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;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665 * dependent on this value.
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
674 struct sched_avg *sa = &se->avg;
676 sa->last_update_time = 0;
678 * sched_avg's period_contrib should be strictly less then 1024, so
679 * we give it 1023 to make sure it is almost a period (1024us), and
680 * will definitely be update (after enqueue).
682 sa->period_contrib = 1023;
683 sa->load_avg = scale_load_down(se->load.weight);
684 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
686 * At this point, util_avg won't be used in select_task_rq_fair anyway
690 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
694 * With new tasks being created, their initial util_avgs are extrapolated
695 * based on the cfs_rq's current util_avg:
697 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
699 * However, in many cases, the above util_avg does not give a desired
700 * value. Moreover, the sum of the util_avgs may be divergent, such
701 * as when the series is a harmonic series.
703 * To solve this problem, we also cap the util_avg of successive tasks to
704 * only 1/2 of the left utilization budget:
706 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
708 * where n denotes the nth task.
710 * For example, a simplest series from the beginning would be like:
712 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
713 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
715 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
716 * if util_avg > util_avg_cap.
718 void post_init_entity_util_avg(struct sched_entity *se)
720 struct cfs_rq *cfs_rq = cfs_rq_of(se);
721 struct sched_avg *sa = &se->avg;
722 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
725 if (cfs_rq->avg.util_avg != 0) {
726 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
727 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
729 if (sa->util_avg > cap)
734 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
738 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq);
739 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq);
741 void init_entity_runnable_average(struct sched_entity *se)
744 void post_init_entity_util_avg(struct sched_entity *se)
750 * Update the current task's runtime statistics.
752 static void update_curr(struct cfs_rq *cfs_rq)
754 struct sched_entity *curr = cfs_rq->curr;
755 u64 now = rq_clock_task(rq_of(cfs_rq));
761 delta_exec = now - curr->exec_start;
762 if (unlikely((s64)delta_exec <= 0))
765 curr->exec_start = now;
767 schedstat_set(curr->statistics.exec_max,
768 max(delta_exec, curr->statistics.exec_max));
770 curr->sum_exec_runtime += delta_exec;
771 schedstat_add(cfs_rq, exec_clock, delta_exec);
773 curr->vruntime += calc_delta_fair(delta_exec, curr);
774 update_min_vruntime(cfs_rq);
776 if (entity_is_task(curr)) {
777 struct task_struct *curtask = task_of(curr);
779 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
780 cpuacct_charge(curtask, delta_exec);
781 account_group_exec_runtime(curtask, delta_exec);
784 account_cfs_rq_runtime(cfs_rq, delta_exec);
787 static void update_curr_fair(struct rq *rq)
789 update_curr(cfs_rq_of(&rq->curr->se));
792 #ifdef CONFIG_SCHEDSTATS
794 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 u64 wait_start = rq_clock(rq_of(cfs_rq));
798 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
799 likely(wait_start > se->statistics.wait_start))
800 wait_start -= se->statistics.wait_start;
802 se->statistics.wait_start = wait_start;
806 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
808 struct task_struct *p;
811 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
813 if (entity_is_task(se)) {
815 if (task_on_rq_migrating(p)) {
817 * Preserve migrating task's wait time so wait_start
818 * time stamp can be adjusted to accumulate wait time
819 * prior to migration.
821 se->statistics.wait_start = delta;
824 trace_sched_stat_wait(p, delta);
827 se->statistics.wait_max = max(se->statistics.wait_max, delta);
828 se->statistics.wait_count++;
829 se->statistics.wait_sum += delta;
830 se->statistics.wait_start = 0;
834 * Task is being enqueued - update stats:
837 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
840 * Are we enqueueing a waiting task? (for current tasks
841 * a dequeue/enqueue event is a NOP)
843 if (se != cfs_rq->curr)
844 update_stats_wait_start(cfs_rq, se);
848 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
851 * Mark the end of the wait period if dequeueing a
854 if (se != cfs_rq->curr)
855 update_stats_wait_end(cfs_rq, se);
857 if (flags & DEQUEUE_SLEEP) {
858 if (entity_is_task(se)) {
859 struct task_struct *tsk = task_of(se);
861 if (tsk->state & TASK_INTERRUPTIBLE)
862 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
863 if (tsk->state & TASK_UNINTERRUPTIBLE)
864 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
871 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
876 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
881 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
886 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
892 * We are picking a new current task - update its stats:
895 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
898 * We are starting a new run period:
900 se->exec_start = rq_clock_task(rq_of(cfs_rq));
903 /**************************************************
904 * Scheduling class queueing methods:
907 #ifdef CONFIG_NUMA_BALANCING
909 * Approximate time to scan a full NUMA task in ms. The task scan period is
910 * calculated based on the tasks virtual memory size and
911 * numa_balancing_scan_size.
913 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
914 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
916 /* Portion of address space to scan in MB */
917 unsigned int sysctl_numa_balancing_scan_size = 256;
919 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
920 unsigned int sysctl_numa_balancing_scan_delay = 1000;
922 static unsigned int task_nr_scan_windows(struct task_struct *p)
924 unsigned long rss = 0;
925 unsigned long nr_scan_pages;
928 * Calculations based on RSS as non-present and empty pages are skipped
929 * by the PTE scanner and NUMA hinting faults should be trapped based
932 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
933 rss = get_mm_rss(p->mm);
937 rss = round_up(rss, nr_scan_pages);
938 return rss / nr_scan_pages;
941 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
942 #define MAX_SCAN_WINDOW 2560
944 static unsigned int task_scan_min(struct task_struct *p)
946 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
947 unsigned int scan, floor;
948 unsigned int windows = 1;
950 if (scan_size < MAX_SCAN_WINDOW)
951 windows = MAX_SCAN_WINDOW / scan_size;
952 floor = 1000 / windows;
954 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
955 return max_t(unsigned int, floor, scan);
958 static unsigned int task_scan_max(struct task_struct *p)
960 unsigned int smin = task_scan_min(p);
963 /* Watch for min being lower than max due to floor calculations */
964 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
965 return max(smin, smax);
968 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
970 rq->nr_numa_running += (p->numa_preferred_nid != -1);
971 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
974 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
976 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
977 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
983 spinlock_t lock; /* nr_tasks, tasks */
989 unsigned long total_faults;
990 unsigned long max_faults_cpu;
992 * Faults_cpu is used to decide whether memory should move
993 * towards the CPU. As a consequence, these stats are weighted
994 * more by CPU use than by memory faults.
996 unsigned long *faults_cpu;
997 unsigned long faults[0];
1000 /* Shared or private faults. */
1001 #define NR_NUMA_HINT_FAULT_TYPES 2
1003 /* Memory and CPU locality */
1004 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1006 /* Averaged statistics, and temporary buffers. */
1007 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1009 pid_t task_numa_group_id(struct task_struct *p)
1011 return p->numa_group ? p->numa_group->gid : 0;
1015 * The averaged statistics, shared & private, memory & cpu,
1016 * occupy the first half of the array. The second half of the
1017 * array is for current counters, which are averaged into the
1018 * first set by task_numa_placement.
1020 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1022 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1025 static inline unsigned long task_faults(struct task_struct *p, int nid)
1027 if (!p->numa_faults)
1030 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1031 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1034 static inline unsigned long group_faults(struct task_struct *p, int nid)
1039 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1040 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1043 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1045 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1046 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1050 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1051 * considered part of a numa group's pseudo-interleaving set. Migrations
1052 * between these nodes are slowed down, to allow things to settle down.
1054 #define ACTIVE_NODE_FRACTION 3
1056 static bool numa_is_active_node(int nid, struct numa_group *ng)
1058 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1061 /* Handle placement on systems where not all nodes are directly connected. */
1062 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1063 int maxdist, bool task)
1065 unsigned long score = 0;
1069 * All nodes are directly connected, and the same distance
1070 * from each other. No need for fancy placement algorithms.
1072 if (sched_numa_topology_type == NUMA_DIRECT)
1076 * This code is called for each node, introducing N^2 complexity,
1077 * which should be ok given the number of nodes rarely exceeds 8.
1079 for_each_online_node(node) {
1080 unsigned long faults;
1081 int dist = node_distance(nid, node);
1084 * The furthest away nodes in the system are not interesting
1085 * for placement; nid was already counted.
1087 if (dist == sched_max_numa_distance || node == nid)
1091 * On systems with a backplane NUMA topology, compare groups
1092 * of nodes, and move tasks towards the group with the most
1093 * memory accesses. When comparing two nodes at distance
1094 * "hoplimit", only nodes closer by than "hoplimit" are part
1095 * of each group. Skip other nodes.
1097 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1101 /* Add up the faults from nearby nodes. */
1103 faults = task_faults(p, node);
1105 faults = group_faults(p, node);
1108 * On systems with a glueless mesh NUMA topology, there are
1109 * no fixed "groups of nodes". Instead, nodes that are not
1110 * directly connected bounce traffic through intermediate
1111 * nodes; a numa_group can occupy any set of nodes.
1112 * The further away a node is, the less the faults count.
1113 * This seems to result in good task placement.
1115 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1116 faults *= (sched_max_numa_distance - dist);
1117 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1127 * These return the fraction of accesses done by a particular task, or
1128 * task group, on a particular numa node. The group weight is given a
1129 * larger multiplier, in order to group tasks together that are almost
1130 * evenly spread out between numa nodes.
1132 static inline unsigned long task_weight(struct task_struct *p, int nid,
1135 unsigned long faults, total_faults;
1137 if (!p->numa_faults)
1140 total_faults = p->total_numa_faults;
1145 faults = task_faults(p, nid);
1146 faults += score_nearby_nodes(p, nid, dist, true);
1148 return 1000 * faults / total_faults;
1151 static inline unsigned long group_weight(struct task_struct *p, int nid,
1154 unsigned long faults, total_faults;
1159 total_faults = p->numa_group->total_faults;
1164 faults = group_faults(p, nid);
1165 faults += score_nearby_nodes(p, nid, dist, false);
1167 return 1000 * faults / total_faults;
1170 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1171 int src_nid, int dst_cpu)
1173 struct numa_group *ng = p->numa_group;
1174 int dst_nid = cpu_to_node(dst_cpu);
1175 int last_cpupid, this_cpupid;
1177 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1180 * Multi-stage node selection is used in conjunction with a periodic
1181 * migration fault to build a temporal task<->page relation. By using
1182 * a two-stage filter we remove short/unlikely relations.
1184 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1185 * a task's usage of a particular page (n_p) per total usage of this
1186 * page (n_t) (in a given time-span) to a probability.
1188 * Our periodic faults will sample this probability and getting the
1189 * same result twice in a row, given these samples are fully
1190 * independent, is then given by P(n)^2, provided our sample period
1191 * is sufficiently short compared to the usage pattern.
1193 * This quadric squishes small probabilities, making it less likely we
1194 * act on an unlikely task<->page relation.
1196 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1197 if (!cpupid_pid_unset(last_cpupid) &&
1198 cpupid_to_nid(last_cpupid) != dst_nid)
1201 /* Always allow migrate on private faults */
1202 if (cpupid_match_pid(p, last_cpupid))
1205 /* A shared fault, but p->numa_group has not been set up yet. */
1210 * Destination node is much more heavily used than the source
1211 * node? Allow migration.
1213 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1214 ACTIVE_NODE_FRACTION)
1218 * Distribute memory according to CPU & memory use on each node,
1219 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1221 * faults_cpu(dst) 3 faults_cpu(src)
1222 * --------------- * - > ---------------
1223 * faults_mem(dst) 4 faults_mem(src)
1225 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1226 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1229 static unsigned long weighted_cpuload(const int cpu);
1230 static unsigned long source_load(int cpu, int type);
1231 static unsigned long target_load(int cpu, int type);
1232 static unsigned long capacity_of(int cpu);
1233 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1235 /* Cached statistics for all CPUs within a node */
1237 unsigned long nr_running;
1240 /* Total compute capacity of CPUs on a node */
1241 unsigned long compute_capacity;
1243 /* Approximate capacity in terms of runnable tasks on a node */
1244 unsigned long task_capacity;
1245 int has_free_capacity;
1249 * XXX borrowed from update_sg_lb_stats
1251 static void update_numa_stats(struct numa_stats *ns, int nid)
1253 int smt, cpu, cpus = 0;
1254 unsigned long capacity;
1256 memset(ns, 0, sizeof(*ns));
1257 for_each_cpu(cpu, cpumask_of_node(nid)) {
1258 struct rq *rq = cpu_rq(cpu);
1260 ns->nr_running += rq->nr_running;
1261 ns->load += weighted_cpuload(cpu);
1262 ns->compute_capacity += capacity_of(cpu);
1268 * If we raced with hotplug and there are no CPUs left in our mask
1269 * the @ns structure is NULL'ed and task_numa_compare() will
1270 * not find this node attractive.
1272 * We'll either bail at !has_free_capacity, or we'll detect a huge
1273 * imbalance and bail there.
1278 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1279 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1280 capacity = cpus / smt; /* cores */
1282 ns->task_capacity = min_t(unsigned, capacity,
1283 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1284 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1287 struct task_numa_env {
1288 struct task_struct *p;
1290 int src_cpu, src_nid;
1291 int dst_cpu, dst_nid;
1293 struct numa_stats src_stats, dst_stats;
1298 struct task_struct *best_task;
1303 static void task_numa_assign(struct task_numa_env *env,
1304 struct task_struct *p, long imp)
1307 put_task_struct(env->best_task);
1310 env->best_imp = imp;
1311 env->best_cpu = env->dst_cpu;
1314 static bool load_too_imbalanced(long src_load, long dst_load,
1315 struct task_numa_env *env)
1318 long orig_src_load, orig_dst_load;
1319 long src_capacity, dst_capacity;
1322 * The load is corrected for the CPU capacity available on each node.
1325 * ------------ vs ---------
1326 * src_capacity dst_capacity
1328 src_capacity = env->src_stats.compute_capacity;
1329 dst_capacity = env->dst_stats.compute_capacity;
1331 /* We care about the slope of the imbalance, not the direction. */
1332 if (dst_load < src_load)
1333 swap(dst_load, src_load);
1335 /* Is the difference below the threshold? */
1336 imb = dst_load * src_capacity * 100 -
1337 src_load * dst_capacity * env->imbalance_pct;
1342 * The imbalance is above the allowed threshold.
1343 * Compare it with the old imbalance.
1345 orig_src_load = env->src_stats.load;
1346 orig_dst_load = env->dst_stats.load;
1348 if (orig_dst_load < orig_src_load)
1349 swap(orig_dst_load, orig_src_load);
1351 old_imb = orig_dst_load * src_capacity * 100 -
1352 orig_src_load * dst_capacity * env->imbalance_pct;
1354 /* Would this change make things worse? */
1355 return (imb > old_imb);
1359 * This checks if the overall compute and NUMA accesses of the system would
1360 * be improved if the source tasks was migrated to the target dst_cpu taking
1361 * into account that it might be best if task running on the dst_cpu should
1362 * be exchanged with the source task
1364 static void task_numa_compare(struct task_numa_env *env,
1365 long taskimp, long groupimp)
1367 struct rq *src_rq = cpu_rq(env->src_cpu);
1368 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1369 struct task_struct *cur;
1370 long src_load, dst_load;
1372 long imp = env->p->numa_group ? groupimp : taskimp;
1374 int dist = env->dist;
1375 bool assigned = false;
1379 raw_spin_lock_irq(&dst_rq->lock);
1382 * No need to move the exiting task or idle task.
1384 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1388 * The task_struct must be protected here to protect the
1389 * p->numa_faults access in the task_weight since the
1390 * numa_faults could already be freed in the following path:
1391 * finish_task_switch()
1392 * --> put_task_struct()
1393 * --> __put_task_struct()
1394 * --> task_numa_free()
1396 get_task_struct(cur);
1399 raw_spin_unlock_irq(&dst_rq->lock);
1402 * Because we have preemption enabled we can get migrated around and
1403 * end try selecting ourselves (current == env->p) as a swap candidate.
1409 * "imp" is the fault differential for the source task between the
1410 * source and destination node. Calculate the total differential for
1411 * the source task and potential destination task. The more negative
1412 * the value is, the more rmeote accesses that would be expected to
1413 * be incurred if the tasks were swapped.
1416 /* Skip this swap candidate if cannot move to the source cpu */
1417 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1421 * If dst and source tasks are in the same NUMA group, or not
1422 * in any group then look only at task weights.
1424 if (cur->numa_group == env->p->numa_group) {
1425 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1426 task_weight(cur, env->dst_nid, dist);
1428 * Add some hysteresis to prevent swapping the
1429 * tasks within a group over tiny differences.
1431 if (cur->numa_group)
1435 * Compare the group weights. If a task is all by
1436 * itself (not part of a group), use the task weight
1439 if (cur->numa_group)
1440 imp += group_weight(cur, env->src_nid, dist) -
1441 group_weight(cur, env->dst_nid, dist);
1443 imp += task_weight(cur, env->src_nid, dist) -
1444 task_weight(cur, env->dst_nid, dist);
1448 if (imp <= env->best_imp && moveimp <= env->best_imp)
1452 /* Is there capacity at our destination? */
1453 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1454 !env->dst_stats.has_free_capacity)
1460 /* Balance doesn't matter much if we're running a task per cpu */
1461 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1462 dst_rq->nr_running == 1)
1466 * In the overloaded case, try and keep the load balanced.
1469 load = task_h_load(env->p);
1470 dst_load = env->dst_stats.load + load;
1471 src_load = env->src_stats.load - load;
1473 if (moveimp > imp && moveimp > env->best_imp) {
1475 * If the improvement from just moving env->p direction is
1476 * better than swapping tasks around, check if a move is
1477 * possible. Store a slightly smaller score than moveimp,
1478 * so an actually idle CPU will win.
1480 if (!load_too_imbalanced(src_load, dst_load, env)) {
1482 put_task_struct(cur);
1488 if (imp <= env->best_imp)
1492 load = task_h_load(cur);
1497 if (load_too_imbalanced(src_load, dst_load, env))
1501 * One idle CPU per node is evaluated for a task numa move.
1502 * Call select_idle_sibling to maybe find a better one.
1505 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1509 task_numa_assign(env, cur, imp);
1513 * The dst_rq->curr isn't assigned. The protection for task_struct is
1516 if (cur && !assigned)
1517 put_task_struct(cur);
1520 static void task_numa_find_cpu(struct task_numa_env *env,
1521 long taskimp, long groupimp)
1525 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1526 /* Skip this CPU if the source task cannot migrate */
1527 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1531 task_numa_compare(env, taskimp, groupimp);
1535 /* Only move tasks to a NUMA node less busy than the current node. */
1536 static bool numa_has_capacity(struct task_numa_env *env)
1538 struct numa_stats *src = &env->src_stats;
1539 struct numa_stats *dst = &env->dst_stats;
1541 if (src->has_free_capacity && !dst->has_free_capacity)
1545 * Only consider a task move if the source has a higher load
1546 * than the destination, corrected for CPU capacity on each node.
1548 * src->load dst->load
1549 * --------------------- vs ---------------------
1550 * src->compute_capacity dst->compute_capacity
1552 if (src->load * dst->compute_capacity * env->imbalance_pct >
1554 dst->load * src->compute_capacity * 100)
1560 static int task_numa_migrate(struct task_struct *p)
1562 struct task_numa_env env = {
1565 .src_cpu = task_cpu(p),
1566 .src_nid = task_node(p),
1568 .imbalance_pct = 112,
1574 struct sched_domain *sd;
1575 unsigned long taskweight, groupweight;
1577 long taskimp, groupimp;
1580 * Pick the lowest SD_NUMA domain, as that would have the smallest
1581 * imbalance and would be the first to start moving tasks about.
1583 * And we want to avoid any moving of tasks about, as that would create
1584 * random movement of tasks -- counter the numa conditions we're trying
1588 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1590 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1594 * Cpusets can break the scheduler domain tree into smaller
1595 * balance domains, some of which do not cross NUMA boundaries.
1596 * Tasks that are "trapped" in such domains cannot be migrated
1597 * elsewhere, so there is no point in (re)trying.
1599 if (unlikely(!sd)) {
1600 p->numa_preferred_nid = task_node(p);
1604 env.dst_nid = p->numa_preferred_nid;
1605 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1606 taskweight = task_weight(p, env.src_nid, dist);
1607 groupweight = group_weight(p, env.src_nid, dist);
1608 update_numa_stats(&env.src_stats, env.src_nid);
1609 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1610 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1611 update_numa_stats(&env.dst_stats, env.dst_nid);
1613 /* Try to find a spot on the preferred nid. */
1614 if (numa_has_capacity(&env))
1615 task_numa_find_cpu(&env, taskimp, groupimp);
1618 * Look at other nodes in these cases:
1619 * - there is no space available on the preferred_nid
1620 * - the task is part of a numa_group that is interleaved across
1621 * multiple NUMA nodes; in order to better consolidate the group,
1622 * we need to check other locations.
1624 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1625 for_each_online_node(nid) {
1626 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1629 dist = node_distance(env.src_nid, env.dst_nid);
1630 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1632 taskweight = task_weight(p, env.src_nid, dist);
1633 groupweight = group_weight(p, env.src_nid, dist);
1636 /* Only consider nodes where both task and groups benefit */
1637 taskimp = task_weight(p, nid, dist) - taskweight;
1638 groupimp = group_weight(p, nid, dist) - groupweight;
1639 if (taskimp < 0 && groupimp < 0)
1644 update_numa_stats(&env.dst_stats, env.dst_nid);
1645 if (numa_has_capacity(&env))
1646 task_numa_find_cpu(&env, taskimp, groupimp);
1651 * If the task is part of a workload that spans multiple NUMA nodes,
1652 * and is migrating into one of the workload's active nodes, remember
1653 * this node as the task's preferred numa node, so the workload can
1655 * A task that migrated to a second choice node will be better off
1656 * trying for a better one later. Do not set the preferred node here.
1658 if (p->numa_group) {
1659 struct numa_group *ng = p->numa_group;
1661 if (env.best_cpu == -1)
1666 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1667 sched_setnuma(p, env.dst_nid);
1670 /* No better CPU than the current one was found. */
1671 if (env.best_cpu == -1)
1675 * Reset the scan period if the task is being rescheduled on an
1676 * alternative node to recheck if the tasks is now properly placed.
1678 p->numa_scan_period = task_scan_min(p);
1680 if (env.best_task == NULL) {
1681 ret = migrate_task_to(p, env.best_cpu);
1683 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1687 ret = migrate_swap(p, env.best_task);
1689 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1690 put_task_struct(env.best_task);
1694 /* Attempt to migrate a task to a CPU on the preferred node. */
1695 static void numa_migrate_preferred(struct task_struct *p)
1697 unsigned long interval = HZ;
1699 /* This task has no NUMA fault statistics yet */
1700 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1703 /* Periodically retry migrating the task to the preferred node */
1704 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1705 p->numa_migrate_retry = jiffies + interval;
1707 /* Success if task is already running on preferred CPU */
1708 if (task_node(p) == p->numa_preferred_nid)
1711 /* Otherwise, try migrate to a CPU on the preferred node */
1712 task_numa_migrate(p);
1716 * Find out how many nodes on the workload is actively running on. Do this by
1717 * tracking the nodes from which NUMA hinting faults are triggered. This can
1718 * be different from the set of nodes where the workload's memory is currently
1721 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1723 unsigned long faults, max_faults = 0;
1724 int nid, active_nodes = 0;
1726 for_each_online_node(nid) {
1727 faults = group_faults_cpu(numa_group, nid);
1728 if (faults > max_faults)
1729 max_faults = faults;
1732 for_each_online_node(nid) {
1733 faults = group_faults_cpu(numa_group, nid);
1734 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1738 numa_group->max_faults_cpu = max_faults;
1739 numa_group->active_nodes = active_nodes;
1743 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1744 * increments. The more local the fault statistics are, the higher the scan
1745 * period will be for the next scan window. If local/(local+remote) ratio is
1746 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1747 * the scan period will decrease. Aim for 70% local accesses.
1749 #define NUMA_PERIOD_SLOTS 10
1750 #define NUMA_PERIOD_THRESHOLD 7
1753 * Increase the scan period (slow down scanning) if the majority of
1754 * our memory is already on our local node, or if the majority of
1755 * the page accesses are shared with other processes.
1756 * Otherwise, decrease the scan period.
1758 static void update_task_scan_period(struct task_struct *p,
1759 unsigned long shared, unsigned long private)
1761 unsigned int period_slot;
1765 unsigned long remote = p->numa_faults_locality[0];
1766 unsigned long local = p->numa_faults_locality[1];
1769 * If there were no record hinting faults then either the task is
1770 * completely idle or all activity is areas that are not of interest
1771 * to automatic numa balancing. Related to that, if there were failed
1772 * migration then it implies we are migrating too quickly or the local
1773 * node is overloaded. In either case, scan slower
1775 if (local + shared == 0 || p->numa_faults_locality[2]) {
1776 p->numa_scan_period = min(p->numa_scan_period_max,
1777 p->numa_scan_period << 1);
1779 p->mm->numa_next_scan = jiffies +
1780 msecs_to_jiffies(p->numa_scan_period);
1786 * Prepare to scale scan period relative to the current period.
1787 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1788 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1789 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1791 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1792 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1793 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1794 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1797 diff = slot * period_slot;
1799 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1802 * Scale scan rate increases based on sharing. There is an
1803 * inverse relationship between the degree of sharing and
1804 * the adjustment made to the scanning period. Broadly
1805 * speaking the intent is that there is little point
1806 * scanning faster if shared accesses dominate as it may
1807 * simply bounce migrations uselessly
1809 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1810 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1813 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1814 task_scan_min(p), task_scan_max(p));
1815 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1819 * Get the fraction of time the task has been running since the last
1820 * NUMA placement cycle. The scheduler keeps similar statistics, but
1821 * decays those on a 32ms period, which is orders of magnitude off
1822 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1823 * stats only if the task is so new there are no NUMA statistics yet.
1825 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1827 u64 runtime, delta, now;
1828 /* Use the start of this time slice to avoid calculations. */
1829 now = p->se.exec_start;
1830 runtime = p->se.sum_exec_runtime;
1832 if (p->last_task_numa_placement) {
1833 delta = runtime - p->last_sum_exec_runtime;
1834 *period = now - p->last_task_numa_placement;
1836 delta = p->se.avg.load_sum / p->se.load.weight;
1837 *period = LOAD_AVG_MAX;
1840 p->last_sum_exec_runtime = runtime;
1841 p->last_task_numa_placement = now;
1847 * Determine the preferred nid for a task in a numa_group. This needs to
1848 * be done in a way that produces consistent results with group_weight,
1849 * otherwise workloads might not converge.
1851 static int preferred_group_nid(struct task_struct *p, int nid)
1856 /* Direct connections between all NUMA nodes. */
1857 if (sched_numa_topology_type == NUMA_DIRECT)
1861 * On a system with glueless mesh NUMA topology, group_weight
1862 * scores nodes according to the number of NUMA hinting faults on
1863 * both the node itself, and on nearby nodes.
1865 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1866 unsigned long score, max_score = 0;
1867 int node, max_node = nid;
1869 dist = sched_max_numa_distance;
1871 for_each_online_node(node) {
1872 score = group_weight(p, node, dist);
1873 if (score > max_score) {
1882 * Finding the preferred nid in a system with NUMA backplane
1883 * interconnect topology is more involved. The goal is to locate
1884 * tasks from numa_groups near each other in the system, and
1885 * untangle workloads from different sides of the system. This requires
1886 * searching down the hierarchy of node groups, recursively searching
1887 * inside the highest scoring group of nodes. The nodemask tricks
1888 * keep the complexity of the search down.
1890 nodes = node_online_map;
1891 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1892 unsigned long max_faults = 0;
1893 nodemask_t max_group = NODE_MASK_NONE;
1896 /* Are there nodes at this distance from each other? */
1897 if (!find_numa_distance(dist))
1900 for_each_node_mask(a, nodes) {
1901 unsigned long faults = 0;
1902 nodemask_t this_group;
1903 nodes_clear(this_group);
1905 /* Sum group's NUMA faults; includes a==b case. */
1906 for_each_node_mask(b, nodes) {
1907 if (node_distance(a, b) < dist) {
1908 faults += group_faults(p, b);
1909 node_set(b, this_group);
1910 node_clear(b, nodes);
1914 /* Remember the top group. */
1915 if (faults > max_faults) {
1916 max_faults = faults;
1917 max_group = this_group;
1919 * subtle: at the smallest distance there is
1920 * just one node left in each "group", the
1921 * winner is the preferred nid.
1926 /* Next round, evaluate the nodes within max_group. */
1934 static void task_numa_placement(struct task_struct *p)
1936 int seq, nid, max_nid = -1, max_group_nid = -1;
1937 unsigned long max_faults = 0, max_group_faults = 0;
1938 unsigned long fault_types[2] = { 0, 0 };
1939 unsigned long total_faults;
1940 u64 runtime, period;
1941 spinlock_t *group_lock = NULL;
1944 * The p->mm->numa_scan_seq field gets updated without
1945 * exclusive access. Use READ_ONCE() here to ensure
1946 * that the field is read in a single access:
1948 seq = READ_ONCE(p->mm->numa_scan_seq);
1949 if (p->numa_scan_seq == seq)
1951 p->numa_scan_seq = seq;
1952 p->numa_scan_period_max = task_scan_max(p);
1954 total_faults = p->numa_faults_locality[0] +
1955 p->numa_faults_locality[1];
1956 runtime = numa_get_avg_runtime(p, &period);
1958 /* If the task is part of a group prevent parallel updates to group stats */
1959 if (p->numa_group) {
1960 group_lock = &p->numa_group->lock;
1961 spin_lock_irq(group_lock);
1964 /* Find the node with the highest number of faults */
1965 for_each_online_node(nid) {
1966 /* Keep track of the offsets in numa_faults array */
1967 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1968 unsigned long faults = 0, group_faults = 0;
1971 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1972 long diff, f_diff, f_weight;
1974 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1975 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1976 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1977 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1979 /* Decay existing window, copy faults since last scan */
1980 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1981 fault_types[priv] += p->numa_faults[membuf_idx];
1982 p->numa_faults[membuf_idx] = 0;
1985 * Normalize the faults_from, so all tasks in a group
1986 * count according to CPU use, instead of by the raw
1987 * number of faults. Tasks with little runtime have
1988 * little over-all impact on throughput, and thus their
1989 * faults are less important.
1991 f_weight = div64_u64(runtime << 16, period + 1);
1992 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1994 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1995 p->numa_faults[cpubuf_idx] = 0;
1997 p->numa_faults[mem_idx] += diff;
1998 p->numa_faults[cpu_idx] += f_diff;
1999 faults += p->numa_faults[mem_idx];
2000 p->total_numa_faults += diff;
2001 if (p->numa_group) {
2003 * safe because we can only change our own group
2005 * mem_idx represents the offset for a given
2006 * nid and priv in a specific region because it
2007 * is at the beginning of the numa_faults array.
2009 p->numa_group->faults[mem_idx] += diff;
2010 p->numa_group->faults_cpu[mem_idx] += f_diff;
2011 p->numa_group->total_faults += diff;
2012 group_faults += p->numa_group->faults[mem_idx];
2016 if (faults > max_faults) {
2017 max_faults = faults;
2021 if (group_faults > max_group_faults) {
2022 max_group_faults = group_faults;
2023 max_group_nid = nid;
2027 update_task_scan_period(p, fault_types[0], fault_types[1]);
2029 if (p->numa_group) {
2030 numa_group_count_active_nodes(p->numa_group);
2031 spin_unlock_irq(group_lock);
2032 max_nid = preferred_group_nid(p, max_group_nid);
2036 /* Set the new preferred node */
2037 if (max_nid != p->numa_preferred_nid)
2038 sched_setnuma(p, max_nid);
2040 if (task_node(p) != p->numa_preferred_nid)
2041 numa_migrate_preferred(p);
2045 static inline int get_numa_group(struct numa_group *grp)
2047 return atomic_inc_not_zero(&grp->refcount);
2050 static inline void put_numa_group(struct numa_group *grp)
2052 if (atomic_dec_and_test(&grp->refcount))
2053 kfree_rcu(grp, rcu);
2056 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2059 struct numa_group *grp, *my_grp;
2060 struct task_struct *tsk;
2062 int cpu = cpupid_to_cpu(cpupid);
2065 if (unlikely(!p->numa_group)) {
2066 unsigned int size = sizeof(struct numa_group) +
2067 4*nr_node_ids*sizeof(unsigned long);
2069 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2073 atomic_set(&grp->refcount, 1);
2074 grp->active_nodes = 1;
2075 grp->max_faults_cpu = 0;
2076 spin_lock_init(&grp->lock);
2078 /* Second half of the array tracks nids where faults happen */
2079 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2082 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2083 grp->faults[i] = p->numa_faults[i];
2085 grp->total_faults = p->total_numa_faults;
2088 rcu_assign_pointer(p->numa_group, grp);
2092 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2094 if (!cpupid_match_pid(tsk, cpupid))
2097 grp = rcu_dereference(tsk->numa_group);
2101 my_grp = p->numa_group;
2106 * Only join the other group if its bigger; if we're the bigger group,
2107 * the other task will join us.
2109 if (my_grp->nr_tasks > grp->nr_tasks)
2113 * Tie-break on the grp address.
2115 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2118 /* Always join threads in the same process. */
2119 if (tsk->mm == current->mm)
2122 /* Simple filter to avoid false positives due to PID collisions */
2123 if (flags & TNF_SHARED)
2126 /* Update priv based on whether false sharing was detected */
2129 if (join && !get_numa_group(grp))
2137 BUG_ON(irqs_disabled());
2138 double_lock_irq(&my_grp->lock, &grp->lock);
2140 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2141 my_grp->faults[i] -= p->numa_faults[i];
2142 grp->faults[i] += p->numa_faults[i];
2144 my_grp->total_faults -= p->total_numa_faults;
2145 grp->total_faults += p->total_numa_faults;
2150 spin_unlock(&my_grp->lock);
2151 spin_unlock_irq(&grp->lock);
2153 rcu_assign_pointer(p->numa_group, grp);
2155 put_numa_group(my_grp);
2163 void task_numa_free(struct task_struct *p)
2165 struct numa_group *grp = p->numa_group;
2166 void *numa_faults = p->numa_faults;
2167 unsigned long flags;
2171 spin_lock_irqsave(&grp->lock, flags);
2172 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2173 grp->faults[i] -= p->numa_faults[i];
2174 grp->total_faults -= p->total_numa_faults;
2177 spin_unlock_irqrestore(&grp->lock, flags);
2178 RCU_INIT_POINTER(p->numa_group, NULL);
2179 put_numa_group(grp);
2182 p->numa_faults = NULL;
2187 * Got a PROT_NONE fault for a page on @node.
2189 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2191 struct task_struct *p = current;
2192 bool migrated = flags & TNF_MIGRATED;
2193 int cpu_node = task_node(current);
2194 int local = !!(flags & TNF_FAULT_LOCAL);
2195 struct numa_group *ng;
2198 if (!static_branch_likely(&sched_numa_balancing))
2201 /* for example, ksmd faulting in a user's mm */
2205 /* Allocate buffer to track faults on a per-node basis */
2206 if (unlikely(!p->numa_faults)) {
2207 int size = sizeof(*p->numa_faults) *
2208 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2210 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2211 if (!p->numa_faults)
2214 p->total_numa_faults = 0;
2215 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2219 * First accesses are treated as private, otherwise consider accesses
2220 * to be private if the accessing pid has not changed
2222 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2225 priv = cpupid_match_pid(p, last_cpupid);
2226 if (!priv && !(flags & TNF_NO_GROUP))
2227 task_numa_group(p, last_cpupid, flags, &priv);
2231 * If a workload spans multiple NUMA nodes, a shared fault that
2232 * occurs wholly within the set of nodes that the workload is
2233 * actively using should be counted as local. This allows the
2234 * scan rate to slow down when a workload has settled down.
2237 if (!priv && !local && ng && ng->active_nodes > 1 &&
2238 numa_is_active_node(cpu_node, ng) &&
2239 numa_is_active_node(mem_node, ng))
2242 task_numa_placement(p);
2245 * Retry task to preferred node migration periodically, in case it
2246 * case it previously failed, or the scheduler moved us.
2248 if (time_after(jiffies, p->numa_migrate_retry))
2249 numa_migrate_preferred(p);
2252 p->numa_pages_migrated += pages;
2253 if (flags & TNF_MIGRATE_FAIL)
2254 p->numa_faults_locality[2] += pages;
2256 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2257 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2258 p->numa_faults_locality[local] += pages;
2261 static void reset_ptenuma_scan(struct task_struct *p)
2264 * We only did a read acquisition of the mmap sem, so
2265 * p->mm->numa_scan_seq is written to without exclusive access
2266 * and the update is not guaranteed to be atomic. That's not
2267 * much of an issue though, since this is just used for
2268 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2269 * expensive, to avoid any form of compiler optimizations:
2271 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2272 p->mm->numa_scan_offset = 0;
2276 * The expensive part of numa migration is done from task_work context.
2277 * Triggered from task_tick_numa().
2279 void task_numa_work(struct callback_head *work)
2281 unsigned long migrate, next_scan, now = jiffies;
2282 struct task_struct *p = current;
2283 struct mm_struct *mm = p->mm;
2284 u64 runtime = p->se.sum_exec_runtime;
2285 struct vm_area_struct *vma;
2286 unsigned long start, end;
2287 unsigned long nr_pte_updates = 0;
2288 long pages, virtpages;
2290 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2292 work->next = work; /* protect against double add */
2294 * Who cares about NUMA placement when they're dying.
2296 * NOTE: make sure not to dereference p->mm before this check,
2297 * exit_task_work() happens _after_ exit_mm() so we could be called
2298 * without p->mm even though we still had it when we enqueued this
2301 if (p->flags & PF_EXITING)
2304 if (!mm->numa_next_scan) {
2305 mm->numa_next_scan = now +
2306 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2310 * Enforce maximal scan/migration frequency..
2312 migrate = mm->numa_next_scan;
2313 if (time_before(now, migrate))
2316 if (p->numa_scan_period == 0) {
2317 p->numa_scan_period_max = task_scan_max(p);
2318 p->numa_scan_period = task_scan_min(p);
2321 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2322 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2326 * Delay this task enough that another task of this mm will likely win
2327 * the next time around.
2329 p->node_stamp += 2 * TICK_NSEC;
2331 start = mm->numa_scan_offset;
2332 pages = sysctl_numa_balancing_scan_size;
2333 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2334 virtpages = pages * 8; /* Scan up to this much virtual space */
2339 down_read(&mm->mmap_sem);
2340 vma = find_vma(mm, start);
2342 reset_ptenuma_scan(p);
2346 for (; vma; vma = vma->vm_next) {
2347 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2348 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2353 * Shared library pages mapped by multiple processes are not
2354 * migrated as it is expected they are cache replicated. Avoid
2355 * hinting faults in read-only file-backed mappings or the vdso
2356 * as migrating the pages will be of marginal benefit.
2359 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2363 * Skip inaccessible VMAs to avoid any confusion between
2364 * PROT_NONE and NUMA hinting ptes
2366 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2370 start = max(start, vma->vm_start);
2371 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2372 end = min(end, vma->vm_end);
2373 nr_pte_updates = change_prot_numa(vma, start, end);
2376 * Try to scan sysctl_numa_balancing_size worth of
2377 * hpages that have at least one present PTE that
2378 * is not already pte-numa. If the VMA contains
2379 * areas that are unused or already full of prot_numa
2380 * PTEs, scan up to virtpages, to skip through those
2384 pages -= (end - start) >> PAGE_SHIFT;
2385 virtpages -= (end - start) >> PAGE_SHIFT;
2388 if (pages <= 0 || virtpages <= 0)
2392 } while (end != vma->vm_end);
2397 * It is possible to reach the end of the VMA list but the last few
2398 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2399 * would find the !migratable VMA on the next scan but not reset the
2400 * scanner to the start so check it now.
2403 mm->numa_scan_offset = start;
2405 reset_ptenuma_scan(p);
2406 up_read(&mm->mmap_sem);
2409 * Make sure tasks use at least 32x as much time to run other code
2410 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2411 * Usually update_task_scan_period slows down scanning enough; on an
2412 * overloaded system we need to limit overhead on a per task basis.
2414 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2415 u64 diff = p->se.sum_exec_runtime - runtime;
2416 p->node_stamp += 32 * diff;
2421 * Drive the periodic memory faults..
2423 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2425 struct callback_head *work = &curr->numa_work;
2429 * We don't care about NUMA placement if we don't have memory.
2431 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2435 * Using runtime rather than walltime has the dual advantage that
2436 * we (mostly) drive the selection from busy threads and that the
2437 * task needs to have done some actual work before we bother with
2440 now = curr->se.sum_exec_runtime;
2441 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2443 if (now > curr->node_stamp + period) {
2444 if (!curr->node_stamp)
2445 curr->numa_scan_period = task_scan_min(curr);
2446 curr->node_stamp += period;
2448 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2449 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2450 task_work_add(curr, work, true);
2455 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2459 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2463 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2466 #endif /* CONFIG_NUMA_BALANCING */
2469 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2471 update_load_add(&cfs_rq->load, se->load.weight);
2472 if (!parent_entity(se))
2473 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2475 if (entity_is_task(se)) {
2476 struct rq *rq = rq_of(cfs_rq);
2478 account_numa_enqueue(rq, task_of(se));
2479 list_add(&se->group_node, &rq->cfs_tasks);
2482 cfs_rq->nr_running++;
2486 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2488 update_load_sub(&cfs_rq->load, se->load.weight);
2489 if (!parent_entity(se))
2490 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2492 if (entity_is_task(se)) {
2493 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2494 list_del_init(&se->group_node);
2497 cfs_rq->nr_running--;
2500 #ifdef CONFIG_FAIR_GROUP_SCHED
2502 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2507 * Use this CPU's real-time load instead of the last load contribution
2508 * as the updating of the contribution is delayed, and we will use the
2509 * the real-time load to calc the share. See update_tg_load_avg().
2511 tg_weight = atomic_long_read(&tg->load_avg);
2512 tg_weight -= cfs_rq->tg_load_avg_contrib;
2513 tg_weight += cfs_rq->load.weight;
2518 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2520 long tg_weight, load, shares;
2522 tg_weight = calc_tg_weight(tg, cfs_rq);
2523 load = cfs_rq->load.weight;
2525 shares = (tg->shares * load);
2527 shares /= tg_weight;
2529 if (shares < MIN_SHARES)
2530 shares = MIN_SHARES;
2531 if (shares > tg->shares)
2532 shares = tg->shares;
2536 # else /* CONFIG_SMP */
2537 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2541 # endif /* CONFIG_SMP */
2542 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2543 unsigned long weight)
2546 /* commit outstanding execution time */
2547 if (cfs_rq->curr == se)
2548 update_curr(cfs_rq);
2549 account_entity_dequeue(cfs_rq, se);
2552 update_load_set(&se->load, weight);
2555 account_entity_enqueue(cfs_rq, se);
2558 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2560 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2562 struct task_group *tg;
2563 struct sched_entity *se;
2567 se = tg->se[cpu_of(rq_of(cfs_rq))];
2568 if (!se || throttled_hierarchy(cfs_rq))
2571 if (likely(se->load.weight == tg->shares))
2574 shares = calc_cfs_shares(cfs_rq, tg);
2576 reweight_entity(cfs_rq_of(se), se, shares);
2578 #else /* CONFIG_FAIR_GROUP_SCHED */
2579 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2582 #endif /* CONFIG_FAIR_GROUP_SCHED */
2585 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2586 static const u32 runnable_avg_yN_inv[] = {
2587 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2588 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2589 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2590 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2591 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2592 0x85aac367, 0x82cd8698,
2596 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2597 * over-estimates when re-combining.
2599 static const u32 runnable_avg_yN_sum[] = {
2600 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2601 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2602 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2606 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2607 * lower integers. See Documentation/scheduler/sched-avg.txt how these
2610 static const u32 __accumulated_sum_N32[] = {
2611 0, 23371, 35056, 40899, 43820, 45281,
2612 46011, 46376, 46559, 46650, 46696, 46719,
2617 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2619 static __always_inline u64 decay_load(u64 val, u64 n)
2621 unsigned int local_n;
2625 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2628 /* after bounds checking we can collapse to 32-bit */
2632 * As y^PERIOD = 1/2, we can combine
2633 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2634 * With a look-up table which covers y^n (n<PERIOD)
2636 * To achieve constant time decay_load.
2638 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2639 val >>= local_n / LOAD_AVG_PERIOD;
2640 local_n %= LOAD_AVG_PERIOD;
2643 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2648 * For updates fully spanning n periods, the contribution to runnable
2649 * average will be: \Sum 1024*y^n
2651 * We can compute this reasonably efficiently by combining:
2652 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2654 static u32 __compute_runnable_contrib(u64 n)
2658 if (likely(n <= LOAD_AVG_PERIOD))
2659 return runnable_avg_yN_sum[n];
2660 else if (unlikely(n >= LOAD_AVG_MAX_N))
2661 return LOAD_AVG_MAX;
2663 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2664 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2665 n %= LOAD_AVG_PERIOD;
2666 contrib = decay_load(contrib, n);
2667 return contrib + runnable_avg_yN_sum[n];
2670 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2673 * We can represent the historical contribution to runnable average as the
2674 * coefficients of a geometric series. To do this we sub-divide our runnable
2675 * history into segments of approximately 1ms (1024us); label the segment that
2676 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2678 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2680 * (now) (~1ms ago) (~2ms ago)
2682 * Let u_i denote the fraction of p_i that the entity was runnable.
2684 * We then designate the fractions u_i as our co-efficients, yielding the
2685 * following representation of historical load:
2686 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2688 * We choose y based on the with of a reasonably scheduling period, fixing:
2691 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2692 * approximately half as much as the contribution to load within the last ms
2695 * When a period "rolls over" and we have new u_0`, multiplying the previous
2696 * sum again by y is sufficient to update:
2697 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2698 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2700 static __always_inline int
2701 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2702 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2704 u64 delta, scaled_delta, periods;
2706 unsigned int delta_w, scaled_delta_w, decayed = 0;
2707 unsigned long scale_freq, scale_cpu;
2709 delta = now - sa->last_update_time;
2711 * This should only happen when time goes backwards, which it
2712 * unfortunately does during sched clock init when we swap over to TSC.
2714 if ((s64)delta < 0) {
2715 sa->last_update_time = now;
2720 * Use 1024ns as the unit of measurement since it's a reasonable
2721 * approximation of 1us and fast to compute.
2726 sa->last_update_time = now;
2728 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2729 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2731 /* delta_w is the amount already accumulated against our next period */
2732 delta_w = sa->period_contrib;
2733 if (delta + delta_w >= 1024) {
2736 /* how much left for next period will start over, we don't know yet */
2737 sa->period_contrib = 0;
2740 * Now that we know we're crossing a period boundary, figure
2741 * out how much from delta we need to complete the current
2742 * period and accrue it.
2744 delta_w = 1024 - delta_w;
2745 scaled_delta_w = cap_scale(delta_w, scale_freq);
2747 sa->load_sum += weight * scaled_delta_w;
2749 cfs_rq->runnable_load_sum +=
2750 weight * scaled_delta_w;
2754 sa->util_sum += scaled_delta_w * scale_cpu;
2758 /* Figure out how many additional periods this update spans */
2759 periods = delta / 1024;
2762 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2764 cfs_rq->runnable_load_sum =
2765 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2767 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2769 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2770 contrib = __compute_runnable_contrib(periods);
2771 contrib = cap_scale(contrib, scale_freq);
2773 sa->load_sum += weight * contrib;
2775 cfs_rq->runnable_load_sum += weight * contrib;
2778 sa->util_sum += contrib * scale_cpu;
2781 /* Remainder of delta accrued against u_0` */
2782 scaled_delta = cap_scale(delta, scale_freq);
2784 sa->load_sum += weight * scaled_delta;
2786 cfs_rq->runnable_load_sum += weight * scaled_delta;
2789 sa->util_sum += scaled_delta * scale_cpu;
2791 sa->period_contrib += delta;
2794 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2796 cfs_rq->runnable_load_avg =
2797 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2799 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2805 #ifdef CONFIG_FAIR_GROUP_SCHED
2807 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2808 * and effective_load (which is not done because it is too costly).
2810 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2812 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2815 * No need to update load_avg for root_task_group as it is not used.
2817 if (cfs_rq->tg == &root_task_group)
2820 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2821 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2822 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2827 * Called within set_task_rq() right before setting a task's cpu. The
2828 * caller only guarantees p->pi_lock is held; no other assumptions,
2829 * including the state of rq->lock, should be made.
2831 void set_task_rq_fair(struct sched_entity *se,
2832 struct cfs_rq *prev, struct cfs_rq *next)
2834 if (!sched_feat(ATTACH_AGE_LOAD))
2838 * We are supposed to update the task to "current" time, then its up to
2839 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2840 * getting what current time is, so simply throw away the out-of-date
2841 * time. This will result in the wakee task is less decayed, but giving
2842 * the wakee more load sounds not bad.
2844 if (se->avg.last_update_time && prev) {
2845 u64 p_last_update_time;
2846 u64 n_last_update_time;
2848 #ifndef CONFIG_64BIT
2849 u64 p_last_update_time_copy;
2850 u64 n_last_update_time_copy;
2853 p_last_update_time_copy = prev->load_last_update_time_copy;
2854 n_last_update_time_copy = next->load_last_update_time_copy;
2858 p_last_update_time = prev->avg.last_update_time;
2859 n_last_update_time = next->avg.last_update_time;
2861 } while (p_last_update_time != p_last_update_time_copy ||
2862 n_last_update_time != n_last_update_time_copy);
2864 p_last_update_time = prev->avg.last_update_time;
2865 n_last_update_time = next->avg.last_update_time;
2867 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2868 &se->avg, 0, 0, NULL);
2869 se->avg.last_update_time = n_last_update_time;
2872 #else /* CONFIG_FAIR_GROUP_SCHED */
2873 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2874 #endif /* CONFIG_FAIR_GROUP_SCHED */
2876 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2878 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2880 struct rq *rq = rq_of(cfs_rq);
2881 int cpu = cpu_of(rq);
2883 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2884 unsigned long max = rq->cpu_capacity_orig;
2887 * There are a few boundary cases this might miss but it should
2888 * get called often enough that that should (hopefully) not be
2889 * a real problem -- added to that it only calls on the local
2890 * CPU, so if we enqueue remotely we'll miss an update, but
2891 * the next tick/schedule should update.
2893 * It will not get called when we go idle, because the idle
2894 * thread is a different class (!fair), nor will the utilization
2895 * number include things like RT tasks.
2897 * As is, the util number is not freq-invariant (we'd have to
2898 * implement arch_scale_freq_capacity() for that).
2902 cpufreq_update_util(rq_clock(rq),
2903 min(cfs_rq->avg.util_avg, max), max);
2908 * Unsigned subtract and clamp on underflow.
2910 * Explicitly do a load-store to ensure the intermediate value never hits
2911 * memory. This allows lockless observations without ever seeing the negative
2914 #define sub_positive(_ptr, _val) do { \
2915 typeof(_ptr) ptr = (_ptr); \
2916 typeof(*ptr) val = (_val); \
2917 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2921 WRITE_ONCE(*ptr, res); \
2924 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2926 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2928 struct sched_avg *sa = &cfs_rq->avg;
2929 int decayed, removed_load = 0, removed_util = 0;
2931 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2932 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2933 sub_positive(&sa->load_avg, r);
2934 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2938 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2939 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2940 sub_positive(&sa->util_avg, r);
2941 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2945 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2946 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2948 #ifndef CONFIG_64BIT
2950 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2953 if (update_freq && (decayed || removed_util))
2954 cfs_rq_util_change(cfs_rq);
2956 return decayed || removed_load;
2959 /* Update task and its cfs_rq load average */
2960 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2962 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2963 u64 now = cfs_rq_clock_task(cfs_rq);
2964 struct rq *rq = rq_of(cfs_rq);
2965 int cpu = cpu_of(rq);
2968 * Track task load average for carrying it to new CPU after migrated, and
2969 * track group sched_entity load average for task_h_load calc in migration
2971 __update_load_avg(now, cpu, &se->avg,
2972 se->on_rq * scale_load_down(se->load.weight),
2973 cfs_rq->curr == se, NULL);
2975 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2976 update_tg_load_avg(cfs_rq, 0);
2979 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2981 if (!sched_feat(ATTACH_AGE_LOAD))
2985 * If we got migrated (either between CPUs or between cgroups) we'll
2986 * have aged the average right before clearing @last_update_time.
2988 if (se->avg.last_update_time) {
2989 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2990 &se->avg, 0, 0, NULL);
2993 * XXX: we could have just aged the entire load away if we've been
2994 * absent from the fair class for too long.
2999 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3000 cfs_rq->avg.load_avg += se->avg.load_avg;
3001 cfs_rq->avg.load_sum += se->avg.load_sum;
3002 cfs_rq->avg.util_avg += se->avg.util_avg;
3003 cfs_rq->avg.util_sum += se->avg.util_sum;
3005 cfs_rq_util_change(cfs_rq);
3008 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3010 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3011 &se->avg, se->on_rq * scale_load_down(se->load.weight),
3012 cfs_rq->curr == se, NULL);
3014 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3015 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3016 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3017 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3019 cfs_rq_util_change(cfs_rq);
3022 /* Add the load generated by se into cfs_rq's load average */
3024 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3026 struct sched_avg *sa = &se->avg;
3027 u64 now = cfs_rq_clock_task(cfs_rq);
3028 int migrated, decayed;
3030 migrated = !sa->last_update_time;
3032 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3033 se->on_rq * scale_load_down(se->load.weight),
3034 cfs_rq->curr == se, NULL);
3037 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3039 cfs_rq->runnable_load_avg += sa->load_avg;
3040 cfs_rq->runnable_load_sum += sa->load_sum;
3043 attach_entity_load_avg(cfs_rq, se);
3045 if (decayed || migrated)
3046 update_tg_load_avg(cfs_rq, 0);
3049 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3051 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3053 update_load_avg(se, 1);
3055 cfs_rq->runnable_load_avg =
3056 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3057 cfs_rq->runnable_load_sum =
3058 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3061 #ifndef CONFIG_64BIT
3062 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3064 u64 last_update_time_copy;
3065 u64 last_update_time;
3068 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3070 last_update_time = cfs_rq->avg.last_update_time;
3071 } while (last_update_time != last_update_time_copy);
3073 return last_update_time;
3076 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3078 return cfs_rq->avg.last_update_time;
3083 * Task first catches up with cfs_rq, and then subtract
3084 * itself from the cfs_rq (task must be off the queue now).
3086 void remove_entity_load_avg(struct sched_entity *se)
3088 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3089 u64 last_update_time;
3092 * Newly created task or never used group entity should not be removed
3093 * from its (source) cfs_rq
3095 if (se->avg.last_update_time == 0)
3098 last_update_time = cfs_rq_last_update_time(cfs_rq);
3100 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3101 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3102 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3105 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3107 return cfs_rq->runnable_load_avg;
3110 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3112 return cfs_rq->avg.load_avg;
3115 static int idle_balance(struct rq *this_rq);
3117 #else /* CONFIG_SMP */
3119 static inline void update_load_avg(struct sched_entity *se, int not_used)
3121 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3122 struct rq *rq = rq_of(cfs_rq);
3124 cpufreq_trigger_update(rq_clock(rq));
3128 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3130 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3131 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3134 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3136 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3138 static inline int idle_balance(struct rq *rq)
3143 #endif /* CONFIG_SMP */
3145 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3147 #ifdef CONFIG_SCHEDSTATS
3148 struct task_struct *tsk = NULL;
3150 if (entity_is_task(se))
3153 if (se->statistics.sleep_start) {
3154 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3159 if (unlikely(delta > se->statistics.sleep_max))
3160 se->statistics.sleep_max = delta;
3162 se->statistics.sleep_start = 0;
3163 se->statistics.sum_sleep_runtime += delta;
3166 account_scheduler_latency(tsk, delta >> 10, 1);
3167 trace_sched_stat_sleep(tsk, delta);
3170 if (se->statistics.block_start) {
3171 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3176 if (unlikely(delta > se->statistics.block_max))
3177 se->statistics.block_max = delta;
3179 se->statistics.block_start = 0;
3180 se->statistics.sum_sleep_runtime += delta;
3183 if (tsk->in_iowait) {
3184 se->statistics.iowait_sum += delta;
3185 se->statistics.iowait_count++;
3186 trace_sched_stat_iowait(tsk, delta);
3189 trace_sched_stat_blocked(tsk, delta);
3192 * Blocking time is in units of nanosecs, so shift by
3193 * 20 to get a milliseconds-range estimation of the
3194 * amount of time that the task spent sleeping:
3196 if (unlikely(prof_on == SLEEP_PROFILING)) {
3197 profile_hits(SLEEP_PROFILING,
3198 (void *)get_wchan(tsk),
3201 account_scheduler_latency(tsk, delta >> 10, 0);
3207 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3209 #ifdef CONFIG_SCHED_DEBUG
3210 s64 d = se->vruntime - cfs_rq->min_vruntime;
3215 if (d > 3*sysctl_sched_latency)
3216 schedstat_inc(cfs_rq, nr_spread_over);
3221 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3223 u64 vruntime = cfs_rq->min_vruntime;
3226 * The 'current' period is already promised to the current tasks,
3227 * however the extra weight of the new task will slow them down a
3228 * little, place the new task so that it fits in the slot that
3229 * stays open at the end.
3231 if (initial && sched_feat(START_DEBIT))
3232 vruntime += sched_vslice(cfs_rq, se);
3234 /* sleeps up to a single latency don't count. */
3236 unsigned long thresh = sysctl_sched_latency;
3239 * Halve their sleep time's effect, to allow
3240 * for a gentler effect of sleepers:
3242 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3248 /* ensure we never gain time by being placed backwards. */
3249 se->vruntime = max_vruntime(se->vruntime, vruntime);
3252 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3254 static inline void check_schedstat_required(void)
3256 #ifdef CONFIG_SCHEDSTATS
3257 if (schedstat_enabled())
3260 /* Force schedstat enabled if a dependent tracepoint is active */
3261 if (trace_sched_stat_wait_enabled() ||
3262 trace_sched_stat_sleep_enabled() ||
3263 trace_sched_stat_iowait_enabled() ||
3264 trace_sched_stat_blocked_enabled() ||
3265 trace_sched_stat_runtime_enabled()) {
3266 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3267 "stat_blocked and stat_runtime require the "
3268 "kernel parameter schedstats=enabled or "
3269 "kernel.sched_schedstats=1\n");
3280 * update_min_vruntime()
3281 * vruntime -= min_vruntime
3285 * update_min_vruntime()
3286 * vruntime += min_vruntime
3288 * this way the vruntime transition between RQs is done when both
3289 * min_vruntime are up-to-date.
3293 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3294 * vruntime -= min_vruntime
3298 * update_min_vruntime()
3299 * vruntime += min_vruntime
3301 * this way we don't have the most up-to-date min_vruntime on the originating
3302 * CPU and an up-to-date min_vruntime on the destination CPU.
3306 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3308 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3309 bool curr = cfs_rq->curr == se;
3312 * If we're the current task, we must renormalise before calling
3316 se->vruntime += cfs_rq->min_vruntime;
3318 update_curr(cfs_rq);
3321 * Otherwise, renormalise after, such that we're placed at the current
3322 * moment in time, instead of some random moment in the past. Being
3323 * placed in the past could significantly boost this task to the
3324 * fairness detriment of existing tasks.
3326 if (renorm && !curr)
3327 se->vruntime += cfs_rq->min_vruntime;
3329 enqueue_entity_load_avg(cfs_rq, se);
3330 account_entity_enqueue(cfs_rq, se);
3331 update_cfs_shares(cfs_rq);
3333 if (flags & ENQUEUE_WAKEUP) {
3334 place_entity(cfs_rq, se, 0);
3335 if (schedstat_enabled())
3336 enqueue_sleeper(cfs_rq, se);
3339 check_schedstat_required();
3340 if (schedstat_enabled()) {
3341 update_stats_enqueue(cfs_rq, se);
3342 check_spread(cfs_rq, se);
3345 __enqueue_entity(cfs_rq, se);
3348 if (cfs_rq->nr_running == 1) {
3349 list_add_leaf_cfs_rq(cfs_rq);
3350 check_enqueue_throttle(cfs_rq);
3354 static void __clear_buddies_last(struct sched_entity *se)
3356 for_each_sched_entity(se) {
3357 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3358 if (cfs_rq->last != se)
3361 cfs_rq->last = NULL;
3365 static void __clear_buddies_next(struct sched_entity *se)
3367 for_each_sched_entity(se) {
3368 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3369 if (cfs_rq->next != se)
3372 cfs_rq->next = NULL;
3376 static void __clear_buddies_skip(struct sched_entity *se)
3378 for_each_sched_entity(se) {
3379 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3380 if (cfs_rq->skip != se)
3383 cfs_rq->skip = NULL;
3387 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3389 if (cfs_rq->last == se)
3390 __clear_buddies_last(se);
3392 if (cfs_rq->next == se)
3393 __clear_buddies_next(se);
3395 if (cfs_rq->skip == se)
3396 __clear_buddies_skip(se);
3399 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3402 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3405 * Update run-time statistics of the 'current'.
3407 update_curr(cfs_rq);
3408 dequeue_entity_load_avg(cfs_rq, se);
3410 if (schedstat_enabled())
3411 update_stats_dequeue(cfs_rq, se, flags);
3413 clear_buddies(cfs_rq, se);
3415 if (se != cfs_rq->curr)
3416 __dequeue_entity(cfs_rq, se);
3418 account_entity_dequeue(cfs_rq, se);
3421 * Normalize the entity after updating the min_vruntime because the
3422 * update can refer to the ->curr item and we need to reflect this
3423 * movement in our normalized position.
3425 if (!(flags & DEQUEUE_SLEEP))
3426 se->vruntime -= cfs_rq->min_vruntime;
3428 /* return excess runtime on last dequeue */
3429 return_cfs_rq_runtime(cfs_rq);
3431 update_min_vruntime(cfs_rq);
3432 update_cfs_shares(cfs_rq);
3436 * Preempt the current task with a newly woken task if needed:
3439 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3441 unsigned long ideal_runtime, delta_exec;
3442 struct sched_entity *se;
3445 ideal_runtime = sched_slice(cfs_rq, curr);
3446 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3447 if (delta_exec > ideal_runtime) {
3448 resched_curr(rq_of(cfs_rq));
3450 * The current task ran long enough, ensure it doesn't get
3451 * re-elected due to buddy favours.
3453 clear_buddies(cfs_rq, curr);
3458 * Ensure that a task that missed wakeup preemption by a
3459 * narrow margin doesn't have to wait for a full slice.
3460 * This also mitigates buddy induced latencies under load.
3462 if (delta_exec < sysctl_sched_min_granularity)
3465 se = __pick_first_entity(cfs_rq);
3466 delta = curr->vruntime - se->vruntime;
3471 if (delta > ideal_runtime)
3472 resched_curr(rq_of(cfs_rq));
3476 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3478 /* 'current' is not kept within the tree. */
3481 * Any task has to be enqueued before it get to execute on
3482 * a CPU. So account for the time it spent waiting on the
3485 if (schedstat_enabled())
3486 update_stats_wait_end(cfs_rq, se);
3487 __dequeue_entity(cfs_rq, se);
3488 update_load_avg(se, 1);
3491 update_stats_curr_start(cfs_rq, se);
3493 #ifdef CONFIG_SCHEDSTATS
3495 * Track our maximum slice length, if the CPU's load is at
3496 * least twice that of our own weight (i.e. dont track it
3497 * when there are only lesser-weight tasks around):
3499 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3500 se->statistics.slice_max = max(se->statistics.slice_max,
3501 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3504 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3508 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3511 * Pick the next process, keeping these things in mind, in this order:
3512 * 1) keep things fair between processes/task groups
3513 * 2) pick the "next" process, since someone really wants that to run
3514 * 3) pick the "last" process, for cache locality
3515 * 4) do not run the "skip" process, if something else is available
3517 static struct sched_entity *
3518 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3520 struct sched_entity *left = __pick_first_entity(cfs_rq);
3521 struct sched_entity *se;
3524 * If curr is set we have to see if its left of the leftmost entity
3525 * still in the tree, provided there was anything in the tree at all.
3527 if (!left || (curr && entity_before(curr, left)))
3530 se = left; /* ideally we run the leftmost entity */
3533 * Avoid running the skip buddy, if running something else can
3534 * be done without getting too unfair.
3536 if (cfs_rq->skip == se) {
3537 struct sched_entity *second;
3540 second = __pick_first_entity(cfs_rq);
3542 second = __pick_next_entity(se);
3543 if (!second || (curr && entity_before(curr, second)))
3547 if (second && wakeup_preempt_entity(second, left) < 1)
3552 * Prefer last buddy, try to return the CPU to a preempted task.
3554 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3558 * Someone really wants this to run. If it's not unfair, run it.
3560 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3563 clear_buddies(cfs_rq, se);
3568 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3570 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3573 * If still on the runqueue then deactivate_task()
3574 * was not called and update_curr() has to be done:
3577 update_curr(cfs_rq);
3579 /* throttle cfs_rqs exceeding runtime */
3580 check_cfs_rq_runtime(cfs_rq);
3582 if (schedstat_enabled()) {
3583 check_spread(cfs_rq, prev);
3585 update_stats_wait_start(cfs_rq, prev);
3589 /* Put 'current' back into the tree. */
3590 __enqueue_entity(cfs_rq, prev);
3591 /* in !on_rq case, update occurred at dequeue */
3592 update_load_avg(prev, 0);
3594 cfs_rq->curr = NULL;
3598 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3601 * Update run-time statistics of the 'current'.
3603 update_curr(cfs_rq);
3606 * Ensure that runnable average is periodically updated.
3608 update_load_avg(curr, 1);
3609 update_cfs_shares(cfs_rq);
3611 #ifdef CONFIG_SCHED_HRTICK
3613 * queued ticks are scheduled to match the slice, so don't bother
3614 * validating it and just reschedule.
3617 resched_curr(rq_of(cfs_rq));
3621 * don't let the period tick interfere with the hrtick preemption
3623 if (!sched_feat(DOUBLE_TICK) &&
3624 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3628 if (cfs_rq->nr_running > 1)
3629 check_preempt_tick(cfs_rq, curr);
3633 /**************************************************
3634 * CFS bandwidth control machinery
3637 #ifdef CONFIG_CFS_BANDWIDTH
3639 #ifdef HAVE_JUMP_LABEL
3640 static struct static_key __cfs_bandwidth_used;
3642 static inline bool cfs_bandwidth_used(void)
3644 return static_key_false(&__cfs_bandwidth_used);
3647 void cfs_bandwidth_usage_inc(void)
3649 static_key_slow_inc(&__cfs_bandwidth_used);
3652 void cfs_bandwidth_usage_dec(void)
3654 static_key_slow_dec(&__cfs_bandwidth_used);
3656 #else /* HAVE_JUMP_LABEL */
3657 static bool cfs_bandwidth_used(void)
3662 void cfs_bandwidth_usage_inc(void) {}
3663 void cfs_bandwidth_usage_dec(void) {}
3664 #endif /* HAVE_JUMP_LABEL */
3667 * default period for cfs group bandwidth.
3668 * default: 0.1s, units: nanoseconds
3670 static inline u64 default_cfs_period(void)
3672 return 100000000ULL;
3675 static inline u64 sched_cfs_bandwidth_slice(void)
3677 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3681 * Replenish runtime according to assigned quota and update expiration time.
3682 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3683 * additional synchronization around rq->lock.
3685 * requires cfs_b->lock
3687 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3691 if (cfs_b->quota == RUNTIME_INF)
3694 now = sched_clock_cpu(smp_processor_id());
3695 cfs_b->runtime = cfs_b->quota;
3696 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3699 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3701 return &tg->cfs_bandwidth;
3704 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3705 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3707 if (unlikely(cfs_rq->throttle_count))
3708 return cfs_rq->throttled_clock_task;
3710 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3713 /* returns 0 on failure to allocate runtime */
3714 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3716 struct task_group *tg = cfs_rq->tg;
3717 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3718 u64 amount = 0, min_amount, expires;
3720 /* note: this is a positive sum as runtime_remaining <= 0 */
3721 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3723 raw_spin_lock(&cfs_b->lock);
3724 if (cfs_b->quota == RUNTIME_INF)
3725 amount = min_amount;
3727 start_cfs_bandwidth(cfs_b);
3729 if (cfs_b->runtime > 0) {
3730 amount = min(cfs_b->runtime, min_amount);
3731 cfs_b->runtime -= amount;
3735 expires = cfs_b->runtime_expires;
3736 raw_spin_unlock(&cfs_b->lock);
3738 cfs_rq->runtime_remaining += amount;
3740 * we may have advanced our local expiration to account for allowed
3741 * spread between our sched_clock and the one on which runtime was
3744 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3745 cfs_rq->runtime_expires = expires;
3747 return cfs_rq->runtime_remaining > 0;
3751 * Note: This depends on the synchronization provided by sched_clock and the
3752 * fact that rq->clock snapshots this value.
3754 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3756 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3758 /* if the deadline is ahead of our clock, nothing to do */
3759 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3762 if (cfs_rq->runtime_remaining < 0)
3766 * If the local deadline has passed we have to consider the
3767 * possibility that our sched_clock is 'fast' and the global deadline
3768 * has not truly expired.
3770 * Fortunately we can check determine whether this the case by checking
3771 * whether the global deadline has advanced. It is valid to compare
3772 * cfs_b->runtime_expires without any locks since we only care about
3773 * exact equality, so a partial write will still work.
3776 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3777 /* extend local deadline, drift is bounded above by 2 ticks */
3778 cfs_rq->runtime_expires += TICK_NSEC;
3780 /* global deadline is ahead, expiration has passed */
3781 cfs_rq->runtime_remaining = 0;
3785 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3787 /* dock delta_exec before expiring quota (as it could span periods) */
3788 cfs_rq->runtime_remaining -= delta_exec;
3789 expire_cfs_rq_runtime(cfs_rq);
3791 if (likely(cfs_rq->runtime_remaining > 0))
3795 * if we're unable to extend our runtime we resched so that the active
3796 * hierarchy can be throttled
3798 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3799 resched_curr(rq_of(cfs_rq));
3802 static __always_inline
3803 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3805 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3808 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3811 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3813 return cfs_bandwidth_used() && cfs_rq->throttled;
3816 /* check whether cfs_rq, or any parent, is throttled */
3817 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3819 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3823 * Ensure that neither of the group entities corresponding to src_cpu or
3824 * dest_cpu are members of a throttled hierarchy when performing group
3825 * load-balance operations.
3827 static inline int throttled_lb_pair(struct task_group *tg,
3828 int src_cpu, int dest_cpu)
3830 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3832 src_cfs_rq = tg->cfs_rq[src_cpu];
3833 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3835 return throttled_hierarchy(src_cfs_rq) ||
3836 throttled_hierarchy(dest_cfs_rq);
3839 /* updated child weight may affect parent so we have to do this bottom up */
3840 static int tg_unthrottle_up(struct task_group *tg, void *data)
3842 struct rq *rq = data;
3843 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3845 cfs_rq->throttle_count--;
3847 if (!cfs_rq->throttle_count) {
3848 /* adjust cfs_rq_clock_task() */
3849 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3850 cfs_rq->throttled_clock_task;
3857 static int tg_throttle_down(struct task_group *tg, void *data)
3859 struct rq *rq = data;
3860 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3862 /* group is entering throttled state, stop time */
3863 if (!cfs_rq->throttle_count)
3864 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3865 cfs_rq->throttle_count++;
3870 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3872 struct rq *rq = rq_of(cfs_rq);
3873 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3874 struct sched_entity *se;
3875 long task_delta, dequeue = 1;
3878 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3880 /* freeze hierarchy runnable averages while throttled */
3882 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3885 task_delta = cfs_rq->h_nr_running;
3886 for_each_sched_entity(se) {
3887 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3888 /* throttled entity or throttle-on-deactivate */
3893 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3894 qcfs_rq->h_nr_running -= task_delta;
3896 if (qcfs_rq->load.weight)
3901 sub_nr_running(rq, task_delta);
3903 cfs_rq->throttled = 1;
3904 cfs_rq->throttled_clock = rq_clock(rq);
3905 raw_spin_lock(&cfs_b->lock);
3906 empty = list_empty(&cfs_b->throttled_cfs_rq);
3909 * Add to the _head_ of the list, so that an already-started
3910 * distribute_cfs_runtime will not see us
3912 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3915 * If we're the first throttled task, make sure the bandwidth
3919 start_cfs_bandwidth(cfs_b);
3921 raw_spin_unlock(&cfs_b->lock);
3924 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3926 struct rq *rq = rq_of(cfs_rq);
3927 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3928 struct sched_entity *se;
3932 se = cfs_rq->tg->se[cpu_of(rq)];
3934 cfs_rq->throttled = 0;
3936 update_rq_clock(rq);
3938 raw_spin_lock(&cfs_b->lock);
3939 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3940 list_del_rcu(&cfs_rq->throttled_list);
3941 raw_spin_unlock(&cfs_b->lock);
3943 /* update hierarchical throttle state */
3944 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3946 if (!cfs_rq->load.weight)
3949 task_delta = cfs_rq->h_nr_running;
3950 for_each_sched_entity(se) {
3954 cfs_rq = cfs_rq_of(se);
3956 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3957 cfs_rq->h_nr_running += task_delta;
3959 if (cfs_rq_throttled(cfs_rq))
3964 add_nr_running(rq, task_delta);
3966 /* determine whether we need to wake up potentially idle cpu */
3967 if (rq->curr == rq->idle && rq->cfs.nr_running)
3971 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3972 u64 remaining, u64 expires)
3974 struct cfs_rq *cfs_rq;
3976 u64 starting_runtime = remaining;
3979 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3981 struct rq *rq = rq_of(cfs_rq);
3983 raw_spin_lock(&rq->lock);
3984 if (!cfs_rq_throttled(cfs_rq))
3987 runtime = -cfs_rq->runtime_remaining + 1;
3988 if (runtime > remaining)
3989 runtime = remaining;
3990 remaining -= runtime;
3992 cfs_rq->runtime_remaining += runtime;
3993 cfs_rq->runtime_expires = expires;
3995 /* we check whether we're throttled above */
3996 if (cfs_rq->runtime_remaining > 0)
3997 unthrottle_cfs_rq(cfs_rq);
4000 raw_spin_unlock(&rq->lock);
4007 return starting_runtime - remaining;
4011 * Responsible for refilling a task_group's bandwidth and unthrottling its
4012 * cfs_rqs as appropriate. If there has been no activity within the last
4013 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4014 * used to track this state.
4016 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4018 u64 runtime, runtime_expires;
4021 /* no need to continue the timer with no bandwidth constraint */
4022 if (cfs_b->quota == RUNTIME_INF)
4023 goto out_deactivate;
4025 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4026 cfs_b->nr_periods += overrun;
4029 * idle depends on !throttled (for the case of a large deficit), and if
4030 * we're going inactive then everything else can be deferred
4032 if (cfs_b->idle && !throttled)
4033 goto out_deactivate;
4035 __refill_cfs_bandwidth_runtime(cfs_b);
4038 /* mark as potentially idle for the upcoming period */
4043 /* account preceding periods in which throttling occurred */
4044 cfs_b->nr_throttled += overrun;
4046 runtime_expires = cfs_b->runtime_expires;
4049 * This check is repeated as we are holding onto the new bandwidth while
4050 * we unthrottle. This can potentially race with an unthrottled group
4051 * trying to acquire new bandwidth from the global pool. This can result
4052 * in us over-using our runtime if it is all used during this loop, but
4053 * only by limited amounts in that extreme case.
4055 while (throttled && cfs_b->runtime > 0) {
4056 runtime = cfs_b->runtime;
4057 raw_spin_unlock(&cfs_b->lock);
4058 /* we can't nest cfs_b->lock while distributing bandwidth */
4059 runtime = distribute_cfs_runtime(cfs_b, runtime,
4061 raw_spin_lock(&cfs_b->lock);
4063 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4065 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4069 * While we are ensured activity in the period following an
4070 * unthrottle, this also covers the case in which the new bandwidth is
4071 * insufficient to cover the existing bandwidth deficit. (Forcing the
4072 * timer to remain active while there are any throttled entities.)
4082 /* a cfs_rq won't donate quota below this amount */
4083 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4084 /* minimum remaining period time to redistribute slack quota */
4085 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4086 /* how long we wait to gather additional slack before distributing */
4087 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4090 * Are we near the end of the current quota period?
4092 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4093 * hrtimer base being cleared by hrtimer_start. In the case of
4094 * migrate_hrtimers, base is never cleared, so we are fine.
4096 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4098 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4101 /* if the call-back is running a quota refresh is already occurring */
4102 if (hrtimer_callback_running(refresh_timer))
4105 /* is a quota refresh about to occur? */
4106 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4107 if (remaining < min_expire)
4113 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4115 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4117 /* if there's a quota refresh soon don't bother with slack */
4118 if (runtime_refresh_within(cfs_b, min_left))
4121 hrtimer_start(&cfs_b->slack_timer,
4122 ns_to_ktime(cfs_bandwidth_slack_period),
4126 /* we know any runtime found here is valid as update_curr() precedes return */
4127 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4129 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4130 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4132 if (slack_runtime <= 0)
4135 raw_spin_lock(&cfs_b->lock);
4136 if (cfs_b->quota != RUNTIME_INF &&
4137 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4138 cfs_b->runtime += slack_runtime;
4140 /* we are under rq->lock, defer unthrottling using a timer */
4141 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4142 !list_empty(&cfs_b->throttled_cfs_rq))
4143 start_cfs_slack_bandwidth(cfs_b);
4145 raw_spin_unlock(&cfs_b->lock);
4147 /* even if it's not valid for return we don't want to try again */
4148 cfs_rq->runtime_remaining -= slack_runtime;
4151 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4153 if (!cfs_bandwidth_used())
4156 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4159 __return_cfs_rq_runtime(cfs_rq);
4163 * This is done with a timer (instead of inline with bandwidth return) since
4164 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4166 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4168 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4171 /* confirm we're still not at a refresh boundary */
4172 raw_spin_lock(&cfs_b->lock);
4173 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4174 raw_spin_unlock(&cfs_b->lock);
4178 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4179 runtime = cfs_b->runtime;
4181 expires = cfs_b->runtime_expires;
4182 raw_spin_unlock(&cfs_b->lock);
4187 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4189 raw_spin_lock(&cfs_b->lock);
4190 if (expires == cfs_b->runtime_expires)
4191 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4192 raw_spin_unlock(&cfs_b->lock);
4196 * When a group wakes up we want to make sure that its quota is not already
4197 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4198 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4200 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4202 if (!cfs_bandwidth_used())
4205 /* Synchronize hierarchical throttle counter: */
4206 if (unlikely(!cfs_rq->throttle_uptodate)) {
4207 struct rq *rq = rq_of(cfs_rq);
4208 struct cfs_rq *pcfs_rq;
4209 struct task_group *tg;
4211 cfs_rq->throttle_uptodate = 1;
4213 /* Get closest up-to-date node, because leaves go first: */
4214 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4215 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4216 if (pcfs_rq->throttle_uptodate)
4220 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4221 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4225 /* an active group must be handled by the update_curr()->put() path */
4226 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4229 /* ensure the group is not already throttled */
4230 if (cfs_rq_throttled(cfs_rq))
4233 /* update runtime allocation */
4234 account_cfs_rq_runtime(cfs_rq, 0);
4235 if (cfs_rq->runtime_remaining <= 0)
4236 throttle_cfs_rq(cfs_rq);
4239 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4240 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4242 if (!cfs_bandwidth_used())
4245 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4249 * it's possible for a throttled entity to be forced into a running
4250 * state (e.g. set_curr_task), in this case we're finished.
4252 if (cfs_rq_throttled(cfs_rq))
4255 throttle_cfs_rq(cfs_rq);
4259 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4261 struct cfs_bandwidth *cfs_b =
4262 container_of(timer, struct cfs_bandwidth, slack_timer);
4264 do_sched_cfs_slack_timer(cfs_b);
4266 return HRTIMER_NORESTART;
4269 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4271 struct cfs_bandwidth *cfs_b =
4272 container_of(timer, struct cfs_bandwidth, period_timer);
4276 raw_spin_lock(&cfs_b->lock);
4278 overrun = hrtimer_forward_now(timer, cfs_b->period);
4282 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4285 cfs_b->period_active = 0;
4286 raw_spin_unlock(&cfs_b->lock);
4288 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4291 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4293 raw_spin_lock_init(&cfs_b->lock);
4295 cfs_b->quota = RUNTIME_INF;
4296 cfs_b->period = ns_to_ktime(default_cfs_period());
4298 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4299 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4300 cfs_b->period_timer.function = sched_cfs_period_timer;
4301 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4302 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4305 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4307 cfs_rq->runtime_enabled = 0;
4308 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4311 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4313 lockdep_assert_held(&cfs_b->lock);
4315 if (!cfs_b->period_active) {
4316 cfs_b->period_active = 1;
4317 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4318 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4322 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4324 /* init_cfs_bandwidth() was not called */
4325 if (!cfs_b->throttled_cfs_rq.next)
4328 hrtimer_cancel(&cfs_b->period_timer);
4329 hrtimer_cancel(&cfs_b->slack_timer);
4332 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4334 struct cfs_rq *cfs_rq;
4336 for_each_leaf_cfs_rq(rq, cfs_rq) {
4337 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4339 raw_spin_lock(&cfs_b->lock);
4340 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4341 raw_spin_unlock(&cfs_b->lock);
4345 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4347 struct cfs_rq *cfs_rq;
4349 for_each_leaf_cfs_rq(rq, cfs_rq) {
4350 if (!cfs_rq->runtime_enabled)
4354 * clock_task is not advancing so we just need to make sure
4355 * there's some valid quota amount
4357 cfs_rq->runtime_remaining = 1;
4359 * Offline rq is schedulable till cpu is completely disabled
4360 * in take_cpu_down(), so we prevent new cfs throttling here.
4362 cfs_rq->runtime_enabled = 0;
4364 if (cfs_rq_throttled(cfs_rq))
4365 unthrottle_cfs_rq(cfs_rq);
4369 #else /* CONFIG_CFS_BANDWIDTH */
4370 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4372 return rq_clock_task(rq_of(cfs_rq));
4375 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4376 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4377 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4378 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4380 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4385 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4390 static inline int throttled_lb_pair(struct task_group *tg,
4391 int src_cpu, int dest_cpu)
4396 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4398 #ifdef CONFIG_FAIR_GROUP_SCHED
4399 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4402 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4406 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4407 static inline void update_runtime_enabled(struct rq *rq) {}
4408 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4410 #endif /* CONFIG_CFS_BANDWIDTH */
4412 /**************************************************
4413 * CFS operations on tasks:
4416 #ifdef CONFIG_SCHED_HRTICK
4417 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4419 struct sched_entity *se = &p->se;
4420 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4422 WARN_ON(task_rq(p) != rq);
4424 if (cfs_rq->nr_running > 1) {
4425 u64 slice = sched_slice(cfs_rq, se);
4426 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4427 s64 delta = slice - ran;
4434 hrtick_start(rq, delta);
4439 * called from enqueue/dequeue and updates the hrtick when the
4440 * current task is from our class and nr_running is low enough
4443 static void hrtick_update(struct rq *rq)
4445 struct task_struct *curr = rq->curr;
4447 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4450 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4451 hrtick_start_fair(rq, curr);
4453 #else /* !CONFIG_SCHED_HRTICK */
4455 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4459 static inline void hrtick_update(struct rq *rq)
4465 * The enqueue_task method is called before nr_running is
4466 * increased. Here we update the fair scheduling stats and
4467 * then put the task into the rbtree:
4470 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4472 struct cfs_rq *cfs_rq;
4473 struct sched_entity *se = &p->se;
4475 for_each_sched_entity(se) {
4478 cfs_rq = cfs_rq_of(se);
4479 enqueue_entity(cfs_rq, se, flags);
4482 * end evaluation on encountering a throttled cfs_rq
4484 * note: in the case of encountering a throttled cfs_rq we will
4485 * post the final h_nr_running increment below.
4487 if (cfs_rq_throttled(cfs_rq))
4489 cfs_rq->h_nr_running++;
4491 flags = ENQUEUE_WAKEUP;
4494 for_each_sched_entity(se) {
4495 cfs_rq = cfs_rq_of(se);
4496 cfs_rq->h_nr_running++;
4498 if (cfs_rq_throttled(cfs_rq))
4501 update_load_avg(se, 1);
4502 update_cfs_shares(cfs_rq);
4506 add_nr_running(rq, 1);
4511 static void set_next_buddy(struct sched_entity *se);
4514 * The dequeue_task method is called before nr_running is
4515 * decreased. We remove the task from the rbtree and
4516 * update the fair scheduling stats:
4518 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4520 struct cfs_rq *cfs_rq;
4521 struct sched_entity *se = &p->se;
4522 int task_sleep = flags & DEQUEUE_SLEEP;
4524 for_each_sched_entity(se) {
4525 cfs_rq = cfs_rq_of(se);
4526 dequeue_entity(cfs_rq, se, flags);
4529 * end evaluation on encountering a throttled cfs_rq
4531 * note: in the case of encountering a throttled cfs_rq we will
4532 * post the final h_nr_running decrement below.
4534 if (cfs_rq_throttled(cfs_rq))
4536 cfs_rq->h_nr_running--;
4538 /* Don't dequeue parent if it has other entities besides us */
4539 if (cfs_rq->load.weight) {
4540 /* Avoid re-evaluating load for this entity: */
4541 se = parent_entity(se);
4543 * Bias pick_next to pick a task from this cfs_rq, as
4544 * p is sleeping when it is within its sched_slice.
4546 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4550 flags |= DEQUEUE_SLEEP;
4553 for_each_sched_entity(se) {
4554 cfs_rq = cfs_rq_of(se);
4555 cfs_rq->h_nr_running--;
4557 if (cfs_rq_throttled(cfs_rq))
4560 update_load_avg(se, 1);
4561 update_cfs_shares(cfs_rq);
4565 sub_nr_running(rq, 1);
4571 #ifdef CONFIG_NO_HZ_COMMON
4573 * per rq 'load' arrray crap; XXX kill this.
4577 * The exact cpuload calculated at every tick would be:
4579 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4581 * If a cpu misses updates for n ticks (as it was idle) and update gets
4582 * called on the n+1-th tick when cpu may be busy, then we have:
4584 * load_n = (1 - 1/2^i)^n * load_0
4585 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
4587 * decay_load_missed() below does efficient calculation of
4589 * load' = (1 - 1/2^i)^n * load
4591 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4592 * This allows us to precompute the above in said factors, thereby allowing the
4593 * reduction of an arbitrary n in O(log_2 n) steps. (See also
4594 * fixed_power_int())
4596 * The calculation is approximated on a 128 point scale.
4598 #define DEGRADE_SHIFT 7
4600 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4601 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4602 { 0, 0, 0, 0, 0, 0, 0, 0 },
4603 { 64, 32, 8, 0, 0, 0, 0, 0 },
4604 { 96, 72, 40, 12, 1, 0, 0, 0 },
4605 { 112, 98, 75, 43, 15, 1, 0, 0 },
4606 { 120, 112, 98, 76, 45, 16, 2, 0 }
4610 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4611 * would be when CPU is idle and so we just decay the old load without
4612 * adding any new load.
4614 static unsigned long
4615 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4619 if (!missed_updates)
4622 if (missed_updates >= degrade_zero_ticks[idx])
4626 return load >> missed_updates;
4628 while (missed_updates) {
4629 if (missed_updates % 2)
4630 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4632 missed_updates >>= 1;
4637 #endif /* CONFIG_NO_HZ_COMMON */
4640 * __cpu_load_update - update the rq->cpu_load[] statistics
4641 * @this_rq: The rq to update statistics for
4642 * @this_load: The current load
4643 * @pending_updates: The number of missed updates
4645 * Update rq->cpu_load[] statistics. This function is usually called every
4646 * scheduler tick (TICK_NSEC).
4648 * This function computes a decaying average:
4650 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4652 * Because of NOHZ it might not get called on every tick which gives need for
4653 * the @pending_updates argument.
4655 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4656 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4657 * = A * (A * load[i]_n-2 + B) + B
4658 * = A * (A * (A * load[i]_n-3 + B) + B) + B
4659 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4660 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4661 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4662 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4664 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4665 * any change in load would have resulted in the tick being turned back on.
4667 * For regular NOHZ, this reduces to:
4669 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
4671 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4674 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4675 unsigned long pending_updates)
4677 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4680 this_rq->nr_load_updates++;
4682 /* Update our load: */
4683 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4684 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4685 unsigned long old_load, new_load;
4687 /* scale is effectively 1 << i now, and >> i divides by scale */
4689 old_load = this_rq->cpu_load[i];
4690 #ifdef CONFIG_NO_HZ_COMMON
4691 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4692 if (tickless_load) {
4693 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4695 * old_load can never be a negative value because a
4696 * decayed tickless_load cannot be greater than the
4697 * original tickless_load.
4699 old_load += tickless_load;
4702 new_load = this_load;
4704 * Round up the averaging division if load is increasing. This
4705 * prevents us from getting stuck on 9 if the load is 10, for
4708 if (new_load > old_load)
4709 new_load += scale - 1;
4711 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4714 sched_avg_update(this_rq);
4717 /* Used instead of source_load when we know the type == 0 */
4718 static unsigned long weighted_cpuload(const int cpu)
4720 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4723 #ifdef CONFIG_NO_HZ_COMMON
4725 * There is no sane way to deal with nohz on smp when using jiffies because the
4726 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4727 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4729 * Therefore we need to avoid the delta approach from the regular tick when
4730 * possible since that would seriously skew the load calculation. This is why we
4731 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4732 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4733 * loop exit, nohz_idle_balance, nohz full exit...)
4735 * This means we might still be one tick off for nohz periods.
4738 static void cpu_load_update_nohz(struct rq *this_rq,
4739 unsigned long curr_jiffies,
4742 unsigned long pending_updates;
4744 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4745 if (pending_updates) {
4746 this_rq->last_load_update_tick = curr_jiffies;
4748 * In the regular NOHZ case, we were idle, this means load 0.
4749 * In the NOHZ_FULL case, we were non-idle, we should consider
4750 * its weighted load.
4752 cpu_load_update(this_rq, load, pending_updates);
4757 * Called from nohz_idle_balance() to update the load ratings before doing the
4760 static void cpu_load_update_idle(struct rq *this_rq)
4763 * bail if there's load or we're actually up-to-date.
4765 if (weighted_cpuload(cpu_of(this_rq)))
4768 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4772 * Record CPU load on nohz entry so we know the tickless load to account
4773 * on nohz exit. cpu_load[0] happens then to be updated more frequently
4774 * than other cpu_load[idx] but it should be fine as cpu_load readers
4775 * shouldn't rely into synchronized cpu_load[*] updates.
4777 void cpu_load_update_nohz_start(void)
4779 struct rq *this_rq = this_rq();
4782 * This is all lockless but should be fine. If weighted_cpuload changes
4783 * concurrently we'll exit nohz. And cpu_load write can race with
4784 * cpu_load_update_idle() but both updater would be writing the same.
4786 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4790 * Account the tickless load in the end of a nohz frame.
4792 void cpu_load_update_nohz_stop(void)
4794 unsigned long curr_jiffies = READ_ONCE(jiffies);
4795 struct rq *this_rq = this_rq();
4798 if (curr_jiffies == this_rq->last_load_update_tick)
4801 load = weighted_cpuload(cpu_of(this_rq));
4802 raw_spin_lock(&this_rq->lock);
4803 update_rq_clock(this_rq);
4804 cpu_load_update_nohz(this_rq, curr_jiffies, load);
4805 raw_spin_unlock(&this_rq->lock);
4807 #else /* !CONFIG_NO_HZ_COMMON */
4808 static inline void cpu_load_update_nohz(struct rq *this_rq,
4809 unsigned long curr_jiffies,
4810 unsigned long load) { }
4811 #endif /* CONFIG_NO_HZ_COMMON */
4813 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4815 #ifdef CONFIG_NO_HZ_COMMON
4816 /* See the mess around cpu_load_update_nohz(). */
4817 this_rq->last_load_update_tick = READ_ONCE(jiffies);
4819 cpu_load_update(this_rq, load, 1);
4823 * Called from scheduler_tick()
4825 void cpu_load_update_active(struct rq *this_rq)
4827 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4829 if (tick_nohz_tick_stopped())
4830 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4832 cpu_load_update_periodic(this_rq, load);
4836 * Return a low guess at the load of a migration-source cpu weighted
4837 * according to the scheduling class and "nice" value.
4839 * We want to under-estimate the load of migration sources, to
4840 * balance conservatively.
4842 static unsigned long source_load(int cpu, int type)
4844 struct rq *rq = cpu_rq(cpu);
4845 unsigned long total = weighted_cpuload(cpu);
4847 if (type == 0 || !sched_feat(LB_BIAS))
4850 return min(rq->cpu_load[type-1], total);
4854 * Return a high guess at the load of a migration-target cpu weighted
4855 * according to the scheduling class and "nice" value.
4857 static unsigned long target_load(int cpu, int type)
4859 struct rq *rq = cpu_rq(cpu);
4860 unsigned long total = weighted_cpuload(cpu);
4862 if (type == 0 || !sched_feat(LB_BIAS))
4865 return max(rq->cpu_load[type-1], total);
4868 static unsigned long capacity_of(int cpu)
4870 return cpu_rq(cpu)->cpu_capacity;
4873 static unsigned long capacity_orig_of(int cpu)
4875 return cpu_rq(cpu)->cpu_capacity_orig;
4878 static unsigned long cpu_avg_load_per_task(int cpu)
4880 struct rq *rq = cpu_rq(cpu);
4881 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4882 unsigned long load_avg = weighted_cpuload(cpu);
4885 return load_avg / nr_running;
4890 #ifdef CONFIG_FAIR_GROUP_SCHED
4892 * effective_load() calculates the load change as seen from the root_task_group
4894 * Adding load to a group doesn't make a group heavier, but can cause movement
4895 * of group shares between cpus. Assuming the shares were perfectly aligned one
4896 * can calculate the shift in shares.
4898 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4899 * on this @cpu and results in a total addition (subtraction) of @wg to the
4900 * total group weight.
4902 * Given a runqueue weight distribution (rw_i) we can compute a shares
4903 * distribution (s_i) using:
4905 * s_i = rw_i / \Sum rw_j (1)
4907 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4908 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4909 * shares distribution (s_i):
4911 * rw_i = { 2, 4, 1, 0 }
4912 * s_i = { 2/7, 4/7, 1/7, 0 }
4914 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4915 * task used to run on and the CPU the waker is running on), we need to
4916 * compute the effect of waking a task on either CPU and, in case of a sync
4917 * wakeup, compute the effect of the current task going to sleep.
4919 * So for a change of @wl to the local @cpu with an overall group weight change
4920 * of @wl we can compute the new shares distribution (s'_i) using:
4922 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4924 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4925 * differences in waking a task to CPU 0. The additional task changes the
4926 * weight and shares distributions like:
4928 * rw'_i = { 3, 4, 1, 0 }
4929 * s'_i = { 3/8, 4/8, 1/8, 0 }
4931 * We can then compute the difference in effective weight by using:
4933 * dw_i = S * (s'_i - s_i) (3)
4935 * Where 'S' is the group weight as seen by its parent.
4937 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4938 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4939 * 4/7) times the weight of the group.
4941 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4943 struct sched_entity *se = tg->se[cpu];
4945 if (!tg->parent) /* the trivial, non-cgroup case */
4948 for_each_sched_entity(se) {
4954 * W = @wg + \Sum rw_j
4956 W = wg + calc_tg_weight(tg, se->my_q);
4961 w = cfs_rq_load_avg(se->my_q) + wl;
4964 * wl = S * s'_i; see (2)
4967 wl = (w * (long)tg->shares) / W;
4972 * Per the above, wl is the new se->load.weight value; since
4973 * those are clipped to [MIN_SHARES, ...) do so now. See
4974 * calc_cfs_shares().
4976 if (wl < MIN_SHARES)
4980 * wl = dw_i = S * (s'_i - s_i); see (3)
4982 wl -= se->avg.load_avg;
4985 * Recursively apply this logic to all parent groups to compute
4986 * the final effective load change on the root group. Since
4987 * only the @tg group gets extra weight, all parent groups can
4988 * only redistribute existing shares. @wl is the shift in shares
4989 * resulting from this level per the above.
4998 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5005 static void record_wakee(struct task_struct *p)
5008 * Only decay a single time; tasks that have less then 1 wakeup per
5009 * jiffy will not have built up many flips.
5011 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5012 current->wakee_flips >>= 1;
5013 current->wakee_flip_decay_ts = jiffies;
5016 if (current->last_wakee != p) {
5017 current->last_wakee = p;
5018 current->wakee_flips++;
5023 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5025 * A waker of many should wake a different task than the one last awakened
5026 * at a frequency roughly N times higher than one of its wakees.
5028 * In order to determine whether we should let the load spread vs consolidating
5029 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5030 * partner, and a factor of lls_size higher frequency in the other.
5032 * With both conditions met, we can be relatively sure that the relationship is
5033 * non-monogamous, with partner count exceeding socket size.
5035 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5036 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5039 static int wake_wide(struct task_struct *p)
5041 unsigned int master = current->wakee_flips;
5042 unsigned int slave = p->wakee_flips;
5043 int factor = this_cpu_read(sd_llc_size);
5046 swap(master, slave);
5047 if (slave < factor || master < slave * factor)
5052 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5054 s64 this_load, load;
5055 s64 this_eff_load, prev_eff_load;
5056 int idx, this_cpu, prev_cpu;
5057 struct task_group *tg;
5058 unsigned long weight;
5062 this_cpu = smp_processor_id();
5063 prev_cpu = task_cpu(p);
5064 load = source_load(prev_cpu, idx);
5065 this_load = target_load(this_cpu, idx);
5068 * If sync wakeup then subtract the (maximum possible)
5069 * effect of the currently running task from the load
5070 * of the current CPU:
5073 tg = task_group(current);
5074 weight = current->se.avg.load_avg;
5076 this_load += effective_load(tg, this_cpu, -weight, -weight);
5077 load += effective_load(tg, prev_cpu, 0, -weight);
5081 weight = p->se.avg.load_avg;
5084 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5085 * due to the sync cause above having dropped this_load to 0, we'll
5086 * always have an imbalance, but there's really nothing you can do
5087 * about that, so that's good too.
5089 * Otherwise check if either cpus are near enough in load to allow this
5090 * task to be woken on this_cpu.
5092 this_eff_load = 100;
5093 this_eff_load *= capacity_of(prev_cpu);
5095 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5096 prev_eff_load *= capacity_of(this_cpu);
5098 if (this_load > 0) {
5099 this_eff_load *= this_load +
5100 effective_load(tg, this_cpu, weight, weight);
5102 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5105 balanced = this_eff_load <= prev_eff_load;
5107 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5112 schedstat_inc(sd, ttwu_move_affine);
5113 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5119 * find_idlest_group finds and returns the least busy CPU group within the
5122 static struct sched_group *
5123 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5124 int this_cpu, int sd_flag)
5126 struct sched_group *idlest = NULL, *group = sd->groups;
5127 unsigned long min_load = ULONG_MAX, this_load = 0;
5128 int load_idx = sd->forkexec_idx;
5129 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5131 if (sd_flag & SD_BALANCE_WAKE)
5132 load_idx = sd->wake_idx;
5135 unsigned long load, avg_load;
5139 /* Skip over this group if it has no CPUs allowed */
5140 if (!cpumask_intersects(sched_group_cpus(group),
5141 tsk_cpus_allowed(p)))
5144 local_group = cpumask_test_cpu(this_cpu,
5145 sched_group_cpus(group));
5147 /* Tally up the load of all CPUs in the group */
5150 for_each_cpu(i, sched_group_cpus(group)) {
5151 /* Bias balancing toward cpus of our domain */
5153 load = source_load(i, load_idx);
5155 load = target_load(i, load_idx);
5160 /* Adjust by relative CPU capacity of the group */
5161 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5164 this_load = avg_load;
5165 } else if (avg_load < min_load) {
5166 min_load = avg_load;
5169 } while (group = group->next, group != sd->groups);
5171 if (!idlest || 100*this_load < imbalance*min_load)
5177 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5180 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5182 unsigned long load, min_load = ULONG_MAX;
5183 unsigned int min_exit_latency = UINT_MAX;
5184 u64 latest_idle_timestamp = 0;
5185 int least_loaded_cpu = this_cpu;
5186 int shallowest_idle_cpu = -1;
5189 /* Traverse only the allowed CPUs */
5190 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5192 struct rq *rq = cpu_rq(i);
5193 struct cpuidle_state *idle = idle_get_state(rq);
5194 if (idle && idle->exit_latency < min_exit_latency) {
5196 * We give priority to a CPU whose idle state
5197 * has the smallest exit latency irrespective
5198 * of any idle timestamp.
5200 min_exit_latency = idle->exit_latency;
5201 latest_idle_timestamp = rq->idle_stamp;
5202 shallowest_idle_cpu = i;
5203 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5204 rq->idle_stamp > latest_idle_timestamp) {
5206 * If equal or no active idle state, then
5207 * the most recently idled CPU might have
5210 latest_idle_timestamp = rq->idle_stamp;
5211 shallowest_idle_cpu = i;
5213 } else if (shallowest_idle_cpu == -1) {
5214 load = weighted_cpuload(i);
5215 if (load < min_load || (load == min_load && i == this_cpu)) {
5217 least_loaded_cpu = i;
5222 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5226 * Try and locate an idle CPU in the sched_domain.
5228 static int select_idle_sibling(struct task_struct *p, int target)
5230 struct sched_domain *sd;
5231 struct sched_group *sg;
5232 int i = task_cpu(p);
5234 if (idle_cpu(target))
5238 * If the prevous cpu is cache affine and idle, don't be stupid.
5240 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5244 * Otherwise, iterate the domains and find an eligible idle cpu.
5246 * A completely idle sched group at higher domains is more
5247 * desirable than an idle group at a lower level, because lower
5248 * domains have smaller groups and usually share hardware
5249 * resources which causes tasks to contend on them, e.g. x86
5250 * hyperthread siblings in the lowest domain (SMT) can contend
5251 * on the shared cpu pipeline.
5253 * However, while we prefer idle groups at higher domains
5254 * finding an idle cpu at the lowest domain is still better than
5255 * returning 'target', which we've already established, isn't
5258 sd = rcu_dereference(per_cpu(sd_llc, target));
5259 for_each_lower_domain(sd) {
5262 if (!cpumask_intersects(sched_group_cpus(sg),
5263 tsk_cpus_allowed(p)))
5266 /* Ensure the entire group is idle */
5267 for_each_cpu(i, sched_group_cpus(sg)) {
5268 if (i == target || !idle_cpu(i))
5273 * It doesn't matter which cpu we pick, the
5274 * whole group is idle.
5276 target = cpumask_first_and(sched_group_cpus(sg),
5277 tsk_cpus_allowed(p));
5281 } while (sg != sd->groups);
5288 * cpu_util returns the amount of capacity of a CPU that is used by CFS
5289 * tasks. The unit of the return value must be the one of capacity so we can
5290 * compare the utilization with the capacity of the CPU that is available for
5291 * CFS task (ie cpu_capacity).
5293 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5294 * recent utilization of currently non-runnable tasks on a CPU. It represents
5295 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5296 * capacity_orig is the cpu_capacity available at the highest frequency
5297 * (arch_scale_freq_capacity()).
5298 * The utilization of a CPU converges towards a sum equal to or less than the
5299 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5300 * the running time on this CPU scaled by capacity_curr.
5302 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5303 * higher than capacity_orig because of unfortunate rounding in
5304 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5305 * the average stabilizes with the new running time. We need to check that the
5306 * utilization stays within the range of [0..capacity_orig] and cap it if
5307 * necessary. Without utilization capping, a group could be seen as overloaded
5308 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5309 * available capacity. We allow utilization to overshoot capacity_curr (but not
5310 * capacity_orig) as it useful for predicting the capacity required after task
5311 * migrations (scheduler-driven DVFS).
5313 static int cpu_util(int cpu)
5315 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5316 unsigned long capacity = capacity_orig_of(cpu);
5318 return (util >= capacity) ? capacity : util;
5322 * select_task_rq_fair: Select target runqueue for the waking task in domains
5323 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5324 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5326 * Balances load by selecting the idlest cpu in the idlest group, or under
5327 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5329 * Returns the target cpu number.
5331 * preempt must be disabled.
5334 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5336 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5337 int cpu = smp_processor_id();
5338 int new_cpu = prev_cpu;
5339 int want_affine = 0;
5340 int sync = wake_flags & WF_SYNC;
5342 if (sd_flag & SD_BALANCE_WAKE) {
5344 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5348 for_each_domain(cpu, tmp) {
5349 if (!(tmp->flags & SD_LOAD_BALANCE))
5353 * If both cpu and prev_cpu are part of this domain,
5354 * cpu is a valid SD_WAKE_AFFINE target.
5356 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5357 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5362 if (tmp->flags & sd_flag)
5364 else if (!want_affine)
5369 sd = NULL; /* Prefer wake_affine over balance flags */
5370 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5375 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5376 new_cpu = select_idle_sibling(p, new_cpu);
5379 struct sched_group *group;
5382 if (!(sd->flags & sd_flag)) {
5387 group = find_idlest_group(sd, p, cpu, sd_flag);
5393 new_cpu = find_idlest_cpu(group, p, cpu);
5394 if (new_cpu == -1 || new_cpu == cpu) {
5395 /* Now try balancing at a lower domain level of cpu */
5400 /* Now try balancing at a lower domain level of new_cpu */
5402 weight = sd->span_weight;
5404 for_each_domain(cpu, tmp) {
5405 if (weight <= tmp->span_weight)
5407 if (tmp->flags & sd_flag)
5410 /* while loop will break here if sd == NULL */
5418 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5419 * cfs_rq_of(p) references at time of call are still valid and identify the
5420 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5422 static void migrate_task_rq_fair(struct task_struct *p)
5425 * As blocked tasks retain absolute vruntime the migration needs to
5426 * deal with this by subtracting the old and adding the new
5427 * min_vruntime -- the latter is done by enqueue_entity() when placing
5428 * the task on the new runqueue.
5430 if (p->state == TASK_WAKING) {
5431 struct sched_entity *se = &p->se;
5432 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5435 #ifndef CONFIG_64BIT
5436 u64 min_vruntime_copy;
5439 min_vruntime_copy = cfs_rq->min_vruntime_copy;
5441 min_vruntime = cfs_rq->min_vruntime;
5442 } while (min_vruntime != min_vruntime_copy);
5444 min_vruntime = cfs_rq->min_vruntime;
5447 se->vruntime -= min_vruntime;
5451 * We are supposed to update the task to "current" time, then its up to date
5452 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5453 * what current time is, so simply throw away the out-of-date time. This
5454 * will result in the wakee task is less decayed, but giving the wakee more
5455 * load sounds not bad.
5457 remove_entity_load_avg(&p->se);
5459 /* Tell new CPU we are migrated */
5460 p->se.avg.last_update_time = 0;
5462 /* We have migrated, no longer consider this task hot */
5463 p->se.exec_start = 0;
5466 static void task_dead_fair(struct task_struct *p)
5468 remove_entity_load_avg(&p->se);
5470 #endif /* CONFIG_SMP */
5472 static unsigned long
5473 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5475 unsigned long gran = sysctl_sched_wakeup_granularity;
5478 * Since its curr running now, convert the gran from real-time
5479 * to virtual-time in his units.
5481 * By using 'se' instead of 'curr' we penalize light tasks, so
5482 * they get preempted easier. That is, if 'se' < 'curr' then
5483 * the resulting gran will be larger, therefore penalizing the
5484 * lighter, if otoh 'se' > 'curr' then the resulting gran will
5485 * be smaller, again penalizing the lighter task.
5487 * This is especially important for buddies when the leftmost
5488 * task is higher priority than the buddy.
5490 return calc_delta_fair(gran, se);
5494 * Should 'se' preempt 'curr'.
5508 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5510 s64 gran, vdiff = curr->vruntime - se->vruntime;
5515 gran = wakeup_gran(curr, se);
5522 static void set_last_buddy(struct sched_entity *se)
5524 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5527 for_each_sched_entity(se)
5528 cfs_rq_of(se)->last = se;
5531 static void set_next_buddy(struct sched_entity *se)
5533 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5536 for_each_sched_entity(se)
5537 cfs_rq_of(se)->next = se;
5540 static void set_skip_buddy(struct sched_entity *se)
5542 for_each_sched_entity(se)
5543 cfs_rq_of(se)->skip = se;
5547 * Preempt the current task with a newly woken task if needed:
5549 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5551 struct task_struct *curr = rq->curr;
5552 struct sched_entity *se = &curr->se, *pse = &p->se;
5553 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5554 int scale = cfs_rq->nr_running >= sched_nr_latency;
5555 int next_buddy_marked = 0;
5557 if (unlikely(se == pse))
5561 * This is possible from callers such as attach_tasks(), in which we
5562 * unconditionally check_prempt_curr() after an enqueue (which may have
5563 * lead to a throttle). This both saves work and prevents false
5564 * next-buddy nomination below.
5566 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5569 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5570 set_next_buddy(pse);
5571 next_buddy_marked = 1;
5575 * We can come here with TIF_NEED_RESCHED already set from new task
5578 * Note: this also catches the edge-case of curr being in a throttled
5579 * group (e.g. via set_curr_task), since update_curr() (in the
5580 * enqueue of curr) will have resulted in resched being set. This
5581 * prevents us from potentially nominating it as a false LAST_BUDDY
5584 if (test_tsk_need_resched(curr))
5587 /* Idle tasks are by definition preempted by non-idle tasks. */
5588 if (unlikely(curr->policy == SCHED_IDLE) &&
5589 likely(p->policy != SCHED_IDLE))
5593 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5594 * is driven by the tick):
5596 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5599 find_matching_se(&se, &pse);
5600 update_curr(cfs_rq_of(se));
5602 if (wakeup_preempt_entity(se, pse) == 1) {
5604 * Bias pick_next to pick the sched entity that is
5605 * triggering this preemption.
5607 if (!next_buddy_marked)
5608 set_next_buddy(pse);
5617 * Only set the backward buddy when the current task is still
5618 * on the rq. This can happen when a wakeup gets interleaved
5619 * with schedule on the ->pre_schedule() or idle_balance()
5620 * point, either of which can * drop the rq lock.
5622 * Also, during early boot the idle thread is in the fair class,
5623 * for obvious reasons its a bad idea to schedule back to it.
5625 if (unlikely(!se->on_rq || curr == rq->idle))
5628 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5632 static struct task_struct *
5633 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5635 struct cfs_rq *cfs_rq = &rq->cfs;
5636 struct sched_entity *se;
5637 struct task_struct *p;
5641 #ifdef CONFIG_FAIR_GROUP_SCHED
5642 if (!cfs_rq->nr_running)
5645 if (prev->sched_class != &fair_sched_class)
5649 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5650 * likely that a next task is from the same cgroup as the current.
5652 * Therefore attempt to avoid putting and setting the entire cgroup
5653 * hierarchy, only change the part that actually changes.
5657 struct sched_entity *curr = cfs_rq->curr;
5660 * Since we got here without doing put_prev_entity() we also
5661 * have to consider cfs_rq->curr. If it is still a runnable
5662 * entity, update_curr() will update its vruntime, otherwise
5663 * forget we've ever seen it.
5667 update_curr(cfs_rq);
5672 * This call to check_cfs_rq_runtime() will do the
5673 * throttle and dequeue its entity in the parent(s).
5674 * Therefore the 'simple' nr_running test will indeed
5677 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5681 se = pick_next_entity(cfs_rq, curr);
5682 cfs_rq = group_cfs_rq(se);
5688 * Since we haven't yet done put_prev_entity and if the selected task
5689 * is a different task than we started out with, try and touch the
5690 * least amount of cfs_rqs.
5693 struct sched_entity *pse = &prev->se;
5695 while (!(cfs_rq = is_same_group(se, pse))) {
5696 int se_depth = se->depth;
5697 int pse_depth = pse->depth;
5699 if (se_depth <= pse_depth) {
5700 put_prev_entity(cfs_rq_of(pse), pse);
5701 pse = parent_entity(pse);
5703 if (se_depth >= pse_depth) {
5704 set_next_entity(cfs_rq_of(se), se);
5705 se = parent_entity(se);
5709 put_prev_entity(cfs_rq, pse);
5710 set_next_entity(cfs_rq, se);
5713 if (hrtick_enabled(rq))
5714 hrtick_start_fair(rq, p);
5721 if (!cfs_rq->nr_running)
5724 put_prev_task(rq, prev);
5727 se = pick_next_entity(cfs_rq, NULL);
5728 set_next_entity(cfs_rq, se);
5729 cfs_rq = group_cfs_rq(se);
5734 if (hrtick_enabled(rq))
5735 hrtick_start_fair(rq, p);
5741 * This is OK, because current is on_cpu, which avoids it being picked
5742 * for load-balance and preemption/IRQs are still disabled avoiding
5743 * further scheduler activity on it and we're being very careful to
5744 * re-start the picking loop.
5746 lockdep_unpin_lock(&rq->lock, cookie);
5747 new_tasks = idle_balance(rq);
5748 lockdep_repin_lock(&rq->lock, cookie);
5750 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5751 * possible for any higher priority task to appear. In that case we
5752 * must re-start the pick_next_entity() loop.
5764 * Account for a descheduled task:
5766 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5768 struct sched_entity *se = &prev->se;
5769 struct cfs_rq *cfs_rq;
5771 for_each_sched_entity(se) {
5772 cfs_rq = cfs_rq_of(se);
5773 put_prev_entity(cfs_rq, se);
5778 * sched_yield() is very simple
5780 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5782 static void yield_task_fair(struct rq *rq)
5784 struct task_struct *curr = rq->curr;
5785 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5786 struct sched_entity *se = &curr->se;
5789 * Are we the only task in the tree?
5791 if (unlikely(rq->nr_running == 1))
5794 clear_buddies(cfs_rq, se);
5796 if (curr->policy != SCHED_BATCH) {
5797 update_rq_clock(rq);
5799 * Update run-time statistics of the 'current'.
5801 update_curr(cfs_rq);
5803 * Tell update_rq_clock() that we've just updated,
5804 * so we don't do microscopic update in schedule()
5805 * and double the fastpath cost.
5807 rq_clock_skip_update(rq, true);
5813 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5815 struct sched_entity *se = &p->se;
5817 /* throttled hierarchies are not runnable */
5818 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5821 /* Tell the scheduler that we'd really like pse to run next. */
5824 yield_task_fair(rq);
5830 /**************************************************
5831 * Fair scheduling class load-balancing methods.
5835 * The purpose of load-balancing is to achieve the same basic fairness the
5836 * per-cpu scheduler provides, namely provide a proportional amount of compute
5837 * time to each task. This is expressed in the following equation:
5839 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5841 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5842 * W_i,0 is defined as:
5844 * W_i,0 = \Sum_j w_i,j (2)
5846 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5847 * is derived from the nice value as per sched_prio_to_weight[].
5849 * The weight average is an exponential decay average of the instantaneous
5852 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5854 * C_i is the compute capacity of cpu i, typically it is the
5855 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5856 * can also include other factors [XXX].
5858 * To achieve this balance we define a measure of imbalance which follows
5859 * directly from (1):
5861 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5863 * We them move tasks around to minimize the imbalance. In the continuous
5864 * function space it is obvious this converges, in the discrete case we get
5865 * a few fun cases generally called infeasible weight scenarios.
5868 * - infeasible weights;
5869 * - local vs global optima in the discrete case. ]
5874 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5875 * for all i,j solution, we create a tree of cpus that follows the hardware
5876 * topology where each level pairs two lower groups (or better). This results
5877 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5878 * tree to only the first of the previous level and we decrease the frequency
5879 * of load-balance at each level inv. proportional to the number of cpus in
5885 * \Sum { --- * --- * 2^i } = O(n) (5)
5887 * `- size of each group
5888 * | | `- number of cpus doing load-balance
5890 * `- sum over all levels
5892 * Coupled with a limit on how many tasks we can migrate every balance pass,
5893 * this makes (5) the runtime complexity of the balancer.
5895 * An important property here is that each CPU is still (indirectly) connected
5896 * to every other cpu in at most O(log n) steps:
5898 * The adjacency matrix of the resulting graph is given by:
5901 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5904 * And you'll find that:
5906 * A^(log_2 n)_i,j != 0 for all i,j (7)
5908 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5909 * The task movement gives a factor of O(m), giving a convergence complexity
5912 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5917 * In order to avoid CPUs going idle while there's still work to do, new idle
5918 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5919 * tree itself instead of relying on other CPUs to bring it work.
5921 * This adds some complexity to both (5) and (8) but it reduces the total idle
5929 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5932 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5937 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5939 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5941 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5944 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5945 * rewrite all of this once again.]
5948 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5950 enum fbq_type { regular, remote, all };
5952 #define LBF_ALL_PINNED 0x01
5953 #define LBF_NEED_BREAK 0x02
5954 #define LBF_DST_PINNED 0x04
5955 #define LBF_SOME_PINNED 0x08
5958 struct sched_domain *sd;
5966 struct cpumask *dst_grpmask;
5968 enum cpu_idle_type idle;
5970 /* The set of CPUs under consideration for load-balancing */
5971 struct cpumask *cpus;
5976 unsigned int loop_break;
5977 unsigned int loop_max;
5979 enum fbq_type fbq_type;
5980 struct list_head tasks;
5984 * Is this task likely cache-hot:
5986 static int task_hot(struct task_struct *p, struct lb_env *env)
5990 lockdep_assert_held(&env->src_rq->lock);
5992 if (p->sched_class != &fair_sched_class)
5995 if (unlikely(p->policy == SCHED_IDLE))
5999 * Buddy candidates are cache hot:
6001 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6002 (&p->se == cfs_rq_of(&p->se)->next ||
6003 &p->se == cfs_rq_of(&p->se)->last))
6006 if (sysctl_sched_migration_cost == -1)
6008 if (sysctl_sched_migration_cost == 0)
6011 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6013 return delta < (s64)sysctl_sched_migration_cost;
6016 #ifdef CONFIG_NUMA_BALANCING
6018 * Returns 1, if task migration degrades locality
6019 * Returns 0, if task migration improves locality i.e migration preferred.
6020 * Returns -1, if task migration is not affected by locality.
6022 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6024 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6025 unsigned long src_faults, dst_faults;
6026 int src_nid, dst_nid;
6028 if (!static_branch_likely(&sched_numa_balancing))
6031 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6034 src_nid = cpu_to_node(env->src_cpu);
6035 dst_nid = cpu_to_node(env->dst_cpu);
6037 if (src_nid == dst_nid)
6040 /* Migrating away from the preferred node is always bad. */
6041 if (src_nid == p->numa_preferred_nid) {
6042 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6048 /* Encourage migration to the preferred node. */
6049 if (dst_nid == p->numa_preferred_nid)
6053 src_faults = group_faults(p, src_nid);
6054 dst_faults = group_faults(p, dst_nid);
6056 src_faults = task_faults(p, src_nid);
6057 dst_faults = task_faults(p, dst_nid);
6060 return dst_faults < src_faults;
6064 static inline int migrate_degrades_locality(struct task_struct *p,
6072 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6075 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6079 lockdep_assert_held(&env->src_rq->lock);
6082 * We do not migrate tasks that are:
6083 * 1) throttled_lb_pair, or
6084 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6085 * 3) running (obviously), or
6086 * 4) are cache-hot on their current CPU.
6088 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6091 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6094 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6096 env->flags |= LBF_SOME_PINNED;
6099 * Remember if this task can be migrated to any other cpu in
6100 * our sched_group. We may want to revisit it if we couldn't
6101 * meet load balance goals by pulling other tasks on src_cpu.
6103 * Also avoid computing new_dst_cpu if we have already computed
6104 * one in current iteration.
6106 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6109 /* Prevent to re-select dst_cpu via env's cpus */
6110 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6111 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6112 env->flags |= LBF_DST_PINNED;
6113 env->new_dst_cpu = cpu;
6121 /* Record that we found atleast one task that could run on dst_cpu */
6122 env->flags &= ~LBF_ALL_PINNED;
6124 if (task_running(env->src_rq, p)) {
6125 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6130 * Aggressive migration if:
6131 * 1) destination numa is preferred
6132 * 2) task is cache cold, or
6133 * 3) too many balance attempts have failed.
6135 tsk_cache_hot = migrate_degrades_locality(p, env);
6136 if (tsk_cache_hot == -1)
6137 tsk_cache_hot = task_hot(p, env);
6139 if (tsk_cache_hot <= 0 ||
6140 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6141 if (tsk_cache_hot == 1) {
6142 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6143 schedstat_inc(p, se.statistics.nr_forced_migrations);
6148 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6153 * detach_task() -- detach the task for the migration specified in env
6155 static void detach_task(struct task_struct *p, struct lb_env *env)
6157 lockdep_assert_held(&env->src_rq->lock);
6159 p->on_rq = TASK_ON_RQ_MIGRATING;
6160 deactivate_task(env->src_rq, p, 0);
6161 set_task_cpu(p, env->dst_cpu);
6165 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6166 * part of active balancing operations within "domain".
6168 * Returns a task if successful and NULL otherwise.
6170 static struct task_struct *detach_one_task(struct lb_env *env)
6172 struct task_struct *p, *n;
6174 lockdep_assert_held(&env->src_rq->lock);
6176 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6177 if (!can_migrate_task(p, env))
6180 detach_task(p, env);
6183 * Right now, this is only the second place where
6184 * lb_gained[env->idle] is updated (other is detach_tasks)
6185 * so we can safely collect stats here rather than
6186 * inside detach_tasks().
6188 schedstat_inc(env->sd, lb_gained[env->idle]);
6194 static const unsigned int sched_nr_migrate_break = 32;
6197 * detach_tasks() -- tries to detach up to imbalance weighted load from
6198 * busiest_rq, as part of a balancing operation within domain "sd".
6200 * Returns number of detached tasks if successful and 0 otherwise.
6202 static int detach_tasks(struct lb_env *env)
6204 struct list_head *tasks = &env->src_rq->cfs_tasks;
6205 struct task_struct *p;
6209 lockdep_assert_held(&env->src_rq->lock);
6211 if (env->imbalance <= 0)
6214 while (!list_empty(tasks)) {
6216 * We don't want to steal all, otherwise we may be treated likewise,
6217 * which could at worst lead to a livelock crash.
6219 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6222 p = list_first_entry(tasks, struct task_struct, se.group_node);
6225 /* We've more or less seen every task there is, call it quits */
6226 if (env->loop > env->loop_max)
6229 /* take a breather every nr_migrate tasks */
6230 if (env->loop > env->loop_break) {
6231 env->loop_break += sched_nr_migrate_break;
6232 env->flags |= LBF_NEED_BREAK;
6236 if (!can_migrate_task(p, env))
6239 load = task_h_load(p);
6241 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6244 if ((load / 2) > env->imbalance)
6247 detach_task(p, env);
6248 list_add(&p->se.group_node, &env->tasks);
6251 env->imbalance -= load;
6253 #ifdef CONFIG_PREEMPT
6255 * NEWIDLE balancing is a source of latency, so preemptible
6256 * kernels will stop after the first task is detached to minimize
6257 * the critical section.
6259 if (env->idle == CPU_NEWLY_IDLE)
6264 * We only want to steal up to the prescribed amount of
6267 if (env->imbalance <= 0)
6272 list_move_tail(&p->se.group_node, tasks);
6276 * Right now, this is one of only two places we collect this stat
6277 * so we can safely collect detach_one_task() stats here rather
6278 * than inside detach_one_task().
6280 schedstat_add(env->sd, lb_gained[env->idle], detached);
6286 * attach_task() -- attach the task detached by detach_task() to its new rq.
6288 static void attach_task(struct rq *rq, struct task_struct *p)
6290 lockdep_assert_held(&rq->lock);
6292 BUG_ON(task_rq(p) != rq);
6293 activate_task(rq, p, 0);
6294 p->on_rq = TASK_ON_RQ_QUEUED;
6295 check_preempt_curr(rq, p, 0);
6299 * attach_one_task() -- attaches the task returned from detach_one_task() to
6302 static void attach_one_task(struct rq *rq, struct task_struct *p)
6304 raw_spin_lock(&rq->lock);
6306 raw_spin_unlock(&rq->lock);
6310 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6313 static void attach_tasks(struct lb_env *env)
6315 struct list_head *tasks = &env->tasks;
6316 struct task_struct *p;
6318 raw_spin_lock(&env->dst_rq->lock);
6320 while (!list_empty(tasks)) {
6321 p = list_first_entry(tasks, struct task_struct, se.group_node);
6322 list_del_init(&p->se.group_node);
6324 attach_task(env->dst_rq, p);
6327 raw_spin_unlock(&env->dst_rq->lock);
6330 #ifdef CONFIG_FAIR_GROUP_SCHED
6331 static void update_blocked_averages(int cpu)
6333 struct rq *rq = cpu_rq(cpu);
6334 struct cfs_rq *cfs_rq;
6335 unsigned long flags;
6337 raw_spin_lock_irqsave(&rq->lock, flags);
6338 update_rq_clock(rq);
6341 * Iterates the task_group tree in a bottom up fashion, see
6342 * list_add_leaf_cfs_rq() for details.
6344 for_each_leaf_cfs_rq(rq, cfs_rq) {
6345 /* throttled entities do not contribute to load */
6346 if (throttled_hierarchy(cfs_rq))
6349 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6350 update_tg_load_avg(cfs_rq, 0);
6352 raw_spin_unlock_irqrestore(&rq->lock, flags);
6356 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6357 * This needs to be done in a top-down fashion because the load of a child
6358 * group is a fraction of its parents load.
6360 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6362 struct rq *rq = rq_of(cfs_rq);
6363 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6364 unsigned long now = jiffies;
6367 if (cfs_rq->last_h_load_update == now)
6370 cfs_rq->h_load_next = NULL;
6371 for_each_sched_entity(se) {
6372 cfs_rq = cfs_rq_of(se);
6373 cfs_rq->h_load_next = se;
6374 if (cfs_rq->last_h_load_update == now)
6379 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6380 cfs_rq->last_h_load_update = now;
6383 while ((se = cfs_rq->h_load_next) != NULL) {
6384 load = cfs_rq->h_load;
6385 load = div64_ul(load * se->avg.load_avg,
6386 cfs_rq_load_avg(cfs_rq) + 1);
6387 cfs_rq = group_cfs_rq(se);
6388 cfs_rq->h_load = load;
6389 cfs_rq->last_h_load_update = now;
6393 static unsigned long task_h_load(struct task_struct *p)
6395 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6397 update_cfs_rq_h_load(cfs_rq);
6398 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6399 cfs_rq_load_avg(cfs_rq) + 1);
6402 static inline void update_blocked_averages(int cpu)
6404 struct rq *rq = cpu_rq(cpu);
6405 struct cfs_rq *cfs_rq = &rq->cfs;
6406 unsigned long flags;
6408 raw_spin_lock_irqsave(&rq->lock, flags);
6409 update_rq_clock(rq);
6410 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6411 raw_spin_unlock_irqrestore(&rq->lock, flags);
6414 static unsigned long task_h_load(struct task_struct *p)
6416 return p->se.avg.load_avg;
6420 /********** Helpers for find_busiest_group ************************/
6429 * sg_lb_stats - stats of a sched_group required for load_balancing
6431 struct sg_lb_stats {
6432 unsigned long avg_load; /*Avg load across the CPUs of the group */
6433 unsigned long group_load; /* Total load over the CPUs of the group */
6434 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6435 unsigned long load_per_task;
6436 unsigned long group_capacity;
6437 unsigned long group_util; /* Total utilization of the group */
6438 unsigned int sum_nr_running; /* Nr tasks running in the group */
6439 unsigned int idle_cpus;
6440 unsigned int group_weight;
6441 enum group_type group_type;
6442 int group_no_capacity;
6443 #ifdef CONFIG_NUMA_BALANCING
6444 unsigned int nr_numa_running;
6445 unsigned int nr_preferred_running;
6450 * sd_lb_stats - Structure to store the statistics of a sched_domain
6451 * during load balancing.
6453 struct sd_lb_stats {
6454 struct sched_group *busiest; /* Busiest group in this sd */
6455 struct sched_group *local; /* Local group in this sd */
6456 unsigned long total_load; /* Total load of all groups in sd */
6457 unsigned long total_capacity; /* Total capacity of all groups in sd */
6458 unsigned long avg_load; /* Average load across all groups in sd */
6460 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6461 struct sg_lb_stats local_stat; /* Statistics of the local group */
6464 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6467 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6468 * local_stat because update_sg_lb_stats() does a full clear/assignment.
6469 * We must however clear busiest_stat::avg_load because
6470 * update_sd_pick_busiest() reads this before assignment.
6472 *sds = (struct sd_lb_stats){
6476 .total_capacity = 0UL,
6479 .sum_nr_running = 0,
6480 .group_type = group_other,
6486 * get_sd_load_idx - Obtain the load index for a given sched domain.
6487 * @sd: The sched_domain whose load_idx is to be obtained.
6488 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6490 * Return: The load index.
6492 static inline int get_sd_load_idx(struct sched_domain *sd,
6493 enum cpu_idle_type idle)
6499 load_idx = sd->busy_idx;
6502 case CPU_NEWLY_IDLE:
6503 load_idx = sd->newidle_idx;
6506 load_idx = sd->idle_idx;
6513 static unsigned long scale_rt_capacity(int cpu)
6515 struct rq *rq = cpu_rq(cpu);
6516 u64 total, used, age_stamp, avg;
6520 * Since we're reading these variables without serialization make sure
6521 * we read them once before doing sanity checks on them.
6523 age_stamp = READ_ONCE(rq->age_stamp);
6524 avg = READ_ONCE(rq->rt_avg);
6525 delta = __rq_clock_broken(rq) - age_stamp;
6527 if (unlikely(delta < 0))
6530 total = sched_avg_period() + delta;
6532 used = div_u64(avg, total);
6534 if (likely(used < SCHED_CAPACITY_SCALE))
6535 return SCHED_CAPACITY_SCALE - used;
6540 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6542 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6543 struct sched_group *sdg = sd->groups;
6545 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6547 capacity *= scale_rt_capacity(cpu);
6548 capacity >>= SCHED_CAPACITY_SHIFT;
6553 cpu_rq(cpu)->cpu_capacity = capacity;
6554 sdg->sgc->capacity = capacity;
6557 void update_group_capacity(struct sched_domain *sd, int cpu)
6559 struct sched_domain *child = sd->child;
6560 struct sched_group *group, *sdg = sd->groups;
6561 unsigned long capacity;
6562 unsigned long interval;
6564 interval = msecs_to_jiffies(sd->balance_interval);
6565 interval = clamp(interval, 1UL, max_load_balance_interval);
6566 sdg->sgc->next_update = jiffies + interval;
6569 update_cpu_capacity(sd, cpu);
6575 if (child->flags & SD_OVERLAP) {
6577 * SD_OVERLAP domains cannot assume that child groups
6578 * span the current group.
6581 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6582 struct sched_group_capacity *sgc;
6583 struct rq *rq = cpu_rq(cpu);
6586 * build_sched_domains() -> init_sched_groups_capacity()
6587 * gets here before we've attached the domains to the
6590 * Use capacity_of(), which is set irrespective of domains
6591 * in update_cpu_capacity().
6593 * This avoids capacity from being 0 and
6594 * causing divide-by-zero issues on boot.
6596 if (unlikely(!rq->sd)) {
6597 capacity += capacity_of(cpu);
6601 sgc = rq->sd->groups->sgc;
6602 capacity += sgc->capacity;
6606 * !SD_OVERLAP domains can assume that child groups
6607 * span the current group.
6610 group = child->groups;
6612 capacity += group->sgc->capacity;
6613 group = group->next;
6614 } while (group != child->groups);
6617 sdg->sgc->capacity = capacity;
6621 * Check whether the capacity of the rq has been noticeably reduced by side
6622 * activity. The imbalance_pct is used for the threshold.
6623 * Return true is the capacity is reduced
6626 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6628 return ((rq->cpu_capacity * sd->imbalance_pct) <
6629 (rq->cpu_capacity_orig * 100));
6633 * Group imbalance indicates (and tries to solve) the problem where balancing
6634 * groups is inadequate due to tsk_cpus_allowed() constraints.
6636 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6637 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6640 * { 0 1 2 3 } { 4 5 6 7 }
6643 * If we were to balance group-wise we'd place two tasks in the first group and
6644 * two tasks in the second group. Clearly this is undesired as it will overload
6645 * cpu 3 and leave one of the cpus in the second group unused.
6647 * The current solution to this issue is detecting the skew in the first group
6648 * by noticing the lower domain failed to reach balance and had difficulty
6649 * moving tasks due to affinity constraints.
6651 * When this is so detected; this group becomes a candidate for busiest; see
6652 * update_sd_pick_busiest(). And calculate_imbalance() and
6653 * find_busiest_group() avoid some of the usual balance conditions to allow it
6654 * to create an effective group imbalance.
6656 * This is a somewhat tricky proposition since the next run might not find the
6657 * group imbalance and decide the groups need to be balanced again. A most
6658 * subtle and fragile situation.
6661 static inline int sg_imbalanced(struct sched_group *group)
6663 return group->sgc->imbalance;
6667 * group_has_capacity returns true if the group has spare capacity that could
6668 * be used by some tasks.
6669 * We consider that a group has spare capacity if the * number of task is
6670 * smaller than the number of CPUs or if the utilization is lower than the
6671 * available capacity for CFS tasks.
6672 * For the latter, we use a threshold to stabilize the state, to take into
6673 * account the variance of the tasks' load and to return true if the available
6674 * capacity in meaningful for the load balancer.
6675 * As an example, an available capacity of 1% can appear but it doesn't make
6676 * any benefit for the load balance.
6679 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6681 if (sgs->sum_nr_running < sgs->group_weight)
6684 if ((sgs->group_capacity * 100) >
6685 (sgs->group_util * env->sd->imbalance_pct))
6692 * group_is_overloaded returns true if the group has more tasks than it can
6694 * group_is_overloaded is not equals to !group_has_capacity because a group
6695 * with the exact right number of tasks, has no more spare capacity but is not
6696 * overloaded so both group_has_capacity and group_is_overloaded return
6700 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6702 if (sgs->sum_nr_running <= sgs->group_weight)
6705 if ((sgs->group_capacity * 100) <
6706 (sgs->group_util * env->sd->imbalance_pct))
6713 group_type group_classify(struct sched_group *group,
6714 struct sg_lb_stats *sgs)
6716 if (sgs->group_no_capacity)
6717 return group_overloaded;
6719 if (sg_imbalanced(group))
6720 return group_imbalanced;
6726 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6727 * @env: The load balancing environment.
6728 * @group: sched_group whose statistics are to be updated.
6729 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6730 * @local_group: Does group contain this_cpu.
6731 * @sgs: variable to hold the statistics for this group.
6732 * @overload: Indicate more than one runnable task for any CPU.
6734 static inline void update_sg_lb_stats(struct lb_env *env,
6735 struct sched_group *group, int load_idx,
6736 int local_group, struct sg_lb_stats *sgs,
6742 memset(sgs, 0, sizeof(*sgs));
6744 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6745 struct rq *rq = cpu_rq(i);
6747 /* Bias balancing toward cpus of our domain */
6749 load = target_load(i, load_idx);
6751 load = source_load(i, load_idx);
6753 sgs->group_load += load;
6754 sgs->group_util += cpu_util(i);
6755 sgs->sum_nr_running += rq->cfs.h_nr_running;
6757 nr_running = rq->nr_running;
6761 #ifdef CONFIG_NUMA_BALANCING
6762 sgs->nr_numa_running += rq->nr_numa_running;
6763 sgs->nr_preferred_running += rq->nr_preferred_running;
6765 sgs->sum_weighted_load += weighted_cpuload(i);
6767 * No need to call idle_cpu() if nr_running is not 0
6769 if (!nr_running && idle_cpu(i))
6773 /* Adjust by relative CPU capacity of the group */
6774 sgs->group_capacity = group->sgc->capacity;
6775 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6777 if (sgs->sum_nr_running)
6778 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6780 sgs->group_weight = group->group_weight;
6782 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6783 sgs->group_type = group_classify(group, sgs);
6787 * update_sd_pick_busiest - return 1 on busiest group
6788 * @env: The load balancing environment.
6789 * @sds: sched_domain statistics
6790 * @sg: sched_group candidate to be checked for being the busiest
6791 * @sgs: sched_group statistics
6793 * Determine if @sg is a busier group than the previously selected
6796 * Return: %true if @sg is a busier group than the previously selected
6797 * busiest group. %false otherwise.
6799 static bool update_sd_pick_busiest(struct lb_env *env,
6800 struct sd_lb_stats *sds,
6801 struct sched_group *sg,
6802 struct sg_lb_stats *sgs)
6804 struct sg_lb_stats *busiest = &sds->busiest_stat;
6806 if (sgs->group_type > busiest->group_type)
6809 if (sgs->group_type < busiest->group_type)
6812 if (sgs->avg_load <= busiest->avg_load)
6815 /* This is the busiest node in its class. */
6816 if (!(env->sd->flags & SD_ASYM_PACKING))
6819 /* No ASYM_PACKING if target cpu is already busy */
6820 if (env->idle == CPU_NOT_IDLE)
6823 * ASYM_PACKING needs to move all the work to the lowest
6824 * numbered CPUs in the group, therefore mark all groups
6825 * higher than ourself as busy.
6827 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6831 /* Prefer to move from highest possible cpu's work */
6832 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6839 #ifdef CONFIG_NUMA_BALANCING
6840 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6842 if (sgs->sum_nr_running > sgs->nr_numa_running)
6844 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6849 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6851 if (rq->nr_running > rq->nr_numa_running)
6853 if (rq->nr_running > rq->nr_preferred_running)
6858 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6863 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6867 #endif /* CONFIG_NUMA_BALANCING */
6870 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6871 * @env: The load balancing environment.
6872 * @sds: variable to hold the statistics for this sched_domain.
6874 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6876 struct sched_domain *child = env->sd->child;
6877 struct sched_group *sg = env->sd->groups;
6878 struct sg_lb_stats tmp_sgs;
6879 int load_idx, prefer_sibling = 0;
6880 bool overload = false;
6882 if (child && child->flags & SD_PREFER_SIBLING)
6885 load_idx = get_sd_load_idx(env->sd, env->idle);
6888 struct sg_lb_stats *sgs = &tmp_sgs;
6891 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6894 sgs = &sds->local_stat;
6896 if (env->idle != CPU_NEWLY_IDLE ||
6897 time_after_eq(jiffies, sg->sgc->next_update))
6898 update_group_capacity(env->sd, env->dst_cpu);
6901 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6908 * In case the child domain prefers tasks go to siblings
6909 * first, lower the sg capacity so that we'll try
6910 * and move all the excess tasks away. We lower the capacity
6911 * of a group only if the local group has the capacity to fit
6912 * these excess tasks. The extra check prevents the case where
6913 * you always pull from the heaviest group when it is already
6914 * under-utilized (possible with a large weight task outweighs
6915 * the tasks on the system).
6917 if (prefer_sibling && sds->local &&
6918 group_has_capacity(env, &sds->local_stat) &&
6919 (sgs->sum_nr_running > 1)) {
6920 sgs->group_no_capacity = 1;
6921 sgs->group_type = group_classify(sg, sgs);
6924 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6926 sds->busiest_stat = *sgs;
6930 /* Now, start updating sd_lb_stats */
6931 sds->total_load += sgs->group_load;
6932 sds->total_capacity += sgs->group_capacity;
6935 } while (sg != env->sd->groups);
6937 if (env->sd->flags & SD_NUMA)
6938 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6940 if (!env->sd->parent) {
6941 /* update overload indicator if we are at root domain */
6942 if (env->dst_rq->rd->overload != overload)
6943 env->dst_rq->rd->overload = overload;
6949 * check_asym_packing - Check to see if the group is packed into the
6952 * This is primarily intended to used at the sibling level. Some
6953 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6954 * case of POWER7, it can move to lower SMT modes only when higher
6955 * threads are idle. When in lower SMT modes, the threads will
6956 * perform better since they share less core resources. Hence when we
6957 * have idle threads, we want them to be the higher ones.
6959 * This packing function is run on idle threads. It checks to see if
6960 * the busiest CPU in this domain (core in the P7 case) has a higher
6961 * CPU number than the packing function is being run on. Here we are
6962 * assuming lower CPU number will be equivalent to lower a SMT thread
6965 * Return: 1 when packing is required and a task should be moved to
6966 * this CPU. The amount of the imbalance is returned in *imbalance.
6968 * @env: The load balancing environment.
6969 * @sds: Statistics of the sched_domain which is to be packed
6971 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6975 if (!(env->sd->flags & SD_ASYM_PACKING))
6978 if (env->idle == CPU_NOT_IDLE)
6984 busiest_cpu = group_first_cpu(sds->busiest);
6985 if (env->dst_cpu > busiest_cpu)
6988 env->imbalance = DIV_ROUND_CLOSEST(
6989 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6990 SCHED_CAPACITY_SCALE);
6996 * fix_small_imbalance - Calculate the minor imbalance that exists
6997 * amongst the groups of a sched_domain, during
6999 * @env: The load balancing environment.
7000 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7003 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7005 unsigned long tmp, capa_now = 0, capa_move = 0;
7006 unsigned int imbn = 2;
7007 unsigned long scaled_busy_load_per_task;
7008 struct sg_lb_stats *local, *busiest;
7010 local = &sds->local_stat;
7011 busiest = &sds->busiest_stat;
7013 if (!local->sum_nr_running)
7014 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7015 else if (busiest->load_per_task > local->load_per_task)
7018 scaled_busy_load_per_task =
7019 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7020 busiest->group_capacity;
7022 if (busiest->avg_load + scaled_busy_load_per_task >=
7023 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7024 env->imbalance = busiest->load_per_task;
7029 * OK, we don't have enough imbalance to justify moving tasks,
7030 * however we may be able to increase total CPU capacity used by
7034 capa_now += busiest->group_capacity *
7035 min(busiest->load_per_task, busiest->avg_load);
7036 capa_now += local->group_capacity *
7037 min(local->load_per_task, local->avg_load);
7038 capa_now /= SCHED_CAPACITY_SCALE;
7040 /* Amount of load we'd subtract */
7041 if (busiest->avg_load > scaled_busy_load_per_task) {
7042 capa_move += busiest->group_capacity *
7043 min(busiest->load_per_task,
7044 busiest->avg_load - scaled_busy_load_per_task);
7047 /* Amount of load we'd add */
7048 if (busiest->avg_load * busiest->group_capacity <
7049 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7050 tmp = (busiest->avg_load * busiest->group_capacity) /
7051 local->group_capacity;
7053 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7054 local->group_capacity;
7056 capa_move += local->group_capacity *
7057 min(local->load_per_task, local->avg_load + tmp);
7058 capa_move /= SCHED_CAPACITY_SCALE;
7060 /* Move if we gain throughput */
7061 if (capa_move > capa_now)
7062 env->imbalance = busiest->load_per_task;
7066 * calculate_imbalance - Calculate the amount of imbalance present within the
7067 * groups of a given sched_domain during load balance.
7068 * @env: load balance environment
7069 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7071 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7073 unsigned long max_pull, load_above_capacity = ~0UL;
7074 struct sg_lb_stats *local, *busiest;
7076 local = &sds->local_stat;
7077 busiest = &sds->busiest_stat;
7079 if (busiest->group_type == group_imbalanced) {
7081 * In the group_imb case we cannot rely on group-wide averages
7082 * to ensure cpu-load equilibrium, look at wider averages. XXX
7084 busiest->load_per_task =
7085 min(busiest->load_per_task, sds->avg_load);
7089 * Avg load of busiest sg can be less and avg load of local sg can
7090 * be greater than avg load across all sgs of sd because avg load
7091 * factors in sg capacity and sgs with smaller group_type are
7092 * skipped when updating the busiest sg:
7094 if (busiest->avg_load <= sds->avg_load ||
7095 local->avg_load >= sds->avg_load) {
7097 return fix_small_imbalance(env, sds);
7101 * If there aren't any idle cpus, avoid creating some.
7103 if (busiest->group_type == group_overloaded &&
7104 local->group_type == group_overloaded) {
7105 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7106 if (load_above_capacity > busiest->group_capacity) {
7107 load_above_capacity -= busiest->group_capacity;
7108 load_above_capacity *= NICE_0_LOAD;
7109 load_above_capacity /= busiest->group_capacity;
7111 load_above_capacity = ~0UL;
7115 * We're trying to get all the cpus to the average_load, so we don't
7116 * want to push ourselves above the average load, nor do we wish to
7117 * reduce the max loaded cpu below the average load. At the same time,
7118 * we also don't want to reduce the group load below the group
7119 * capacity. Thus we look for the minimum possible imbalance.
7121 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7123 /* How much load to actually move to equalise the imbalance */
7124 env->imbalance = min(
7125 max_pull * busiest->group_capacity,
7126 (sds->avg_load - local->avg_load) * local->group_capacity
7127 ) / SCHED_CAPACITY_SCALE;
7130 * if *imbalance is less than the average load per runnable task
7131 * there is no guarantee that any tasks will be moved so we'll have
7132 * a think about bumping its value to force at least one task to be
7135 if (env->imbalance < busiest->load_per_task)
7136 return fix_small_imbalance(env, sds);
7139 /******* find_busiest_group() helpers end here *********************/
7142 * find_busiest_group - Returns the busiest group within the sched_domain
7143 * if there is an imbalance.
7145 * Also calculates the amount of weighted load which should be moved
7146 * to restore balance.
7148 * @env: The load balancing environment.
7150 * Return: - The busiest group if imbalance exists.
7152 static struct sched_group *find_busiest_group(struct lb_env *env)
7154 struct sg_lb_stats *local, *busiest;
7155 struct sd_lb_stats sds;
7157 init_sd_lb_stats(&sds);
7160 * Compute the various statistics relavent for load balancing at
7163 update_sd_lb_stats(env, &sds);
7164 local = &sds.local_stat;
7165 busiest = &sds.busiest_stat;
7167 /* ASYM feature bypasses nice load balance check */
7168 if (check_asym_packing(env, &sds))
7171 /* There is no busy sibling group to pull tasks from */
7172 if (!sds.busiest || busiest->sum_nr_running == 0)
7175 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7176 / sds.total_capacity;
7179 * If the busiest group is imbalanced the below checks don't
7180 * work because they assume all things are equal, which typically
7181 * isn't true due to cpus_allowed constraints and the like.
7183 if (busiest->group_type == group_imbalanced)
7186 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7187 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7188 busiest->group_no_capacity)
7192 * If the local group is busier than the selected busiest group
7193 * don't try and pull any tasks.
7195 if (local->avg_load >= busiest->avg_load)
7199 * Don't pull any tasks if this group is already above the domain
7202 if (local->avg_load >= sds.avg_load)
7205 if (env->idle == CPU_IDLE) {
7207 * This cpu is idle. If the busiest group is not overloaded
7208 * and there is no imbalance between this and busiest group
7209 * wrt idle cpus, it is balanced. The imbalance becomes
7210 * significant if the diff is greater than 1 otherwise we
7211 * might end up to just move the imbalance on another group
7213 if ((busiest->group_type != group_overloaded) &&
7214 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7218 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7219 * imbalance_pct to be conservative.
7221 if (100 * busiest->avg_load <=
7222 env->sd->imbalance_pct * local->avg_load)
7227 /* Looks like there is an imbalance. Compute it */
7228 calculate_imbalance(env, &sds);
7237 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7239 static struct rq *find_busiest_queue(struct lb_env *env,
7240 struct sched_group *group)
7242 struct rq *busiest = NULL, *rq;
7243 unsigned long busiest_load = 0, busiest_capacity = 1;
7246 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7247 unsigned long capacity, wl;
7251 rt = fbq_classify_rq(rq);
7254 * We classify groups/runqueues into three groups:
7255 * - regular: there are !numa tasks
7256 * - remote: there are numa tasks that run on the 'wrong' node
7257 * - all: there is no distinction
7259 * In order to avoid migrating ideally placed numa tasks,
7260 * ignore those when there's better options.
7262 * If we ignore the actual busiest queue to migrate another
7263 * task, the next balance pass can still reduce the busiest
7264 * queue by moving tasks around inside the node.
7266 * If we cannot move enough load due to this classification
7267 * the next pass will adjust the group classification and
7268 * allow migration of more tasks.
7270 * Both cases only affect the total convergence complexity.
7272 if (rt > env->fbq_type)
7275 capacity = capacity_of(i);
7277 wl = weighted_cpuload(i);
7280 * When comparing with imbalance, use weighted_cpuload()
7281 * which is not scaled with the cpu capacity.
7284 if (rq->nr_running == 1 && wl > env->imbalance &&
7285 !check_cpu_capacity(rq, env->sd))
7289 * For the load comparisons with the other cpu's, consider
7290 * the weighted_cpuload() scaled with the cpu capacity, so
7291 * that the load can be moved away from the cpu that is
7292 * potentially running at a lower capacity.
7294 * Thus we're looking for max(wl_i / capacity_i), crosswise
7295 * multiplication to rid ourselves of the division works out
7296 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
7297 * our previous maximum.
7299 if (wl * busiest_capacity > busiest_load * capacity) {
7301 busiest_capacity = capacity;
7310 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7311 * so long as it is large enough.
7313 #define MAX_PINNED_INTERVAL 512
7315 /* Working cpumask for load_balance and load_balance_newidle. */
7316 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7318 static int need_active_balance(struct lb_env *env)
7320 struct sched_domain *sd = env->sd;
7322 if (env->idle == CPU_NEWLY_IDLE) {
7325 * ASYM_PACKING needs to force migrate tasks from busy but
7326 * higher numbered CPUs in order to pack all tasks in the
7327 * lowest numbered CPUs.
7329 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7334 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7335 * It's worth migrating the task if the src_cpu's capacity is reduced
7336 * because of other sched_class or IRQs if more capacity stays
7337 * available on dst_cpu.
7339 if ((env->idle != CPU_NOT_IDLE) &&
7340 (env->src_rq->cfs.h_nr_running == 1)) {
7341 if ((check_cpu_capacity(env->src_rq, sd)) &&
7342 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7346 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7349 static int active_load_balance_cpu_stop(void *data);
7351 static int should_we_balance(struct lb_env *env)
7353 struct sched_group *sg = env->sd->groups;
7354 struct cpumask *sg_cpus, *sg_mask;
7355 int cpu, balance_cpu = -1;
7358 * In the newly idle case, we will allow all the cpu's
7359 * to do the newly idle load balance.
7361 if (env->idle == CPU_NEWLY_IDLE)
7364 sg_cpus = sched_group_cpus(sg);
7365 sg_mask = sched_group_mask(sg);
7366 /* Try to find first idle cpu */
7367 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7368 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7375 if (balance_cpu == -1)
7376 balance_cpu = group_balance_cpu(sg);
7379 * First idle cpu or the first cpu(busiest) in this sched group
7380 * is eligible for doing load balancing at this and above domains.
7382 return balance_cpu == env->dst_cpu;
7386 * Check this_cpu to ensure it is balanced within domain. Attempt to move
7387 * tasks if there is an imbalance.
7389 static int load_balance(int this_cpu, struct rq *this_rq,
7390 struct sched_domain *sd, enum cpu_idle_type idle,
7391 int *continue_balancing)
7393 int ld_moved, cur_ld_moved, active_balance = 0;
7394 struct sched_domain *sd_parent = sd->parent;
7395 struct sched_group *group;
7397 unsigned long flags;
7398 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7400 struct lb_env env = {
7402 .dst_cpu = this_cpu,
7404 .dst_grpmask = sched_group_cpus(sd->groups),
7406 .loop_break = sched_nr_migrate_break,
7409 .tasks = LIST_HEAD_INIT(env.tasks),
7413 * For NEWLY_IDLE load_balancing, we don't need to consider
7414 * other cpus in our group
7416 if (idle == CPU_NEWLY_IDLE)
7417 env.dst_grpmask = NULL;
7419 cpumask_copy(cpus, cpu_active_mask);
7421 schedstat_inc(sd, lb_count[idle]);
7424 if (!should_we_balance(&env)) {
7425 *continue_balancing = 0;
7429 group = find_busiest_group(&env);
7431 schedstat_inc(sd, lb_nobusyg[idle]);
7435 busiest = find_busiest_queue(&env, group);
7437 schedstat_inc(sd, lb_nobusyq[idle]);
7441 BUG_ON(busiest == env.dst_rq);
7443 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7445 env.src_cpu = busiest->cpu;
7446 env.src_rq = busiest;
7449 if (busiest->nr_running > 1) {
7451 * Attempt to move tasks. If find_busiest_group has found
7452 * an imbalance but busiest->nr_running <= 1, the group is
7453 * still unbalanced. ld_moved simply stays zero, so it is
7454 * correctly treated as an imbalance.
7456 env.flags |= LBF_ALL_PINNED;
7457 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
7460 raw_spin_lock_irqsave(&busiest->lock, flags);
7463 * cur_ld_moved - load moved in current iteration
7464 * ld_moved - cumulative load moved across iterations
7466 cur_ld_moved = detach_tasks(&env);
7469 * We've detached some tasks from busiest_rq. Every
7470 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7471 * unlock busiest->lock, and we are able to be sure
7472 * that nobody can manipulate the tasks in parallel.
7473 * See task_rq_lock() family for the details.
7476 raw_spin_unlock(&busiest->lock);
7480 ld_moved += cur_ld_moved;
7483 local_irq_restore(flags);
7485 if (env.flags & LBF_NEED_BREAK) {
7486 env.flags &= ~LBF_NEED_BREAK;
7491 * Revisit (affine) tasks on src_cpu that couldn't be moved to
7492 * us and move them to an alternate dst_cpu in our sched_group
7493 * where they can run. The upper limit on how many times we
7494 * iterate on same src_cpu is dependent on number of cpus in our
7497 * This changes load balance semantics a bit on who can move
7498 * load to a given_cpu. In addition to the given_cpu itself
7499 * (or a ilb_cpu acting on its behalf where given_cpu is
7500 * nohz-idle), we now have balance_cpu in a position to move
7501 * load to given_cpu. In rare situations, this may cause
7502 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7503 * _independently_ and at _same_ time to move some load to
7504 * given_cpu) causing exceess load to be moved to given_cpu.
7505 * This however should not happen so much in practice and
7506 * moreover subsequent load balance cycles should correct the
7507 * excess load moved.
7509 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7511 /* Prevent to re-select dst_cpu via env's cpus */
7512 cpumask_clear_cpu(env.dst_cpu, env.cpus);
7514 env.dst_rq = cpu_rq(env.new_dst_cpu);
7515 env.dst_cpu = env.new_dst_cpu;
7516 env.flags &= ~LBF_DST_PINNED;
7518 env.loop_break = sched_nr_migrate_break;
7521 * Go back to "more_balance" rather than "redo" since we
7522 * need to continue with same src_cpu.
7528 * We failed to reach balance because of affinity.
7531 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7533 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7534 *group_imbalance = 1;
7537 /* All tasks on this runqueue were pinned by CPU affinity */
7538 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7539 cpumask_clear_cpu(cpu_of(busiest), cpus);
7540 if (!cpumask_empty(cpus)) {
7542 env.loop_break = sched_nr_migrate_break;
7545 goto out_all_pinned;
7550 schedstat_inc(sd, lb_failed[idle]);
7552 * Increment the failure counter only on periodic balance.
7553 * We do not want newidle balance, which can be very
7554 * frequent, pollute the failure counter causing
7555 * excessive cache_hot migrations and active balances.
7557 if (idle != CPU_NEWLY_IDLE)
7558 sd->nr_balance_failed++;
7560 if (need_active_balance(&env)) {
7561 raw_spin_lock_irqsave(&busiest->lock, flags);
7563 /* don't kick the active_load_balance_cpu_stop,
7564 * if the curr task on busiest cpu can't be
7567 if (!cpumask_test_cpu(this_cpu,
7568 tsk_cpus_allowed(busiest->curr))) {
7569 raw_spin_unlock_irqrestore(&busiest->lock,
7571 env.flags |= LBF_ALL_PINNED;
7572 goto out_one_pinned;
7576 * ->active_balance synchronizes accesses to
7577 * ->active_balance_work. Once set, it's cleared
7578 * only after active load balance is finished.
7580 if (!busiest->active_balance) {
7581 busiest->active_balance = 1;
7582 busiest->push_cpu = this_cpu;
7585 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7587 if (active_balance) {
7588 stop_one_cpu_nowait(cpu_of(busiest),
7589 active_load_balance_cpu_stop, busiest,
7590 &busiest->active_balance_work);
7593 /* We've kicked active balancing, force task migration. */
7594 sd->nr_balance_failed = sd->cache_nice_tries+1;
7597 sd->nr_balance_failed = 0;
7599 if (likely(!active_balance)) {
7600 /* We were unbalanced, so reset the balancing interval */
7601 sd->balance_interval = sd->min_interval;
7604 * If we've begun active balancing, start to back off. This
7605 * case may not be covered by the all_pinned logic if there
7606 * is only 1 task on the busy runqueue (because we don't call
7609 if (sd->balance_interval < sd->max_interval)
7610 sd->balance_interval *= 2;
7617 * We reach balance although we may have faced some affinity
7618 * constraints. Clear the imbalance flag if it was set.
7621 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7623 if (*group_imbalance)
7624 *group_imbalance = 0;
7629 * We reach balance because all tasks are pinned at this level so
7630 * we can't migrate them. Let the imbalance flag set so parent level
7631 * can try to migrate them.
7633 schedstat_inc(sd, lb_balanced[idle]);
7635 sd->nr_balance_failed = 0;
7638 /* tune up the balancing interval */
7639 if (((env.flags & LBF_ALL_PINNED) &&
7640 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7641 (sd->balance_interval < sd->max_interval))
7642 sd->balance_interval *= 2;
7649 static inline unsigned long
7650 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7652 unsigned long interval = sd->balance_interval;
7655 interval *= sd->busy_factor;
7657 /* scale ms to jiffies */
7658 interval = msecs_to_jiffies(interval);
7659 interval = clamp(interval, 1UL, max_load_balance_interval);
7665 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7667 unsigned long interval, next;
7669 interval = get_sd_balance_interval(sd, cpu_busy);
7670 next = sd->last_balance + interval;
7672 if (time_after(*next_balance, next))
7673 *next_balance = next;
7677 * idle_balance is called by schedule() if this_cpu is about to become
7678 * idle. Attempts to pull tasks from other CPUs.
7680 static int idle_balance(struct rq *this_rq)
7682 unsigned long next_balance = jiffies + HZ;
7683 int this_cpu = this_rq->cpu;
7684 struct sched_domain *sd;
7685 int pulled_task = 0;
7689 * We must set idle_stamp _before_ calling idle_balance(), such that we
7690 * measure the duration of idle_balance() as idle time.
7692 this_rq->idle_stamp = rq_clock(this_rq);
7694 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7695 !this_rq->rd->overload) {
7697 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7699 update_next_balance(sd, 0, &next_balance);
7705 raw_spin_unlock(&this_rq->lock);
7707 update_blocked_averages(this_cpu);
7709 for_each_domain(this_cpu, sd) {
7710 int continue_balancing = 1;
7711 u64 t0, domain_cost;
7713 if (!(sd->flags & SD_LOAD_BALANCE))
7716 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7717 update_next_balance(sd, 0, &next_balance);
7721 if (sd->flags & SD_BALANCE_NEWIDLE) {
7722 t0 = sched_clock_cpu(this_cpu);
7724 pulled_task = load_balance(this_cpu, this_rq,
7726 &continue_balancing);
7728 domain_cost = sched_clock_cpu(this_cpu) - t0;
7729 if (domain_cost > sd->max_newidle_lb_cost)
7730 sd->max_newidle_lb_cost = domain_cost;
7732 curr_cost += domain_cost;
7735 update_next_balance(sd, 0, &next_balance);
7738 * Stop searching for tasks to pull if there are
7739 * now runnable tasks on this rq.
7741 if (pulled_task || this_rq->nr_running > 0)
7746 raw_spin_lock(&this_rq->lock);
7748 if (curr_cost > this_rq->max_idle_balance_cost)
7749 this_rq->max_idle_balance_cost = curr_cost;
7752 * While browsing the domains, we released the rq lock, a task could
7753 * have been enqueued in the meantime. Since we're not going idle,
7754 * pretend we pulled a task.
7756 if (this_rq->cfs.h_nr_running && !pulled_task)
7760 /* Move the next balance forward */
7761 if (time_after(this_rq->next_balance, next_balance))
7762 this_rq->next_balance = next_balance;
7764 /* Is there a task of a high priority class? */
7765 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7769 this_rq->idle_stamp = 0;
7775 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7776 * running tasks off the busiest CPU onto idle CPUs. It requires at
7777 * least 1 task to be running on each physical CPU where possible, and
7778 * avoids physical / logical imbalances.
7780 static int active_load_balance_cpu_stop(void *data)
7782 struct rq *busiest_rq = data;
7783 int busiest_cpu = cpu_of(busiest_rq);
7784 int target_cpu = busiest_rq->push_cpu;
7785 struct rq *target_rq = cpu_rq(target_cpu);
7786 struct sched_domain *sd;
7787 struct task_struct *p = NULL;
7789 raw_spin_lock_irq(&busiest_rq->lock);
7791 /* make sure the requested cpu hasn't gone down in the meantime */
7792 if (unlikely(busiest_cpu != smp_processor_id() ||
7793 !busiest_rq->active_balance))
7796 /* Is there any task to move? */
7797 if (busiest_rq->nr_running <= 1)
7801 * This condition is "impossible", if it occurs
7802 * we need to fix it. Originally reported by
7803 * Bjorn Helgaas on a 128-cpu setup.
7805 BUG_ON(busiest_rq == target_rq);
7807 /* Search for an sd spanning us and the target CPU. */
7809 for_each_domain(target_cpu, sd) {
7810 if ((sd->flags & SD_LOAD_BALANCE) &&
7811 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7816 struct lb_env env = {
7818 .dst_cpu = target_cpu,
7819 .dst_rq = target_rq,
7820 .src_cpu = busiest_rq->cpu,
7821 .src_rq = busiest_rq,
7825 schedstat_inc(sd, alb_count);
7827 p = detach_one_task(&env);
7829 schedstat_inc(sd, alb_pushed);
7830 /* Active balancing done, reset the failure counter. */
7831 sd->nr_balance_failed = 0;
7833 schedstat_inc(sd, alb_failed);
7838 busiest_rq->active_balance = 0;
7839 raw_spin_unlock(&busiest_rq->lock);
7842 attach_one_task(target_rq, p);
7849 static inline int on_null_domain(struct rq *rq)
7851 return unlikely(!rcu_dereference_sched(rq->sd));
7854 #ifdef CONFIG_NO_HZ_COMMON
7856 * idle load balancing details
7857 * - When one of the busy CPUs notice that there may be an idle rebalancing
7858 * needed, they will kick the idle load balancer, which then does idle
7859 * load balancing for all the idle CPUs.
7862 cpumask_var_t idle_cpus_mask;
7864 unsigned long next_balance; /* in jiffy units */
7865 } nohz ____cacheline_aligned;
7867 static inline int find_new_ilb(void)
7869 int ilb = cpumask_first(nohz.idle_cpus_mask);
7871 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7878 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7879 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7880 * CPU (if there is one).
7882 static void nohz_balancer_kick(void)
7886 nohz.next_balance++;
7888 ilb_cpu = find_new_ilb();
7890 if (ilb_cpu >= nr_cpu_ids)
7893 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7896 * Use smp_send_reschedule() instead of resched_cpu().
7897 * This way we generate a sched IPI on the target cpu which
7898 * is idle. And the softirq performing nohz idle load balance
7899 * will be run before returning from the IPI.
7901 smp_send_reschedule(ilb_cpu);
7905 void nohz_balance_exit_idle(unsigned int cpu)
7907 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7909 * Completely isolated CPUs don't ever set, so we must test.
7911 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7912 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7913 atomic_dec(&nohz.nr_cpus);
7915 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7919 static inline void set_cpu_sd_state_busy(void)
7921 struct sched_domain *sd;
7922 int cpu = smp_processor_id();
7925 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7927 if (!sd || !sd->nohz_idle)
7931 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7936 void set_cpu_sd_state_idle(void)
7938 struct sched_domain *sd;
7939 int cpu = smp_processor_id();
7942 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7944 if (!sd || sd->nohz_idle)
7948 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7954 * This routine will record that the cpu is going idle with tick stopped.
7955 * This info will be used in performing idle load balancing in the future.
7957 void nohz_balance_enter_idle(int cpu)
7960 * If this cpu is going down, then nothing needs to be done.
7962 if (!cpu_active(cpu))
7965 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7969 * If we're a completely isolated CPU, we don't play.
7971 if (on_null_domain(cpu_rq(cpu)))
7974 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7975 atomic_inc(&nohz.nr_cpus);
7976 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7980 static DEFINE_SPINLOCK(balancing);
7983 * Scale the max load_balance interval with the number of CPUs in the system.
7984 * This trades load-balance latency on larger machines for less cross talk.
7986 void update_max_interval(void)
7988 max_load_balance_interval = HZ*num_online_cpus()/10;
7992 * It checks each scheduling domain to see if it is due to be balanced,
7993 * and initiates a balancing operation if so.
7995 * Balancing parameters are set up in init_sched_domains.
7997 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7999 int continue_balancing = 1;
8001 unsigned long interval;
8002 struct sched_domain *sd;
8003 /* Earliest time when we have to do rebalance again */
8004 unsigned long next_balance = jiffies + 60*HZ;
8005 int update_next_balance = 0;
8006 int need_serialize, need_decay = 0;
8009 update_blocked_averages(cpu);
8012 for_each_domain(cpu, sd) {
8014 * Decay the newidle max times here because this is a regular
8015 * visit to all the domains. Decay ~1% per second.
8017 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8018 sd->max_newidle_lb_cost =
8019 (sd->max_newidle_lb_cost * 253) / 256;
8020 sd->next_decay_max_lb_cost = jiffies + HZ;
8023 max_cost += sd->max_newidle_lb_cost;
8025 if (!(sd->flags & SD_LOAD_BALANCE))
8029 * Stop the load balance at this level. There is another
8030 * CPU in our sched group which is doing load balancing more
8033 if (!continue_balancing) {
8039 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8041 need_serialize = sd->flags & SD_SERIALIZE;
8042 if (need_serialize) {
8043 if (!spin_trylock(&balancing))
8047 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8048 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8050 * The LBF_DST_PINNED logic could have changed
8051 * env->dst_cpu, so we can't know our idle
8052 * state even if we migrated tasks. Update it.
8054 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8056 sd->last_balance = jiffies;
8057 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8060 spin_unlock(&balancing);
8062 if (time_after(next_balance, sd->last_balance + interval)) {
8063 next_balance = sd->last_balance + interval;
8064 update_next_balance = 1;
8069 * Ensure the rq-wide value also decays but keep it at a
8070 * reasonable floor to avoid funnies with rq->avg_idle.
8072 rq->max_idle_balance_cost =
8073 max((u64)sysctl_sched_migration_cost, max_cost);
8078 * next_balance will be updated only when there is a need.
8079 * When the cpu is attached to null domain for ex, it will not be
8082 if (likely(update_next_balance)) {
8083 rq->next_balance = next_balance;
8085 #ifdef CONFIG_NO_HZ_COMMON
8087 * If this CPU has been elected to perform the nohz idle
8088 * balance. Other idle CPUs have already rebalanced with
8089 * nohz_idle_balance() and nohz.next_balance has been
8090 * updated accordingly. This CPU is now running the idle load
8091 * balance for itself and we need to update the
8092 * nohz.next_balance accordingly.
8094 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8095 nohz.next_balance = rq->next_balance;
8100 #ifdef CONFIG_NO_HZ_COMMON
8102 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8103 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8105 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8107 int this_cpu = this_rq->cpu;
8110 /* Earliest time when we have to do rebalance again */
8111 unsigned long next_balance = jiffies + 60*HZ;
8112 int update_next_balance = 0;
8114 if (idle != CPU_IDLE ||
8115 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8118 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8119 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8123 * If this cpu gets work to do, stop the load balancing
8124 * work being done for other cpus. Next load
8125 * balancing owner will pick it up.
8130 rq = cpu_rq(balance_cpu);
8133 * If time for next balance is due,
8136 if (time_after_eq(jiffies, rq->next_balance)) {
8137 raw_spin_lock_irq(&rq->lock);
8138 update_rq_clock(rq);
8139 cpu_load_update_idle(rq);
8140 raw_spin_unlock_irq(&rq->lock);
8141 rebalance_domains(rq, CPU_IDLE);
8144 if (time_after(next_balance, rq->next_balance)) {
8145 next_balance = rq->next_balance;
8146 update_next_balance = 1;
8151 * next_balance will be updated only when there is a need.
8152 * When the CPU is attached to null domain for ex, it will not be
8155 if (likely(update_next_balance))
8156 nohz.next_balance = next_balance;
8158 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8162 * Current heuristic for kicking the idle load balancer in the presence
8163 * of an idle cpu in the system.
8164 * - This rq has more than one task.
8165 * - This rq has at least one CFS task and the capacity of the CPU is
8166 * significantly reduced because of RT tasks or IRQs.
8167 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8168 * multiple busy cpu.
8169 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8170 * domain span are idle.
8172 static inline bool nohz_kick_needed(struct rq *rq)
8174 unsigned long now = jiffies;
8175 struct sched_domain *sd;
8176 struct sched_group_capacity *sgc;
8177 int nr_busy, cpu = rq->cpu;
8180 if (unlikely(rq->idle_balance))
8184 * We may be recently in ticked or tickless idle mode. At the first
8185 * busy tick after returning from idle, we will update the busy stats.
8187 set_cpu_sd_state_busy();
8188 nohz_balance_exit_idle(cpu);
8191 * None are in tickless mode and hence no need for NOHZ idle load
8194 if (likely(!atomic_read(&nohz.nr_cpus)))
8197 if (time_before(now, nohz.next_balance))
8200 if (rq->nr_running >= 2)
8204 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8206 sgc = sd->groups->sgc;
8207 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8216 sd = rcu_dereference(rq->sd);
8218 if ((rq->cfs.h_nr_running >= 1) &&
8219 check_cpu_capacity(rq, sd)) {
8225 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8226 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8227 sched_domain_span(sd)) < cpu)) {
8237 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8241 * run_rebalance_domains is triggered when needed from the scheduler tick.
8242 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8244 static void run_rebalance_domains(struct softirq_action *h)
8246 struct rq *this_rq = this_rq();
8247 enum cpu_idle_type idle = this_rq->idle_balance ?
8248 CPU_IDLE : CPU_NOT_IDLE;
8251 * If this cpu has a pending nohz_balance_kick, then do the
8252 * balancing on behalf of the other idle cpus whose ticks are
8253 * stopped. Do nohz_idle_balance *before* rebalance_domains to
8254 * give the idle cpus a chance to load balance. Else we may
8255 * load balance only within the local sched_domain hierarchy
8256 * and abort nohz_idle_balance altogether if we pull some load.
8258 nohz_idle_balance(this_rq, idle);
8259 rebalance_domains(this_rq, idle);
8263 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8265 void trigger_load_balance(struct rq *rq)
8267 /* Don't need to rebalance while attached to NULL domain */
8268 if (unlikely(on_null_domain(rq)))
8271 if (time_after_eq(jiffies, rq->next_balance))
8272 raise_softirq(SCHED_SOFTIRQ);
8273 #ifdef CONFIG_NO_HZ_COMMON
8274 if (nohz_kick_needed(rq))
8275 nohz_balancer_kick();
8279 static void rq_online_fair(struct rq *rq)
8283 update_runtime_enabled(rq);
8286 static void rq_offline_fair(struct rq *rq)
8290 /* Ensure any throttled groups are reachable by pick_next_task */
8291 unthrottle_offline_cfs_rqs(rq);
8294 #endif /* CONFIG_SMP */
8297 * scheduler tick hitting a task of our scheduling class:
8299 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
8301 struct cfs_rq *cfs_rq;
8302 struct sched_entity *se = &curr->se;
8304 for_each_sched_entity(se) {
8305 cfs_rq = cfs_rq_of(se);
8306 entity_tick(cfs_rq, se, queued);
8309 if (static_branch_unlikely(&sched_numa_balancing))
8310 task_tick_numa(rq, curr);
8314 * called on fork with the child task as argument from the parent's context
8315 * - child not yet on the tasklist
8316 * - preemption disabled
8318 static void task_fork_fair(struct task_struct *p)
8320 struct cfs_rq *cfs_rq;
8321 struct sched_entity *se = &p->se, *curr;
8322 int this_cpu = smp_processor_id();
8323 struct rq *rq = this_rq();
8324 unsigned long flags;
8326 raw_spin_lock_irqsave(&rq->lock, flags);
8328 update_rq_clock(rq);
8330 cfs_rq = task_cfs_rq(current);
8331 curr = cfs_rq->curr;
8334 * Not only the cpu but also the task_group of the parent might have
8335 * been changed after parent->se.parent,cfs_rq were copied to
8336 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
8337 * of child point to valid ones.
8340 __set_task_cpu(p, this_cpu);
8343 update_curr(cfs_rq);
8346 se->vruntime = curr->vruntime;
8347 place_entity(cfs_rq, se, 1);
8349 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
8351 * Upon rescheduling, sched_class::put_prev_task() will place
8352 * 'current' within the tree based on its new key value.
8354 swap(curr->vruntime, se->vruntime);
8358 se->vruntime -= cfs_rq->min_vruntime;
8360 raw_spin_unlock_irqrestore(&rq->lock, flags);
8364 * Priority of the task has changed. Check to see if we preempt
8368 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
8370 if (!task_on_rq_queued(p))
8374 * Reschedule if we are currently running on this runqueue and
8375 * our priority decreased, or if we are not currently running on
8376 * this runqueue and our priority is higher than the current's
8378 if (rq->curr == p) {
8379 if (p->prio > oldprio)
8382 check_preempt_curr(rq, p, 0);
8385 static inline bool vruntime_normalized(struct task_struct *p)
8387 struct sched_entity *se = &p->se;
8390 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
8391 * the dequeue_entity(.flags=0) will already have normalized the
8398 * When !on_rq, vruntime of the task has usually NOT been normalized.
8399 * But there are some cases where it has already been normalized:
8401 * - A forked child which is waiting for being woken up by
8402 * wake_up_new_task().
8403 * - A task which has been woken up by try_to_wake_up() and
8404 * waiting for actually being woken up by sched_ttwu_pending().
8406 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
8412 static void detach_task_cfs_rq(struct task_struct *p)
8414 struct sched_entity *se = &p->se;
8415 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8417 if (!vruntime_normalized(p)) {
8419 * Fix up our vruntime so that the current sleep doesn't
8420 * cause 'unlimited' sleep bonus.
8422 place_entity(cfs_rq, se, 0);
8423 se->vruntime -= cfs_rq->min_vruntime;
8426 /* Catch up with the cfs_rq and remove our load when we leave */
8427 detach_entity_load_avg(cfs_rq, se);
8430 static void attach_task_cfs_rq(struct task_struct *p)
8432 struct sched_entity *se = &p->se;
8433 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8435 #ifdef CONFIG_FAIR_GROUP_SCHED
8437 * Since the real-depth could have been changed (only FAIR
8438 * class maintain depth value), reset depth properly.
8440 se->depth = se->parent ? se->parent->depth + 1 : 0;
8443 /* Synchronize task with its cfs_rq */
8444 attach_entity_load_avg(cfs_rq, se);
8446 if (!vruntime_normalized(p))
8447 se->vruntime += cfs_rq->min_vruntime;
8450 static void switched_from_fair(struct rq *rq, struct task_struct *p)
8452 detach_task_cfs_rq(p);
8455 static void switched_to_fair(struct rq *rq, struct task_struct *p)
8457 attach_task_cfs_rq(p);
8459 if (task_on_rq_queued(p)) {
8461 * We were most likely switched from sched_rt, so
8462 * kick off the schedule if running, otherwise just see
8463 * if we can still preempt the current task.
8468 check_preempt_curr(rq, p, 0);
8472 /* Account for a task changing its policy or group.
8474 * This routine is mostly called to set cfs_rq->curr field when a task
8475 * migrates between groups/classes.
8477 static void set_curr_task_fair(struct rq *rq)
8479 struct sched_entity *se = &rq->curr->se;
8481 for_each_sched_entity(se) {
8482 struct cfs_rq *cfs_rq = cfs_rq_of(se);
8484 set_next_entity(cfs_rq, se);
8485 /* ensure bandwidth has been allocated on our new cfs_rq */
8486 account_cfs_rq_runtime(cfs_rq, 0);
8490 void init_cfs_rq(struct cfs_rq *cfs_rq)
8492 cfs_rq->tasks_timeline = RB_ROOT;
8493 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8494 #ifndef CONFIG_64BIT
8495 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8498 atomic_long_set(&cfs_rq->removed_load_avg, 0);
8499 atomic_long_set(&cfs_rq->removed_util_avg, 0);
8503 #ifdef CONFIG_FAIR_GROUP_SCHED
8504 static void task_move_group_fair(struct task_struct *p)
8506 detach_task_cfs_rq(p);
8507 set_task_rq(p, task_cpu(p));
8510 /* Tell se's cfs_rq has been changed -- migrated */
8511 p->se.avg.last_update_time = 0;
8513 attach_task_cfs_rq(p);
8516 void free_fair_sched_group(struct task_group *tg)
8520 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8522 for_each_possible_cpu(i) {
8524 kfree(tg->cfs_rq[i]);
8533 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8535 struct sched_entity *se;
8536 struct cfs_rq *cfs_rq;
8540 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8543 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8547 tg->shares = NICE_0_LOAD;
8549 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8551 for_each_possible_cpu(i) {
8554 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8555 GFP_KERNEL, cpu_to_node(i));
8559 se = kzalloc_node(sizeof(struct sched_entity),
8560 GFP_KERNEL, cpu_to_node(i));
8564 init_cfs_rq(cfs_rq);
8565 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8566 init_entity_runnable_average(se);
8568 raw_spin_lock_irq(&rq->lock);
8569 post_init_entity_util_avg(se);
8570 raw_spin_unlock_irq(&rq->lock);
8581 void unregister_fair_sched_group(struct task_group *tg)
8583 unsigned long flags;
8587 for_each_possible_cpu(cpu) {
8589 remove_entity_load_avg(tg->se[cpu]);
8592 * Only empty task groups can be destroyed; so we can speculatively
8593 * check on_list without danger of it being re-added.
8595 if (!tg->cfs_rq[cpu]->on_list)
8600 raw_spin_lock_irqsave(&rq->lock, flags);
8601 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8602 raw_spin_unlock_irqrestore(&rq->lock, flags);
8606 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8607 struct sched_entity *se, int cpu,
8608 struct sched_entity *parent)
8610 struct rq *rq = cpu_rq(cpu);
8614 init_cfs_rq_runtime(cfs_rq);
8616 tg->cfs_rq[cpu] = cfs_rq;
8619 /* se could be NULL for root_task_group */
8624 se->cfs_rq = &rq->cfs;
8627 se->cfs_rq = parent->my_q;
8628 se->depth = parent->depth + 1;
8632 /* guarantee group entities always have weight */
8633 update_load_set(&se->load, NICE_0_LOAD);
8634 se->parent = parent;
8637 static DEFINE_MUTEX(shares_mutex);
8639 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8642 unsigned long flags;
8645 * We can't change the weight of the root cgroup.
8650 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8652 mutex_lock(&shares_mutex);
8653 if (tg->shares == shares)
8656 tg->shares = shares;
8657 for_each_possible_cpu(i) {
8658 struct rq *rq = cpu_rq(i);
8659 struct sched_entity *se;
8662 /* Propagate contribution to hierarchy */
8663 raw_spin_lock_irqsave(&rq->lock, flags);
8665 /* Possible calls to update_curr() need rq clock */
8666 update_rq_clock(rq);
8667 for_each_sched_entity(se)
8668 update_cfs_shares(group_cfs_rq(se));
8669 raw_spin_unlock_irqrestore(&rq->lock, flags);
8673 mutex_unlock(&shares_mutex);
8676 #else /* CONFIG_FAIR_GROUP_SCHED */
8678 void free_fair_sched_group(struct task_group *tg) { }
8680 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8685 void unregister_fair_sched_group(struct task_group *tg) { }
8687 #endif /* CONFIG_FAIR_GROUP_SCHED */
8690 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8692 struct sched_entity *se = &task->se;
8693 unsigned int rr_interval = 0;
8696 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8699 if (rq->cfs.load.weight)
8700 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8706 * All the scheduling class methods:
8708 const struct sched_class fair_sched_class = {
8709 .next = &idle_sched_class,
8710 .enqueue_task = enqueue_task_fair,
8711 .dequeue_task = dequeue_task_fair,
8712 .yield_task = yield_task_fair,
8713 .yield_to_task = yield_to_task_fair,
8715 .check_preempt_curr = check_preempt_wakeup,
8717 .pick_next_task = pick_next_task_fair,
8718 .put_prev_task = put_prev_task_fair,
8721 .select_task_rq = select_task_rq_fair,
8722 .migrate_task_rq = migrate_task_rq_fair,
8724 .rq_online = rq_online_fair,
8725 .rq_offline = rq_offline_fair,
8727 .task_dead = task_dead_fair,
8728 .set_cpus_allowed = set_cpus_allowed_common,
8731 .set_curr_task = set_curr_task_fair,
8732 .task_tick = task_tick_fair,
8733 .task_fork = task_fork_fair,
8735 .prio_changed = prio_changed_fair,
8736 .switched_from = switched_from_fair,
8737 .switched_to = switched_to_fair,
8739 .get_rr_interval = get_rr_interval_fair,
8741 .update_curr = update_curr_fair,
8743 #ifdef CONFIG_FAIR_GROUP_SCHED
8744 .task_move_group = task_move_group_fair,
8748 #ifdef CONFIG_SCHED_DEBUG
8749 void print_cfs_stats(struct seq_file *m, int cpu)
8751 struct cfs_rq *cfs_rq;
8754 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8755 print_cfs_rq(m, cpu, cfs_rq);
8759 #ifdef CONFIG_NUMA_BALANCING
8760 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8763 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8765 for_each_online_node(node) {
8766 if (p->numa_faults) {
8767 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8768 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8770 if (p->numa_group) {
8771 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8772 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8774 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8777 #endif /* CONFIG_NUMA_BALANCING */
8778 #endif /* CONFIG_SCHED_DEBUG */
8780 __init void init_sched_fair_class(void)
8783 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8785 #ifdef CONFIG_NO_HZ_COMMON
8786 nohz.next_balance = jiffies;
8787 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);