacity)
1631 * Only consider a task move if the source has a higher load
1632 * than the destination, corrected for CPU capacity on each node.
1634 * src->load dst->load
1635 * --------------------- vs ---------------------
1636 * src->compute_capacity dst->compute_capacity
1638 if (src->load * dst->compute_capacity * env->imbalance_pct >
1640 dst->load * src->compute_capacity * 100)
1646 static int task_numa_migrate(struct task_struct *p)
1648 struct task_numa_env env = {
1651 .src_cpu = task_cpu(p),
1652 .src_nid = task_node(p),
1654 .imbalance_pct = 112,
1660 struct sched_domain *sd;
1661 unsigned long taskweight, groupweight;
1663 long taskimp, groupimp;
1666 * Pick the lowest SD_NUMA domain, as that would have the smallest
1667 * imbalance and would be the first to start moving tasks about.
1669 * And we want to avoid any moving of tasks about, as that would create
1670 * random movement of tasks -- counter the numa conditions we're trying
1674 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1676 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1680 * Cpusets can break the scheduler domain tree into smaller
1681 * balance domains, some of which do not cross NUMA boundaries.
1682 * Tasks that are "trapped" in such domains cannot be migrated
1683 * elsewhere, so there is no point in (re)trying.
1685 if (unlikely(!sd)) {
1686 p->numa_preferred_nid = task_node(p);
1690 env.dst_nid = p->numa_preferred_nid;
1691 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1692 taskweight = task_weight(p, env.src_nid, dist);
1693 groupweight = group_weight(p, env.src_nid, dist);
1694 update_numa_stats(&env.src_stats, env.src_nid);
1695 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1696 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1697 update_numa_stats(&env.dst_stats, env.dst_nid);
1699 /* Try to find a spot on the preferred nid. */
1700 if (numa_has_capacity(&env))
1701 task_numa_find_cpu(&env, taskimp, groupimp);
1704 * Look at other nodes in these cases:
1705 * - there is no space available on the preferred_nid
1706 * - the task is part of a numa_group that is interleaved across
1707 * multiple NUMA nodes; in order to better consolidate the group,
1708 * we need to check other locations.
1710 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1711 for_each_online_node(nid) {
1712 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1715 dist = node_distance(env.src_nid, env.dst_nid);
1716 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1718 taskweight = task_weight(p, env.src_nid, dist);
1719 groupweight = group_weight(p, env.src_nid, dist);
1722 /* Only consider nodes where both task and groups benefit */
1723 taskimp = task_weight(p, nid, dist) - taskweight;
1724 groupimp = group_weight(p, nid, dist) - groupweight;
1725 if (taskimp < 0 && groupimp < 0)
1730 update_numa_stats(&env.dst_stats, env.dst_nid);
1731 if (numa_has_capacity(&env))
1732 task_numa_find_cpu(&env, taskimp, groupimp);
1737 * If the task is part of a workload that spans multiple NUMA nodes,
1738 * and is migrating into one of the workload's active nodes, remember
1739 * this node as the task's preferred numa node, so the workload can
1741 * A task that migrated to a second choice node will be better off
1742 * trying for a better one later. Do not set the preferred node here.
1744 if (p->numa_group) {
1745 struct numa_group *ng = p->numa_group;
1747 if (env.best_cpu == -1)
1752 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1753 sched_setnuma(p, env.dst_nid);
1756 /* No better CPU than the current one was found. */
1757 if (env.best_cpu == -1)
1761 * Reset the scan period if the task is being rescheduled on an
1762 * alternative node to recheck if the tasks is now properly placed.
1764 p->numa_scan_period = task_scan_min(p);
1766 if (env.best_task == NULL) {
1767 ret = migrate_task_to(p, env.best_cpu);
1769 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1773 ret = migrate_swap(p, env.best_task);
1775 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1776 put_task_struct(env.best_task);
1780 /* Attempt to migrate a task to a CPU on the preferred node. */
1781 static void numa_migrate_preferred(struct task_struct *p)
1783 unsigned long interval = HZ;
1785 /* This task has no NUMA fault statistics yet */
1786 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1789 /* Periodically retry migrating the task to the preferred node */
1790 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1791 p->numa_migrate_retry = jiffies + interval;
1793 /* Success if task is already running on preferred CPU */
1794 if (task_node(p) == p->numa_preferred_nid)
1797 /* Otherwise, try migrate to a CPU on the preferred node */
1798 task_numa_migrate(p);
1802 * Find out how many nodes on the workload is actively running on. Do this by
1803 * tracking the nodes from which NUMA hinting faults are triggered. This can
1804 * be different from the set of nodes where the workload's memory is currently
1807 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1809 unsigned long faults, max_faults = 0;
1810 int nid, active_nodes = 0;
1812 for_each_online_node(nid) {
1813 faults = group_faults_cpu(numa_group, nid);
1814 if (faults > max_faults)
1815 max_faults = faults;
1818 for_each_online_node(nid) {
1819 faults = group_faults_cpu(numa_group, nid);
1820 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1824 numa_group->max_faults_cpu = max_faults;
1825 numa_group->active_nodes = active_nodes;
1829 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1830 * increments. The more local the fault statistics are, the higher the scan
1831 * period will be for the next scan window. If local/(local+remote) ratio is
1832 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1833 * the scan period will decrease. Aim for 70% local accesses.
1835 #define NUMA_PERIOD_SLOTS 10
1836 #define NUMA_PERIOD_THRESHOLD 7
1839 * Increase the scan period (slow down scanning) if the majority of
1840 * our memory is already on our local node, or if the majority of
1841 * the page accesses are shared with other processes.
1842 * Otherwise, decrease the scan period.
1844 static void update_task_scan_period(struct task_struct *p,
1845 unsigned long shared, unsigned long private)
1847 unsigned int period_slot;
1851 unsigned long remote = p->numa_faults_locality[0];
1852 unsigned long local = p->numa_faults_locality[1];
1855 * If there were no record hinting faults then either the task is
1856 * completely idle or all activity is areas that are not of interest
1857 * to automatic numa balancing. Related to that, if there were failed
1858 * migration then it implies we are migrating too quickly or the local
1859 * node is overloaded. In either case, scan slower
1861 if (local + shared == 0 || p->numa_faults_locality[2]) {
1862 p->numa_scan_period = min(p->numa_scan_period_max,
1863 p->numa_scan_period << 1);
1865 p->mm->numa_next_scan = jiffies +
1866 msecs_to_jiffies(p->numa_scan_period);
1872 * Prepare to scale scan period relative to the current period.
1873 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1874 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1875 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1877 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1878 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1879 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1880 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1883 diff = slot * period_slot;
1885 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1888 * Scale scan rate increases based on sharing. There is an
1889 * inverse relationship between the degree of sharing and
1890 * the adjustment made to the scanning period. Broadly
1891 * speaking the intent is that there is little point
1892 * scanning faster if shared accesses dominate as it may
1893 * simply bounce migrations uselessly
1895 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1896 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1899 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1900 task_scan_min(p), task_scan_max(p));
1901 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1905 * Get the fraction of time the task has been running since the last
1906 * NUMA placement cycle. The scheduler keeps similar statistics, but
1907 * decays those on a 32ms period, which is orders of magnitude off
1908 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1909 * stats only if the task is so new there are no NUMA statistics yet.
1911 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1913 u64 runtime, delta, now;
1914 /* Use the start of this time slice to avoid calculations. */
1915 now = p->se.exec_start;
1916 runtime = p->se.sum_exec_runtime;
1918 if (p->last_task_numa_placement) {
1919 delta = runtime - p->last_sum_exec_runtime;
1920 *period = now - p->last_task_numa_placement;
1922 delta = p->se.avg.load_sum / p->se.load.weight;
1923 *period = LOAD_AVG_MAX;
1926 p->last_sum_exec_runtime = runtime;
1927 p->last_task_numa_placement = now;
1933 * Determine the preferred nid for a task in a numa_group. This needs to
1934 * be done in a way that produces consistent results with group_weight,
1935 * otherwise workloads might not converge.
1937 static int preferred_group_nid(struct task_struct *p, int nid)
1942 /* Direct connections between all NUMA nodes. */
1943 if (sched_numa_topology_type == NUMA_DIRECT)
1947 * On a system with glueless mesh NUMA topology, group_weight
1948 * scores nodes according to the number of NUMA hinting faults on
1949 * both the node itself, and on nearby nodes.
1951 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1952 unsigned long score, max_score = 0;
1953 int node, max_node = nid;
1955 dist = sched_max_numa_distance;
1957 for_each_online_node(node) {
1958 score = group_weight(p, node, dist);
1959 if (score > max_score) {
1968 * Finding the preferred nid in a system with NUMA backplane
1969 * interconnect topology is more involved. The goal is to locate
1970 * tasks from numa_groups near each other in the system, and
1971 * untangle workloads from different sides of the system. This requires
1972 * searching down the hierarchy of node groups, recursively searching
1973 * inside the highest scoring group of nodes. The nodemask tricks
1974 * keep the complexity of the search down.
1976 nodes = node_online_map;
1977 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1978 unsigned long max_faults = 0;
1979 nodemask_t max_group = NODE_MASK_NONE;
1982 /* Are there nodes at this distance from each other? */
1983 if (!find_numa_distance(dist))
1986 for_each_node_mask(a, nodes) {
1987 unsigned long faults = 0;
1988 nodemask_t this_group;
1989 nodes_clear(this_group);
1991 /* Sum group's NUMA faults; includes a==b case. */
1992 for_each_node_mask(b, nodes) {
1993 if (node_distance(a, b) < dist) {
1994 faults += group_faults(p, b);
1995 node_set(b, this_group);
1996 node_clear(b, nodes);
2000 /* Remember the top group. */
2001 if (faults > max_faults) {
2002 max_faults = faults;
2003 max_group = this_group;
2005 * subtle: at the smallest distance there is
2006 * just one node left in each "group", the
2007 * winner is the preferred nid.
2012 /* Next round, evaluate the nodes within max_group. */
2020 static void task_numa_placement(struct task_struct *p)
2022 int seq, nid, max_nid = -1, max_group_nid = -1;
2023 unsigned long max_faults = 0, max_group_faults = 0;
2024 unsigned long fault_types[2] = { 0, 0 };
2025 unsigned long total_faults;
2026 u64 runtime, period;
2027 spinlock_t *group_lock = NULL;
2030 * The p->mm->numa_scan_seq field gets updated without
2031 * exclusive access. Use READ_ONCE() here to ensure
2032 * that the field is read in a single access:
2034 seq = READ_ONCE(p->mm->numa_scan_seq);
2035 if (p->numa_scan_seq == seq)
2037 p->numa_scan_seq = seq;
2038 p->numa_scan_period_max = task_scan_max(p);
2040 total_faults = p->numa_faults_locality[0] +
2041 p->numa_faults_locality[1];
2042 runtime = numa_get_avg_runtime(p, &period);
2044 /* If the task is part of a group prevent parallel updates to group stats */
2045 if (p->numa_group) {
2046 group_lock = &p->numa_group->lock;
2047 spin_lock_irq(group_lock);
2050 /* Find the node with the highest number of faults */
2051 for_each_online_node(nid) {
2052 /* Keep track of the offsets in numa_faults array */
2053 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2054 unsigned long faults = 0, group_faults = 0;
2057 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2058 long diff, f_diff, f_weight;
2060 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2061 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2062 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2063 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2065 /* Decay existing window, copy faults since last scan */
2066 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2067 fault_types[priv] += p->numa_faults[membuf_idx];
2068 p->numa_faults[membuf_idx] = 0;
2071 * Normalize the faults_from, so all tasks in a group
2072 * count according to CPU use, instead of by the raw
2073 * number of faults. Tasks with little runtime have
2074 * little over-all impact on throughput, and thus their
2075 * faults are less important.
2077 f_weight = div64_u64(runtime << 16, period + 1);
2078 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2080 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2081 p->numa_faults[cpubuf_idx] = 0;
2083 p->numa_faults[mem_idx] += diff;
2084 p->numa_faults[cpu_idx] += f_diff;
2085 faults += p->numa_faults[mem_idx];
2086 p->total_numa_faults += diff;
2087 if (p->numa_group) {
2089 * safe because we can only change our own group
2091 * mem_idx represents the offset for a given
2092 * nid and priv in a specific region because it
2093 * is at the beginning of the numa_faults array.
2095 p->numa_group->faults[mem_idx] += diff;
2096 p->numa_group->faults_cpu[mem_idx] += f_diff;
2097 p->numa_group->total_faults += diff;
2098 group_faults += p->numa_group->faults[mem_idx];
2102 if (faults > max_faults) {
2103 max_faults = faults;
2107 if (group_faults > max_group_faults) {
2108 max_group_faults = group_faults;
2109 max_group_nid = nid;
2113 update_task_scan_period(p, fault_types[0], fault_types[1]);
2115 if (p->numa_group) {
2116 numa_group_count_active_nodes(p->numa_group);
2117 spin_unlock_irq(group_lock);
2118 max_nid = preferred_group_nid(p, max_group_nid);
2122 /* Set the new preferred node */
2123 if (max_nid != p->numa_preferred_nid)
2124 sched_setnuma(p, max_nid);
2126 if (task_node(p) != p->numa_preferred_nid)
2127 numa_migrate_preferred(p);
2131 static inline int get_numa_group(struct numa_group *grp)
2133 return atomic_inc_not_zero(&grp->refcount);
2136 static inline void put_numa_group(struct numa_group *grp)
2138 if (atomic_dec_and_test(&grp->refcount))
2139 kfree_rcu(grp, rcu);
2142 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2145 struct numa_group *grp, *my_grp;
2146 struct task_struct *tsk;
2148 int cpu = cpupid_to_cpu(cpupid);
2151 if (unlikely(!p->numa_group)) {
2152 unsigned int size = sizeof(struct numa_group) +
2153 4*nr_node_ids*sizeof(unsigned long);
2155 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2159 atomic_set(&grp->refcount, 1);
2160 grp->active_nodes = 1;
2161 grp->max_faults_cpu = 0;
2162 spin_lock_init(&grp->lock);
2164 /* Second half of the array tracks nids where faults happen */
2165 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2168 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2169 grp->faults[i] = p->numa_faults[i];
2171 grp->total_faults = p->total_numa_faults;
2174 rcu_assign_pointer(p->numa_group, grp);
2178 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2180 if (!cpupid_match_pid(tsk, cpupid))
2183 grp = rcu_dereference(tsk->numa_group);
2187 my_grp = p->numa_group;
2192 * Only join the other group if its bigger; if we're the bigger group,
2193 * the other task will join us.
2195 if (my_grp->nr_tasks > grp->nr_tasks)
2199 * Tie-break on the grp address.
2201 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2204 /* Always join threads in the same process. */
2205 if (tsk->mm == current->mm)
2208 /* Simple filter to avoid false positives due to PID collisions */
2209 if (flags & TNF_SHARED)
2212 /* Update priv based on whether false sharing was detected */
2215 if (join && !get_numa_group(grp))
2223 BUG_ON(irqs_disabled());
2224 double_lock_irq(&my_grp->lock, &grp->lock);
2226 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2227 my_grp->faults[i] -= p->numa_faults[i];
2228 grp->faults[i] += p->numa_faults[i];
2230 my_grp->total_faults -= p->total_numa_faults;
2231 grp->total_faults += p->total_numa_faults;
2236 spin_unlock(&my_grp->lock);
2237 spin_unlock_irq(&grp->lock);
2239 rcu_assign_pointer(p->numa_group, grp);
2241 put_numa_group(my_grp);
2249 void task_numa_free(struct task_struct *p)
2251 struct numa_group *grp = p->numa_group;
2252 void *numa_faults = p->numa_faults;
2253 unsigned long flags;
2257 spin_lock_irqsave(&grp->lock, flags);
2258 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2259 grp->faults[i] -= p->numa_faults[i];
2260 grp->total_faults -= p->total_numa_faults;
2263 spin_unlock_irqrestore(&grp->lock, flags);
2264 RCU_INIT_POINTER(p->numa_group, NULL);
2265 put_numa_group(grp);
2268 p->numa_faults = NULL;
2273 * Got a PROT_NONE fault for a page on @node.
2275 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2277 struct task_struct *p = current;
2278 bool migrated = flags & TNF_MIGRATED;
2279 int cpu_node = task_node(current);
2280 int local = !!(flags & TNF_FAULT_LOCAL);
2281 struct numa_group *ng;
2284 if (!static_branch_likely(&sched_numa_balancing))
2287 /* for example, ksmd faulting in a user's mm */
2291 /* Allocate buffer to track faults on a per-node basis */
2292 if (unlikely(!p->numa_faults)) {
2293 int size = sizeof(*p->numa_faults) *
2294 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2296 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2297 if (!p->numa_faults)
2300 p->total_numa_faults = 0;
2301 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2305 * First accesses are treated as private, otherwise consider accesses
2306 * to be private if the accessing pid has not changed
2308 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2311 priv = cpupid_match_pid(p, last_cpupid);
2312 if (!priv && !(flags & TNF_NO_GROUP))
2313 task_numa_group(p, last_cpupid, flags, &priv);
2317 * If a workload spans multiple NUMA nodes, a shared fault that
2318 * occurs wholly within the set of nodes that the workload is
2319 * actively using should be counted as local. This allows the
2320 * scan rate to slow down when a workload has settled down.
2323 if (!priv && !local && ng && ng->active_nodes > 1 &&
2324 numa_is_active_node(cpu_node, ng) &&
2325 numa_is_active_node(mem_node, ng))
2328 task_numa_placement(p);
2331 * Retry task to preferred node migration periodically, in case it
2332 * case it previously failed, or the scheduler moved us.
2334 if (time_after(jiffies, p->numa_migrate_retry))
2335 numa_migrate_preferred(p);
2338 p->numa_pages_migrated += pages;
2339 if (flags & TNF_MIGRATE_FAIL)
2340 p->numa_faults_locality[2] += pages;
2342 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2343 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2344 p->numa_faults_locality[local] += pages;
2347 static void reset_ptenuma_scan(struct task_struct *p)
2350 * We only did a read acquisition of the mmap sem, so
2351 * p->mm->numa_scan_seq is written to without exclusive access
2352 * and the update is not guaranteed to be atomic. That's not
2353 * much of an issue though, since this is just used for
2354 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2355 * expensive, to avoid any form of compiler optimizations:
2357 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2358 p->mm->numa_scan_offset = 0;
2362 * The expensive part of numa migration is done from task_work context.
2363 * Triggered from task_tick_numa().
2365 void task_numa_work(struct callback_head *work)
2367 unsigned long migrate, next_scan, now = jiffies;
2368 struct task_struct *p = current;
2369 struct mm_struct *mm = p->mm;
2370 u64 runtime = p->se.sum_exec_runtime;
2371 struct vm_area_struct *vma;
2372 unsigned long start, end;
2373 unsigned long nr_pte_updates = 0;
2374 long pages, virtpages;
2376 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2378 work->next = work; /* protect against double add */
2380 * Who cares about NUMA placement when they're dying.
2382 * NOTE: make sure not to dereference p->mm before this check,
2383 * exit_task_work() happens _after_ exit_mm() so we could be called
2384 * without p->mm even though we still had it when we enqueued this
2387 if (p->flags & PF_EXITING)
2390 if (!mm->numa_next_scan) {
2391 mm->numa_next_scan = now +
2392 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2396 * Enforce maximal scan/migration frequency..
2398 migrate = mm->numa_next_scan;
2399 if (time_before(now, migrate))
2402 if (p->numa_scan_period == 0) {
2403 p->numa_scan_period_max = task_scan_max(p);
2404 p->numa_scan_period = task_scan_min(p);
2407 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2408 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2412 * Delay this task enough that another task of this mm will likely win
2413 * the next time around.
2415 p->node_stamp += 2 * TICK_NSEC;
2417 start = mm->numa_scan_offset;
2418 pages = sysctl_numa_balancing_scan_size;
2419 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2420 virtpages = pages * 8; /* Scan up to this much virtual space */
2425 down_read(&mm->mmap_sem);
2426 vma = find_vma(mm, start);
2428 reset_ptenuma_scan(p);
2432 for (; vma; vma = vma->vm_next) {
2433 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2434 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2439 * Shared library pages mapped by multiple processes are not
2440 * migrated as it is expected they are cache replicated. Avoid
2441 * hinting faults in read-only file-backed mappings or the vdso
2442 * as migrating the pages will be of marginal benefit.
2445 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2449 * Skip inaccessible VMAs to avoid any confusion between
2450 * PROT_NONE and NUMA hinting ptes
2452 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2456 start = max(start, vma->vm_start);
2457 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2458 end = min(end, vma->vm_end);
2459 nr_pte_updates = change_prot_numa(vma, start, end);
2462 * Try to scan sysctl_numa_balancing_size worth of
2463 * hpages that have at least one present PTE that
2464 * is not already pte-numa. If the VMA contains
2465 * areas that are unused or already full of prot_numa
2466 * PTEs, scan up to virtpages, to skip through those
2470 pages -= (end - start) >> PAGE_SHIFT;
2471 virtpages -= (end - start) >> PAGE_SHIFT;
2474 if (pages <= 0 || virtpages <= 0)
2478 } while (end != vma->vm_end);
2483 * It is possible to reach the end of the VMA list but the last few
2484 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2485 * would find the !migratable VMA on the next scan but not reset the
2486 * scanner to the start so check it now.
2489 mm->numa_scan_offset = start;
2491 reset_ptenuma_scan(p);
2492 up_read(&mm->mmap_sem);
2495 * Make sure tasks use at least 32x as much time to run other code
2496 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2497 * Usually update_task_scan_period slows down scanning enough; on an
2498 * overloaded system we need to limit overhead on a per task basis.
2500 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2501 u64 diff = p->se.sum_exec_runtime - runtime;
2502 p->node_stamp += 32 * diff;
2507 * Drive the periodic memory faults..
2509 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2511 struct callback_head *work = &curr->numa_work;
2515 * We don't care about NUMA placement if we don't have memory.
2517 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2521 * Using runtime rather than walltime has the dual advantage that
2522 * we (mostly) drive the selection from busy threads and that the
2523 * task needs to have done some actual work before we bother with
2526 now = curr->se.sum_exec_runtime;
2527 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2529 if (now > curr->node_stamp + period) {
2530 if (!curr->node_stamp)
2531 curr->numa_scan_period = task_scan_min(curr);
2532 curr->node_stamp += period;
2534 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2535 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2536 task_work_add(curr, work, true);
2541 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2545 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2549 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2552 #endif /* CONFIG_NUMA_BALANCING */
2555 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2557 update_load_add(&cfs_rq->load, se->load.weight);
2558 if (!parent_entity(se))
2559 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2561 if (entity_is_task(se)) {
2562 struct rq *rq = rq_of(cfs_rq);
2564 account_numa_enqueue(rq, task_of(se));
2565 list_add(&se->group_node, &rq->cfs_tasks);
2568 cfs_rq->nr_running++;
2572 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2574 update_load_sub(&cfs_rq->load, se->load.weight);
2575 if (!parent_entity(se))
2576 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2578 if (entity_is_task(se)) {
2579 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2580 list_del_init(&se->group_node);
2583 cfs_rq->nr_running--;
2586 #ifdef CONFIG_FAIR_GROUP_SCHED
2588 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2590 long tg_weight, load, shares;
2593 * This really should be: cfs_rq->avg.load_avg, but instead we use
2594 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2595 * the shares for small weight interactive tasks.
2597 load = scale_load_down(cfs_rq->load.weight);
2599 tg_weight = atomic_long_read(&tg->load_avg);
2601 /* Ensure tg_weight >= load */
2602 tg_weight -= cfs_rq->tg_load_avg_contrib;
2605 shares = (tg->shares * load);
2607 shares /= tg_weight;
2609 if (shares < MIN_SHARES)
2610 shares = MIN_SHARES;
2611 if (shares > tg->shares)
2612 shares = tg->shares;
2616 # else /* CONFIG_SMP */
2617 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2621 # endif /* CONFIG_SMP */
2623 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2624 unsigned long weight)
2627 /* commit outstanding execution time */
2628 if (cfs_rq->curr == se)
2629 update_curr(cfs_rq);
2630 account_entity_dequeue(cfs_rq, se);
2633 update_load_set(&se->load, weight);
2636 account_entity_enqueue(cfs_rq, se);
2639 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2641 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2643 struct task_group *tg;
2644 struct sched_entity *se;
2648 se = tg->se[cpu_of(rq_of(cfs_rq))];
2649 if (!se || throttled_hierarchy(cfs_rq))
2652 if (likely(se->load.weight == tg->shares))
2655 shares = calc_cfs_shares(cfs_rq, tg);
2657 reweight_entity(cfs_rq_of(se), se, shares);
2659 #else /* CONFIG_FAIR_GROUP_SCHED */
2660 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2663 #endif /* CONFIG_FAIR_GROUP_SCHED */
2666 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2667 static const u32 runnable_avg_yN_inv[] = {
2668 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2669 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2670 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2671 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2672 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2673 0x85aac367, 0x82cd8698,
2677 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2678 * over-estimates when re-combining.
2680 static const u32 runnable_avg_yN_sum[] = {
2681 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2682 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2683 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2687 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2688 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2691 static const u32 __accumulated_sum_N32[] = {
2692 0, 23371, 35056, 40899, 43820, 45281,
2693 46011, 46376, 46559, 46650, 46696, 46719,
2698 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2700 static __always_inline u64 decay_load(u64 val, u64 n)
2702 unsigned int local_n;
2706 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2709 /* after bounds checking we can collapse to 32-bit */
2713 * As y^PERIOD = 1/2, we can combine
2714 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2715 * With a look-up table which covers y^n (n<PERIOD)
2717 * To achieve constant time decay_load.
2719 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2720 val >>= local_n / LOAD_AVG_PERIOD;
2721 local_n %= LOAD_AVG_PERIOD;
2724 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2729 * For updates fully spanning n periods, the contribution to runnable
2730 * average will be: \Sum 1024*y^n
2732 * We can compute this reasonably efficiently by combining:
2733 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2735 static u32 __compute_runnable_contrib(u64 n)
2739 if (likely(n <= LOAD_AVG_PERIOD))
2740 return runnable_avg_yN_sum[n];
2741 else if (unlikely(n >= LOAD_AVG_MAX_N))
2742 return LOAD_AVG_MAX;
2744 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2745 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2746 n %= LOAD_AVG_PERIOD;
2747 contrib = decay_load(contrib, n);
2748 return contrib + runnable_avg_yN_sum[n];
2751 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2754 * We can represent the historical contribution to runnable average as the
2755 * coefficients of a geometric series. To do this we sub-divide our runnable
2756 * history into segments of approximately 1ms (1024us); label the segment that
2757 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2759 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2761 * (now) (~1ms ago) (~2ms ago)
2763 * Let u_i denote the fraction of p_i that the entity was runnable.
2765 * We then designate the fractions u_i as our co-efficients, yielding the
2766 * following representation of historical load:
2767 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2769 * We choose y based on the with of a reasonably scheduling period, fixing:
2772 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2773 * approximately half as much as the contribution to load within the last ms
2776 * When a period "rolls over" and we have new u_0`, multiplying the previous
2777 * sum again by y is sufficient to update:
2778 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2779 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2781 static __always_inline int
2782 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2783 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2785 u64 delta, scaled_delta, periods;
2787 unsigned int delta_w, scaled_delta_w, decayed = 0;
2788 unsigned long scale_freq, scale_cpu;
2790 delta = now - sa->last_update_time;
2792 * This should only happen when time goes backwards, which it
2793 * unfortunately does during sched clock init when we swap over to TSC.
2795 if ((s64)delta < 0) {
2796 sa->last_update_time = now;
2801 * Use 1024ns as the unit of measurement since it's a reasonable
2802 * approximation of 1us and fast to compute.
2807 sa->last_update_time = now;
2809 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2810 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2812 /* delta_w is the amount already accumulated against our next period */
2813 delta_w = sa->period_contrib;
2814 if (delta + delta_w >= 1024) {
2817 /* how much left for next period will start over, we don't know yet */
2818 sa->period_contrib = 0;
2821 * Now that we know we're crossing a period boundary, figure
2822 * out how much from delta we need to complete the current
2823 * period and accrue it.
2825 delta_w = 1024 - delta_w;
2826 scaled_delta_w = cap_scale(delta_w, scale_freq);
2828 sa->load_sum += weight * scaled_delta_w;
2830 cfs_rq->runnable_load_sum +=
2831 weight * scaled_delta_w;
2835 sa->util_sum += scaled_delta_w * scale_cpu;
2839 /* Figure out how many additional periods this update spans */
2840 periods = delta / 1024;
2843 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2845 cfs_rq->runnable_load_sum =
2846 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2848 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2850 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2851 contrib = __compute_runnable_contrib(periods);
2852 contrib = cap_scale(contrib, scale_freq);
2854 sa->load_sum += weight * contrib;
2856 cfs_rq->runnable_load_sum += weight * contrib;
2859 sa->util_sum += contrib * scale_cpu;
2862 /* Remainder of delta accrued against u_0` */
2863 scaled_delta = cap_scale(delta, scale_freq);
2865 sa->load_sum += weight * scaled_delta;
2867 cfs_rq->runnable_load_sum += weight * scaled_delta;
2870 sa->util_sum += scaled_delta * scale_cpu;
2872 sa->period_contrib += delta;
2875 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2877 cfs_rq->runnable_load_avg =
2878 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2880 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2886 #ifdef CONFIG_FAIR_GROUP_SCHED
2888 * update_tg_load_avg - update the tg's load avg
2889 * @cfs_rq: the cfs_rq whose avg changed
2890 * @force: update regardless of how small the difference
2892 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2893 * However, because tg->load_avg is a global value there are performance
2896 * In order to avoid having to look at the other cfs_rq's, we use a
2897 * differential update where we store the last value we propagated. This in
2898 * turn allows skipping updates if the differential is 'small'.
2900 * Updating tg's load_avg is necessary before update_cfs_share() (which is
2901 * done) and effective_load() (which is not done because it is too costly).
2903 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2905 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2908 * No need to update load_avg for root_task_group as it is not used.
2910 if (cfs_rq->tg == &root_task_group)
2913 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2914 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2915 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2920 * Called within set_task_rq() right before setting a task's cpu. The
2921 * caller only guarantees p->pi_lock is held; no other assumptions,
2922 * including the state of rq->lock, should be made.
2924 void set_task_rq_fair(struct sched_entity *se,
2925 struct cfs_rq *prev, struct cfs_rq *next)
2927 if (!sched_feat(ATTACH_AGE_LOAD))
2931 * We are supposed to update the task to "current" time, then its up to
2932 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2933 * getting what current time is, so simply throw away the out-of-date
2934 * time. This will result in the wakee task is less decayed, but giving
2935 * the wakee more load sounds not bad.
2937 if (se->avg.last_update_time && prev) {
2938 u64 p_last_update_time;
2939 u64 n_last_update_time;
2941 #ifndef CONFIG_64BIT
2942 u64 p_last_update_time_copy;
2943 u64 n_last_update_time_copy;
2946 p_last_update_time_copy = prev->load_last_update_time_copy;
2947 n_last_update_time_copy = next->load_last_update_time_copy;
2951 p_last_update_time = prev->avg.last_update_time;
2952 n_last_update_time = next->avg.last_update_time;
2954 } while (p_last_update_time != p_last_update_time_copy ||
2955 n_last_update_time != n_last_update_time_copy);
2957 p_last_update_time = prev->avg.last_update_time;
2958 n_last_update_time = next->avg.last_update_time;
2960 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2961 &se->avg, 0, 0, NULL);
2962 se->avg.last_update_time = n_last_update_time;
2965 #else /* CONFIG_FAIR_GROUP_SCHED */
2966 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2967 #endif /* CONFIG_FAIR_GROUP_SCHED */
2969 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2971 if (&this_rq()->cfs == cfs_rq) {
2973 * There are a few boundary cases this might miss but it should
2974 * get called often enough that that should (hopefully) not be
2975 * a real problem -- added to that it only calls on the local
2976 * CPU, so if we enqueue remotely we'll miss an update, but
2977 * the next tick/schedule should update.
2979 * It will not get called when we go idle, because the idle
2980 * thread is a different class (!fair), nor will the utilization
2981 * number include things like RT tasks.
2983 * As is, the util number is not freq-invariant (we'd have to
2984 * implement arch_scale_freq_capacity() for that).
2988 cpufreq_update_util(rq_of(cfs_rq), 0);
2993 * Unsigned subtract and clamp on underflow.
2995 * Explicitly do a load-store to ensure the intermediate value never hits
2996 * memory. This allows lockless observations without ever seeing the negative
2999 #define sub_positive(_ptr, _val) do { \
3000 typeof(_ptr) ptr = (_ptr); \
3001 typeof(*ptr) val = (_val); \
3002 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3006 WRITE_ONCE(*ptr, res); \
3010 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3011 * @now: current time, as per cfs_rq_clock_task()
3012 * @cfs_rq: cfs_rq to update
3013 * @update_freq: should we call cfs_rq_util_change() or will the call do so
3015 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3016 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3017 * post_init_entity_util_avg().
3019 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3021 * Returns true if the load decayed or we removed load.
3023 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3024 * call update_tg_load_avg() when this function returns true.
3027 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3029 struct sched_avg *sa = &cfs_rq->avg;
3030 int decayed, removed_load = 0, removed_util = 0;
3032 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3033 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3034 sub_positive(&sa->load_avg, r);
3035 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3039 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3040 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3041 sub_positive(&sa->util_avg, r);
3042 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3046 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3047 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3049 #ifndef CONFIG_64BIT
3051 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3054 if (update_freq && (decayed || removed_util))
3055 cfs_rq_util_change(cfs_rq);
3057 return decayed || removed_load;
3060 /* Update task and its cfs_rq load average */
3061 static inline void update_load_avg(struct sched_entity *se, int update_tg)
3063 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3064 u64 now = cfs_rq_clock_task(cfs_rq);
3065 struct rq *rq = rq_of(cfs_rq);
3066 int cpu = cpu_of(rq);
3069 * Track task load average for carrying it to new CPU after migrated, and
3070 * track group sched_entity load average for task_h_load calc in migration
3072 __update_load_avg(now, cpu, &se->avg,
3073 se->on_rq * scale_load_down(se->load.weight),
3074 cfs_rq->curr == se, NULL);
3076 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
3077 update_tg_load_avg(cfs_rq, 0);
3081 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3082 * @cfs_rq: cfs_rq to attach to
3083 * @se: sched_entity to attach
3085 * Must call update_cfs_rq_load_avg() before this, since we rely on
3086 * cfs_rq->avg.last_update_time being current.
3088 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3090 if (!sched_feat(ATTACH_AGE_LOAD))
3094 * If we got migrated (either between CPUs or between cgroups) we'll
3095 * have aged the average right before clearing @last_update_time.
3097 * Or we're fresh through post_init_entity_util_avg().
3099 if (se->avg.last_update_time) {
3100 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3101 &se->avg, 0, 0, NULL);
3104 * XXX: we could have just aged the entire load away if we've been
3105 * absent from the fair class for too long.
3110 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3111 cfs_rq->avg.load_avg += se->avg.load_avg;
3112 cfs_rq->avg.load_sum += se->avg.load_sum;
3113 cfs_rq->avg.util_avg += se->avg.util_avg;
3114 cfs_rq->avg.util_sum += se->avg.util_sum;
3116 cfs_rq_util_change(cfs_rq);
3120 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3121 * @cfs_rq: cfs_rq to detach from
3122 * @se: sched_entity to detach
3124 * Must call update_cfs_rq_load_avg() before this, since we rely on
3125 * cfs_rq->avg.last_update_time being current.
3127 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3129 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3130 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3131 cfs_rq->curr == se, NULL);
3133 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3134 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3135 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3136 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3138 cfs_rq_util_change(cfs_rq);
3141 /* Add the load generated by se into cfs_rq's load average */
3143 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3145 struct sched_avg *sa = &se->avg;
3146 u64 now = cfs_rq_clock_task(cfs_rq);
3147 int migrated, decayed;
3149 migrated = !sa->last_update_time;
3151 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3152 se->on_rq * scale_load_down(se->load.weight),
3153 cfs_rq->curr == se, NULL);
3156 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3158 cfs_rq->runnable_load_avg += sa->load_avg;
3159 cfs_rq->runnable_load_sum += sa->load_sum;
3162 attach_entity_load_avg(cfs_rq, se);
3164 if (decayed || migrated)
3165 update_tg_load_avg(cfs_rq, 0);
3168 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3170 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3172 update_load_avg(se, 1);
3174 cfs_rq->runnable_load_avg =
3175 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3176 cfs_rq->runnable_load_sum =
3177 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3180 #ifndef CONFIG_64BIT
3181 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3183 u64 last_update_time_copy;
3184 u64 last_update_time;
3187 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3189 last_update_time = cfs_rq->avg.last_update_time;
3190 } while (last_update_time != last_update_time_copy);
3192 return last_update_time;
3195 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3197 return cfs_rq->avg.last_update_time;
3202 * Task first catches up with cfs_rq, and then subtract
3203 * itself from the cfs_rq (task must be off the queue now).
3205 void remove_entity_load_avg(struct sched_entity *se)
3207 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3208 u64 last_update_time;
3211 * tasks cannot exit without having gone through wake_up_new_task() ->
3212 * post_init_entity_util_avg() which will have added things to the
3213 * cfs_rq, so we can remove unconditionally.
3215 * Similarly for groups, they will have passed through
3216 * post_init_entity_util_avg() before unregister_sched_fair_group()
3220 last_update_time = cfs_rq_last_update_time(cfs_rq);
3222 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3223 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3224 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3227 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3229 return cfs_rq->runnable_load_avg;
3232 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3234 return cfs_rq->avg.load_avg;
3237 static int idle_balance(struct rq *this_rq);
3239 #else /* CONFIG_SMP */
3242 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3247 static inline void update_load_avg(struct sched_entity *se, int not_used)
3249 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3253 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3255 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3256 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3259 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3261 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3263 static inline int idle_balance(struct rq *rq)
3268 #endif /* CONFIG_SMP */
3270 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3272 #ifdef CONFIG_SCHED_DEBUG
3273 s64 d = se->vruntime - cfs_rq->min_vruntime;
3278 if (d > 3*sysctl_sched_latency)
3279 schedstat_inc(cfs_rq->nr_spread_over);
3284 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3286 u64 vruntime = cfs_rq->min_vruntime;
3289 * The 'current' period is already promised to the current tasks,
3290 * however the extra weight of the new task will slow them down a
3291 * little, place the new task so that it fits in the slot that
3292 * stays open at the end.
3294 if (initial && sched_feat(START_DEBIT))
3295 vruntime += sched_vslice(cfs_rq, se);
3297 /* sleeps up to a single latency don't count. */
3299 unsigned long thresh = sysctl_sched_latency;
3302 * Halve their sleep time's effect, to allow
3303 * for a gentler effect of sleepers:
3305 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3311 /* ensure we never gain time by being placed backwards. */
3312 se->vruntime = max_vruntime(se->vruntime, vruntime);
3315 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3317 static inline void check_schedstat_required(void)
3319 #ifdef CONFIG_SCHEDSTATS
3320 if (schedstat_enabled())
3323 /* Force schedstat enabled if a dependent tracepoint is active */
3324 if (trace_sched_stat_wait_enabled() ||
3325 trace_sched_stat_sleep_enabled() ||
3326 trace_sched_stat_iowait_enabled() ||
3327 trace_sched_stat_blocked_enabled() ||
3328 trace_sched_stat_runtime_enabled()) {
3329 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3330 "stat_blocked and stat_runtime require the "
3331 "kernel parameter schedstats=enabled or "
3332 "kernel.sched_schedstats=1\n");
3343 * update_min_vruntime()
3344 * vruntime -= min_vruntime
3348 * update_min_vruntime()
3349 * vruntime += min_vruntime
3351 * this way the vruntime transition between RQs is done when both
3352 * min_vruntime are up-to-date.
3356 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3357 * vruntime -= min_vruntime
3361 * update_min_vruntime()
3362 * vruntime += min_vruntime
3364 * this way we don't have the most up-to-date min_vruntime on the originating
3365 * CPU and an up-to-date min_vruntime on the destination CPU.
3369 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3371 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3372 bool curr = cfs_rq->curr == se;
3375 * If we're the current task, we must renormalise before calling
3379 se->vruntime += cfs_rq->min_vruntime;
3381 update_curr(cfs_rq);
3384 * Otherwise, renormalise after, such that we're placed at the current
3385 * moment in time, instead of some random moment in the past. Being
3386 * placed in the past could significantly boost this task to the
3387 * fairness detriment of existing tasks.
3389 if (renorm && !curr)
3390 se->vruntime += cfs_rq->min_vruntime;
3392 enqueue_entity_load_avg(cfs_rq, se);
3393 account_entity_enqueue(cfs_rq, se);
3394 update_cfs_shares(cfs_rq);
3396 if (flags & ENQUEUE_WAKEUP)
3397 place_entity(cfs_rq, se, 0);
3399 check_schedstat_required();
3400 update_stats_enqueue(cfs_rq, se, flags);
3401 check_spread(cfs_rq, se);
3403 __enqueue_entity(cfs_rq, se);
3406 if (cfs_rq->nr_running == 1) {
3407 list_add_leaf_cfs_rq(cfs_rq);
3408 check_enqueue_throttle(cfs_rq);
3412 static void __clear_buddies_last(struct sched_entity *se)
3414 for_each_sched_entity(se) {
3415 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3416 if (cfs_rq->last != se)
3419 cfs_rq->last = NULL;
3423 static void __clear_buddies_next(struct sched_entity *se)
3425 for_each_sched_entity(se) {
3426 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3427 if (cfs_rq->next != se)
3430 cfs_rq->next = NULL;
3434 static void __clear_buddies_skip(struct sched_entity *se)
3436 for_each_sched_entity(se) {
3437 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3438 if (cfs_rq->skip != se)
3441 cfs_rq->skip = NULL;
3445 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3447 if (cfs_rq->last == se)
3448 __clear_buddies_last(se);
3450 if (cfs_rq->next == se)
3451 __clear_buddies_next(se);
3453 if (cfs_rq->skip == se)
3454 __clear_buddies_skip(se);
3457 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3460 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3463 * Update run-time statistics of the 'current'.
3465 update_curr(cfs_rq);
3466 dequeue_entity_load_avg(cfs_rq, se);
3468 update_stats_dequeue(cfs_rq, se, flags);
3470 clear_buddies(cfs_rq, se);
3472 if (se != cfs_rq->curr)
3473 __dequeue_entity(cfs_rq, se);
3475 account_entity_dequeue(cfs_rq, se);
3478 * Normalize after update_curr(); which will also have moved
3479 * min_vruntime if @se is the one holding it back. But before doing
3480 * update_min_vruntime() again, which will discount @se's position and
3481 * can move min_vruntime forward still more.
3483 if (!(flags & DEQUEUE_SLEEP))
3484 se->vruntime -= cfs_rq->min_vruntime;
3486 /* return excess runtime on last dequeue */
3487 return_cfs_rq_runtime(cfs_rq);
3489 update_cfs_shares(cfs_rq);
3492 * Now advance min_vruntime if @se was the entity holding it back,
3493 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3494 * put back on, and if we advance min_vruntime, we'll be placed back
3495 * further than we started -- ie. we'll be penalized.
3497 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3498 update_min_vruntime(cfs_rq);
3502 * Preempt the current task with a newly woken task if needed:
3505 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3507 unsigned long ideal_runtime, delta_exec;
3508 struct sched_entity *se;
3511 ideal_runtime = sched_slice(cfs_rq, curr);
3512 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3513 if (delta_exec > ideal_runtime) {
3514 resched_curr(rq_of(cfs_rq));
3516 * The current task ran long enough, ensure it doesn't get
3517 * re-elected due to buddy favours.
3519 clear_buddies(cfs_rq, curr);
3524 * Ensure that a task that missed wakeup preemption by a
3525 * narrow margin doesn't have to wait for a full slice.
3526 * This also mitigates buddy induced latencies under load.
3528 if (delta_exec < sysctl_sched_min_granularity)
3531 se = __pick_first_entity(cfs_rq);
3532 delta = curr->vruntime - se->vruntime;
3537 if (delta > ideal_runtime)
3538 resched_curr(rq_of(cfs_rq));
3542 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3544 /* 'current' is not kept within the tree. */
3547 * Any task has to be enqueued before it get to execute on
3548 * a CPU. So account for the time it spent waiting on the
3551 update_stats_wait_end(cfs_rq, se);
3552 __dequeue_entity(cfs_rq, se);
3553 update_load_avg(se, 1);
3556 update_stats_curr_start(cfs_rq, se);
3560 * Track our maximum slice length, if the CPU's load is at
3561 * least twice that of our own weight (i.e. dont track it
3562 * when there are only lesser-weight tasks around):
3564 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3565 schedstat_set(se->statistics.slice_max,
3566 max((u64)schedstat_val(se->statistics.slice_max),
3567 se->sum_exec_runtime - se->prev_sum_exec_runtime));
3570 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3574 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3577 * Pick the next process, keeping these things in mind, in this order:
3578 * 1) keep things fair between processes/task groups
3579 * 2) pick the "next" process, since someone really wants that to run
3580 * 3) pick the "last" process, for cache locality
3581 * 4) do not run the "skip" process, if something else is available
3583 static struct sched_entity *
3584 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3586 struct sched_entity *left = __pick_first_entity(cfs_rq);
3587 struct sched_entity *se;
3590 * If curr is set we have to see if its left of the leftmost entity
3591 * still in the tree, provided there was anything in the tree at all.
3593 if (!left || (curr && entity_before(curr, left)))
3596 se = left; /* ideally we run the leftmost entity */
3599 * Avoid running the skip buddy, if running something else can
3600 * be done without getting too unfair.
3602 if (cfs_rq->skip == se) {
3603 struct sched_entity *second;
3606 second = __pick_first_entity(cfs_rq);
3608 second = __pick_next_entity(se);
3609 if (!second || (curr && entity_before(curr, second)))
3613 if (second && wakeup_preempt_entity(second, left) < 1)
3618 * Prefer last buddy, try to return the CPU to a preempted task.
3620 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3624 * Someone really wants this to run. If it's not unfair, run it.
3626 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3629 clear_buddies(cfs_rq, se);
3634 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3636 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3639 * If still on the runqueue then deactivate_task()
3640 * was not called and update_curr() has to be done:
3643 update_curr(cfs_rq);
3645 /* throttle cfs_rqs exceeding runtime */
3646 check_cfs_rq_runtime(cfs_rq);
3648 check_spread(cfs_rq, prev);
3651 update_stats_wait_start(cfs_rq, prev);
3652 /* Put 'current' back into the tree. */
3653 __enqueue_entity(cfs_rq, prev);
3654 /* in !on_rq case, update occurred at dequeue */
3655 update_load_avg(prev, 0);
3657 cfs_rq->curr = NULL;
3661 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3664 * Update run-time statistics of the 'current'.
3666 update_curr(cfs_rq);
3669 * Ensure that runnable average is periodically updated.
3671 update_load_avg(curr, 1);
3672 update_cfs_shares(cfs_rq);
3674 #ifdef CONFIG_SCHED_HRTICK
3676 * queued ticks are scheduled to match the slice, so don't bother
3677 * validating it and just reschedule.
3680 resched_curr(rq_of(cfs_rq));
3684 * don't let the period tick interfere with the hrtick preemption
3686 if (!sched_feat(DOUBLE_TICK) &&
3687 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3691 if (cfs_rq->nr_running > 1)
3692 check_preempt_tick(cfs_rq, curr);
3696 /**************************************************
3697 * CFS bandwidth control machinery
3700 #ifdef CONFIG_CFS_BANDWIDTH
3702 #ifdef HAVE_JUMP_LABEL
3703 static struct static_key __cfs_bandwidth_used;
3705 static inline bool cfs_bandwidth_used(void)
3707 return static_key_false(&__cfs_bandwidth_used);
3710 void cfs_bandwidth_usage_inc(void)
3712 static_key_slow_inc(&__cfs_bandwidth_used);
3715 void cfs_bandwidth_usage_dec(void)
3717 static_key_slow_dec(&__cfs_bandwidth_used);
3719 #else /* HAVE_JUMP_LABEL */
3720 static bool cfs_bandwidth_used(void)
3725 void cfs_bandwidth_usage_inc(void) {}
3726 void cfs_bandwidth_usage_dec(void) {}
3727 #endif /* HAVE_JUMP_LABEL */
3730 * default period for cfs group bandwidth.
3731 * default: 0.1s, units: nanoseconds
3733 static inline u64 default_cfs_period(void)
3735 return 100000000ULL;
3738 static inline u64 sched_cfs_bandwidth_slice(void)
3740 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3744 * Replenish runtime according to assigned quota and update expiration time.
3745 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3746 * additional synchronization around rq->lock.
3748 * requires cfs_b->lock
3750 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3754 if (cfs_b->quota == RUNTIME_INF)
3757 now = sched_clock_cpu(smp_processor_id());
3758 cfs_b->runtime = cfs_b->quota;
3759 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3762 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3764 return &tg->cfs_bandwidth;
3767 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3768 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3770 if (unlikely(cfs_rq->throttle_count))
3771 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3773 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3776 /* returns 0 on failure to allocate runtime */
3777 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3779 struct task_group *tg = cfs_rq->tg;
3780 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3781 u64 amount = 0, min_amount, expires;
3783 /* note: this is a positive sum as runtime_remaining <= 0 */
3784 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3786 raw_spin_lock(&cfs_b->lock);
3787 if (cfs_b->quota == RUNTIME_INF)
3788 amount = min_amount;
3790 start_cfs_bandwidth(cfs_b);
3792 if (cfs_b->runtime > 0) {
3793 amount = min(cfs_b->runtime, min_amount);
3794 cfs_b->runtime -= amount;
3798 expires = cfs_b->runtime_expires;
3799 raw_spin_unlock(&cfs_b->lock);
3801 cfs_rq->runtime_remaining += amount;
3803 * we may have advanced our local expiration to account for allowed
3804 * spread between our sched_clock and the one on which runtime was
3807 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3808 cfs_rq->runtime_expires = expires;
3810 return cfs_rq->runtime_remaining > 0;
3814 * Note: This depends on the synchronization provided by sched_clock and the
3815 * fact that rq->clock snapshots this value.
3817 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3819 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3821 /* if the deadline is ahead of our clock, nothing to do */
3822 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3825 if (cfs_rq->runtime_remaining < 0)
3829 * If the local deadline has passed we have to consider the
3830 * possibility that our sched_clock is 'fast' and the global deadline
3831 * has not truly expired.
3833 * Fortunately we can check determine whether this the case by checking
3834 * whether the global deadline has advanced. It is valid to compare
3835 * cfs_b->runtime_expires without any locks since we only care about
3836 * exact equality, so a partial write will still work.
3839 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3840 /* extend local deadline, drift is bounded above by 2 ticks */
3841 cfs_rq->runtime_expires += TICK_NSEC;
3843 /* global deadline is ahead, expiration has passed */
3844 cfs_rq->runtime_remaining = 0;
3848 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3850 /* dock delta_exec before expiring quota (as it could span periods) */
3851 cfs_rq->runtime_remaining -= delta_exec;
3852 expire_cfs_rq_runtime(cfs_rq);
3854 if (likely(cfs_rq->runtime_remaining > 0))
3858 * if we're unable to extend our runtime we resched so that the active
3859 * hierarchy can be throttled
3861 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3862 resched_curr(rq_of(cfs_rq));
3865 static __always_inline
3866 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3868 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3871 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3874 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3876 return cfs_bandwidth_used() && cfs_rq->throttled;
3879 /* check whether cfs_rq, or any parent, is throttled */
3880 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3882 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3886 * Ensure that neither of the group entities corresponding to src_cpu or
3887 * dest_cpu are members of a throttled hierarchy when performing group
3888 * load-balance operations.
3890 static inline int throttled_lb_pair(struct task_group *tg,
3891 int src_cpu, int dest_cpu)
3893 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3895 src_cfs_rq = tg->cfs_rq[src_cpu];
3896 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3898 return throttled_hierarchy(src_cfs_rq) ||
3899 throttled_hierarchy(dest_cfs_rq);
3902 /* updated child weight may affect parent so we have to do this bottom up */
3903 static int tg_unthrottle_up(struct task_group *tg, void *data)
3905 struct rq *rq = data;
3906 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3908 cfs_rq->throttle_count--;
3909 if (!cfs_rq->throttle_count) {
3910 /* adjust cfs_rq_clock_task() */
3911 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3912 cfs_rq->throttled_clock_task;
3918 static int tg_throttle_down(struct task_group *tg, void *data)
3920 struct rq *rq = data;
3921 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3923 /* group is entering throttled state, stop time */
3924 if (!cfs_rq->throttle_count)
3925 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3926 cfs_rq->throttle_count++;
3931 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3933 struct rq *rq = rq_of(cfs_rq);
3934 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3935 struct sched_entity *se;
3936 long task_delta, dequeue = 1;
3939 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3941 /* freeze hierarchy runnable averages while throttled */
3943 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3946 task_delta = cfs_rq->h_nr_running;
3947 for_each_sched_entity(se) {
3948 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3949 /* throttled entity or throttle-on-deactivate */
3954 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3955 qcfs_rq->h_nr_running -= task_delta;
3957 if (qcfs_rq->load.weight)
3962 sub_nr_running(rq, task_delta);
3964 cfs_rq->throttled = 1;
3965 cfs_rq->throttled_clock = rq_clock(rq);
3966 raw_spin_lock(&cfs_b->lock);
3967 empty = list_empty(&cfs_b->throttled_cfs_rq);
3970 * Add to the _head_ of the list, so that an already-started
3971 * distribute_cfs_runtime will not see us
3973 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3976 * If we're the first throttled task, make sure the bandwidth
3980 start_cfs_bandwidth(cfs_b);
3982 raw_spin_unlock(&cfs_b->lock);
3985 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3987 struct rq *rq = rq_of(cfs_rq);
3988 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3989 struct sched_entity *se;
3993 se = cfs_rq->tg->se[cpu_of(rq)];
3995 cfs_rq->throttled = 0;
3997 update_rq_clock(rq);
3999 raw_spin_lock(&cfs_b->lock);
4000 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4001 list_del_rcu(&cfs_rq->throttled_list);
4002 raw_spin_unlock(&cfs_b->lock);
4004 /* update hierarchical throttle state */
4005 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4007 if (!cfs_rq->load.weight)
4010 task_delta = cfs_rq->h_nr_running;
4011 for_each_sched_entity(se) {
4015 cfs_rq = cfs_rq_of(se);
4017 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4018 cfs_rq->h_nr_running += task_delta;
4020 if (cfs_rq_throttled(cfs_rq))
4025 add_nr_running(rq, task_delta);
4027 /* determine whether we need to wake up potentially idle cpu */
4028 if (rq->curr == rq->idle && rq->cfs.nr_running)
4032 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4033 u64 remaining, u64 expires)
4035 struct cfs_rq *cfs_rq;
4037 u64 starting_runtime = remaining;
4040 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4042 struct rq *rq = rq_of(cfs_rq);
4044 raw_spin_lock(&rq->lock);
4045 if (!cfs_rq_throttled(cfs_rq))
4048 runtime = -cfs_rq->runtime_remaining + 1;
4049 if (runtime > remaining)
4050 runtime = remaining;
4051 remaining -= runtime;
4053 cfs_rq->runtime_remaining += runtime;
4054 cfs_rq->runtime_expires = expires;
4056 /* we check whether we're throttled above */
4057 if (cfs_rq->runtime_remaining > 0)
4058 unthrottle_cfs_rq(cfs_rq);
4061 raw_spin_unlock(&rq->lock);
4068 return starting_runtime - remaining;
4072 * Responsible for refilling a task_group's bandwidth and unthrottling its
4073 * cfs_rqs as appropriate. If there has been no activity within the last
4074 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4075 * used to track this state.
4077 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4079 u64 runtime, runtime_expires;
4082 /* no need to continue the timer with no bandwidth constraint */
4083 if (cfs_b->quota == RUNTIME_INF)
4084 goto out_deactivate;
4086 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4087 cfs_b->nr_periods += overrun;
4090 * idle depends on !throttled (for the case of a large deficit), and if
4091 * we're going inactive then everything else can be deferred
4093 if (cfs_b->idle && !throttled)
4094 goto out_deactivate;
4096 __refill_cfs_bandwidth_runtime(cfs_b);
4099 /* mark as potentially idle for the upcoming period */
4104 /* account preceding periods in which throttling occurred */
4105 cfs_b->nr_throttled += overrun;
4107 runtime_expires = cfs_b->runtime_expires;
4110 * This check is repeated as we are holding onto the new bandwidth while
4111 * we unthrottle. This can potentially race with an unthrottled group
4112 * trying to acquire new bandwidth from the global pool. This can result
4113 * in us over-using our runtime if it is all used during this loop, but
4114 * only by limited amounts in that extreme case.
4116 while (throttled && cfs_b->runtime > 0) {
4117 runtime = cfs_b->runtime;
4118 raw_spin_unlock(&cfs_b->lock);
4119 /* we can't nest cfs_b->lock while distributing bandwidth */
4120 runtime = distribute_cfs_runtime(cfs_b, runtime,
4122 raw_spin_lock(&cfs_b->lock);
4124 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4126 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4130 * While we are ensured activity in the period following an
4131 * unthrottle, this also covers the case in which the new bandwidth is
4132 * insufficient to cover the existing bandwidth deficit. (Forcing the
4133 * timer to remain active while there are any throttled entities.)
4143 /* a cfs_rq won't donate quota below this amount */
4144 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4145 /* minimum remaining period time to redistribute slack quota */
4146 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4147 /* how long we wait to gather additional slack before distributing */
4148 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4151 * Are we near the end of the current quota period?
4153 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4154 * hrtimer base being cleared by hrtimer_start. In the case of
4155 * migrate_hrtimers, base is never cleared, so we are fine.
4157 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4159 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4162 /* if the call-back is running a quota refresh is already occurring */
4163 if (hrtimer_callback_running(refresh_timer))
4166 /* is a quota refresh about to occur? */
4167 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4168 if (remaining < min_expire)
4174 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4176 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4178 /* if there's a quota refresh soon don't bother with slack */
4179 if (runtime_refresh_within(cfs_b, min_left))
4182 hrtimer_start(&cfs_b->slack_timer,
4183 ns_to_ktime(cfs_bandwidth_slack_period),
4187 /* we know any runtime found here is valid as update_curr() precedes return */
4188 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4190 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4191 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4193 if (slack_runtime <= 0)
4196 raw_spin_lock(&cfs_b->lock);
4197 if (cfs_b->quota != RUNTIME_INF &&
4198 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4199 cfs_b->runtime += slack_runtime;
4201 /* we are under rq->lock, defer unthrottling using a timer */
4202 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4203 !list_empty(&cfs_b->throttled_cfs_rq))
4204 start_cfs_slack_bandwidth(cfs_b);
4206 raw_spin_unlock(&cfs_b->lock);
4208 /* even if it's not valid for return we don't want to try again */
4209 cfs_rq->runtime_remaining -= slack_runtime;
4212 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4214 if (!cfs_bandwidth_used())
4217 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4220 __return_cfs_rq_runtime(cfs_rq);
4224 * This is done with a timer (instead of inline with bandwidth return) since
4225 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4227 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4229 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4232 /* confirm we're still not at a refresh boundary */
4233 raw_spin_lock(&cfs_b->lock);
4234 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4235 raw_spin_unlock(&cfs_b->lock);
4239 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4240 runtime = cfs_b->runtime;
4242 expires = cfs_b->runtime_expires;
4243 raw_spin_unlock(&cfs_b->lock);
4248 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4250 raw_spin_lock(&cfs_b->lock);
4251 if (expires == cfs_b->runtime_expires)
4252 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4253 raw_spin_unlock(&cfs_b->lock);
4257 * When a group wakes up we want to make sure that its quota is not already
4258 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4259 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4261 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4263 if (!cfs_bandwidth_used())
4266 /* an active group must be handled by the update_curr()->put() path */
4267 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4270 /* ensure the group is not already throttled */
4271 if (cfs_rq_throttled(cfs_rq))
4274 /* update runtime allocation */
4275 account_cfs_rq_runtime(cfs_rq, 0);
4276 if (cfs_rq->runtime_remaining <= 0)
4277 throttle_cfs_rq(cfs_rq);
4280 static void sync_throttle(struct task_group *tg, int cpu)
4282 struct cfs_rq *pcfs_rq, *cfs_rq;
4284 if (!cfs_bandwidth_used())
4290 cfs_rq = tg->cfs_rq[cpu];
4291 pcfs_rq = tg->parent->cfs_rq[cpu];
4293 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4294 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4297 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4298 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4300 if (!cfs_bandwidth_used())
4303 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4307 * it's possible for a throttled entity to be forced into a running
4308 * state (e.g. set_curr_task), in this case we're finished.
4310 if (cfs_rq_throttled(cfs_rq))
4313 throttle_cfs_rq(cfs_rq);
4317 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4319 struct cfs_bandwidth *cfs_b =
4320 container_of(timer, struct cfs_bandwidth, slack_timer);
4322 do_sched_cfs_slack_timer(cfs_b);
4324 return HRTIMER_NORESTART;
4327 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4329 struct cfs_bandwidth *cfs_b =
4330 container_of(timer, struct cfs_bandwidth, period_timer);
4334 raw_spin_lock(&cfs_b->lock);
4336 overrun = hrtimer_forward_now(timer, cfs_b->period);
4340 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4343 cfs_b->period_active = 0;
4344 raw_spin_unlock(&cfs_b->lock);
4346 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4349 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4351 raw_spin_lock_init(&cfs_b->lock);
4353 cfs_b->quota = RUNTIME_INF;
4354 cfs_b->period = ns_to_ktime(default_cfs_period());
4356 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4357 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4358 cfs_b->period_timer.function = sched_cfs_period_timer;
4359 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4360 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4363 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4365 cfs_rq->runtime_enabled = 0;
4366 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4369 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4371 lockdep_assert_held(&cfs_b->lock);
4373 if (!cfs_b->period_active) {
4374 cfs_b->period_active = 1;
4375 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4376 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4380 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4382 /* init_cfs_bandwidth() was not called */
4383 if (!cfs_b->throttled_cfs_rq.next)
4386 hrtimer_cancel(&cfs_b->period_timer);
4387 hrtimer_cancel(&cfs_b->slack_timer);
4390 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4392 struct cfs_rq *cfs_rq;
4394 for_each_leaf_cfs_rq(rq, cfs_rq) {
4395 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4397 raw_spin_lock(&cfs_b->lock);
4398 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4399 raw_spin_unlock(&cfs_b->lock);
4403 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4405 struct cfs_rq *cfs_rq;
4407 for_each_leaf_cfs_rq(rq, cfs_rq) {
4408 if (!cfs_rq->runtime_enabled)
4412 * clock_task is not advancing so we just need to make sure
4413 * there's some valid quota amount
4415 cfs_rq->runtime_remaining = 1;
4417 * Offline rq is schedulable till cpu is completely disabled
4418 * in take_cpu_down(), so we prevent new cfs throttling here.
4420 cfs_rq->runtime_enabled = 0;
4422 if (cfs_rq_throttled(cfs_rq))
4423 unthrottle_cfs_rq(cfs_rq);
4427 #else /* CONFIG_CFS_BANDWIDTH */
4428 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4430 return rq_clock_task(rq_of(cfs_rq));
4433 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4434 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4435 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4436 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4437 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4439 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4444 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4449 static inline int throttled_lb_pair(struct task_group *tg,
4450 int src_cpu, int dest_cpu)
4455 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4457 #ifdef CONFIG_FAIR_GROUP_SCHED
4458 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4461 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4465 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4466 static inline void update_runtime_enabled(struct rq *rq) {}
4467 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4469 #endif /* CONFIG_CFS_BANDWIDTH */
4471 /**************************************************
4472 * CFS operations on tasks:
4475 #ifdef CONFIG_SCHED_HRTICK
4476 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4478 struct sched_entity *se = &p->se;
4479 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4481 SCHED_WARN_ON(task_rq(p) != rq);
4483 if (rq->cfs.h_nr_running > 1) {
4484 u64 slice = sched_slice(cfs_rq, se);
4485 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4486 s64 delta = slice - ran;
4493 hrtick_start(rq, delta);
4498 * called from enqueue/dequeue and updates the hrtick when the
4499 * current task is from our class and nr_running is low enough
4502 static void hrtick_update(struct rq *rq)
4504 struct task_struct *curr = rq->curr;
4506 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4509 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4510 hrtick_start_fair(rq, curr);
4512 #else /* !CONFIG_SCHED_HRTICK */
4514 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4518 static inline void hrtick_update(struct rq *rq)
4524 * The enqueue_task method is called before nr_running is
4525 * increased. Here we update the fair scheduling stats and
4526 * then put the task into the rbtree:
4529 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4531 struct cfs_rq *cfs_rq;
4532 struct sched_entity *se = &p->se;
4535 * If in_iowait is set, the code below may not trigger any cpufreq
4536 * utilization updates, so do it here explicitly with the IOWAIT flag
4540 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4542 for_each_sched_entity(se) {
4545 cfs_rq = cfs_rq_of(se);
4546 enqueue_entity(cfs_rq, se, flags);
4549 * end evaluation on encountering a throttled cfs_rq
4551 * note: in the case of encountering a throttled cfs_rq we will
4552 * post the final h_nr_running increment below.
4554 if (cfs_rq_throttled(cfs_rq))
4556 cfs_rq->h_nr_running++;
4558 flags = ENQUEUE_WAKEUP;
4561 for_each_sched_entity(se) {
4562 cfs_rq = cfs_rq_of(se);
4563 cfs_rq->h_nr_running++;
4565 if (cfs_rq_throttled(cfs_rq))
4568 update_load_avg(se, 1);
4569 update_cfs_shares(cfs_rq);
4573 add_nr_running(rq, 1);
4578 static void set_next_buddy(struct sched_entity *se);
4581 * The dequeue_task method is called before nr_running is
4582 * decreased. We remove the task from the rbtree and
4583 * update the fair scheduling stats:
4585 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4587 struct cfs_rq *cfs_rq;
4588 struct sched_entity *se = &p->se;
4589 int task_sleep = flags & DEQUEUE_SLEEP;
4591 for_each_sched_entity(se) {
4592 cfs_rq = cfs_rq_of(se);
4593 dequeue_entity(cfs_rq, se, flags);
4596 * end evaluation on encountering a throttled cfs_rq
4598 * note: in the case of encountering a throttled cfs_rq we will
4599 * post the final h_nr_running decrement below.
4601 if (cfs_rq_throttled(cfs_rq))
4603 cfs_rq->h_nr_running--;
4605 /* Don't dequeue parent if it has other entities besides us */
4606 if (cfs_rq->load.weight) {
4607 /* Avoid re-evaluating load for this entity: */
4608 se = parent_entity(se);
4610 * Bias pick_next to pick a task from this cfs_rq, as
4611 * p is sleeping when it is within its sched_slice.
4613 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4617 flags |= DEQUEUE_SLEEP;
4620 for_each_sched_entity(se) {
4621 cfs_rq = cfs_rq_of(se);
4622 cfs_rq->h_nr_running--;
4624 if (cfs_rq_throttled(cfs_rq))
4627 update_load_avg(se, 1);
4628 update_cfs_shares(cfs_rq);
4632 sub_nr_running(rq, 1);
4639 /* Working cpumask for: load_balance, load_balance_newidle. */
4640 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4641 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4643 #ifdef CONFIG_NO_HZ_COMMON
4645 * per rq 'load' arrray crap; XXX kill this.
4649 * The exact cpuload calculated at every tick would be:
4651 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4653 * If a cpu misses updates for n ticks (as it was idle) and update gets
4654 * called on the n+1-th tick when cpu may be busy, then we have:
4656 * load_n = (1 - 1/2^i)^n * load_0
4657 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4659 * decay_load_missed() below does efficient calculation of
4661 * load' = (1 - 1/2^i)^n * load
4663 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4664 * This allows us to precompute the above in said factors, thereby allowing the
4665 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4666 * fixed_power_int())
4668 * The calculation is approximated on a 128 point scale.
4670 #define DEGRADE_SHIFT 7
4672 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4673 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4674 { 0, 0, 0, 0, 0, 0, 0, 0 },
4675 { 64, 32, 8, 0, 0, 0, 0, 0 },
4676 { 96, 72, 40, 12, 1, 0, 0, 0 },
4677 { 112, 98, 75, 43, 15, 1, 0, 0 },
4678 { 120, 112, 98, 76, 45, 16, 2, 0 }
4682 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4683 * would be when CPU is idle and so we just decay the old load without
4684 * adding any new load.
4686 static unsigned long
4687 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4691 if (!missed_updates)
4694 if (missed_updates >= degrade_zero_ticks[idx])
4698 return load >> missed_updates;
4700 while (missed_updates) {
4701 if (missed_updates % 2)
4702 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4704 missed_updates >>= 1;
4709 #endif /* CONFIG_NO_HZ_COMMON */
4712 * __cpu_load_update - update the rq->cpu_load[] statistics
4713 * @this_rq: The rq to update statistics for
4714 * @this_load: The current load
4715 * @pending_updates: The number of missed updates
4717 * Update rq->cpu_load[] statistics. This function is usually called every
4718 * scheduler tick (TICK_NSEC).
4720 * This function computes a decaying average:
4722 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4724 * Because of NOHZ it might not get called on every tick which gives need for
4725 * the @pending_updates argument.
4727 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4728 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4729 * = A * (A * load[i]_n-2 + B) + B
4730 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4731 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4732 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4733 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4734 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4736 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4737 * any change in load would have resulted in the tick being turned back on.
4739 * For regular NOHZ, this reduces to:
4741 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4743 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4746 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4747 unsigned long pending_updates)
4749 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4752 this_rq->nr_load_updates++;
4754 /* Update our load: */
4755 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4756 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4757 unsigned long old_load, new_load;
4759 /* scale is effectively 1 << i now, and >> i divides by scale */
4761 old_load = this_rq->cpu_load[i];
4762 #ifdef CONFIG_NO_HZ_COMMON
4763 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4764 if (tickless_load) {
4765 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4767 * old_load can never be a negative value because a
4768 * decayed tickless_load cannot be greater than the
4769 * original tickless_load.
4771 old_load += tickless_load;
4774 new_load = this_load;
4776 * Round up the averaging division if load is increasing. This
4777 * prevents us from getting stuck on 9 if the load is 10, for
4780 if (new_load > old_load)
4781 new_load += scale - 1;
4783 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4786 sched_avg_update(this_rq);
4789 /* Used instead of source_load when we know the type == 0 */
4790 static unsigned long weighted_cpuload(const int cpu)
4792 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4795 #ifdef CONFIG_NO_HZ_COMMON
4797 * There is no sane way to deal with nohz on smp when using jiffies because the
4798 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4799 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4801 * Therefore we need to avoid the delta approach from the regular tick when
4802 * possible since that would seriously skew the load calculation. This is why we
4803 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4804 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4805 * loop exit, nohz_idle_balance, nohz full exit...)
4807 * This means we might still be one tick off for nohz periods.
4810 static void cpu_load_update_nohz(struct rq *this_rq,
4811 unsigned long curr_jiffies,
4814 unsigned long pending_updates;
4816 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4817 if (pending_updates) {
4818 this_rq->last_load_update_tick = curr_jiffies;
4820 * In the regular NOHZ case, we were idle, this means load 0.
4821 * In the NOHZ_FULL case, we were non-idle, we should consider
4822 * its weighted load.
4824 cpu_load_update(this_rq, load, pending_updates);
4829 * Called from nohz_idle_balance() to update the load ratings before doing the
4832 static void cpu_load_update_idle(struct rq *this_rq)
4835 * bail if there's load or we're actually up-to-date.
4837 if (weighted_cpuload(cpu_of(this_rq)))
4840 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4844 * Record CPU load on nohz entry so we know the tickless load to account
4845 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4846 * than other cpu_load[idx] but it should be fine as cpu_load readers
4847 * shouldn't rely into synchronized cpu_load[*] updates.
4849 void cpu_load_update_nohz_start(void)
4851 struct rq *this_rq = this_rq();
4854 * This is all lockless but should be fine. If weighted_cpuload changes
4855 * concurrently we'll exit nohz. And cpu_load write can race with
4856 * cpu_load_update_idle() but both updater would be writing the same.
4858 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4862 * Account the tickless load in the end of a nohz frame.
4864 void cpu_load_update_nohz_stop(void)
4866 unsigned long curr_jiffies = READ_ONCE(jiffies);
4867 struct rq *this_rq = this_rq();
4870 if (curr_jiffies == this_rq->last_load_update_tick)
4873 load = weighted_cpuload(cpu_of(this_rq));
4874 raw_spin_lock(&this_rq->lock);
4875 update_rq_clock(this_rq);
4876 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4877 raw_spin_unlock(&this_rq->lock);
4879 #else /* !CONFIG_NO_HZ_COMMON */
4880 static inline void cpu_load_update_nohz(struct rq *this_rq,
4881 unsigned long curr_jiffies,
4882 unsigned long load) { }
4883 #endif /* CONFIG_NO_HZ_COMMON */
4885 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4887 #ifdef CONFIG_NO_HZ_COMMON
4888 /* See the mess around cpu_load_update_nohz(). */
4889 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4891 cpu_load_update(this_rq, load, 1);
4895 * Called from scheduler_tick()
4897 void cpu_load_update_active(struct rq *this_rq)
4899 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4901 if (tick_nohz_tick_stopped())
4902 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4904 cpu_load_update_periodic(this_rq, load);
4908 * Return a low guess at the load of a migration-source cpu weighted
4909 * according to the scheduling class and "nice" value.
4911 * We want to under-estimate the load of migration sources, to
4912 * balance conservatively.
4914 static unsigned long source_load(int cpu, int type)
4916 struct rq *rq = cpu_rq(cpu);
4917 unsigned long total = weighted_cpuload(cpu);
4919 if (type == 0 || !sched_feat(LB_BIAS))
4922 return min(rq->cpu_load[type-1], total);
4926 * Return a high guess at the load of a migration-target cpu weighted
4927 * according to the scheduling class and "nice" value.
4929 static unsigned long target_load(int cpu, int type)
4931 struct rq *rq = cpu_rq(cpu);
4932 unsigned long total = weighted_cpuload(cpu);
4934 if (type == 0 || !sched_feat(LB_BIAS))
4937 return max(rq->cpu_load[type-1], total);
4940 static unsigned long capacity_of(int cpu)
4942 return cpu_rq(cpu)->cpu_capacity;
4945 static unsigned long capacity_orig_of(int cpu)
4947 return cpu_rq(cpu)->cpu_capacity_orig;
4950 static unsigned long cpu_avg_load_per_task(int cpu)
4952 struct rq *rq = cpu_rq(cpu);
4953 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4954 unsigned long load_avg = weighted_cpuload(cpu);
4957 return load_avg / nr_running;
4962 #ifdef CONFIG_FAIR_GROUP_SCHED
4964 * effective_load() calculates the load change as seen from the root_task_group
4966 * Adding load to a group doesn't make a group heavier, but can cause movement
4967 * of group shares between cpus. Assuming the shares were perfectly aligned one
4968 * can calculate the shift in shares.
4970 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4971 * on this @cpu and results in a total addition (subtraction) of @wg to the
4972 * total group weight.
4974 * Given a runqueue weight distribution (rw_i) we can compute a shares
4975 * distribution (s_i) using:
4977 * s_i = rw_i / \Sum rw_j (1)
4979 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4980 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4981 * shares distribution (s_i):
4983 * rw_i = { 2, 4, 1, 0 }
4984 * s_i = { 2/7, 4/7, 1/7, 0 }
4986 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4987 * task used to run on and the CPU the waker is running on), we need to
4988 * compute the effect of waking a task on either CPU and, in case of a sync
4989 * wakeup, compute the effect of the current task going to sleep.
4991 * So for a change of @wl to the local @cpu with an overall group weight change
4992 * of @wl we can compute the new shares distribution (s'_i) using:
4994 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4996 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4997 * differences in waking a task to CPU 0. The additional task changes the
4998 * weight and shares distributions like:
5000 * rw'_i = { 3, 4, 1, 0 }
5001 * s'_i = { 3/8, 4/8, 1/8, 0 }
5003 * We can then compute the difference in effective weight by using:
5005 * dw_i = S * (s'_i - s_i) (3)
5007 * Where 'S' is the group weight as seen by its parent.
5009 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5010 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5011 * 4/7) times the weight of the group.
5013 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5015 struct sched_entity *se = tg->se[cpu];
5017 if (!tg->parent) /* the trivial, non-cgroup case */
5020 for_each_sched_entity(se) {
5021 struct cfs_rq *cfs_rq = se->my_q;
5022 long W, w = cfs_rq_load_avg(cfs_rq);
5027 * W = @wg + \Sum rw_j
5029 W = wg + atomic_long_read(&tg->load_avg);
5031 /* Ensure \Sum rw_j >= rw_i */
5032 W -= cfs_rq->tg_load_avg_contrib;
5041 * wl = S * s'_i; see (2)
5044 wl = (w * (long)scale_load_down(tg->shares)) / W;
5046 wl = scale_load_down(tg->shares);
5049 * Per the above, wl is the new se->load.weight value; since
5050 * those are clipped to [MIN_SHARES, ...) do so now. See
5051 * calc_cfs_shares().
5053 if (wl < MIN_SHARES)
5057 * wl = dw_i = S * (s'_i - s_i); see (3)
5059 wl -= se->avg.load_avg;
5062 * Recursively apply this logic to all parent groups to compute
5063 * the final effective load change on the root group. Since
5064 * only the @tg group gets extra weight, all parent groups can
5065 * only redistribute existing shares. @wl is the shift in shares
5066 * resulting from this level per the above.
5075 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5082 static void record_wakee(struct task_struct *p)
5085 * Only decay a single time; tasks that have less then 1 wakeup per
5086 * jiffy will not have built up many flips.
5088 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5089 current->wakee_flips >>= 1;
5090 current->wakee_flip_decay_ts = jiffies;
5093 if (current->last_wakee != p) {
5094 current->last_wakee = p;
5095 current->wakee_flips++;
5100 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5102 * A waker of many should wake a different task than the one last awakened
5103 * at a frequency roughly N times higher than one of its wakees.
5105 * In order to determine whether we should let the load spread vs consolidating
5106 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5107 * partner, and a factor of lls_size higher frequency in the other.
5109 * With both conditions met, we can be relatively sure that the relationship is
5110 * non-monogamous, with partner count exceeding socket size.
5112 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5113 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5116 static int wake_wide(struct task_struct *p)
5118 unsigned int master = current->wakee_flips;
5119 unsigned int slave = p->wakee_flips;
5120 int factor = this_cpu_read(sd_llc_size);
5123 swap(master, slave);
5124 if (slave < factor || master < slave * factor)
5129 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5130 int prev_cpu, int sync)
5132 s64 this_load, load;
5133 s64 this_eff_load, prev_eff_load;
5135 struct task_group *tg;
5136 unsigned long weight;
5140 this_cpu = smp_processor_id();
5141 load = source_load(prev_cpu, idx);
5142 this_load = target_load(this_cpu, idx);
5145 * If sync wakeup then subtract the (maximum possible)
5146 * effect of the currently running task from the load
5147 * of the current CPU:
5150 tg = task_group(current);
5151 weight = current->se.avg.load_avg;
5153 this_load += effective_load(tg, this_cpu, -weight, -weight);
5154 load += effective_load(tg, prev_cpu, 0, -weight);
5158 weight = p->se.avg.load_avg;
5161 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5162 * due to the sync cause above having dropped this_load to 0, we'll
5163 * always have an imbalance, but there's really nothing you can do
5164 * about that, so that's good too.
5166 * Otherwise check if either cpus are near enough in load to allow this
5167 * task to be woken on this_cpu.
5169 this_eff_load = 100;
5170 this_eff_load *= capacity_of(prev_cpu);
5172 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5173 prev_eff_load *= capacity_of(this_cpu);
5175 if (this_load > 0) {
5176 this_eff_load *= this_load +
5177 effective_load(tg, this_cpu, weight, weight);
5179 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5182 balanced = this_eff_load <= prev_eff_load;
5184 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5189 schedstat_inc(sd->ttwu_move_affine);
5190 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5196 * find_idlest_group finds and returns the least busy CPU group within the
5199 static struct sched_group *
5200 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5201 int this_cpu, int sd_flag)
5203 struct sched_group *idlest = NULL, *group = sd->groups;
5204 unsigned long min_load = ULONG_MAX, this_load = 0;
5205 int load_idx = sd->forkexec_idx;
5206 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5208 if (sd_flag & SD_BALANCE_WAKE)
5209 load_idx = sd->wake_idx;
5212 unsigned long load, avg_load;
5216 /* Skip over this group if it has no CPUs allowed */
5217 if (!cpumask_intersects(sched_group_cpus(group),
5218 tsk_cpus_allowed(p)))
5221 local_group = cpumask_test_cpu(this_cpu,
5222 sched_group_cpus(group));
5224 /* Tally up the load of all CPUs in the group */
5227 for_each_cpu(i, sched_group_cpus(group)) {
5228 /* Bias balancing toward cpus of our domain */
5230 load = source_load(i, load_idx);
5232 load = target_load(i, load_idx);
5237 /* Adjust by relative CPU capacity of the group */
5238 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5241 this_load = avg_load;
5242 } else if (avg_load < min_load) {
5243 min_load = avg_load;
5246 } while (group = group->next, group != sd->groups);
5248 if (!idlest || 100*this_load < imbalance*min_load)
5254 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5257 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5259 unsigned long load, min_load = ULONG_MAX;
5260 unsigned int min_exit_latency = UINT_MAX;
5261 u64 latest_idle_timestamp = 0;
5262 int least_loaded_cpu = this_cpu;
5263 int shallowest_idle_cpu = -1;
5266 /* Check if we have any choice: */
5267 if (group->group_weight == 1)
5268 return cpumask_first(sched_group_cpus(group));
5270 /* Traverse only the allowed CPUs */
5271 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5273 struct rq *rq = cpu_rq(i);
5274 struct cpuidle_state *idle = idle_get_state(rq);
5275 if (idle && idle->exit_latency < min_exit_latency) {
5277 * We give priority to a CPU whose idle state
5278 * has the smallest exit latency irrespective
5279 * of any idle timestamp.
5281 min_exit_latency = idle->exit_latency;
5282 latest_idle_timestamp = rq->idle_stamp;
5283 shallowest_idle_cpu = i;
5284 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5285 rq->idle_stamp > latest_idle_timestamp) {
5287 * If equal or no active idle state, then
5288 * the most recently idled CPU might have
5291 latest_idle_timestamp = rq->idle_stamp;
5292 shallowest_idle_cpu = i;
5294 } else if (shallowest_idle_cpu == -1) {
5295 load = weighted_cpuload(i);
5296 if (load < min_load || (load == min_load && i == this_cpu)) {
5298 least_loaded_cpu = i;
5303 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5307 * Implement a for_each_cpu() variant that starts the scan at a given cpu
5308 * (@start), and wraps around.
5310 * This is used to scan for idle CPUs; such that not all CPUs looking for an
5311 * idle CPU find the same CPU. The down-side is that tasks tend to cycle
5312 * through the LLC domain.
5314 * Especially tbench is found sensitive to this.
5317 static int cpumask_next_wrap(int n, const struct cpumask *mask, int start, int *wrapped)
5322 next = find_next_bit(cpumask_bits(mask), nr_cpumask_bits, n+1);
5326 return nr_cpumask_bits;
5328 if (next >= nr_cpumask_bits) {
5338 #define for_each_cpu_wrap(cpu, mask, start, wrap) \
5339 for ((wrap) = 0, (cpu) = (start)-1; \
5340 (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)), \
5341 (cpu) < nr_cpumask_bits; )
5343 #ifdef CONFIG_SCHED_SMT
5345 static inline void set_idle_cores(int cpu, int val)
5347 struct sched_domain_shared *sds;
5349 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5351 WRITE_ONCE(sds->has_idle_cores, val);
5354 static inline bool test_idle_cores(int cpu, bool def)
5356 struct sched_domain_shared *sds;
5358 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5360 return READ_ONCE(sds->has_idle_cores);
5366 * Scans the local SMT mask to see if the entire core is idle, and records this
5367 * information in sd_llc_shared->has_idle_cores.
5369 * Since SMT siblings share all cache levels, inspecting this limited remote
5370 * state should be fairly cheap.
5372 void __update_idle_core(struct rq *rq)
5374 int core = cpu_of(rq);
5378 if (test_idle_cores(core, true))
5381 for_each_cpu(cpu, cpu_smt_mask(core)) {
5389 set_idle_cores(core, 1);
5395 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5396 * there are no idle cores left in the system; tracked through
5397 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5399 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5401 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5402 int core, cpu, wrap;
5404 if (!static_branch_likely(&sched_smt_present))
5407 if (!test_idle_cores(target, false))
5410 cpumask_and(cpus, sched_domain_span(sd), tsk_cpus_allowed(p));
5412 for_each_cpu_wrap(core, cpus, target, wrap) {
5415 for_each_cpu(cpu, cpu_smt_mask(core)) {
5416 cpumask_clear_cpu(cpu, cpus);
5426 * Failed to find an idle core; stop looking for one.
5428 set_idle_cores(target, 0);
5434 * Scan the local SMT mask for idle CPUs.
5436 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5440 if (!static_branch_likely(&sched_smt_present))
5443 for_each_cpu(cpu, cpu_smt_mask(target)) {
5444 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5453 #else /* CONFIG_SCHED_SMT */
5455 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5460 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5465 #endif /* CONFIG_SCHED_SMT */
5468 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5469 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5470 * average idle time for this rq (as found in rq->avg_idle).
5472 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5474 struct sched_domain *this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5475 u64 avg_idle = this_rq()->avg_idle;
5476 u64 avg_cost = this_sd->avg_scan_cost;
5482 * Due to large variance we need a large fuzz factor; hackbench in
5483 * particularly is sensitive here.
5485 if ((avg_idle / 512) < avg_cost)
5488 time = local_clock();
5490 for_each_cpu_wrap(cpu, sched_domain_span(sd), target, wrap) {
5491 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5497 time = local_clock() - time;
5498 cost = this_sd->avg_scan_cost;
5499 delta = (s64)(time - cost) / 8;
5500 this_sd->avg_scan_cost += delta;
5506 * Try and locate an idle core/thread in the LLC cache domain.
5508 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5510 struct sched_domain *sd;
5513 if (idle_cpu(target))
5517 * If the previous cpu is cache affine and idle, don't be stupid.
5519 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5522 sd = rcu_dereference(per_cpu(sd_llc, target));
5526 i = select_idle_core(p, sd, target);
5527 if ((unsigned)i < nr_cpumask_bits)
5530 i = select_idle_cpu(p, sd, target);
5531 if ((unsigned)i < nr_cpumask_bits)
5534 i = select_idle_smt(p, sd, target);
5535 if ((unsigned)i < nr_cpumask_bits)
5542 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5543 * tasks. The unit of the return value must be the one of capacity so we can
5544 * compare the utilization with the capacity of the CPU that is available for
5545 * CFS task (ie cpu_capacity).
5547 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5548 * recent utilization of currently non-runnable tasks on a CPU. It represents
5549 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5550 * capacity_orig is the cpu_capacity available at the highest frequency
5551 * (arch_scale_freq_capacity()).
5552 * The utilization of a CPU converges towards a sum equal to or less than the
5553 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5554 * the running time on this CPU scaled by capacity_curr.
5556 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5557 * higher than capacity_orig because of unfortunate rounding in
5558 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5559 * the average stabilizes with the new running time. We need to check that the
5560 * utilization stays within the range of [0..capacity_orig] and cap it if
5561 * necessary. Without utilization capping, a group could be seen as overloaded
5562 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5563 * available capacity. We allow utilization to overshoot capacity_curr (but not
5564 * capacity_orig) as it useful for predicting the capacity required after task
5565 * migrations (scheduler-driven DVFS).
5567 static int cpu_util(int cpu)
5569 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5570 unsigned long capacity = capacity_orig_of(cpu);
5572 return (util >= capacity) ? capacity : util;
5575 static inline int task_util(struct task_struct *p)
5577 return p->se.avg.util_avg;
5581 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5582 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5584 * In that case WAKE_AFFINE doesn't make sense and we'll let
5585 * BALANCE_WAKE sort things out.
5587 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5589 long min_cap, max_cap;
5591 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5592 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5594 /* Minimum capacity is close to max, no need to abort wake_affine */
5595 if (max_cap - min_cap < max_cap >> 3)
5598 return min_cap * 1024 < task_util(p) * capacity_margin;
5602 * select_task_rq_fair: Select target runqueue for the waking task in domains
5603 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5604 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5606 * Balances load by selecting the idlest cpu in the idlest group, or under
5607 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5609 * Returns the target cpu number.
5611 * preempt must be disabled.
5614 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5616 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5617 int cpu = smp_processor_id();
5618 int new_cpu = prev_cpu;
5619 int want_affine = 0;
5620 int sync = wake_flags & WF_SYNC;
5622 if (sd_flag & SD_BALANCE_WAKE) {
5624 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5625 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5629 for_each_domain(cpu, tmp) {
5630 if (!(tmp->flags & SD_LOAD_BALANCE))
5634 * If both cpu and prev_cpu are part of this domain,
5635 * cpu is a valid SD_WAKE_AFFINE target.
5637 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5638 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5643 if (tmp->flags & sd_flag)
5645 else if (!want_affine)
5650 sd = NULL; /* Prefer wake_affine over balance flags */
5651 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5656 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5657 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5660 struct sched_group *group;
5663 if (!(sd->flags & sd_flag)) {
5668 group = find_idlest_group(sd, p, cpu, sd_flag);
5674 new_cpu = find_idlest_cpu(group, p, cpu);
5675 if (new_cpu == -1 || new_cpu == cpu) {
5676 /* Now try balancing at a lower domain level of cpu */
5681 /* Now try balancing at a lower domain level of new_cpu */
5683 weight = sd->span_weight;
5685 for_each_domain(cpu, tmp) {
5686 if (weight <= tmp->span_weight)
5688 if (tmp->flags & sd_flag)
5691 /* while loop will break here if sd == NULL */
5699 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5700 * cfs_rq_of(p) references at time of call are still valid and identify the
5701 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5703 static void migrate_task_rq_fair(struct task_struct *p)
5706 * As blocked tasks retain absolute vruntime the migration needs to
5707 * deal with this by subtracting the old and adding the new
5708 * min_vruntime -- the latter is done by enqueue_entity() when placing
5709 * the task on the new runqueue.
5711 if (p->state == TASK_WAKING) {
5712 struct sched_entity *se = &p->se;
5713 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5716 #ifndef CONFIG_64BIT
5717 u64 min_vruntime_copy;
5720 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5722 min_vruntime = cfs_rq->min_vruntime;
5723 } while (min_vruntime != min_vruntime_copy);
5725 min_vruntime = cfs_rq->min_vruntime;
5728 se->vruntime -= min_vruntime;
5732 * We are supposed to update the task to "current" time, then its up to date
5733 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5734 * what current time is, so simply throw away the out-of-date time. This
5735 * will result in the wakee task is less decayed, but giving the wakee more
5736 * load sounds not bad.
5738 remove_entity_load_avg(&p->se);
5740 /* Tell new CPU we are migrated */
5741 p->se.avg.last_update_time = 0;
5743 /* We have migrated, no longer consider this task hot */
5744 p->se.exec_start = 0;
5747 static void task_dead_fair(struct task_struct *p)
5749 remove_entity_load_avg(&p->se);
5751 #endif /* CONFIG_SMP */
5753 static unsigned long
5754 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5756 unsigned long gran = sysctl_sched_wakeup_granularity;
5759 * Since its curr running now, convert the gran from real-time
5760 * to virtual-time in his units.
5762 * By using 'se' instead of 'curr' we penalize light tasks, so
5763 * they get preempted easier. That is, if 'se' < 'curr' then
5764 * the resulting gran will be larger, therefore penalizing the
5765 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5766 * be smaller, again penalizing the lighter task.
5768 * This is especially important for buddies when the leftmost
5769 * task is higher priority than the buddy.
5771 return calc_delta_fair(gran, se);
5775 * Should 'se' preempt 'curr'.
5789 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5791 s64 gran, vdiff = curr->vruntime - se->vruntime;
5796 gran = wakeup_gran(curr, se);
5803 static void set_last_buddy(struct sched_entity *se)
5805 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5808 for_each_sched_entity(se)
5809 cfs_rq_of(se)->last = se;
5812 static void set_next_buddy(struct sched_entity *se)
5814 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5817 for_each_sched_entity(se)
5818 cfs_rq_of(se)->next = se;
5821 static void set_skip_buddy(struct sched_entity *se)
5823 for_each_sched_entity(se)
5824 cfs_rq_of(se)->skip = se;
5828 * Preempt the current task with a newly woken task if needed:
5830 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5832 struct task_struct *curr = rq->curr;
5833 struct sched_entity *se = &curr->se, *pse = &p->se;
5834 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5835 int scale = cfs_rq->nr_running >= sched_nr_latency;
5836 int next_buddy_marked = 0;
5838 if (unlikely(se == pse))
5842 * This is possible from callers such as attach_tasks(), in which we
5843 * unconditionally check_prempt_curr() after an enqueue (which may have
5844 * lead to a throttle). This both saves work and prevents false
5845 * next-buddy nomination below.
5847 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5850 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5851 set_next_buddy(pse);
5852 next_buddy_marked = 1;
5856 * We can come here with TIF_NEED_RESCHED already set from new task
5859 * Note: this also catches the edge-case of curr being in a throttled
5860 * group (e.g. via set_curr_task), since update_curr() (in the
5861 * enqueue of curr) will have resulted in resched being set. This
5862 * prevents us from potentially nominating it as a false LAST_BUDDY
5865 if (test_tsk_need_resched(curr))
5868 /* Idle tasks are by definition preempted by non-idle tasks. */
5869 if (unlikely(curr->policy == SCHED_IDLE) &&
5870 likely(p->policy != SCHED_IDLE))
5874 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5875 * is driven by the tick):
5877 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5880 find_matching_se(&se, &pse);
5881 update_curr(cfs_rq_of(se));
5883 if (wakeup_preempt_entity(se, pse) == 1) {
5885 * Bias pick_next to pick the sched entity that is
5886 * triggering this preemption.
5888 if (!next_buddy_marked)
5889 set_next_buddy(pse);
5898 * Only set the backward buddy when the current task is still
5899 * on the rq. This can happen when a wakeup gets interleaved
5900 * with schedule on the ->pre_schedule() or idle_balance()
5901 * point, either of which can * drop the rq lock.
5903 * Also, during early boot the idle thread is in the fair class,
5904 * for obvious reasons its a bad idea to schedule back to it.
5906 if (unlikely(!se->on_rq || curr == rq->idle))
5909 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5913 static struct task_struct *
5914 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5916 struct cfs_rq *cfs_rq = &rq->cfs;
5917 struct sched_entity *se;
5918 struct task_struct *p;
5922 #ifdef CONFIG_FAIR_GROUP_SCHED
5923 if (!cfs_rq->nr_running)
5926 if (prev->sched_class != &fair_sched_class)
5930 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5931 * likely that a next task is from the same cgroup as the current.
5933 * Therefore attempt to avoid putting and setting the entire cgroup
5934 * hierarchy, only change the part that actually changes.
5938 struct sched_entity *curr = cfs_rq->curr;
5941 * Since we got here without doing put_prev_entity() we also
5942 * have to consider cfs_rq->curr. If it is still a runnable
5943 * entity, update_curr() will update its vruntime, otherwise
5944 * forget we've ever seen it.
5948 update_curr(cfs_rq);
5953 * This call to check_cfs_rq_runtime() will do the
5954 * throttle and dequeue its entity in the parent(s).
5955 * Therefore the 'simple' nr_running test will indeed
5958 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5962 se = pick_next_entity(cfs_rq, curr);
5963 cfs_rq = group_cfs_rq(se);
5969 * Since we haven't yet done put_prev_entity and if the selected task
5970 * is a different task than we started out with, try and touch the
5971 * least amount of cfs_rqs.
5974 struct sched_entity *pse = &prev->se;
5976 while (!(cfs_rq = is_same_group(se, pse))) {
5977 int se_depth = se->depth;
5978 int pse_depth = pse->depth;
5980 if (se_depth <= pse_depth) {
5981 put_prev_entity(cfs_rq_of(pse), pse);
5982 pse = parent_entity(pse);
5984 if (se_depth >= pse_depth) {
5985 set_next_entity(cfs_rq_of(se), se);
5986 se = parent_entity(se);
5990 put_prev_entity(cfs_rq, pse);
5991 set_next_entity(cfs_rq, se);
5994 if (hrtick_enabled(rq))
5995 hrtick_start_fair(rq, p);
6002 if (!cfs_rq->nr_running)
6005 put_prev_task(rq, prev);
6008 se = pick_next_entity(cfs_rq, NULL);
6009 set_next_entity(cfs_rq, se);
6010 cfs_rq = group_cfs_rq(se);
6015 if (hrtick_enabled(rq))
6016 hrtick_start_fair(rq, p);
6022 * This is OK, because current is on_cpu, which avoids it being picked
6023 * for load-balance and preemption/IRQs are still disabled avoiding
6024 * further scheduler activity on it and we're being very careful to
6025 * re-start the picking loop.
6027 lockdep_unpin_lock(&rq->lock, cookie);
6028 new_tasks = idle_balance(rq);
6029 lockdep_repin_lock(&rq->lock, cookie);
6031 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6032 * possible for any higher priority task to appear. In that case we
6033 * must re-start the pick_next_entity() loop.
6045 * Account for a descheduled task:
6047 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6049 struct sched_entity *se = &prev->se;
6050 struct cfs_rq *cfs_rq;
6052 for_each_sched_entity(se) {
6053 cfs_rq = cfs_rq_of(se);
6054 put_prev_entity(cfs_rq, se);
6059 * sched_yield() is very simple
6061 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6063 static void yield_task_fair(struct rq *rq)
6065 struct task_struct *curr = rq->curr;
6066 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6067 struct sched_entity *se = &curr->se;
6070 * Are we the only task in the tree?
6072 if (unlikely(rq->nr_running == 1))
6075 clear_buddies(cfs_rq, se);
6077 if (curr->policy != SCHED_BATCH) {
6078 update_rq_clock(rq);
6080 * Update run-time statistics of the 'current'.
6082 update_curr(cfs_rq);
6084 * Tell update_rq_clock() that we've just updated,
6085 * so we don't do microscopic update in schedule()
6086 * and double the fastpath cost.
6088 rq_clock_skip_update(rq, true);
6094 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6096 struct sched_entity *se = &p->se;
6098 /* throttled hierarchies are not runnable */
6099 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6102 /* Tell the scheduler that we'd really like pse to run next. */
6105 yield_task_fair(rq);
6111 /**************************************************
6112 * Fair scheduling class load-balancing methods.
6116 * The purpose of load-balancing is to achieve the same basic fairness the
6117 * per-cpu scheduler provides, namely provide a proportional amount of compute
6118 * time to each task. This is expressed in the following equation:
6120 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6122 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6123 * W_i,0 is defined as:
6125 * W_i,0 = \Sum_j w_i,j (2)
6127 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6128 * is derived from the nice value as per sched_prio_to_weight[].
6130 * The weight average is an exponential decay average of the instantaneous
6133 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6135 * C_i is the compute capacity of cpu i, typically it is the
6136 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6137 * can also include other factors [XXX].
6139 * To achieve this balance we define a measure of imbalance which follows
6140 * directly from (1):
6142 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6144 * We them move tasks around to minimize the imbalance. In the continuous
6145 * function space it is obvious this converges, in the discrete case we get
6146 * a few fun cases generally called infeasible weight scenarios.
6149 * - infeasible weights;
6150 * - local vs global optima in the discrete case. ]
6155 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6156 * for all i,j solution, we create a tree of cpus that follows the hardware
6157 * topology where each level pairs two lower groups (or better). This results
6158 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6159 * tree to only the first of the previous level and we decrease the frequency
6160 * of load-balance at each level inv. proportional to the number of cpus in
6166 * \Sum { --- * --- * 2^i } = O(n) (5)
6168 * `- size of each group
6169 * | | `- number of cpus doing load-balance
6171 * `- sum over all levels
6173 * Coupled with a limit on how many tasks we can migrate every balance pass,
6174 * this makes (5) the runtime complexity of the balancer.
6176 * An important property here is that each CPU is still (indirectly) connected
6177 * to every other cpu in at most O(log n) steps:
6179 * The adjacency matrix of the resulting graph is given by:
6182 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6185 * And you'll find that:
6187 * A^(log_2 n)_i,j != 0 for all i,j (7)
6189 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6190 * The task movement gives a factor of O(m), giving a convergence complexity
6193 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6198 * In order to avoid CPUs going idle while there's still work to do, new idle
6199 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6200 * tree itself instead of relying on other CPUs to bring it work.
6202 * This adds some complexity to both (5) and (8) but it reduces the total idle
6210 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6213 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6218 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6220 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6222 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6225 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6226 * rewrite all of this once again.]
6229 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6231 enum fbq_type { regular, remote, all };
6233 #define LBF_ALL_PINNED 0x01
6234 #define LBF_NEED_BREAK 0x02
6235 #define LBF_DST_PINNED 0x04
6236 #define LBF_SOME_PINNED 0x08
6239 struct sched_domain *sd;
6247 struct cpumask *dst_grpmask;
6249 enum cpu_idle_type idle;
6251 /* The set of CPUs under consideration for load-balancing */
6252 struct cpumask *cpus;
6257 unsigned int loop_break;
6258 unsigned int loop_max;
6260 enum fbq_type fbq_type;
6261 struct list_head tasks;
6265 * Is this task likely cache-hot:
6267 static int task_hot(struct task_struct *p, struct lb_env *env)
6271 lockdep_assert_held(&env->src_rq->lock);
6273 if (p->sched_class != &fair_sched_class)
6276 if (unlikely(p->policy == SCHED_IDLE))
6280 * Buddy candidates are cache hot:
6282 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6283 (&p->se == cfs_rq_of(&p->se)->next ||
6284 &p->se == cfs_rq_of(&p->se)->last))
6287 if (sysctl_sched_migration_cost == -1)
6289 if (sysctl_sched_migration_cost == 0)
6292 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6294 return delta < (s64)sysctl_sched_migration_cost;
6297 #ifdef CONFIG_NUMA_BALANCING
6299 * Returns 1, if task migration degrades locality
6300 * Returns 0, if task migration improves locality i.e migration preferred.
6301 * Returns -1, if task migration is not affected by locality.
6303 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6305 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6306 unsigned long src_faults, dst_faults;
6307 int src_nid, dst_nid;
6309 if (!static_branch_likely(&sched_numa_balancing))
6312 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6315 src_nid = cpu_to_node(env->src_cpu);
6316 dst_nid = cpu_to_node(env->dst_cpu);
6318 if (src_nid == dst_nid)
6321 /* Migrating away from the preferred node is always bad. */
6322 if (src_nid == p->numa_preferred_nid) {
6323 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6329 /* Encourage migration to the preferred node. */
6330 if (dst_nid == p->numa_preferred_nid)
6334 src_faults = group_faults(p, src_nid);
6335 dst_faults = group_faults(p, dst_nid);
6337 src_faults = task_faults(p, src_nid);
6338 dst_faults = task_faults(p, dst_nid);
6341 return dst_faults < src_faults;
6345 static inline int migrate_degrades_locality(struct task_struct *p,
6353 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6356 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6360 lockdep_assert_held(&env->src_rq->lock);
6363 * We do not migrate tasks that are:
6364 * 1) throttled_lb_pair, or
6365 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6366 * 3) running (obviously), or
6367 * 4) are cache-hot on their current CPU.
6369 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6372 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6375 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6377 env->flags |= LBF_SOME_PINNED;
6380 * Remember if this task can be migrated to any other cpu in
6381 * our sched_group. We may want to revisit it if we couldn't
6382 * meet load balance goals by pulling other tasks on src_cpu.
6384 * Also avoid computing new_dst_cpu if we have already computed
6385 * one in current iteration.
6387 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6390 /* Prevent to re-select dst_cpu via env's cpus */
6391 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6392 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6393 env->flags |= LBF_DST_PINNED;
6394 env->new_dst_cpu = cpu;
6402 /* Record that we found atleast one task that could run on dst_cpu */
6403 env->flags &= ~LBF_ALL_PINNED;
6405 if (task_running(env->src_rq, p)) {
6406 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6411 * Aggressive migration if:
6412 * 1) destination numa is preferred
6413 * 2) task is cache cold, or
6414 * 3) too many balance attempts have failed.
6416 tsk_cache_hot = migrate_degrades_locality(p, env);
6417 if (tsk_cache_hot == -1)
6418 tsk_cache_hot = task_hot(p, env);
6420 if (tsk_cache_hot <= 0 ||
6421 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6422 if (tsk_cache_hot == 1) {
6423 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6424 schedstat_inc(p->se.statistics.nr_forced_migrations);
6429 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6434 * detach_task() -- detach the task for the migration specified in env
6436 static void detach_task(struct task_struct *p, struct lb_env *env)
6438 lockdep_assert_held(&env->src_rq->lock);
6440 p->on_rq = TASK_ON_RQ_MIGRATING;
6441 deactivate_task(env->src_rq, p, 0);
6442 set_task_cpu(p, env->dst_cpu);
6446 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6447 * part of active balancing operations within "domain".
6449 * Returns a task if successful and NULL otherwise.
6451 static struct task_struct *detach_one_task(struct lb_env *env)
6453 struct task_struct *p, *n;
6455 lockdep_assert_held(&env->src_rq->lock);
6457 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6458 if (!can_migrate_task(p, env))
6461 detach_task(p, env);
6464 * Right now, this is only the second place where
6465 * lb_gained[env->idle] is updated (other is detach_tasks)
6466 * so we can safely collect stats here rather than
6467 * inside detach_tasks().
6469 schedstat_inc(env->sd->lb_gained[env->idle]);
6475 static const unsigned int sched_nr_migrate_break = 32;
6478 * detach_tasks() -- tries to detach up to imbalance weighted load from
6479 * busiest_rq, as part of a balancing operation within domain "sd".
6481 * Returns number of detached tasks if successful and 0 otherwise.
6483 static int detach_tasks(struct lb_env *env)
6485 struct list_head *tasks = &env->src_rq->cfs_tasks;
6486 struct task_struct *p;
6490 lockdep_assert_held(&env->src_rq->lock);
6492 if (env->imbalance <= 0)
6495 while (!list_empty(tasks)) {
6497 * We don't want to steal all, otherwise we may be treated likewise,
6498 * which could at worst lead to a livelock crash.
6500 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6503 p = list_first_entry(tasks, struct task_struct, se.group_node);
6506 /* We've more or less seen every task there is, call it quits */
6507 if (env->loop > env->loop_max)
6510 /* take a breather every nr_migrate tasks */
6511 if (env->loop > env->loop_break) {
6512 env->loop_break += sched_nr_migrate_break;
6513 env->flags |= LBF_NEED_BREAK;
6517 if (!can_migrate_task(p, env))
6520 load = task_h_load(p);
6522 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6525 if ((load / 2) > env->imbalance)
6528 detach_task(p, env);
6529 list_add(&p->se.group_node, &env->tasks);
6532 env->imbalance -= load;
6534 #ifdef CONFIG_PREEMPT
6536 * NEWIDLE balancing is a source of latency, so preemptible
6537 * kernels will stop after the first task is detached to minimize
6538 * the critical section.
6540 if (env->idle == CPU_NEWLY_IDLE)
6545 * We only want to steal up to the prescribed amount of
6548 if (env->imbalance <= 0)
6553 list_move_tail(&p->se.group_node, tasks);
6557 * Right now, this is one of only two places we collect this stat
6558 * so we can safely collect detach_one_task() stats here rather
6559 * than inside detach_one_task().
6561 schedstat_add(env->sd->lb_gained[env->idle], detached);
6567 * attach_task() -- attach the task detached by detach_task() to its new rq.
6569 static void attach_task(struct rq *rq, struct task_struct *p)
6571 lockdep_assert_held(&rq->lock);
6573 BUG_ON(task_rq(p) != rq);
6574 activate_task(rq, p, 0);
6575 p->on_rq = TASK_ON_RQ_QUEUED;
6576 check_preempt_curr(rq, p, 0);
6580 * attach_one_task() -- attaches the task returned from detach_one_task() to
6583 static void attach_one_task(struct rq *rq, struct task_struct *p)
6585 raw_spin_lock(&rq->lock);
6587 raw_spin_unlock(&rq->lock);
6591 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6594 static void attach_tasks(struct lb_env *env)
6596 struct list_head *tasks = &env->tasks;
6597 struct task_struct *p;
6599 raw_spin_lock(&env->dst_rq->lock);
6601 while (!list_empty(tasks)) {
6602 p = list_first_entry(tasks, struct task_struct, se.group_node);
6603 list_del_init(&p->se.group_node);
6605 attach_task(env->dst_rq, p);
6608 raw_spin_unlock(&env->dst_rq->lock);
6611 #ifdef CONFIG_FAIR_GROUP_SCHED
6612 static void update_blocked_averages(int cpu)
6614 struct rq *rq = cpu_rq(cpu);
6615 struct cfs_rq *cfs_rq;
6616 unsigned long flags;
6618 raw_spin_lock_irqsave(&rq->lock, flags);
6619 update_rq_clock(rq);
6622 * Iterates the task_group tree in a bottom up fashion, see
6623 * list_add_leaf_cfs_rq() for details.
6625 for_each_leaf_cfs_rq(rq, cfs_rq) {
6626 /* throttled entities do not contribute to load */
6627 if (throttled_hierarchy(cfs_rq))
6630 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6631 update_tg_load_avg(cfs_rq, 0);
6633 raw_spin_unlock_irqrestore(&rq->lock, flags);
6637 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6638 * This needs to be done in a top-down fashion because the load of a child
6639 * group is a fraction of its parents load.
6641 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6643 struct rq *rq = rq_of(cfs_rq);
6644 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6645 unsigned long now = jiffies;
6648 if (cfs_rq->last_h_load_update == now)
6651 cfs_rq->h_load_next = NULL;
6652 for_each_sched_entity(se) {
6653 cfs_rq = cfs_rq_of(se);
6654 cfs_rq->h_load_next = se;
6655 if (cfs_rq->last_h_load_update == now)
6660 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6661 cfs_rq->last_h_load_update = now;
6664 while ((se = cfs_rq->h_load_next) != NULL) {
6665 load = cfs_rq->h_load;
6666 load = div64_ul(load * se->avg.load_avg,
6667 cfs_rq_load_avg(cfs_rq) + 1);
6668 cfs_rq = group_cfs_rq(se);
6669 cfs_rq->h_load = load;
6670 cfs_rq->last_h_load_update = now;
6674 static unsigned long task_h_load(struct task_struct *p)
6676 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6678 update_cfs_rq_h_load(cfs_rq);
6679 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6680 cfs_rq_load_avg(cfs_rq) + 1);
6683 static inline void update_blocked_averages(int cpu)
6685 struct rq *rq = cpu_rq(cpu);
6686 struct cfs_rq *cfs_rq = &rq->cfs;
6687 unsigned long flags;
6689 raw_spin_lock_irqsave(&rq->lock, flags);
6690 update_rq_clock(rq);
6691 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6692 raw_spin_unlock_irqrestore(&rq->lock, flags);
6695 static unsigned long task_h_load(struct task_struct *p)
6697 return p->se.avg.load_avg;
6701 /********** Helpers for find_busiest_group ************************/
6710 * sg_lb_stats - stats of a sched_group required for load_balancing
6712 struct sg_lb_stats {
6713 unsigned long avg_load; /*Avg load across the CPUs of the group */
6714 unsigned long group_load; /* Total load over the CPUs of the group */
6715 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6716 unsigned long load_per_task;
6717 unsigned long group_capacity;
6718 unsigned long group_util; /* Total utilization of the group */
6719 unsigned int sum_nr_running; /* Nr tasks running in the group */
6720 unsigned int idle_cpus;
6721 unsigned int group_weight;
6722 enum group_type group_type;
6723 int group_no_capacity;
6724 #ifdef CONFIG_NUMA_BALANCING
6725 unsigned int nr_numa_running;
6726 unsigned int nr_preferred_running;
6731 * sd_lb_stats - Structure to store the statistics of a sched_domain
6732 * during load balancing.
6734 struct sd_lb_stats {
6735 struct sched_group *busiest; /* Busiest group in this sd */
6736 struct sched_group *local; /* Local group in this sd */
6737 unsigned long total_load; /* Total load of all groups in sd */
6738 unsigned long total_capacity; /* Total capacity of all groups in sd */
6739 unsigned long avg_load; /* Average load across all groups in sd */
6741 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6742 struct sg_lb_stats local_stat; /* Statistics of the local group */
6745 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6748 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6749 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6750 * We must however clear busiest_stat::avg_load because
6751 * update_sd_pick_busiest() reads this before assignment.
6753 *sds = (struct sd_lb_stats){
6757 .total_capacity = 0UL,
6760 .sum_nr_running = 0,
6761 .group_type = group_other,
6767 * get_sd_load_idx - Obtain the load index for a given sched domain.
6768 * @sd: The sched_domain whose load_idx is to be obtained.
6769 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6771 * Return: The load index.
6773 static inline int get_sd_load_idx(struct sched_domain *sd,
6774 enum cpu_idle_type idle)
6780 load_idx = sd->busy_idx;
6783 case CPU_NEWLY_IDLE:
6784 load_idx = sd->newidle_idx;
6787 load_idx = sd->idle_idx;
6794 static unsigned long scale_rt_capacity(int cpu)
6796 struct rq *rq = cpu_rq(cpu);
6797 u64 total, used, age_stamp, avg;
6801 * Since we're reading these variables without serialization make sure
6802 * we read them once before doing sanity checks on them.
6804 age_stamp = READ_ONCE(rq->age_stamp);
6805 avg = READ_ONCE(rq->rt_avg);
6806 delta = __rq_clock_broken(rq) - age_stamp;
6808 if (unlikely(delta < 0))
6811 total = sched_avg_period() + delta;
6813 used = div_u64(avg, total);
6815 if (likely(used < SCHED_CAPACITY_SCALE))
6816 return SCHED_CAPACITY_SCALE - used;
6821 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6823 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6824 struct sched_group *sdg = sd->groups;
6826 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6828 capacity *= scale_rt_capacity(cpu);
6829 capacity >>= SCHED_CAPACITY_SHIFT;
6834 cpu_rq(cpu)->cpu_capacity = capacity;
6835 sdg->sgc->capacity = capacity;
6838 void update_group_capacity(struct sched_domain *sd, int cpu)
6840 struct sched_domain *child = sd->child;
6841 struct sched_group *group, *sdg = sd->groups;
6842 unsigned long capacity;
6843 unsigned long interval;
6845 interval = msecs_to_jiffies(sd->balance_interval);
6846 interval = clamp(interval, 1UL, max_load_balance_interval);
6847 sdg->sgc->next_update = jiffies + interval;
6850 update_cpu_capacity(sd, cpu);
6856 if (child->flags & SD_OVERLAP) {
6858 * SD_OVERLAP domains cannot assume that child groups
6859 * span the current group.
6862 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6863 struct sched_group_capacity *sgc;
6864 struct rq *rq = cpu_rq(cpu);
6867 * build_sched_domains() -> init_sched_groups_capacity()
6868 * gets here before we've attached the domains to the
6871 * Use capacity_of(), which is set irrespective of domains
6872 * in update_cpu_capacity().
6874 * This avoids capacity from being 0 and
6875 * causing divide-by-zero issues on boot.
6877 if (unlikely(!rq->sd)) {
6878 capacity += capacity_of(cpu);
6882 sgc = rq->sd->groups->sgc;
6883 capacity += sgc->capacity;
6887 * !SD_OVERLAP domains can assume that child groups
6888 * span the current group.
6891 group = child->groups;
6893 capacity += group->sgc->capacity;
6894 group = group->next;
6895 } while (group != child->groups);
6898 sdg->sgc->capacity = capacity;
6902 * Check whether the capacity of the rq has been noticeably reduced by side
6903 * activity. The imbalance_pct is used for the threshold.
6904 * Return true is the capacity is reduced
6907 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6909 return ((rq->cpu_capacity * sd->imbalance_pct) <
6910 (rq->cpu_capacity_orig * 100));
6914 * Group imbalance indicates (and tries to solve) the problem where balancing
6915 * groups is inadequate due to tsk_cpus_allowed() constraints.
6917 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6918 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6921 * { 0 1 2 3 } { 4 5 6 7 }
6924 * If we were to balance group-wise we'd place two tasks in the first group and
6925 * two tasks in the second group. Clearly this is undesired as it will overload
6926 * cpu 3 and leave one of the cpus in the second group unused.
6928 * The current solution to this issue is detecting the skew in the first group
6929 * by noticing the lower domain failed to reach balance and had difficulty
6930 * moving tasks due to affinity constraints.
6932 * When this is so detected; this group becomes a candidate for busiest; see
6933 * update_sd_pick_busiest(). And calculate_imbalance() and
6934 * find_busiest_group() avoid some of the usual balance conditions to allow it
6935 * to create an effective group imbalance.
6937 * This is a somewhat tricky proposition since the next run might not find the
6938 * group imbalance and decide the groups need to be balanced again. A most
6939 * subtle and fragile situation.
6942 static inline int sg_imbalanced(struct sched_group *group)
6944 return group->sgc->imbalance;
6948 * group_has_capacity returns true if the group has spare capacity that could
6949 * be used by some tasks.
6950 * We consider that a group has spare capacity if the * number of task is
6951 * smaller than the number of CPUs or if the utilization is lower than the
6952 * available capacity for CFS tasks.
6953 * For the latter, we use a threshold to stabilize the state, to take into
6954 * account the variance of the tasks' load and to return true if the available
6955 * capacity in meaningful for the load balancer.
6956 * As an example, an available capacity of 1% can appear but it doesn't make
6957 * any benefit for the load balance.
6960 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6962 if (sgs->sum_nr_running < sgs->group_weight)
6965 if ((sgs->group_capacity * 100) >
6966 (sgs->group_util * env->sd->imbalance_pct))
6973 * group_is_overloaded returns true if the group has more tasks than it can
6975 * group_is_overloaded is not equals to !group_has_capacity because a group
6976 * with the exact right number of tasks, has no more spare capacity but is not
6977 * overloaded so both group_has_capacity and group_is_overloaded return
6981 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6983 if (sgs->sum_nr_running <= sgs->group_weight)
6986 if ((sgs->group_capacity * 100) <
6987 (sgs->group_util * env->sd->imbalance_pct))
6994 group_type group_classify(struct sched_group *group,
6995 struct sg_lb_stats *sgs)
6997 if (sgs->group_no_capacity)
6998 return group_overloaded;
7000 if (sg_imbalanced(group))
7001 return group_imbalanced;
7007 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7008 * @env: The load balancing environment.
7009 * @group: sched_group whose statistics are to be updated.
7010 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7011 * @local_group: Does group contain this_cpu.
7012 * @sgs: variable to hold the statistics for this group.
7013 * @overload: Indicate more than one runnable task for any CPU.
7015 static inline void update_sg_lb_stats(struct lb_env *env,
7016 struct sched_group *group, int load_idx,
7017 int local_group, struct sg_lb_stats *sgs,
7023 memset(sgs, 0, sizeof(*sgs));
7025 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7026 struct rq *rq = cpu_rq(i);
7028 /* Bias balancing toward cpus of our domain */
7030 load = target_load(i, load_idx);
7032 load = source_load(i, load_idx);
7034 sgs->group_load += load;
7035 sgs->group_util += cpu_util(i);
7036 sgs->sum_nr_running += rq->cfs.h_nr_running;
7038 nr_running = rq->nr_running;
7042 #ifdef CONFIG_NUMA_BALANCING
7043 sgs->nr_numa_running += rq->nr_numa_running;
7044 sgs->nr_preferred_running += rq->nr_preferred_running;
7046 sgs->sum_weighted_load += weighted_cpuload(i);
7048 * No need to call idle_cpu() if nr_running is not 0
7050 if (!nr_running && idle_cpu(i))
7054 /* Adjust by relative CPU capacity of the group */
7055 sgs->group_capacity = group->sgc->capacity;
7056 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7058 if (sgs->sum_nr_running)
7059 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7061 sgs->group_weight = group->group_weight;
7063 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7064 sgs->group_type = group_classify(group, sgs);
7068 * update_sd_pick_busiest - return 1 on busiest group
7069 * @env: The load balancing environment.
7070 * @sds: sched_domain statistics
7071 * @sg: sched_group candidate to be checked for being the busiest
7072 * @sgs: sched_group statistics
7074 * Determine if @sg is a busier group than the previously selected
7077 * Return: %true if @sg is a busier group than the previously selected
7078 * busiest group. %false otherwise.
7080 static bool update_sd_pick_busiest(struct lb_env *env,
7081 struct sd_lb_stats *sds,
7082 struct sched_group *sg,
7083 struct sg_lb_stats *sgs)
7085 struct sg_lb_stats *busiest = &sds->busiest_stat;
7087 if (sgs->group_type > busiest->group_type)
7090 if (sgs->group_type < busiest->group_type)
7093 if (sgs->avg_load <= busiest->avg_load)
7096 /* This is the busiest node in its class. */
7097 if (!(env->sd->flags & SD_ASYM_PACKING))
7100 /* No ASYM_PACKING if target cpu is already busy */
7101 if (env->idle == CPU_NOT_IDLE)
7104 * ASYM_PACKING needs to move all the work to the lowest
7105 * numbered CPUs in the group, therefore mark all groups
7106 * higher than ourself as busy.
7108 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7112 /* Prefer to move from highest possible cpu's work */
7113 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
7120 #ifdef CONFIG_NUMA_BALANCING
7121 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7123 if (sgs->sum_nr_running > sgs->nr_numa_running)
7125 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7130 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7132 if (rq->nr_running > rq->nr_numa_running)
7134 if (rq->nr_running > rq->nr_preferred_running)
7139 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7144 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7148 #endif /* CONFIG_NUMA_BALANCING */
7151 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7152 * @env: The load balancing environment.
7153 * @sds: variable to hold the statistics for this sched_domain.
7155 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7157 struct sched_domain *child = env->sd->child;
7158 struct sched_group *sg = env->sd->groups;
7159 struct sg_lb_stats tmp_sgs;
7160 int load_idx, prefer_sibling = 0;
7161 bool overload = false;
7163 if (child && child->flags & SD_PREFER_SIBLING)
7166 load_idx = get_sd_load_idx(env->sd, env->idle);
7169 struct sg_lb_stats *sgs = &tmp_sgs;
7172 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7175 sgs = &sds->local_stat;
7177 if (env->idle != CPU_NEWLY_IDLE ||
7178 time_after_eq(jiffies, sg->sgc->next_update))
7179 update_group_capacity(env->sd, env->dst_cpu);
7182 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7189 * In case the child domain prefers tasks go to siblings
7190 * first, lower the sg capacity so that we'll try
7191 * and move all the excess tasks away. We lower the capacity
7192 * of a group only if the local group has the capacity to fit
7193 * these excess tasks. The extra check prevents the case where
7194 * you always pull from the heaviest group when it is already
7195 * under-utilized (possible with a large weight task outweighs
7196 * the tasks on the system).
7198 if (prefer_sibling && sds->local &&
7199 group_has_capacity(env, &sds->local_stat) &&
7200 (sgs->sum_nr_running > 1)) {
7201 sgs->group_no_capacity = 1;
7202 sgs->group_type = group_classify(sg, sgs);
7205 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7207 sds->busiest_stat = *sgs;
7211 /* Now, start updating sd_lb_stats */
7212 sds->total_load += sgs->group_load;
7213 sds->total_capacity += sgs->group_capacity;
7216 } while (sg != env->sd->groups);
7218 if (env->sd->flags & SD_NUMA)
7219 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7221 if (!env->sd->parent) {
7222 /* update overload indicator if we are at root domain */
7223 if (env->dst_rq->rd->overload != overload)
7224 env->dst_rq->rd->overload = overload;
7230 * check_asym_packing - Check to see if the group is packed into the
7233 * This is primarily intended to used at the sibling level. Some
7234 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7235 * case of POWER7, it can move to lower SMT modes only when higher
7236 * threads are idle. When in lower SMT modes, the threads will
7237 * perform better since they share less core resources. Hence when we
7238 * have idle threads, we want them to be the higher ones.
7240 * This packing function is run on idle threads. It checks to see if
7241 * the busiest CPU in this domain (core in the P7 case) has a higher
7242 * CPU number than the packing function is being run on. Here we are
7243 * assuming lower CPU number will be equivalent to lower a SMT thread
7246 * Return: 1 when packing is required and a task should be moved to
7247 * this CPU. The amount of the imbalance is returned in *imbalance.
7249 * @env: The load balancing environment.
7250 * @sds: Statistics of the sched_domain which is to be packed
7252 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7256 if (!(env->sd->flags & SD_ASYM_PACKING))
7259 if (env->idle == CPU_NOT_IDLE)
7265 busiest_cpu = group_first_cpu(sds->busiest);
7266 if (env->dst_cpu > busiest_cpu)
7269 env->imbalance = DIV_ROUND_CLOSEST(
7270 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7271 SCHED_CAPACITY_SCALE);
7277 * fix_small_imbalance - Calculate the minor imbalance that exists
7278 * amongst the groups of a sched_domain, during
7280 * @env: The load balancing environment.
7281 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7284 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7286 unsigned long tmp, capa_now = 0, capa_move = 0;
7287 unsigned int imbn = 2;
7288 unsigned long scaled_busy_load_per_task;
7289 struct sg_lb_stats *local, *busiest;
7291 local = &sds->local_stat;
7292 busiest = &sds->busiest_stat;
7294 if (!local->sum_nr_running)
7295 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7296 else if (busiest->load_per_task > local->load_per_task)
7299 scaled_busy_load_per_task =
7300 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7301 busiest->group_capacity;
7303 if (busiest->avg_load + scaled_busy_load_per_task >=
7304 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7305 env->imbalance = busiest->load_per_task;
7310 * OK, we don't have enough imbalance to justify moving tasks,
7311 * however we may be able to increase total CPU capacity used by
7315 capa_now += busiest->group_capacity *
7316 min(busiest->load_per_task, busiest->avg_load);
7317 capa_now += local->group_capacity *
7318 min(local->load_per_task, local->avg_load);
7319 capa_now /= SCHED_CAPACITY_SCALE;
7321 /* Amount of load we'd subtract */
7322 if (busiest->avg_load > scaled_busy_load_per_task) {
7323 capa_move += busiest->group_capacity *
7324 min(busiest->load_per_task,
7325 busiest->avg_load - scaled_busy_load_per_task);
7328 /* Amount of load we'd add */
7329 if (busiest->avg_load * busiest->group_capacity <
7330 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7331 tmp = (busiest->avg_load * busiest->group_capacity) /
7332 local->group_capacity;
7334 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7335 local->group_capacity;
7337 capa_move += local->group_capacity *
7338 min(local->load_per_task, local->avg_load + tmp);
7339 capa_move /= SCHED_CAPACITY_SCALE;
7341 /* Move if we gain throughput */
7342 if (capa_move > capa_now)
7343 env->imbalance = busiest->load_per_task;
7347 * calculate_imbalance - Calculate the amount of imbalance present within the
7348 * groups of a given sched_domain during load balance.
7349 * @env: load balance environment
7350 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7352 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7354 unsigned long max_pull, load_above_capacity = ~0UL;
7355 struct sg_lb_stats *local, *busiest;
7357 local = &sds->local_stat;
7358 busiest = &sds->busiest_stat;
7360 if (busiest->group_type == group_imbalanced) {
7362 * In the group_imb case we cannot rely on group-wide averages
7363 * to ensure cpu-load equilibrium, look at wider averages. XXX
7365 busiest->load_per_task =
7366 min(busiest->load_per_task, sds->avg_load);
7370 * Avg load of busiest sg can be less and avg load of local sg can
7371 * be greater than avg load across all sgs of sd because avg load
7372 * factors in sg capacity and sgs with smaller group_type are
7373 * skipped when updating the busiest sg:
7375 if (busiest->avg_load <= sds->avg_load ||
7376 local->avg_load >= sds->avg_load) {
7378 return fix_small_imbalance(env, sds);
7382 * If there aren't any idle cpus, avoid creating some.
7384 if (busiest->group_type == group_overloaded &&
7385 local->group_type == group_overloaded) {
7386 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7387 if (load_above_capacity > busiest->group_capacity) {
7388 load_above_capacity -= busiest->group_capacity;
7389 load_above_capacity *= scale_load_down(NICE_0_LOAD);
7390 load_above_capacity /= busiest->group_capacity;
7392 load_above_capacity = ~0UL;
7396 * We're trying to get all the cpus to the average_load, so we don't
7397 * want to push ourselves above the average load, nor do we wish to
7398 * reduce the max loaded cpu below the average load. At the same time,
7399 * we also don't want to reduce the group load below the group
7400 * capacity. Thus we look for the minimum possible imbalance.
7402 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7404 /* How much load to actually move to equalise the imbalance */
7405 env->imbalance = min(
7406 max_pull * busiest->group_capacity,
7407 (sds->avg_load - local->avg_load) * local->group_capacity
7408 ) / SCHED_CAPACITY_SCALE;
7411 * if *imbalance is less than the average load per runnable task
7412 * there is no guarantee that any tasks will be moved so we'll have
7413 * a think about bumping its value to force at least one task to be
7416 if (env->imbalance < busiest->load_per_task)
7417 return fix_small_imbalance(env, sds);
7420 /******* find_busiest_group() helpers end here *********************/
7423 * find_busiest_group - Returns the busiest group within the sched_domain
7424 * if there is an imbalance.
7426 * Also calculates the amount of weighted load which should be moved
7427 * to restore balance.
7429 * @env: The load balancing environment.
7431 * Return: - The busiest group if imbalance exists.
7433 static struct sched_group *find_busiest_group(struct lb_env *env)
7435 struct sg_lb_stats *local, *busiest;
7436 struct sd_lb_stats sds;
7438 init_sd_lb_stats(&sds);
7441 * Compute the various statistics relavent for load balancing at
7444 update_sd_lb_stats(env, &sds);
7445 local = &sds.local_stat;
7446 busiest = &sds.busiest_stat;
7448 /* ASYM feature bypasses nice load balance check */
7449 if (check_asym_packing(env, &sds))
7452 /* There is no busy sibling group to pull tasks from */
7453 if (!sds.busiest || busiest->sum_nr_running == 0)
7456 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7457 / sds.total_capacity;
7460 * If the busiest group is imbalanced the below checks don't
7461 * work because they assume all things are equal, which typically
7462 * isn't true due to cpus_allowed constraints and the like.
7464 if (busiest->group_type == group_imbalanced)
7467 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7468 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7469 busiest->group_no_capacity)
7473 * If the local group is busier than the selected busiest group
7474 * don't try and pull any tasks.
7476 if (local->avg_load >= busiest->avg_load)
7480 * Don't pull any tasks if this group is already above the domain
7483 if (local->avg_load >= sds.avg_load)
7486 if (env->idle == CPU_IDLE) {
7488 * This cpu is idle. If the busiest group is not overloaded
7489 * and there is no imbalance between this and busiest group
7490 * wrt idle cpus, it is balanced. The imbalance becomes
7491 * significant if the diff is greater than 1 otherwise we
7492 * might end up to just move the imbalance on another group
7494 if ((busiest->group_type != group_overloaded) &&
7495 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7499 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7500 * imbalance_pct to be conservative.
7502 if (100 * busiest->avg_load <=
7503 env->sd->imbalance_pct * local->avg_load)
7508 /* Looks like there is an imbalance. Compute it */
7509 calculate_imbalance(env, &sds);
7518 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7520 static struct rq *find_busiest_queue(struct lb_env *env,
7521 struct sched_group *group)
7523 struct rq *busiest = NULL, *rq;
7524 unsigned long busiest_load = 0, busiest_capacity = 1;
7527 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7528 unsigned long capacity, wl;
7532 rt = fbq_classify_rq(rq);
7535 * We classify groups/runqueues into three groups:
7536 * - regular: there are !numa tasks
7537 * - remote: there are numa tasks that run on the 'wrong' node
7538 * - all: there is no distinction
7540 * In order to avoid migrating ideally placed numa tasks,
7541 * ignore those when there's better options.
7543 * If we ignore the actual busiest queue to migrate another
7544 * task, the next balance pass can still reduce the busiest
7545 * queue by moving tasks around inside the node.
7547 * If we cannot move enough load due to this classification
7548 * the next pass will adjust the group classification and
7549 * allow migration of more tasks.
7551 * Both cases only affect the total convergence complexity.
7553 if (rt > env->fbq_type)
7556 capacity = capacity_of(i);
7558 wl = weighted_cpuload(i);
7561 * When comparing with imbalance, use weighted_cpuload()
7562 * which is not scaled with the cpu capacity.
7565 if (rq->nr_running == 1 && wl > env->imbalance &&
7566 !check_cpu_capacity(rq, env->sd))
7570 * For the load comparisons with the other cpu's, consider
7571 * the weighted_cpuload() scaled with the cpu capacity, so
7572 * that the load can be moved away from the cpu that is
7573 * potentially running at a lower capacity.
7575 * Thus we're looking for max(wl_i / capacity_i), crosswise
7576 * multiplication to rid ourselves of the division works out
7577 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7578 * our previous maximum.
7580 if (wl * busiest_capacity > busiest_load * capacity) {
7582 busiest_capacity = capacity;
7591 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7592 * so long as it is large enough.
7594 #define MAX_PINNED_INTERVAL 512
7596 static int need_active_balance(struct lb_env *env)
7598 struct sched_domain *sd = env->sd;
7600 if (env->idle == CPU_NEWLY_IDLE) {
7603 * ASYM_PACKING needs to force migrate tasks from busy but
7604 * higher numbered CPUs in order to pack all tasks in the
7605 * lowest numbered CPUs.
7607 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7612 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7613 * It's worth migrating the task if the src_cpu's capacity is reduced
7614 * because of other sched_class or IRQs if more capacity stays
7615 * available on dst_cpu.
7617 if ((env->idle != CPU_NOT_IDLE) &&
7618 (env->src_rq->cfs.h_nr_running == 1)) {
7619 if ((check_cpu_capacity(env->src_rq, sd)) &&
7620 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7624 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7627 static int active_load_balance_cpu_stop(void *data);
7629 static int should_we_balance(struct lb_env *env)
7631 struct sched_group *sg = env->sd->groups;
7632 struct cpumask *sg_cpus, *sg_mask;
7633 int cpu, balance_cpu = -1;
7636 * In the newly idle case, we will allow all the cpu's
7637 * to do the newly idle load balance.
7639 if (env->idle == CPU_NEWLY_IDLE)
7642 sg_cpus = sched_group_cpus(sg);
7643 sg_mask = sched_group_mask(sg);
7644 /* Try to find first idle cpu */
7645 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7646 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7653 if (balance_cpu == -1)
7654 balance_cpu = group_balance_cpu(sg);
7657 * First idle cpu or the first cpu(busiest) in this sched group
7658 * is eligible for doing load balancing at this and above domains.
7660 return balance_cpu == env->dst_cpu;
7664 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7665 * tasks if there is an imbalance.
7667 static int load_balance(int this_cpu, struct rq *this_rq,
7668 struct sched_domain *sd, enum cpu_idle_type idle,
7669 int *continue_balancing)
7671 int ld_moved, cur_ld_moved, active_balance = 0;
7672 struct sched_domain *sd_parent = sd->parent;
7673 struct sched_group *group;
7675 unsigned long flags;
7676 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7678 struct lb_env env = {
7680 .dst_cpu = this_cpu,
7682 .dst_grpmask = sched_group_cpus(sd->groups),
7684 .loop_break = sched_nr_migrate_break,
7687 .tasks = LIST_HEAD_INIT(env.tasks),
7691 * For NEWLY_IDLE load_balancing, we don't need to consider
7692 * other cpus in our group
7694 if (idle == CPU_NEWLY_IDLE)
7695 env.dst_grpmask = NULL;
7697 cpumask_copy(cpus, cpu_active_mask);
7699 schedstat_inc(sd->lb_count[idle]);
7702 if (!should_we_balance(&env)) {
7703 *continue_balancing = 0;
7707 group = find_busiest_group(&env);
7709 schedstat_inc(sd->lb_nobusyg[idle]);
7713 busiest = find_busiest_queue(&env, group);
7715 schedstat_inc(sd->lb_nobusyq[idle]);
7719 BUG_ON(busiest == env.dst_rq);
7721 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
7723 env.src_cpu = busiest->cpu;
7724 env.src_rq = busiest;
7727 if (busiest->nr_running > 1) {
7729 * Attempt to move tasks. If find_busiest_group has found
7730 * an imbalance but busiest->nr_running <= 1, the group is
7731 * still unbalanced. ld_moved simply stays zero, so it is
7732 * correctly treated as an imbalance.
7734 env.flags |= LBF_ALL_PINNED;
7735 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7738 raw_spin_lock_irqsave(&busiest->lock, flags);
7741 * cur_ld_moved - load moved in current iteration
7742 * ld_moved - cumulative load moved across iterations
7744 cur_ld_moved = detach_tasks(&env);
7747 * We've detached some tasks from busiest_rq. Every
7748 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7749 * unlock busiest->lock, and we are able to be sure
7750 * that nobody can manipulate the tasks in parallel.
7751 * See task_rq_lock() family for the details.
7754 raw_spin_unlock(&busiest->lock);
7758 ld_moved += cur_ld_moved;
7761 local_irq_restore(flags);
7763 if (env.flags & LBF_NEED_BREAK) {
7764 env.flags &= ~LBF_NEED_BREAK;
7769 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7770 * us and move them to an alternate dst_cpu in our sched_group
7771 * where they can run. The upper limit on how many times we
7772 * iterate on same src_cpu is dependent on number of cpus in our
7775 * This changes load balance semantics a bit on who can move
7776 * load to a given_cpu. In addition to the given_cpu itself
7777 * (or a ilb_cpu acting on its behalf where given_cpu is
7778 * nohz-idle), we now have balance_cpu in a position to move
7779 * load to given_cpu. In rare situations, this may cause
7780 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7781 * _independently_ and at _same_ time to move some load to
7782 * given_cpu) causing exceess load to be moved to given_cpu.
7783 * This however should not happen so much in practice and
7784 * moreover subsequent load balance cycles should correct the
7785 * excess load moved.
7787 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7789 /* Prevent to re-select dst_cpu via env's cpus */
7790 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7792 env.dst_rq = cpu_rq(env.new_dst_cpu);
7793 env.dst_cpu = env.new_dst_cpu;
7794 env.flags &= ~LBF_DST_PINNED;
7796 env.loop_break = sched_nr_migrate_break;
7799 * Go back to "more_balance" rather than "redo" since we
7800 * need to continue with same src_cpu.
7806 * We failed to reach balance because of affinity.
7809 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7811 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7812 *group_imbalance = 1;
7815 /* All tasks on this runqueue were pinned by CPU affinity */
7816 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7817 cpumask_clear_cpu(cpu_of(busiest), cpus);
7818 if (!cpumask_empty(cpus)) {
7820 env.loop_break = sched_nr_migrate_break;
7823 goto out_all_pinned;
7828 schedstat_inc(sd->lb_failed[idle]);
7830 * Increment the failure counter only on periodic balance.
7831 * We do not want newidle balance, which can be very
7832 * frequent, pollute the failure counter causing
7833 * excessive cache_hot migrations and active balances.
7835 if (idle != CPU_NEWLY_IDLE)
7836 sd->nr_balance_failed++;
7838 if (need_active_balance(&env)) {
7839 raw_spin_lock_irqsave(&busiest->lock, flags);
7841 /* don't kick the active_load_balance_cpu_stop,
7842 * if the curr task on busiest cpu can't be
7845 if (!cpumask_test_cpu(this_cpu,
7846 tsk_cpus_allowed(busiest->curr))) {
7847 raw_spin_unlock_irqrestore(&busiest->lock,
7849 env.flags |= LBF_ALL_PINNED;
7850 goto out_one_pinned;
7854 * ->active_balance synchronizes accesses to
7855 * ->active_balance_work. Once set, it's cleared
7856 * only after active load balance is finished.
7858 if (!busiest->active_balance) {
7859 busiest->active_balance = 1;
7860 busiest->push_cpu = this_cpu;
7863 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7865 if (active_balance) {
7866 stop_one_cpu_nowait(cpu_of(busiest),
7867 active_load_balance_cpu_stop, busiest,
7868 &busiest->active_balance_work);
7871 /* We've kicked active balancing, force task migration. */
7872 sd->nr_balance_failed = sd->cache_nice_tries+1;
7875 sd->nr_balance_failed = 0;
7877 if (likely(!active_balance)) {
7878 /* We were unbalanced, so reset the balancing interval */
7879 sd->balance_interval = sd->min_interval;
7882 * If we've begun active balancing, start to back off. This
7883 * case may not be covered by the all_pinned logic if there
7884 * is only 1 task on the busy runqueue (because we don't call
7887 if (sd->balance_interval < sd->max_interval)
7888 sd->balance_interval *= 2;
7895 * We reach balance although we may have faced some affinity
7896 * constraints. Clear the imbalance flag if it was set.
7899 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7901 if (*group_imbalance)
7902 *group_imbalance = 0;
7907 * We reach balance because all tasks are pinned at this level so
7908 * we can't migrate them. Let the imbalance flag set so parent level
7909 * can try to migrate them.
7911 schedstat_inc(sd->lb_balanced[idle]);
7913 sd->nr_balance_failed = 0;
7916 /* tune up the balancing interval */
7917 if (((env.flags & LBF_ALL_PINNED) &&
7918 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7919 (sd->balance_interval < sd->max_interval))
7920 sd->balance_interval *= 2;
7927 static inline unsigned long
7928 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7930 unsigned long interval = sd->balance_interval;
7933 interval *= sd->busy_factor;
7935 /* scale ms to jiffies */
7936 interval = msecs_to_jiffies(interval);
7937 interval = clamp(interval, 1UL, max_load_balance_interval);
7943 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
7945 unsigned long interval, next;
7947 /* used by idle balance, so cpu_busy = 0 */
7948 interval = get_sd_balance_interval(sd, 0);
7949 next = sd->last_balance + interval;
7951 if (time_after(*next_balance, next))
7952 *next_balance = next;
7956 * idle_balance is called by schedule() if this_cpu is about to become
7957 * idle. Attempts to pull tasks from other CPUs.
7959 static int idle_balance(struct rq *this_rq)
7961 unsigned long next_balance = jiffies + HZ;
7962 int this_cpu = this_rq->cpu;
7963 struct sched_domain *sd;
7964 int pulled_task = 0;
7968 * We must set idle_stamp _before_ calling idle_balance(), such that we
7969 * measure the duration of idle_balance() as idle time.
7971 this_rq->idle_stamp = rq_clock(this_rq);
7973 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7974 !this_rq->rd->overload) {
7976 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7978 update_next_balance(sd, &next_balance);
7984 raw_spin_unlock(&this_rq->lock);
7986 update_blocked_averages(this_cpu);
7988 for_each_domain(this_cpu, sd) {
7989 int continue_balancing = 1;
7990 u64 t0, domain_cost;
7992 if (!(sd->flags & SD_LOAD_BALANCE))
7995 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7996 update_next_balance(sd, &next_balance);
8000 if (sd->flags & SD_BALANCE_NEWIDLE) {
8001 t0 = sched_clock_cpu(this_cpu);
8003 pulled_task = load_balance(this_cpu, this_rq,
8005 &continue_balancing);
8007 domain_cost = sched_clock_cpu(this_cpu) - t0;
8008 if (domain_cost > sd->max_newidle_lb_cost)
8009 sd->max_newidle_lb_cost = domain_cost;
8011 curr_cost += domain_cost;
8014 update_next_balance(sd, &next_balance);
8017 * Stop searching for tasks to pull if there are
8018 * now runnable tasks on this rq.
8020 if (pulled_task || this_rq->nr_running > 0)
8025 raw_spin_lock(&this_rq->lock);
8027 if (curr_cost > this_rq->max_idle_balance_cost)
8028 this_rq->max_idle_balance_cost = curr_cost;
8031 * While browsing the domains, we released the rq lock, a task could
8032 * have been enqueued in the meantime. Since we're not going idle,
8033 * pretend we pulled a task.
8035 if (this_rq->cfs.h_nr_running && !pulled_task)
8039 /* Move the next balance forward */
8040 if (time_after(this_rq->next_balance, next_balance))
8041 this_rq->next_balance = next_balance;
8043 /* Is there a task of a high priority class? */
8044 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8048 this_rq->idle_stamp = 0;
8054 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8055 * running tasks off the busiest CPU onto idle CPUs. It requires at
8056 * least 1 task to be running on each physical CPU where possible, and
8057 * avoids physical / logical imbalances.
8059 static int active_load_balance_cpu_stop(void *data)
8061 struct rq *busiest_rq = data;
8062 int busiest_cpu = cpu_of(busiest_rq);
8063 int target_cpu = busiest_rq->push_cpu;
8064 struct rq *target_rq = cpu_rq(target_cpu);
8065 struct sched_domain *sd;
8066 struct task_struct *p = NULL;
8068 raw_spin_lock_irq(&busiest_rq->lock);
8070 /* make sure the requested cpu hasn't gone down in the meantime */
8071 if (unlikely(busiest_cpu != smp_processor_id() ||
8072 !busiest_rq->active_balance))
8075 /* Is there any task to move? */
8076 if (busiest_rq->nr_running <= 1)
8080 * This condition is "impossible", if it occurs
8081 * we need to fix it. Originally reported by
8082 * Bjorn Helgaas on a 128-cpu setup.
8084 BUG_ON(busiest_rq == target_rq);
8086 /* Search for an sd spanning us and the target CPU. */
8088 for_each_domain(target_cpu, sd) {
8089 if ((sd->flags & SD_LOAD_BALANCE) &&
8090 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8095 struct lb_env env = {
8097 .dst_cpu = target_cpu,
8098 .dst_rq = target_rq,
8099 .src_cpu = busiest_rq->cpu,
8100 .src_rq = busiest_rq,
8104 schedstat_inc(sd->alb_count);
8106 p = detach_one_task(&env);
8108 schedstat_inc(sd->alb_pushed);
8109 /* Active balancing done, reset the failure counter. */
8110 sd->nr_balance_failed = 0;
8112 schedstat_inc(sd->alb_failed);
8117 busiest_rq->active_balance = 0;
8118 raw_spin_unlock(&busiest_rq->lock);
8121 attach_one_task(target_rq, p);
8128 static inline int on_null_domain(struct rq *rq)
8130 return unlikely(!rcu_dereference_sched(rq->sd));
8133 #ifdef CONFIG_NO_HZ_COMMON
8135 * idle load balancing details
8136 * - When one of the busy CPUs notice that there may be an idle rebalancing
8137 * needed, they will kick the idle load balancer, which then does idle
8138 * load balancing for all the idle CPUs.
8141 cpumask_var_t idle_cpus_mask;
8143 unsigned long next_balance; /* in jiffy units */
8144 } nohz ____cacheline_aligned;
8146 static inline int find_new_ilb(void)
8148 int ilb = cpumask_first(nohz.idle_cpus_mask);
8150 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8157 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8158 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8159 * CPU (if there is one).
8161 static void nohz_balancer_kick(void)
8165 nohz.next_balance++;
8167 ilb_cpu = find_new_ilb();
8169 if (ilb_cpu >= nr_cpu_ids)
8172 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8175 * Use smp_send_reschedule() instead of resched_cpu().
8176 * This way we generate a sched IPI on the target cpu which
8177 * is idle. And the softirq performing nohz idle load balance
8178 * will be run before returning from the IPI.
8180 smp_send_reschedule(ilb_cpu);
8184 void nohz_balance_exit_idle(unsigned int cpu)
8186 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8188 * Completely isolated CPUs don't ever set, so we must test.
8190 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8191 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8192 atomic_dec(&nohz.nr_cpus);
8194 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8198 static inline void set_cpu_sd_state_busy(void)
8200 struct sched_domain *sd;
8201 int cpu = smp_processor_id();
8204 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8206 if (!sd || !sd->nohz_idle)
8210 atomic_inc(&sd->shared->nr_busy_cpus);
8215 void set_cpu_sd_state_idle(void)
8217 struct sched_domain *sd;
8218 int cpu = smp_processor_id();
8221 sd = rcu_dereference(per_cpu(sd_llc, cpu));
8223 if (!sd || sd->nohz_idle)
8227 atomic_dec(&sd->shared->nr_busy_cpus);
8233 * This routine will record that the cpu is going idle with tick stopped.
8234 * This info will be used in performing idle load balancing in the future.
8236 void nohz_balance_enter_idle(int cpu)
8239 * If this cpu is going down, then nothing needs to be done.
8241 if (!cpu_active(cpu))
8244 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8248 * If we're a completely isolated CPU, we don't play.
8250 if (on_null_domain(cpu_rq(cpu)))
8253 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8254 atomic_inc(&nohz.nr_cpus);
8255 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8259 static DEFINE_SPINLOCK(balancing);
8262 * Scale the max load_balance interval with the number of CPUs in the system.
8263 * This trades load-balance latency on larger machines for less cross talk.
8265 void update_max_interval(void)
8267 max_load_balance_interval = HZ*num_online_cpus()/10;
8271 * It checks each scheduling domain to see if it is due to be balanced,
8272 * and initiates a balancing operation if so.
8274 * Balancing parameters are set up in init_sched_domains.
8276 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8278 int continue_balancing = 1;
8280 unsigned long interval;
8281 struct sched_domain *sd;
8282 /* Earliest time when we have to do rebalance again */
8283 unsigned long next_balance = jiffies + 60*HZ;
8284 int update_next_balance = 0;
8285 int need_serialize, need_decay = 0;
8288 update_blocked_averages(cpu);
8291 for_each_domain(cpu, sd) {
8293 * Decay the newidle max times here because this is a regular
8294 * visit to all the domains. Decay ~1% per second.
8296 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8297 sd->max_newidle_lb_cost =
8298 (sd->max_newidle_lb_cost * 253) / 256;
8299 sd->next_decay_max_lb_cost = jiffies + HZ;
8302 max_cost += sd->max_newidle_lb_cost;
8304 if (!(sd->flags & SD_LOAD_BALANCE))
8308 * Stop the load balance at this level. There is another
8309 * CPU in our sched group which is doing load balancing more
8312 if (!continue_balancing) {
8318 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8320 need_serialize = sd->flags & SD_SERIALIZE;
8321 if (need_serialize) {
8322 if (!spin_trylock(&balancing))
8326 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8327 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8329 * The LBF_DST_PINNED logic could have changed
8330 * env->dst_cpu, so we can't know our idle
8331 * state even if we migrated tasks. Update it.
8333 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8335 sd->last_balance = jiffies;
8336 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8339 spin_unlock(&balancing);
8341 if (time_after(next_balance, sd->last_balance + interval)) {
8342 next_balance = sd->last_balance + interval;
8343 update_next_balance = 1;
8348 * Ensure the rq-wide value also decays but keep it at a
8349 * reasonable floor to avoid funnies with rq->avg_idle.
8351 rq->max_idle_balance_cost =
8352 max((u64)sysctl_sched_migration_cost, max_cost);
8357 * next_balance will be updated only when there is a need.
8358 * When the cpu is attached to null domain for ex, it will not be
8361 if (likely(update_next_balance)) {
8362 rq->next_balance = next_balance;
8364 #ifdef CONFIG_NO_HZ_COMMON
8366 * If this CPU has been elected to perform the nohz idle
8367 * balance. Other idle CPUs have already rebalanced with
8368 * nohz_idle_balance() and nohz.next_balance has been
8369 * updated accordingly. This CPU is now running the idle load
8370 * balance for itself and we need to update the
8371 * nohz.next_balance accordingly.
8373 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8374 nohz.next_balance = rq->next_balance;
8379 #ifdef CONFIG_NO_HZ_COMMON
8381 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8382 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8384 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8386 int this_cpu = this_rq->cpu;
8389 /* Earliest time when we have to do rebalance again */
8390 unsigned long next_balance = jiffies + 60*HZ;
8391 int update_next_balance = 0;
8393 if (idle != CPU_IDLE ||
8394 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8397 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8398 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8402 * If this cpu gets work to do, stop the load balancing
8403 * work being done for other cpus. Next load
8404 * balancing owner will pick it up.
8409 rq = cpu_rq(balance_cpu);
8412 * If time for next balance is due,
8415 if (time_after_eq(jiffies, rq->next_balance)) {
8416 raw_spin_lock_irq(&rq->lock);
8417 update_rq_clock(rq);
8418 cpu_load_update_idle(rq);
8419 raw_spin_unlock_irq(&rq->lock);
8420 rebalance_domains(rq, CPU_IDLE);
8423 if (time_after(next_balance, rq->next_balance)) {
8424 next_balance = rq->next_balance;
8425 update_next_balance = 1;
8430 * next_balance will be updated only when there is a need.
8431 * When the CPU is attached to null domain for ex, it will not be
8434 if (likely(update_next_balance))
8435 nohz.next_balance = next_balance;
8437 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8441 * Current heuristic for kicking the idle load balancer in the presence
8442 * of an idle cpu in the system.
8443 * - This rq has more than one task.
8444 * - This rq has at least one CFS task and the capacity of the CPU is
8445 * significantly reduced because of RT tasks or IRQs.
8446 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8447 * multiple busy cpu.
8448 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8449 * domain span are idle.
8451 static inline bool nohz_kick_needed(struct rq *rq)
8453 unsigned long now = jiffies;
8454 struct sched_domain_shared *sds;
8455 struct sched_domain *sd;
8456 int nr_busy, cpu = rq->cpu;
8459 if (unlikely(rq->idle_balance))
8463 * We may be recently in ticked or tickless idle mode. At the first
8464 * busy tick after returning from idle, we will update the busy stats.
8466 set_cpu_sd_state_busy();
8467 nohz_balance_exit_idle(cpu);
8470 * None are in tickless mode and hence no need for NOHZ idle load
8473 if (likely(!atomic_read(&nohz.nr_cpus)))
8476 if (time_before(now, nohz.next_balance))
8479 if (rq->nr_running >= 2)
8483 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8486 * XXX: write a coherent comment on why we do this.
8487 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8489 nr_busy = atomic_read(&sds->nr_busy_cpus);
8497 sd = rcu_dereference(rq->sd);
8499 if ((rq->cfs.h_nr_running >= 1) &&
8500 check_cpu_capacity(rq, sd)) {
8506 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8507 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8508 sched_domain_span(sd)) < cpu)) {
8518 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8522 * run_rebalance_domains is triggered when needed from the scheduler tick.
8523 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8525 static void run_rebalance_domains(struct softirq_action *h)
8527 struct rq *this_rq = this_rq();
8528 enum cpu_idle_type idle = this_rq->idle_balance ?
8529 CPU_IDLE : CPU_NOT_IDLE;
8532 * If this cpu has a pending nohz_balance_kick, then do the
8533 * balancing on behalf of the other idle cpus whose ticks are
8534 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8535 * give the idle cpus a chance to load balance. Else we may
8536 * load balance only within the local sched_domain hierarchy
8537 * and abort nohz_idle_balance altogether if we pull some load.
8539 nohz_idle_balance(this_rq, idle);
8540 rebalance_domains(this_rq, idle);
8544 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8546 void trigger_load_balance(struct rq *rq)
8548 /* Don't need to rebalance while attached to NULL domain */
8549 if (unlikely(on_null_domain(rq)))
8552 if (time_after_eq(jiffies, rq->next_balance))
8553 raise_softirq(SCHED_SOFTIRQ);
8554 #ifdef CONFIG_NO_HZ_COMMON
8555 if (nohz_kick_needed(rq))
8556 nohz_balancer_kick();
8560 static void rq_online_fair(struct rq *rq)
8564 update_runtime_enabled(rq);
8567 static void rq_offline_fair(struct rq *rq)
8571 /* Ensure any throttled groups are reachable by pick_next_task */
8572 unthrottle_offline_cfs_rqs(rq);
8575 #endif /* CONFIG_SMP */
8578 * scheduler tick hitting a task of our scheduling class:
8580 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8582 struct cfs_rq *cfs_rq;
8583 struct sched_entity *se = &curr->se;
8585 for_each_sched_entity(se) {
8586 cfs_rq = cfs_rq_of(se);
8587 entity_tick(cfs_rq, se, queued);
8590 if (static_branch_unlikely(&sched_numa_balancing))
8591 task_tick_numa(rq, curr);
8595 * called on fork with the child task as argument from the parent's context
8596 * - child not yet on the tasklist
8597 * - preemption disabled
8599 static void task_fork_fair(struct task_struct *p)
8601 struct cfs_rq *cfs_rq;
8602 struct sched_entity *se = &p->se, *curr;
8603 struct rq *rq = this_rq();
8605 raw_spin_lock(&rq->lock);
8606 update_rq_clock(rq);
8608 cfs_rq = task_cfs_rq(current);
8609 curr = cfs_rq->curr;
8611 update_curr(cfs_rq);
8612 se->vruntime = curr->vruntime;
8614 place_entity(cfs_rq, se, 1);
8616 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8618 * Upon rescheduling, sched_class::put_prev_task() will place
8619 * 'current' within the tree based on its new key value.
8621 swap(curr->vruntime, se->vruntime);
8625 se->vruntime -= cfs_rq->min_vruntime;
8626 raw_spin_unlock(&rq->lock);
8630 * Priority of the task has changed. Check to see if we preempt
8634 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8636 if (!task_on_rq_queued(p))
8640 * Reschedule if we are currently running on this runqueue and
8641 * our priority decreased, or if we are not currently running on
8642 * this runqueue and our priority is higher than the current's
8644 if (rq->curr == p) {
8645 if (p->prio > oldprio)
8648 check_preempt_curr(rq, p, 0);
8651 static inline bool vruntime_normalized(struct task_struct *p)
8653 struct sched_entity *se = &p->se;
8656 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8657 * the dequeue_entity(.flags=0) will already have normalized the
8664 * When !on_rq, vruntime of the task has usually NOT been normalized.
8665 * But there are some cases where it has already been normalized:
8667 * - A forked child which is waiting for being woken up by
8668 * wake_up_new_task().
8669 * - A task which has been woken up by try_to_wake_up() and
8670 * waiting for actually being woken up by sched_ttwu_pending().
8672 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8678 static void detach_task_cfs_rq(struct task_struct *p)
8680 struct sched_entity *se = &p->se;
8681 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8682 u64 now = cfs_rq_clock_task(cfs_rq);
8684 if (!vruntime_normalized(p)) {
8686 * Fix up our vruntime so that the current sleep doesn't
8687 * cause 'unlimited' sleep bonus.
8689 place_entity(cfs_rq, se, 0);
8690 se->vruntime -= cfs_rq->min_vruntime;
8693 /* Catch up with the cfs_rq and remove our load when we leave */
8694 update_cfs_rq_load_avg(now, cfs_rq, false);
8695 detach_entity_load_avg(cfs_rq, se);
8696 update_tg_load_avg(cfs_rq, false);
8699 static void attach_task_cfs_rq(struct task_struct *p)
8701 struct sched_entity *se = &p->se;
8702 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8703 u64 now = cfs_rq_clock_task(cfs_rq);
8705 #ifdef CONFIG_FAIR_GROUP_SCHED
8707 * Since the real-depth could have been changed (only FAIR
8708 * class maintain depth value), reset depth properly.
8710 se->depth = se->parent ? se->parent->depth + 1 : 0;
8713 /* Synchronize task with its cfs_rq */
8714 update_cfs_rq_load_avg(now, cfs_rq, false);
8715 attach_entity_load_avg(cfs_rq, se);
8716 update_tg_load_avg(cfs_rq, false);
8718 if (!vruntime_normalized(p))
8719 se->vruntime += cfs_rq->min_vruntime;
8722 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8724 detach_task_cfs_rq(p);
8727 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8729 attach_task_cfs_rq(p);
8731 if (task_on_rq_queued(p)) {
8733 * We were most likely switched from sched_rt, so
8734 * kick off the schedule if running, otherwise just see
8735 * if we can still preempt the current task.
8740 check_preempt_curr(rq, p, 0);
8744 /* Account for a task changing its policy or group.
8746 * This routine is mostly called to set cfs_rq->curr field when a task
8747 * migrates between groups/classes.
8749 static void set_curr_task_fair(struct rq *rq)
8751 struct sched_entity *se = &rq->curr->se;
8753 for_each_sched_entity(se) {
8754 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8756 set_next_entity(cfs_rq, se);
8757 /* ensure bandwidth has been allocated on our new cfs_rq */
8758 account_cfs_rq_runtime(cfs_rq, 0);
8762 void init_cfs_rq(struct cfs_rq *cfs_rq)
8764 cfs_rq->tasks_timeline = RB_ROOT;
8765 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8766 #ifndef CONFIG_64BIT
8767 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8770 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8771 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8775 #ifdef CONFIG_FAIR_GROUP_SCHED
8776 static void task_set_group_fair(struct task_struct *p)
8778 struct sched_entity *se = &p->se;
8780 set_task_rq(p, task_cpu(p));
8781 se->depth = se->parent ? se->parent->depth + 1 : 0;
8784 static void task_move_group_fair(struct task_struct *p)
8786 detach_task_cfs_rq(p);
8787 set_task_rq(p, task_cpu(p));
8790 /* Tell se's cfs_rq has been changed -- migrated */
8791 p->se.avg.last_update_time = 0;
8793 attach_task_cfs_rq(p);
8796 static void task_change_group_fair(struct task_struct *p, int type)
8799 case TASK_SET_GROUP:
8800 task_set_group_fair(p);
8803 case TASK_MOVE_GROUP:
8804 task_move_group_fair(p);
8809 void free_fair_sched_group(struct task_group *tg)
8813 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8815 for_each_possible_cpu(i) {
8817 kfree(tg->cfs_rq[i]);
8826 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8828 struct sched_entity *se;
8829 struct cfs_rq *cfs_rq;
8833 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8836 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8840 tg->shares = NICE_0_LOAD;
8842 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8844 for_each_possible_cpu(i) {
8847 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8848 GFP_KERNEL, cpu_to_node(i));
8852 se = kzalloc_node(sizeof(struct sched_entity),
8853 GFP_KERNEL, cpu_to_node(i));
8857 init_cfs_rq(cfs_rq);
8858 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8859 init_entity_runnable_average(se);
8870 void online_fair_sched_group(struct task_group *tg)
8872 struct sched_entity *se;
8876 for_each_possible_cpu(i) {
8880 raw_spin_lock_irq(&rq->lock);
8881 post_init_entity_util_avg(se);
8882 sync_throttle(tg, i);
8883 raw_spin_unlock_irq(&rq->lock);
8887 void unregister_fair_sched_group(struct task_group *tg)
8889 unsigned long flags;
8893 for_each_possible_cpu(cpu) {
8895 remove_entity_load_avg(tg->se[cpu]);
8898 * Only empty task groups can be destroyed; so we can speculatively
8899 * check on_list without danger of it being re-added.
8901 if (!tg->cfs_rq[cpu]->on_list)
8906 raw_spin_lock_irqsave(&rq->lock, flags);
8907 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8908 raw_spin_unlock_irqrestore(&rq->lock, flags);
8912 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8913 struct sched_entity *se, int cpu,
8914 struct sched_entity *parent)
8916 struct rq *rq = cpu_rq(cpu);
8920 init_cfs_rq_runtime(cfs_rq);
8922 tg->cfs_rq[cpu] = cfs_rq;
8925 /* se could be NULL for root_task_group */
8930 se->cfs_rq = &rq->cfs;
8933 se->cfs_rq = parent->my_q;
8934 se->depth = parent->depth + 1;
8938 /* guarantee group entities always have weight */
8939 update_load_set(&se->load, NICE_0_LOAD);
8940 se->parent = parent;
8943 static DEFINE_MUTEX(shares_mutex);
8945 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8948 unsigned long flags;
8951 * We can't change the weight of the root cgroup.
8956 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8958 mutex_lock(&shares_mutex);
8959 if (tg->shares == shares)
8962 tg->shares = shares;
8963 for_each_possible_cpu(i) {
8964 struct rq *rq = cpu_rq(i);
8965 struct sched_entity *se;
8968 /* Propagate contribution to hierarchy */
8969 raw_spin_lock_irqsave(&rq->lock, flags);
8971 /* Possible calls to update_curr() need rq clock */
8972 update_rq_clock(rq);
8973 for_each_sched_entity(se)
8974 update_cfs_shares(group_cfs_rq(se));
8975 raw_spin_unlock_irqrestore(&rq->lock, flags);
8979 mutex_unlock(&shares_mutex);
8982 #else /* CONFIG_FAIR_GROUP_SCHED */
8984 void free_fair_sched_group(struct task_group *tg) { }
8986 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8991 void online_fair_sched_group(struct task_group *tg) { }
8993 void unregister_fair_sched_group(struct task_group *tg) { }
8995 #endif /* CONFIG_FAIR_GROUP_SCHED */
8998 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9000 struct sched_entity *se = &task->se;
9001 unsigned int rr_interval = 0;
9004 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9007 if (rq->cfs.load.weight)
9008 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9014 * All the scheduling class methods:
9016 const struct sched_class fair_sched_class = {
9017 .next = &idle_sched_class,
9018 .enqueue_task = enqueue_task_fair,
9019 .dequeue_task = dequeue_task_fair,
9020 .yield_task = yield_task_fair,
9021 .yield_to_task = yield_to_task_fair,
9023 .check_preempt_curr = check_preempt_wakeup,
9025 .pick_next_task = pick_next_task_fair,
9026 .put_prev_task = put_prev_task_fair,
9029 .select_task_rq = select_task_rq_fair,
9030 .migrate_task_rq = migrate_task_rq_fair,
9032 .rq_online = rq_online_fair,
9033 .rq_offline = rq_offline_fair,
9035 .task_dead = task_dead_fair,
9036 .set_cpus_allowed = set_cpus_allowed_common,
9039 .set_curr_task = set_curr_task_fair,
9040 .task_tick = task_tick_fair,
9041 .task_fork = task_fork_fair,
9043 .prio_changed = prio_changed_fair,
9044 .switched_from = switched_from_fair,
9045 .switched_to = switched_to_fair,
9047 .get_rr_interval = get_rr_interval_fair,
9049 .update_curr = update_curr_fair,
9051 #ifdef CONFIG_FAIR_GROUP_SCHED
9052 .task_change_group = task_change_group_fair,
9056 #ifdef CONFIG_SCHED_DEBUG
9057 void print_cfs_stats(struct seq_file *m, int cpu)
9059 struct cfs_rq *cfs_rq;
9062 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9063 print_cfs_rq(m, cpu, cfs_rq);
9067 #ifdef CONFIG_NUMA_BALANCING
9068 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9071 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9073 for_each_online_node(node) {
9074 if (p->numa_faults) {
9075 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9076 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9078 if (p->numa_group) {
9079 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9080 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9082 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9085 #endif /* CONFIG_NUMA_BALANCING */
9086 #endif /* CONFIG_SCHED_DEBUG */
9088 __init void init_sched_fair_class(void)
9091 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9093 #ifdef CONFIG_NO_HZ_COMMON
9094 nohz.next_balance = jiffies;
9095 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);