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/latencytop.h>
24 #include <linux/sched.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>
33 #include <linux/module.h>
35 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
56 unsigned int sysctl_sched_is_big_little = 0;
57 unsigned int sysctl_sched_sync_hint_enable = 1;
58 unsigned int sysctl_sched_initial_task_util = 0;
59 unsigned int sysctl_sched_cstate_aware = 1;
61 #ifdef CONFIG_SCHED_WALT
62 unsigned int sysctl_sched_use_walt_cpu_util = 1;
63 unsigned int sysctl_sched_use_walt_task_util = 1;
64 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
68 * The initial- and re-scaling of tunables is configurable
69 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
72 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
73 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
74 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
76 enum sched_tunable_scaling sysctl_sched_tunable_scaling
77 = SCHED_TUNABLESCALING_LOG;
80 * Minimal preemption granularity for CPU-bound tasks:
81 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 unsigned int sysctl_sched_min_granularity = 750000ULL;
84 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
87 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
89 static unsigned int sched_nr_latency = 8;
92 * After fork, child runs first. If set to 0 (default) then
93 * parent will (try to) run first.
95 unsigned int sysctl_sched_child_runs_first __read_mostly;
98 * SCHED_OTHER wake-up granularity.
99 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
101 * This option delays the preemption effects of decoupled workloads
102 * and reduces their over-scheduling. Synchronous workloads will still
103 * have immediate wakeup/sleep latencies.
105 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
106 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
108 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
111 * The exponential sliding window over which load is averaged for shares
115 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
117 #ifdef CONFIG_CFS_BANDWIDTH
119 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
120 * each time a cfs_rq requests quota.
122 * Note: in the case that the slice exceeds the runtime remaining (either due
123 * to consumption or the quota being specified to be smaller than the slice)
124 * we will always only issue the remaining available time.
126 * default: 5 msec, units: microseconds
128 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
137 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
143 static inline void update_load_set(struct load_weight *lw, unsigned long w)
150 * Increase the granularity value when there are more CPUs,
151 * because with more CPUs the 'effective latency' as visible
152 * to users decreases. But the relationship is not linear,
153 * so pick a second-best guess by going with the log2 of the
156 * This idea comes from the SD scheduler of Con Kolivas:
158 static unsigned int get_update_sysctl_factor(void)
160 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
163 switch (sysctl_sched_tunable_scaling) {
164 case SCHED_TUNABLESCALING_NONE:
167 case SCHED_TUNABLESCALING_LINEAR:
170 case SCHED_TUNABLESCALING_LOG:
172 factor = 1 + ilog2(cpus);
179 static void update_sysctl(void)
181 unsigned int factor = get_update_sysctl_factor();
183 #define SET_SYSCTL(name) \
184 (sysctl_##name = (factor) * normalized_sysctl_##name)
185 SET_SYSCTL(sched_min_granularity);
186 SET_SYSCTL(sched_latency);
187 SET_SYSCTL(sched_wakeup_granularity);
191 void sched_init_granularity(void)
196 #define WMULT_CONST (~0U)
197 #define WMULT_SHIFT 32
199 static void __update_inv_weight(struct load_weight *lw)
203 if (likely(lw->inv_weight))
206 w = scale_load_down(lw->weight);
208 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
210 else if (unlikely(!w))
211 lw->inv_weight = WMULT_CONST;
213 lw->inv_weight = WMULT_CONST / w;
217 * delta_exec * weight / lw.weight
219 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
221 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
222 * we're guaranteed shift stays positive because inv_weight is guaranteed to
223 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
225 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
226 * weight/lw.weight <= 1, and therefore our shift will also be positive.
228 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
230 u64 fact = scale_load_down(weight);
231 int shift = WMULT_SHIFT;
233 __update_inv_weight(lw);
235 if (unlikely(fact >> 32)) {
242 /* hint to use a 32x32->64 mul */
243 fact = (u64)(u32)fact * lw->inv_weight;
250 return mul_u64_u32_shr(delta_exec, fact, shift);
254 const struct sched_class fair_sched_class;
256 /**************************************************************
257 * CFS operations on generic schedulable entities:
260 #ifdef CONFIG_FAIR_GROUP_SCHED
262 /* cpu runqueue to which this cfs_rq is attached */
263 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
268 /* An entity is a task if it doesn't "own" a runqueue */
269 #define entity_is_task(se) (!se->my_q)
271 static inline struct task_struct *task_of(struct sched_entity *se)
273 #ifdef CONFIG_SCHED_DEBUG
274 WARN_ON_ONCE(!entity_is_task(se));
276 return container_of(se, struct task_struct, se);
279 /* Walk up scheduling entities hierarchy */
280 #define for_each_sched_entity(se) \
281 for (; se; se = se->parent)
283 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
288 /* runqueue on which this entity is (to be) queued */
289 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
294 /* runqueue "owned" by this group */
295 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
300 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
302 if (!cfs_rq->on_list) {
304 * Ensure we either appear before our parent (if already
305 * enqueued) or force our parent to appear after us when it is
306 * enqueued. The fact that we always enqueue bottom-up
307 * reduces this to two cases.
309 if (cfs_rq->tg->parent &&
310 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
311 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
312 &rq_of(cfs_rq)->leaf_cfs_rq_list);
314 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
315 &rq_of(cfs_rq)->leaf_cfs_rq_list);
322 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
324 if (cfs_rq->on_list) {
325 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
330 /* Iterate thr' all leaf cfs_rq's on a runqueue */
331 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
332 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
334 /* Do the two (enqueued) entities belong to the same group ? */
335 static inline struct cfs_rq *
336 is_same_group(struct sched_entity *se, struct sched_entity *pse)
338 if (se->cfs_rq == pse->cfs_rq)
344 static inline struct sched_entity *parent_entity(struct sched_entity *se)
350 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
352 int se_depth, pse_depth;
355 * preemption test can be made between sibling entities who are in the
356 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
357 * both tasks until we find their ancestors who are siblings of common
361 /* First walk up until both entities are at same depth */
362 se_depth = (*se)->depth;
363 pse_depth = (*pse)->depth;
365 while (se_depth > pse_depth) {
367 *se = parent_entity(*se);
370 while (pse_depth > se_depth) {
372 *pse = parent_entity(*pse);
375 while (!is_same_group(*se, *pse)) {
376 *se = parent_entity(*se);
377 *pse = parent_entity(*pse);
381 #else /* !CONFIG_FAIR_GROUP_SCHED */
383 static inline struct task_struct *task_of(struct sched_entity *se)
385 return container_of(se, struct task_struct, se);
388 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
390 return container_of(cfs_rq, struct rq, cfs);
393 #define entity_is_task(se) 1
395 #define for_each_sched_entity(se) \
396 for (; se; se = NULL)
398 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
400 return &task_rq(p)->cfs;
403 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
405 struct task_struct *p = task_of(se);
406 struct rq *rq = task_rq(p);
411 /* runqueue "owned" by this group */
412 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
417 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
421 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
425 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
426 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 static inline struct sched_entity *parent_entity(struct sched_entity *se)
434 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
438 #endif /* CONFIG_FAIR_GROUP_SCHED */
440 static __always_inline
441 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
443 /**************************************************************
444 * Scheduling class tree data structure manipulation methods:
447 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
449 s64 delta = (s64)(vruntime - max_vruntime);
451 max_vruntime = vruntime;
456 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
458 s64 delta = (s64)(vruntime - min_vruntime);
460 min_vruntime = vruntime;
465 static inline int entity_before(struct sched_entity *a,
466 struct sched_entity *b)
468 return (s64)(a->vruntime - b->vruntime) < 0;
471 static void update_min_vruntime(struct cfs_rq *cfs_rq)
473 u64 vruntime = cfs_rq->min_vruntime;
476 vruntime = cfs_rq->curr->vruntime;
478 if (cfs_rq->rb_leftmost) {
479 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
484 vruntime = se->vruntime;
486 vruntime = min_vruntime(vruntime, se->vruntime);
489 /* ensure we never gain time by being placed backwards. */
490 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
493 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
498 * Enqueue an entity into the rb-tree:
500 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
502 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
503 struct rb_node *parent = NULL;
504 struct sched_entity *entry;
508 * Find the right place in the rbtree:
512 entry = rb_entry(parent, struct sched_entity, run_node);
514 * We dont care about collisions. Nodes with
515 * the same key stay together.
517 if (entity_before(se, entry)) {
518 link = &parent->rb_left;
520 link = &parent->rb_right;
526 * Maintain a cache of leftmost tree entries (it is frequently
530 cfs_rq->rb_leftmost = &se->run_node;
532 rb_link_node(&se->run_node, parent, link);
533 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
536 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
538 if (cfs_rq->rb_leftmost == &se->run_node) {
539 struct rb_node *next_node;
541 next_node = rb_next(&se->run_node);
542 cfs_rq->rb_leftmost = next_node;
545 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
548 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
550 struct rb_node *left = cfs_rq->rb_leftmost;
555 return rb_entry(left, struct sched_entity, run_node);
558 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
560 struct rb_node *next = rb_next(&se->run_node);
565 return rb_entry(next, struct sched_entity, run_node);
568 #ifdef CONFIG_SCHED_DEBUG
569 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
571 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
576 return rb_entry(last, struct sched_entity, run_node);
579 /**************************************************************
580 * Scheduling class statistics methods:
583 int sched_proc_update_handler(struct ctl_table *table, int write,
584 void __user *buffer, size_t *lenp,
587 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
588 unsigned int factor = get_update_sysctl_factor();
593 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
594 sysctl_sched_min_granularity);
596 #define WRT_SYSCTL(name) \
597 (normalized_sysctl_##name = sysctl_##name / (factor))
598 WRT_SYSCTL(sched_min_granularity);
599 WRT_SYSCTL(sched_latency);
600 WRT_SYSCTL(sched_wakeup_granularity);
610 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
612 if (unlikely(se->load.weight != NICE_0_LOAD))
613 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
619 * The idea is to set a period in which each task runs once.
621 * When there are too many tasks (sched_nr_latency) we have to stretch
622 * this period because otherwise the slices get too small.
624 * p = (nr <= nl) ? l : l*nr/nl
626 static u64 __sched_period(unsigned long nr_running)
628 if (unlikely(nr_running > sched_nr_latency))
629 return nr_running * sysctl_sched_min_granularity;
631 return sysctl_sched_latency;
635 * We calculate the wall-time slice from the period by taking a part
636 * proportional to the weight.
640 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
642 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
644 for_each_sched_entity(se) {
645 struct load_weight *load;
646 struct load_weight lw;
648 cfs_rq = cfs_rq_of(se);
649 load = &cfs_rq->load;
651 if (unlikely(!se->on_rq)) {
654 update_load_add(&lw, se->load.weight);
657 slice = __calc_delta(slice, se->load.weight, load);
663 * We calculate the vruntime slice of a to-be-inserted task.
667 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
669 return calc_delta_fair(sched_slice(cfs_rq, se), se);
673 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
674 static unsigned long task_h_load(struct task_struct *p);
677 * We choose a half-life close to 1 scheduling period.
678 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
679 * dependent on this value.
681 #define LOAD_AVG_PERIOD 32
682 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
683 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
685 /* Give new sched_entity start runnable values to heavy its load in infant time */
686 void init_entity_runnable_average(struct sched_entity *se)
688 struct sched_avg *sa = &se->avg;
690 sa->last_update_time = 0;
692 * sched_avg's period_contrib should be strictly less then 1024, so
693 * we give it 1023 to make sure it is almost a period (1024us), and
694 * will definitely be update (after enqueue).
696 sa->period_contrib = 1023;
697 sa->load_avg = scale_load_down(se->load.weight);
698 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
699 sa->util_avg = sched_freq() ?
700 sysctl_sched_initial_task_util :
701 scale_load_down(SCHED_LOAD_SCALE);
702 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
703 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
707 void init_entity_runnable_average(struct sched_entity *se)
713 * Update the current task's runtime statistics.
715 static void update_curr(struct cfs_rq *cfs_rq)
717 struct sched_entity *curr = cfs_rq->curr;
718 u64 now = rq_clock_task(rq_of(cfs_rq));
724 delta_exec = now - curr->exec_start;
725 if (unlikely((s64)delta_exec <= 0))
728 curr->exec_start = now;
730 schedstat_set(curr->statistics.exec_max,
731 max(delta_exec, curr->statistics.exec_max));
733 curr->sum_exec_runtime += delta_exec;
734 schedstat_add(cfs_rq, exec_clock, delta_exec);
736 curr->vruntime += calc_delta_fair(delta_exec, curr);
737 update_min_vruntime(cfs_rq);
739 if (entity_is_task(curr)) {
740 struct task_struct *curtask = task_of(curr);
742 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
743 cpuacct_charge(curtask, delta_exec);
744 account_group_exec_runtime(curtask, delta_exec);
747 account_cfs_rq_runtime(cfs_rq, delta_exec);
750 static void update_curr_fair(struct rq *rq)
752 update_curr(cfs_rq_of(&rq->curr->se));
756 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
758 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
762 * Task is being enqueued - update stats:
764 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 * Are we enqueueing a waiting task? (for current tasks
768 * a dequeue/enqueue event is a NOP)
770 if (se != cfs_rq->curr)
771 update_stats_wait_start(cfs_rq, se);
775 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
777 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
778 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
779 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
780 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
782 #ifdef CONFIG_SCHEDSTATS
783 if (entity_is_task(se)) {
784 trace_sched_stat_wait(task_of(se),
785 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
788 schedstat_set(se->statistics.wait_start, 0);
792 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
795 * Mark the end of the wait period if dequeueing a
798 if (se != cfs_rq->curr)
799 update_stats_wait_end(cfs_rq, se);
803 * We are picking a new current task - update its stats:
806 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
809 * We are starting a new run period:
811 se->exec_start = rq_clock_task(rq_of(cfs_rq));
814 /**************************************************
815 * Scheduling class queueing methods:
818 #ifdef CONFIG_NUMA_BALANCING
820 * Approximate time to scan a full NUMA task in ms. The task scan period is
821 * calculated based on the tasks virtual memory size and
822 * numa_balancing_scan_size.
824 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
825 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 /* Portion of address space to scan in MB */
828 unsigned int sysctl_numa_balancing_scan_size = 256;
830 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
831 unsigned int sysctl_numa_balancing_scan_delay = 1000;
833 static unsigned int task_nr_scan_windows(struct task_struct *p)
835 unsigned long rss = 0;
836 unsigned long nr_scan_pages;
839 * Calculations based on RSS as non-present and empty pages are skipped
840 * by the PTE scanner and NUMA hinting faults should be trapped based
843 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
844 rss = get_mm_rss(p->mm);
848 rss = round_up(rss, nr_scan_pages);
849 return rss / nr_scan_pages;
852 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
853 #define MAX_SCAN_WINDOW 2560
855 static unsigned int task_scan_min(struct task_struct *p)
857 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
858 unsigned int scan, floor;
859 unsigned int windows = 1;
861 if (scan_size < MAX_SCAN_WINDOW)
862 windows = MAX_SCAN_WINDOW / scan_size;
863 floor = 1000 / windows;
865 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
866 return max_t(unsigned int, floor, scan);
869 static unsigned int task_scan_max(struct task_struct *p)
871 unsigned int smin = task_scan_min(p);
874 /* Watch for min being lower than max due to floor calculations */
875 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
876 return max(smin, smax);
879 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
881 rq->nr_numa_running += (p->numa_preferred_nid != -1);
882 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
885 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
887 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
888 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
894 spinlock_t lock; /* nr_tasks, tasks */
899 nodemask_t active_nodes;
900 unsigned long total_faults;
902 * Faults_cpu is used to decide whether memory should move
903 * towards the CPU. As a consequence, these stats are weighted
904 * more by CPU use than by memory faults.
906 unsigned long *faults_cpu;
907 unsigned long faults[0];
910 /* Shared or private faults. */
911 #define NR_NUMA_HINT_FAULT_TYPES 2
913 /* Memory and CPU locality */
914 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
916 /* Averaged statistics, and temporary buffers. */
917 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
919 pid_t task_numa_group_id(struct task_struct *p)
921 return p->numa_group ? p->numa_group->gid : 0;
925 * The averaged statistics, shared & private, memory & cpu,
926 * occupy the first half of the array. The second half of the
927 * array is for current counters, which are averaged into the
928 * first set by task_numa_placement.
930 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
932 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
935 static inline unsigned long task_faults(struct task_struct *p, int nid)
940 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
941 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
944 static inline unsigned long group_faults(struct task_struct *p, int nid)
949 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
950 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
953 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
955 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
956 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
959 /* Handle placement on systems where not all nodes are directly connected. */
960 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
961 int maxdist, bool task)
963 unsigned long score = 0;
967 * All nodes are directly connected, and the same distance
968 * from each other. No need for fancy placement algorithms.
970 if (sched_numa_topology_type == NUMA_DIRECT)
974 * This code is called for each node, introducing N^2 complexity,
975 * which should be ok given the number of nodes rarely exceeds 8.
977 for_each_online_node(node) {
978 unsigned long faults;
979 int dist = node_distance(nid, node);
982 * The furthest away nodes in the system are not interesting
983 * for placement; nid was already counted.
985 if (dist == sched_max_numa_distance || node == nid)
989 * On systems with a backplane NUMA topology, compare groups
990 * of nodes, and move tasks towards the group with the most
991 * memory accesses. When comparing two nodes at distance
992 * "hoplimit", only nodes closer by than "hoplimit" are part
993 * of each group. Skip other nodes.
995 if (sched_numa_topology_type == NUMA_BACKPLANE &&
999 /* Add up the faults from nearby nodes. */
1001 faults = task_faults(p, node);
1003 faults = group_faults(p, node);
1006 * On systems with a glueless mesh NUMA topology, there are
1007 * no fixed "groups of nodes". Instead, nodes that are not
1008 * directly connected bounce traffic through intermediate
1009 * nodes; a numa_group can occupy any set of nodes.
1010 * The further away a node is, the less the faults count.
1011 * This seems to result in good task placement.
1013 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1014 faults *= (sched_max_numa_distance - dist);
1015 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1025 * These return the fraction of accesses done by a particular task, or
1026 * task group, on a particular numa node. The group weight is given a
1027 * larger multiplier, in order to group tasks together that are almost
1028 * evenly spread out between numa nodes.
1030 static inline unsigned long task_weight(struct task_struct *p, int nid,
1033 unsigned long faults, total_faults;
1035 if (!p->numa_faults)
1038 total_faults = p->total_numa_faults;
1043 faults = task_faults(p, nid);
1044 faults += score_nearby_nodes(p, nid, dist, true);
1046 return 1000 * faults / total_faults;
1049 static inline unsigned long group_weight(struct task_struct *p, int nid,
1052 unsigned long faults, total_faults;
1057 total_faults = p->numa_group->total_faults;
1062 faults = group_faults(p, nid);
1063 faults += score_nearby_nodes(p, nid, dist, false);
1065 return 1000 * faults / total_faults;
1068 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1069 int src_nid, int dst_cpu)
1071 struct numa_group *ng = p->numa_group;
1072 int dst_nid = cpu_to_node(dst_cpu);
1073 int last_cpupid, this_cpupid;
1075 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1078 * Multi-stage node selection is used in conjunction with a periodic
1079 * migration fault to build a temporal task<->page relation. By using
1080 * a two-stage filter we remove short/unlikely relations.
1082 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1083 * a task's usage of a particular page (n_p) per total usage of this
1084 * page (n_t) (in a given time-span) to a probability.
1086 * Our periodic faults will sample this probability and getting the
1087 * same result twice in a row, given these samples are fully
1088 * independent, is then given by P(n)^2, provided our sample period
1089 * is sufficiently short compared to the usage pattern.
1091 * This quadric squishes small probabilities, making it less likely we
1092 * act on an unlikely task<->page relation.
1094 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1095 if (!cpupid_pid_unset(last_cpupid) &&
1096 cpupid_to_nid(last_cpupid) != dst_nid)
1099 /* Always allow migrate on private faults */
1100 if (cpupid_match_pid(p, last_cpupid))
1103 /* A shared fault, but p->numa_group has not been set up yet. */
1108 * Do not migrate if the destination is not a node that
1109 * is actively used by this numa group.
1111 if (!node_isset(dst_nid, ng->active_nodes))
1115 * Source is a node that is not actively used by this
1116 * numa group, while the destination is. Migrate.
1118 if (!node_isset(src_nid, ng->active_nodes))
1122 * Both source and destination are nodes in active
1123 * use by this numa group. Maximize memory bandwidth
1124 * by migrating from more heavily used groups, to less
1125 * heavily used ones, spreading the load around.
1126 * Use a 1/4 hysteresis to avoid spurious page movement.
1128 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1131 static unsigned long weighted_cpuload(const int cpu);
1132 static unsigned long source_load(int cpu, int type);
1133 static unsigned long target_load(int cpu, int type);
1134 static unsigned long capacity_of(int cpu);
1135 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1137 /* Cached statistics for all CPUs within a node */
1139 unsigned long nr_running;
1142 /* Total compute capacity of CPUs on a node */
1143 unsigned long compute_capacity;
1145 /* Approximate capacity in terms of runnable tasks on a node */
1146 unsigned long task_capacity;
1147 int has_free_capacity;
1151 * XXX borrowed from update_sg_lb_stats
1153 static void update_numa_stats(struct numa_stats *ns, int nid)
1155 int smt, cpu, cpus = 0;
1156 unsigned long capacity;
1158 memset(ns, 0, sizeof(*ns));
1159 for_each_cpu(cpu, cpumask_of_node(nid)) {
1160 struct rq *rq = cpu_rq(cpu);
1162 ns->nr_running += rq->nr_running;
1163 ns->load += weighted_cpuload(cpu);
1164 ns->compute_capacity += capacity_of(cpu);
1170 * If we raced with hotplug and there are no CPUs left in our mask
1171 * the @ns structure is NULL'ed and task_numa_compare() will
1172 * not find this node attractive.
1174 * We'll either bail at !has_free_capacity, or we'll detect a huge
1175 * imbalance and bail there.
1180 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1181 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1182 capacity = cpus / smt; /* cores */
1184 ns->task_capacity = min_t(unsigned, capacity,
1185 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1186 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1189 struct task_numa_env {
1190 struct task_struct *p;
1192 int src_cpu, src_nid;
1193 int dst_cpu, dst_nid;
1195 struct numa_stats src_stats, dst_stats;
1200 struct task_struct *best_task;
1205 static void task_numa_assign(struct task_numa_env *env,
1206 struct task_struct *p, long imp)
1209 put_task_struct(env->best_task);
1212 env->best_imp = imp;
1213 env->best_cpu = env->dst_cpu;
1216 static bool load_too_imbalanced(long src_load, long dst_load,
1217 struct task_numa_env *env)
1220 long orig_src_load, orig_dst_load;
1221 long src_capacity, dst_capacity;
1224 * The load is corrected for the CPU capacity available on each node.
1227 * ------------ vs ---------
1228 * src_capacity dst_capacity
1230 src_capacity = env->src_stats.compute_capacity;
1231 dst_capacity = env->dst_stats.compute_capacity;
1233 /* We care about the slope of the imbalance, not the direction. */
1234 if (dst_load < src_load)
1235 swap(dst_load, src_load);
1237 /* Is the difference below the threshold? */
1238 imb = dst_load * src_capacity * 100 -
1239 src_load * dst_capacity * env->imbalance_pct;
1244 * The imbalance is above the allowed threshold.
1245 * Compare it with the old imbalance.
1247 orig_src_load = env->src_stats.load;
1248 orig_dst_load = env->dst_stats.load;
1250 if (orig_dst_load < orig_src_load)
1251 swap(orig_dst_load, orig_src_load);
1253 old_imb = orig_dst_load * src_capacity * 100 -
1254 orig_src_load * dst_capacity * env->imbalance_pct;
1256 /* Would this change make things worse? */
1257 return (imb > old_imb);
1261 * This checks if the overall compute and NUMA accesses of the system would
1262 * be improved if the source tasks was migrated to the target dst_cpu taking
1263 * into account that it might be best if task running on the dst_cpu should
1264 * be exchanged with the source task
1266 static void task_numa_compare(struct task_numa_env *env,
1267 long taskimp, long groupimp)
1269 struct rq *src_rq = cpu_rq(env->src_cpu);
1270 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1271 struct task_struct *cur;
1272 long src_load, dst_load;
1274 long imp = env->p->numa_group ? groupimp : taskimp;
1276 int dist = env->dist;
1277 bool assigned = false;
1281 raw_spin_lock_irq(&dst_rq->lock);
1284 * No need to move the exiting task or idle task.
1286 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1290 * The task_struct must be protected here to protect the
1291 * p->numa_faults access in the task_weight since the
1292 * numa_faults could already be freed in the following path:
1293 * finish_task_switch()
1294 * --> put_task_struct()
1295 * --> __put_task_struct()
1296 * --> task_numa_free()
1298 get_task_struct(cur);
1301 raw_spin_unlock_irq(&dst_rq->lock);
1304 * Because we have preemption enabled we can get migrated around and
1305 * end try selecting ourselves (current == env->p) as a swap candidate.
1311 * "imp" is the fault differential for the source task between the
1312 * source and destination node. Calculate the total differential for
1313 * the source task and potential destination task. The more negative
1314 * the value is, the more rmeote accesses that would be expected to
1315 * be incurred if the tasks were swapped.
1318 /* Skip this swap candidate if cannot move to the source cpu */
1319 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1323 * If dst and source tasks are in the same NUMA group, or not
1324 * in any group then look only at task weights.
1326 if (cur->numa_group == env->p->numa_group) {
1327 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1328 task_weight(cur, env->dst_nid, dist);
1330 * Add some hysteresis to prevent swapping the
1331 * tasks within a group over tiny differences.
1333 if (cur->numa_group)
1337 * Compare the group weights. If a task is all by
1338 * itself (not part of a group), use the task weight
1341 if (cur->numa_group)
1342 imp += group_weight(cur, env->src_nid, dist) -
1343 group_weight(cur, env->dst_nid, dist);
1345 imp += task_weight(cur, env->src_nid, dist) -
1346 task_weight(cur, env->dst_nid, dist);
1350 if (imp <= env->best_imp && moveimp <= env->best_imp)
1354 /* Is there capacity at our destination? */
1355 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1356 !env->dst_stats.has_free_capacity)
1362 /* Balance doesn't matter much if we're running a task per cpu */
1363 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1364 dst_rq->nr_running == 1)
1368 * In the overloaded case, try and keep the load balanced.
1371 load = task_h_load(env->p);
1372 dst_load = env->dst_stats.load + load;
1373 src_load = env->src_stats.load - load;
1375 if (moveimp > imp && moveimp > env->best_imp) {
1377 * If the improvement from just moving env->p direction is
1378 * better than swapping tasks around, check if a move is
1379 * possible. Store a slightly smaller score than moveimp,
1380 * so an actually idle CPU will win.
1382 if (!load_too_imbalanced(src_load, dst_load, env)) {
1384 put_task_struct(cur);
1390 if (imp <= env->best_imp)
1394 load = task_h_load(cur);
1399 if (load_too_imbalanced(src_load, dst_load, env))
1403 * One idle CPU per node is evaluated for a task numa move.
1404 * Call select_idle_sibling to maybe find a better one.
1407 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1412 task_numa_assign(env, cur, imp);
1416 * The dst_rq->curr isn't assigned. The protection for task_struct is
1419 if (cur && !assigned)
1420 put_task_struct(cur);
1423 static void task_numa_find_cpu(struct task_numa_env *env,
1424 long taskimp, long groupimp)
1428 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1429 /* Skip this CPU if the source task cannot migrate */
1430 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1434 task_numa_compare(env, taskimp, groupimp);
1438 /* Only move tasks to a NUMA node less busy than the current node. */
1439 static bool numa_has_capacity(struct task_numa_env *env)
1441 struct numa_stats *src = &env->src_stats;
1442 struct numa_stats *dst = &env->dst_stats;
1444 if (src->has_free_capacity && !dst->has_free_capacity)
1448 * Only consider a task move if the source has a higher load
1449 * than the destination, corrected for CPU capacity on each node.
1451 * src->load dst->load
1452 * --------------------- vs ---------------------
1453 * src->compute_capacity dst->compute_capacity
1455 if (src->load * dst->compute_capacity * env->imbalance_pct >
1457 dst->load * src->compute_capacity * 100)
1463 static int task_numa_migrate(struct task_struct *p)
1465 struct task_numa_env env = {
1468 .src_cpu = task_cpu(p),
1469 .src_nid = task_node(p),
1471 .imbalance_pct = 112,
1477 struct sched_domain *sd;
1478 unsigned long taskweight, groupweight;
1480 long taskimp, groupimp;
1483 * Pick the lowest SD_NUMA domain, as that would have the smallest
1484 * imbalance and would be the first to start moving tasks about.
1486 * And we want to avoid any moving of tasks about, as that would create
1487 * random movement of tasks -- counter the numa conditions we're trying
1491 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1493 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1497 * Cpusets can break the scheduler domain tree into smaller
1498 * balance domains, some of which do not cross NUMA boundaries.
1499 * Tasks that are "trapped" in such domains cannot be migrated
1500 * elsewhere, so there is no point in (re)trying.
1502 if (unlikely(!sd)) {
1503 p->numa_preferred_nid = task_node(p);
1507 env.dst_nid = p->numa_preferred_nid;
1508 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1509 taskweight = task_weight(p, env.src_nid, dist);
1510 groupweight = group_weight(p, env.src_nid, dist);
1511 update_numa_stats(&env.src_stats, env.src_nid);
1512 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1513 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1514 update_numa_stats(&env.dst_stats, env.dst_nid);
1516 /* Try to find a spot on the preferred nid. */
1517 if (numa_has_capacity(&env))
1518 task_numa_find_cpu(&env, taskimp, groupimp);
1521 * Look at other nodes in these cases:
1522 * - there is no space available on the preferred_nid
1523 * - the task is part of a numa_group that is interleaved across
1524 * multiple NUMA nodes; in order to better consolidate the group,
1525 * we need to check other locations.
1527 if (env.best_cpu == -1 || (p->numa_group &&
1528 nodes_weight(p->numa_group->active_nodes) > 1)) {
1529 for_each_online_node(nid) {
1530 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1533 dist = node_distance(env.src_nid, env.dst_nid);
1534 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1536 taskweight = task_weight(p, env.src_nid, dist);
1537 groupweight = group_weight(p, env.src_nid, dist);
1540 /* Only consider nodes where both task and groups benefit */
1541 taskimp = task_weight(p, nid, dist) - taskweight;
1542 groupimp = group_weight(p, nid, dist) - groupweight;
1543 if (taskimp < 0 && groupimp < 0)
1548 update_numa_stats(&env.dst_stats, env.dst_nid);
1549 if (numa_has_capacity(&env))
1550 task_numa_find_cpu(&env, taskimp, groupimp);
1555 * If the task is part of a workload that spans multiple NUMA nodes,
1556 * and is migrating into one of the workload's active nodes, remember
1557 * this node as the task's preferred numa node, so the workload can
1559 * A task that migrated to a second choice node will be better off
1560 * trying for a better one later. Do not set the preferred node here.
1562 if (p->numa_group) {
1563 if (env.best_cpu == -1)
1568 if (node_isset(nid, p->numa_group->active_nodes))
1569 sched_setnuma(p, env.dst_nid);
1572 /* No better CPU than the current one was found. */
1573 if (env.best_cpu == -1)
1577 * Reset the scan period if the task is being rescheduled on an
1578 * alternative node to recheck if the tasks is now properly placed.
1580 p->numa_scan_period = task_scan_min(p);
1582 if (env.best_task == NULL) {
1583 ret = migrate_task_to(p, env.best_cpu);
1585 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1589 ret = migrate_swap(p, env.best_task);
1591 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1592 put_task_struct(env.best_task);
1596 /* Attempt to migrate a task to a CPU on the preferred node. */
1597 static void numa_migrate_preferred(struct task_struct *p)
1599 unsigned long interval = HZ;
1601 /* This task has no NUMA fault statistics yet */
1602 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1605 /* Periodically retry migrating the task to the preferred node */
1606 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1607 p->numa_migrate_retry = jiffies + interval;
1609 /* Success if task is already running on preferred CPU */
1610 if (task_node(p) == p->numa_preferred_nid)
1613 /* Otherwise, try migrate to a CPU on the preferred node */
1614 task_numa_migrate(p);
1618 * Find the nodes on which the workload is actively running. We do this by
1619 * tracking the nodes from which NUMA hinting faults are triggered. This can
1620 * be different from the set of nodes where the workload's memory is currently
1623 * The bitmask is used to make smarter decisions on when to do NUMA page
1624 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1625 * are added when they cause over 6/16 of the maximum number of faults, but
1626 * only removed when they drop below 3/16.
1628 static void update_numa_active_node_mask(struct numa_group *numa_group)
1630 unsigned long faults, max_faults = 0;
1633 for_each_online_node(nid) {
1634 faults = group_faults_cpu(numa_group, nid);
1635 if (faults > max_faults)
1636 max_faults = faults;
1639 for_each_online_node(nid) {
1640 faults = group_faults_cpu(numa_group, nid);
1641 if (!node_isset(nid, numa_group->active_nodes)) {
1642 if (faults > max_faults * 6 / 16)
1643 node_set(nid, numa_group->active_nodes);
1644 } else if (faults < max_faults * 3 / 16)
1645 node_clear(nid, numa_group->active_nodes);
1650 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1651 * increments. The more local the fault statistics are, the higher the scan
1652 * period will be for the next scan window. If local/(local+remote) ratio is
1653 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1654 * the scan period will decrease. Aim for 70% local accesses.
1656 #define NUMA_PERIOD_SLOTS 10
1657 #define NUMA_PERIOD_THRESHOLD 7
1660 * Increase the scan period (slow down scanning) if the majority of
1661 * our memory is already on our local node, or if the majority of
1662 * the page accesses are shared with other processes.
1663 * Otherwise, decrease the scan period.
1665 static void update_task_scan_period(struct task_struct *p,
1666 unsigned long shared, unsigned long private)
1668 unsigned int period_slot;
1672 unsigned long remote = p->numa_faults_locality[0];
1673 unsigned long local = p->numa_faults_locality[1];
1676 * If there were no record hinting faults then either the task is
1677 * completely idle or all activity is areas that are not of interest
1678 * to automatic numa balancing. Related to that, if there were failed
1679 * migration then it implies we are migrating too quickly or the local
1680 * node is overloaded. In either case, scan slower
1682 if (local + shared == 0 || p->numa_faults_locality[2]) {
1683 p->numa_scan_period = min(p->numa_scan_period_max,
1684 p->numa_scan_period << 1);
1686 p->mm->numa_next_scan = jiffies +
1687 msecs_to_jiffies(p->numa_scan_period);
1693 * Prepare to scale scan period relative to the current period.
1694 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1695 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1696 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1698 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1699 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1700 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1701 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1704 diff = slot * period_slot;
1706 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1709 * Scale scan rate increases based on sharing. There is an
1710 * inverse relationship between the degree of sharing and
1711 * the adjustment made to the scanning period. Broadly
1712 * speaking the intent is that there is little point
1713 * scanning faster if shared accesses dominate as it may
1714 * simply bounce migrations uselessly
1716 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1717 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1720 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1721 task_scan_min(p), task_scan_max(p));
1722 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1726 * Get the fraction of time the task has been running since the last
1727 * NUMA placement cycle. The scheduler keeps similar statistics, but
1728 * decays those on a 32ms period, which is orders of magnitude off
1729 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1730 * stats only if the task is so new there are no NUMA statistics yet.
1732 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1734 u64 runtime, delta, now;
1735 /* Use the start of this time slice to avoid calculations. */
1736 now = p->se.exec_start;
1737 runtime = p->se.sum_exec_runtime;
1739 if (p->last_task_numa_placement) {
1740 delta = runtime - p->last_sum_exec_runtime;
1741 *period = now - p->last_task_numa_placement;
1743 delta = p->se.avg.load_sum / p->se.load.weight;
1744 *period = LOAD_AVG_MAX;
1747 p->last_sum_exec_runtime = runtime;
1748 p->last_task_numa_placement = now;
1754 * Determine the preferred nid for a task in a numa_group. This needs to
1755 * be done in a way that produces consistent results with group_weight,
1756 * otherwise workloads might not converge.
1758 static int preferred_group_nid(struct task_struct *p, int nid)
1763 /* Direct connections between all NUMA nodes. */
1764 if (sched_numa_topology_type == NUMA_DIRECT)
1768 * On a system with glueless mesh NUMA topology, group_weight
1769 * scores nodes according to the number of NUMA hinting faults on
1770 * both the node itself, and on nearby nodes.
1772 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1773 unsigned long score, max_score = 0;
1774 int node, max_node = nid;
1776 dist = sched_max_numa_distance;
1778 for_each_online_node(node) {
1779 score = group_weight(p, node, dist);
1780 if (score > max_score) {
1789 * Finding the preferred nid in a system with NUMA backplane
1790 * interconnect topology is more involved. The goal is to locate
1791 * tasks from numa_groups near each other in the system, and
1792 * untangle workloads from different sides of the system. This requires
1793 * searching down the hierarchy of node groups, recursively searching
1794 * inside the highest scoring group of nodes. The nodemask tricks
1795 * keep the complexity of the search down.
1797 nodes = node_online_map;
1798 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1799 unsigned long max_faults = 0;
1800 nodemask_t max_group = NODE_MASK_NONE;
1803 /* Are there nodes at this distance from each other? */
1804 if (!find_numa_distance(dist))
1807 for_each_node_mask(a, nodes) {
1808 unsigned long faults = 0;
1809 nodemask_t this_group;
1810 nodes_clear(this_group);
1812 /* Sum group's NUMA faults; includes a==b case. */
1813 for_each_node_mask(b, nodes) {
1814 if (node_distance(a, b) < dist) {
1815 faults += group_faults(p, b);
1816 node_set(b, this_group);
1817 node_clear(b, nodes);
1821 /* Remember the top group. */
1822 if (faults > max_faults) {
1823 max_faults = faults;
1824 max_group = this_group;
1826 * subtle: at the smallest distance there is
1827 * just one node left in each "group", the
1828 * winner is the preferred nid.
1833 /* Next round, evaluate the nodes within max_group. */
1841 static void task_numa_placement(struct task_struct *p)
1843 int seq, nid, max_nid = -1, max_group_nid = -1;
1844 unsigned long max_faults = 0, max_group_faults = 0;
1845 unsigned long fault_types[2] = { 0, 0 };
1846 unsigned long total_faults;
1847 u64 runtime, period;
1848 spinlock_t *group_lock = NULL;
1851 * The p->mm->numa_scan_seq field gets updated without
1852 * exclusive access. Use READ_ONCE() here to ensure
1853 * that the field is read in a single access:
1855 seq = READ_ONCE(p->mm->numa_scan_seq);
1856 if (p->numa_scan_seq == seq)
1858 p->numa_scan_seq = seq;
1859 p->numa_scan_period_max = task_scan_max(p);
1861 total_faults = p->numa_faults_locality[0] +
1862 p->numa_faults_locality[1];
1863 runtime = numa_get_avg_runtime(p, &period);
1865 /* If the task is part of a group prevent parallel updates to group stats */
1866 if (p->numa_group) {
1867 group_lock = &p->numa_group->lock;
1868 spin_lock_irq(group_lock);
1871 /* Find the node with the highest number of faults */
1872 for_each_online_node(nid) {
1873 /* Keep track of the offsets in numa_faults array */
1874 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1875 unsigned long faults = 0, group_faults = 0;
1878 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1879 long diff, f_diff, f_weight;
1881 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1882 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1883 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1884 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1886 /* Decay existing window, copy faults since last scan */
1887 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1888 fault_types[priv] += p->numa_faults[membuf_idx];
1889 p->numa_faults[membuf_idx] = 0;
1892 * Normalize the faults_from, so all tasks in a group
1893 * count according to CPU use, instead of by the raw
1894 * number of faults. Tasks with little runtime have
1895 * little over-all impact on throughput, and thus their
1896 * faults are less important.
1898 f_weight = div64_u64(runtime << 16, period + 1);
1899 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1901 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1902 p->numa_faults[cpubuf_idx] = 0;
1904 p->numa_faults[mem_idx] += diff;
1905 p->numa_faults[cpu_idx] += f_diff;
1906 faults += p->numa_faults[mem_idx];
1907 p->total_numa_faults += diff;
1908 if (p->numa_group) {
1910 * safe because we can only change our own group
1912 * mem_idx represents the offset for a given
1913 * nid and priv in a specific region because it
1914 * is at the beginning of the numa_faults array.
1916 p->numa_group->faults[mem_idx] += diff;
1917 p->numa_group->faults_cpu[mem_idx] += f_diff;
1918 p->numa_group->total_faults += diff;
1919 group_faults += p->numa_group->faults[mem_idx];
1923 if (faults > max_faults) {
1924 max_faults = faults;
1928 if (group_faults > max_group_faults) {
1929 max_group_faults = group_faults;
1930 max_group_nid = nid;
1934 update_task_scan_period(p, fault_types[0], fault_types[1]);
1936 if (p->numa_group) {
1937 update_numa_active_node_mask(p->numa_group);
1938 spin_unlock_irq(group_lock);
1939 max_nid = preferred_group_nid(p, max_group_nid);
1943 /* Set the new preferred node */
1944 if (max_nid != p->numa_preferred_nid)
1945 sched_setnuma(p, max_nid);
1947 if (task_node(p) != p->numa_preferred_nid)
1948 numa_migrate_preferred(p);
1952 static inline int get_numa_group(struct numa_group *grp)
1954 return atomic_inc_not_zero(&grp->refcount);
1957 static inline void put_numa_group(struct numa_group *grp)
1959 if (atomic_dec_and_test(&grp->refcount))
1960 kfree_rcu(grp, rcu);
1963 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1966 struct numa_group *grp, *my_grp;
1967 struct task_struct *tsk;
1969 int cpu = cpupid_to_cpu(cpupid);
1972 if (unlikely(!p->numa_group)) {
1973 unsigned int size = sizeof(struct numa_group) +
1974 4*nr_node_ids*sizeof(unsigned long);
1976 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1980 atomic_set(&grp->refcount, 1);
1981 spin_lock_init(&grp->lock);
1983 /* Second half of the array tracks nids where faults happen */
1984 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1987 node_set(task_node(current), grp->active_nodes);
1989 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1990 grp->faults[i] = p->numa_faults[i];
1992 grp->total_faults = p->total_numa_faults;
1995 rcu_assign_pointer(p->numa_group, grp);
1999 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2001 if (!cpupid_match_pid(tsk, cpupid))
2004 grp = rcu_dereference(tsk->numa_group);
2008 my_grp = p->numa_group;
2013 * Only join the other group if its bigger; if we're the bigger group,
2014 * the other task will join us.
2016 if (my_grp->nr_tasks > grp->nr_tasks)
2020 * Tie-break on the grp address.
2022 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2025 /* Always join threads in the same process. */
2026 if (tsk->mm == current->mm)
2029 /* Simple filter to avoid false positives due to PID collisions */
2030 if (flags & TNF_SHARED)
2033 /* Update priv based on whether false sharing was detected */
2036 if (join && !get_numa_group(grp))
2044 BUG_ON(irqs_disabled());
2045 double_lock_irq(&my_grp->lock, &grp->lock);
2047 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2048 my_grp->faults[i] -= p->numa_faults[i];
2049 grp->faults[i] += p->numa_faults[i];
2051 my_grp->total_faults -= p->total_numa_faults;
2052 grp->total_faults += p->total_numa_faults;
2057 spin_unlock(&my_grp->lock);
2058 spin_unlock_irq(&grp->lock);
2060 rcu_assign_pointer(p->numa_group, grp);
2062 put_numa_group(my_grp);
2070 void task_numa_free(struct task_struct *p)
2072 struct numa_group *grp = p->numa_group;
2073 void *numa_faults = p->numa_faults;
2074 unsigned long flags;
2078 spin_lock_irqsave(&grp->lock, flags);
2079 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2080 grp->faults[i] -= p->numa_faults[i];
2081 grp->total_faults -= p->total_numa_faults;
2084 spin_unlock_irqrestore(&grp->lock, flags);
2085 RCU_INIT_POINTER(p->numa_group, NULL);
2086 put_numa_group(grp);
2089 p->numa_faults = NULL;
2094 * Got a PROT_NONE fault for a page on @node.
2096 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2098 struct task_struct *p = current;
2099 bool migrated = flags & TNF_MIGRATED;
2100 int cpu_node = task_node(current);
2101 int local = !!(flags & TNF_FAULT_LOCAL);
2104 if (!static_branch_likely(&sched_numa_balancing))
2107 /* for example, ksmd faulting in a user's mm */
2111 /* Allocate buffer to track faults on a per-node basis */
2112 if (unlikely(!p->numa_faults)) {
2113 int size = sizeof(*p->numa_faults) *
2114 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2116 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2117 if (!p->numa_faults)
2120 p->total_numa_faults = 0;
2121 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2125 * First accesses are treated as private, otherwise consider accesses
2126 * to be private if the accessing pid has not changed
2128 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2131 priv = cpupid_match_pid(p, last_cpupid);
2132 if (!priv && !(flags & TNF_NO_GROUP))
2133 task_numa_group(p, last_cpupid, flags, &priv);
2137 * If a workload spans multiple NUMA nodes, a shared fault that
2138 * occurs wholly within the set of nodes that the workload is
2139 * actively using should be counted as local. This allows the
2140 * scan rate to slow down when a workload has settled down.
2142 if (!priv && !local && p->numa_group &&
2143 node_isset(cpu_node, p->numa_group->active_nodes) &&
2144 node_isset(mem_node, p->numa_group->active_nodes))
2147 task_numa_placement(p);
2150 * Retry task to preferred node migration periodically, in case it
2151 * case it previously failed, or the scheduler moved us.
2153 if (time_after(jiffies, p->numa_migrate_retry))
2154 numa_migrate_preferred(p);
2157 p->numa_pages_migrated += pages;
2158 if (flags & TNF_MIGRATE_FAIL)
2159 p->numa_faults_locality[2] += pages;
2161 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2162 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2163 p->numa_faults_locality[local] += pages;
2166 static void reset_ptenuma_scan(struct task_struct *p)
2169 * We only did a read acquisition of the mmap sem, so
2170 * p->mm->numa_scan_seq is written to without exclusive access
2171 * and the update is not guaranteed to be atomic. That's not
2172 * much of an issue though, since this is just used for
2173 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2174 * expensive, to avoid any form of compiler optimizations:
2176 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2177 p->mm->numa_scan_offset = 0;
2181 * The expensive part of numa migration is done from task_work context.
2182 * Triggered from task_tick_numa().
2184 void task_numa_work(struct callback_head *work)
2186 unsigned long migrate, next_scan, now = jiffies;
2187 struct task_struct *p = current;
2188 struct mm_struct *mm = p->mm;
2189 struct vm_area_struct *vma;
2190 unsigned long start, end;
2191 unsigned long nr_pte_updates = 0;
2192 long pages, virtpages;
2194 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2196 work->next = work; /* protect against double add */
2198 * Who cares about NUMA placement when they're dying.
2200 * NOTE: make sure not to dereference p->mm before this check,
2201 * exit_task_work() happens _after_ exit_mm() so we could be called
2202 * without p->mm even though we still had it when we enqueued this
2205 if (p->flags & PF_EXITING)
2208 if (!mm->numa_next_scan) {
2209 mm->numa_next_scan = now +
2210 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2214 * Enforce maximal scan/migration frequency..
2216 migrate = mm->numa_next_scan;
2217 if (time_before(now, migrate))
2220 if (p->numa_scan_period == 0) {
2221 p->numa_scan_period_max = task_scan_max(p);
2222 p->numa_scan_period = task_scan_min(p);
2225 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2226 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2230 * Delay this task enough that another task of this mm will likely win
2231 * the next time around.
2233 p->node_stamp += 2 * TICK_NSEC;
2235 start = mm->numa_scan_offset;
2236 pages = sysctl_numa_balancing_scan_size;
2237 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2238 virtpages = pages * 8; /* Scan up to this much virtual space */
2243 down_read(&mm->mmap_sem);
2244 vma = find_vma(mm, start);
2246 reset_ptenuma_scan(p);
2250 for (; vma; vma = vma->vm_next) {
2251 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2252 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2257 * Shared library pages mapped by multiple processes are not
2258 * migrated as it is expected they are cache replicated. Avoid
2259 * hinting faults in read-only file-backed mappings or the vdso
2260 * as migrating the pages will be of marginal benefit.
2263 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2267 * Skip inaccessible VMAs to avoid any confusion between
2268 * PROT_NONE and NUMA hinting ptes
2270 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2274 start = max(start, vma->vm_start);
2275 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2276 end = min(end, vma->vm_end);
2277 nr_pte_updates = change_prot_numa(vma, start, end);
2280 * Try to scan sysctl_numa_balancing_size worth of
2281 * hpages that have at least one present PTE that
2282 * is not already pte-numa. If the VMA contains
2283 * areas that are unused or already full of prot_numa
2284 * PTEs, scan up to virtpages, to skip through those
2288 pages -= (end - start) >> PAGE_SHIFT;
2289 virtpages -= (end - start) >> PAGE_SHIFT;
2292 if (pages <= 0 || virtpages <= 0)
2296 } while (end != vma->vm_end);
2301 * It is possible to reach the end of the VMA list but the last few
2302 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2303 * would find the !migratable VMA on the next scan but not reset the
2304 * scanner to the start so check it now.
2307 mm->numa_scan_offset = start;
2309 reset_ptenuma_scan(p);
2310 up_read(&mm->mmap_sem);
2314 * Drive the periodic memory faults..
2316 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2318 struct callback_head *work = &curr->numa_work;
2322 * We don't care about NUMA placement if we don't have memory.
2324 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2328 * Using runtime rather than walltime has the dual advantage that
2329 * we (mostly) drive the selection from busy threads and that the
2330 * task needs to have done some actual work before we bother with
2333 now = curr->se.sum_exec_runtime;
2334 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2336 if (now > curr->node_stamp + period) {
2337 if (!curr->node_stamp)
2338 curr->numa_scan_period = task_scan_min(curr);
2339 curr->node_stamp += period;
2341 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2342 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2343 task_work_add(curr, work, true);
2348 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2352 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2356 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2359 #endif /* CONFIG_NUMA_BALANCING */
2362 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2364 update_load_add(&cfs_rq->load, se->load.weight);
2365 if (!parent_entity(se))
2366 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2368 if (entity_is_task(se)) {
2369 struct rq *rq = rq_of(cfs_rq);
2371 account_numa_enqueue(rq, task_of(se));
2372 list_add(&se->group_node, &rq->cfs_tasks);
2375 cfs_rq->nr_running++;
2379 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2381 update_load_sub(&cfs_rq->load, se->load.weight);
2382 if (!parent_entity(se))
2383 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2384 if (entity_is_task(se)) {
2385 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2386 list_del_init(&se->group_node);
2388 cfs_rq->nr_running--;
2391 #ifdef CONFIG_FAIR_GROUP_SCHED
2393 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2398 * Use this CPU's real-time load instead of the last load contribution
2399 * as the updating of the contribution is delayed, and we will use the
2400 * the real-time load to calc the share. See update_tg_load_avg().
2402 tg_weight = atomic_long_read(&tg->load_avg);
2403 tg_weight -= cfs_rq->tg_load_avg_contrib;
2404 tg_weight += cfs_rq->load.weight;
2409 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2411 long tg_weight, load, shares;
2413 tg_weight = calc_tg_weight(tg, cfs_rq);
2414 load = cfs_rq->load.weight;
2416 shares = (tg->shares * load);
2418 shares /= tg_weight;
2420 if (shares < MIN_SHARES)
2421 shares = MIN_SHARES;
2422 if (shares > tg->shares)
2423 shares = tg->shares;
2427 # else /* CONFIG_SMP */
2428 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2432 # endif /* CONFIG_SMP */
2433 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2434 unsigned long weight)
2437 /* commit outstanding execution time */
2438 if (cfs_rq->curr == se)
2439 update_curr(cfs_rq);
2440 account_entity_dequeue(cfs_rq, se);
2443 update_load_set(&se->load, weight);
2446 account_entity_enqueue(cfs_rq, se);
2449 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2451 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2453 struct task_group *tg;
2454 struct sched_entity *se;
2458 se = tg->se[cpu_of(rq_of(cfs_rq))];
2459 if (!se || throttled_hierarchy(cfs_rq))
2462 if (likely(se->load.weight == tg->shares))
2465 shares = calc_cfs_shares(cfs_rq, tg);
2467 reweight_entity(cfs_rq_of(se), se, shares);
2469 #else /* CONFIG_FAIR_GROUP_SCHED */
2470 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2473 #endif /* CONFIG_FAIR_GROUP_SCHED */
2476 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2477 static const u32 runnable_avg_yN_inv[] = {
2478 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2479 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2480 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2481 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2482 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2483 0x85aac367, 0x82cd8698,
2487 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2488 * over-estimates when re-combining.
2490 static const u32 runnable_avg_yN_sum[] = {
2491 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2492 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2493 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2498 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2500 static __always_inline u64 decay_load(u64 val, u64 n)
2502 unsigned int local_n;
2506 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2509 /* after bounds checking we can collapse to 32-bit */
2513 * As y^PERIOD = 1/2, we can combine
2514 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2515 * With a look-up table which covers y^n (n<PERIOD)
2517 * To achieve constant time decay_load.
2519 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2520 val >>= local_n / LOAD_AVG_PERIOD;
2521 local_n %= LOAD_AVG_PERIOD;
2524 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2529 * For updates fully spanning n periods, the contribution to runnable
2530 * average will be: \Sum 1024*y^n
2532 * We can compute this reasonably efficiently by combining:
2533 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2535 static u32 __compute_runnable_contrib(u64 n)
2539 if (likely(n <= LOAD_AVG_PERIOD))
2540 return runnable_avg_yN_sum[n];
2541 else if (unlikely(n >= LOAD_AVG_MAX_N))
2542 return LOAD_AVG_MAX;
2544 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2546 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2547 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2549 n -= LOAD_AVG_PERIOD;
2550 } while (n > LOAD_AVG_PERIOD);
2552 contrib = decay_load(contrib, n);
2553 return contrib + runnable_avg_yN_sum[n];
2556 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2557 #error "load tracking assumes 2^10 as unit"
2560 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2563 * We can represent the historical contribution to runnable average as the
2564 * coefficients of a geometric series. To do this we sub-divide our runnable
2565 * history into segments of approximately 1ms (1024us); label the segment that
2566 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2568 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2570 * (now) (~1ms ago) (~2ms ago)
2572 * Let u_i denote the fraction of p_i that the entity was runnable.
2574 * We then designate the fractions u_i as our co-efficients, yielding the
2575 * following representation of historical load:
2576 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2578 * We choose y based on the with of a reasonably scheduling period, fixing:
2581 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2582 * approximately half as much as the contribution to load within the last ms
2585 * When a period "rolls over" and we have new u_0`, multiplying the previous
2586 * sum again by y is sufficient to update:
2587 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2588 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2590 static __always_inline int
2591 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2592 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2594 u64 delta, scaled_delta, periods;
2596 unsigned int delta_w, scaled_delta_w, decayed = 0;
2597 unsigned long scale_freq, scale_cpu;
2599 delta = now - sa->last_update_time;
2601 * This should only happen when time goes backwards, which it
2602 * unfortunately does during sched clock init when we swap over to TSC.
2604 if ((s64)delta < 0) {
2605 sa->last_update_time = now;
2610 * Use 1024ns as the unit of measurement since it's a reasonable
2611 * approximation of 1us and fast to compute.
2616 sa->last_update_time = now;
2618 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2619 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2620 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2622 /* delta_w is the amount already accumulated against our next period */
2623 delta_w = sa->period_contrib;
2624 if (delta + delta_w >= 1024) {
2627 /* how much left for next period will start over, we don't know yet */
2628 sa->period_contrib = 0;
2631 * Now that we know we're crossing a period boundary, figure
2632 * out how much from delta we need to complete the current
2633 * period and accrue it.
2635 delta_w = 1024 - delta_w;
2636 scaled_delta_w = cap_scale(delta_w, scale_freq);
2638 sa->load_sum += weight * scaled_delta_w;
2640 cfs_rq->runnable_load_sum +=
2641 weight * scaled_delta_w;
2645 sa->util_sum += scaled_delta_w * scale_cpu;
2649 /* Figure out how many additional periods this update spans */
2650 periods = delta / 1024;
2653 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2655 cfs_rq->runnable_load_sum =
2656 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2658 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2660 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2661 contrib = __compute_runnable_contrib(periods);
2662 contrib = cap_scale(contrib, scale_freq);
2664 sa->load_sum += weight * contrib;
2666 cfs_rq->runnable_load_sum += weight * contrib;
2669 sa->util_sum += contrib * scale_cpu;
2672 /* Remainder of delta accrued against u_0` */
2673 scaled_delta = cap_scale(delta, scale_freq);
2675 sa->load_sum += weight * scaled_delta;
2677 cfs_rq->runnable_load_sum += weight * scaled_delta;
2680 sa->util_sum += scaled_delta * scale_cpu;
2682 sa->period_contrib += delta;
2685 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2687 cfs_rq->runnable_load_avg =
2688 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2690 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2696 #ifdef CONFIG_FAIR_GROUP_SCHED
2698 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2699 * and effective_load (which is not done because it is too costly).
2701 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2703 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2705 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2706 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2707 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2711 #else /* CONFIG_FAIR_GROUP_SCHED */
2712 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2713 #endif /* CONFIG_FAIR_GROUP_SCHED */
2715 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2717 if (&this_rq()->cfs == cfs_rq) {
2719 * There are a few boundary cases this might miss but it should
2720 * get called often enough that that should (hopefully) not be
2721 * a real problem -- added to that it only calls on the local
2722 * CPU, so if we enqueue remotely we'll miss an update, but
2723 * the next tick/schedule should update.
2725 * It will not get called when we go idle, because the idle
2726 * thread is a different class (!fair), nor will the utilization
2727 * number include things like RT tasks.
2729 * As is, the util number is not freq-invariant (we'd have to
2730 * implement arch_scale_freq_capacity() for that).
2734 cpufreq_update_util(rq_of(cfs_rq), 0);
2738 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2741 * Unsigned subtract and clamp on underflow.
2743 * Explicitly do a load-store to ensure the intermediate value never hits
2744 * memory. This allows lockless observations without ever seeing the negative
2747 #define sub_positive(_ptr, _val) do { \
2748 typeof(_ptr) ptr = (_ptr); \
2749 typeof(*ptr) val = (_val); \
2750 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2754 WRITE_ONCE(*ptr, res); \
2757 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2758 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq,
2761 struct sched_avg *sa = &cfs_rq->avg;
2762 int decayed, removed = 0, removed_util = 0;
2764 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2765 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2766 sub_positive(&sa->load_avg, r);
2767 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2771 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2772 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2773 sub_positive(&sa->util_avg, r);
2774 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2778 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2779 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2781 #ifndef CONFIG_64BIT
2783 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2786 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2787 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2788 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2790 if (update_freq && (decayed || removed_util))
2791 cfs_rq_util_change(cfs_rq);
2793 return decayed || removed;
2796 /* Update task and its cfs_rq load average */
2797 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2799 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2800 u64 now = cfs_rq_clock_task(cfs_rq);
2801 int cpu = cpu_of(rq_of(cfs_rq));
2804 * Track task load average for carrying it to new CPU after migrated, and
2805 * track group sched_entity load average for task_h_load calc in migration
2807 __update_load_avg(now, cpu, &se->avg,
2808 se->on_rq * scale_load_down(se->load.weight),
2809 cfs_rq->curr == se, NULL);
2811 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2812 update_tg_load_avg(cfs_rq, 0);
2814 if (entity_is_task(se))
2815 trace_sched_load_avg_task(task_of(se), &se->avg);
2818 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2820 if (!sched_feat(ATTACH_AGE_LOAD))
2824 * If we got migrated (either between CPUs or between cgroups) we'll
2825 * have aged the average right before clearing @last_update_time.
2827 if (se->avg.last_update_time) {
2828 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2829 &se->avg, 0, 0, NULL);
2832 * XXX: we could have just aged the entire load away if we've been
2833 * absent from the fair class for too long.
2838 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2839 cfs_rq->avg.load_avg += se->avg.load_avg;
2840 cfs_rq->avg.load_sum += se->avg.load_sum;
2841 cfs_rq->avg.util_avg += se->avg.util_avg;
2842 cfs_rq->avg.util_sum += se->avg.util_sum;
2844 cfs_rq_util_change(cfs_rq);
2847 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2849 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2850 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2851 cfs_rq->curr == se, NULL);
2853 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2854 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2855 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2856 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2858 cfs_rq_util_change(cfs_rq);
2861 /* Add the load generated by se into cfs_rq's load average */
2863 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2865 struct sched_avg *sa = &se->avg;
2866 u64 now = cfs_rq_clock_task(cfs_rq);
2867 int migrated, decayed;
2869 migrated = !sa->last_update_time;
2871 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2872 se->on_rq * scale_load_down(se->load.weight),
2873 cfs_rq->curr == se, NULL);
2876 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
2878 cfs_rq->runnable_load_avg += sa->load_avg;
2879 cfs_rq->runnable_load_sum += sa->load_sum;
2882 attach_entity_load_avg(cfs_rq, se);
2884 if (decayed || migrated)
2885 update_tg_load_avg(cfs_rq, 0);
2888 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2890 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2892 update_load_avg(se, 1);
2894 cfs_rq->runnable_load_avg =
2895 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2896 cfs_rq->runnable_load_sum =
2897 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2900 #ifndef CONFIG_64BIT
2901 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2903 u64 last_update_time_copy;
2904 u64 last_update_time;
2907 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2909 last_update_time = cfs_rq->avg.last_update_time;
2910 } while (last_update_time != last_update_time_copy);
2912 return last_update_time;
2915 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2917 return cfs_rq->avg.last_update_time;
2922 * Task first catches up with cfs_rq, and then subtract
2923 * itself from the cfs_rq (task must be off the queue now).
2925 void remove_entity_load_avg(struct sched_entity *se)
2927 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2928 u64 last_update_time;
2931 * Newly created task or never used group entity should not be removed
2932 * from its (source) cfs_rq
2934 if (se->avg.last_update_time == 0)
2937 last_update_time = cfs_rq_last_update_time(cfs_rq);
2939 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2940 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2941 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2945 * Update the rq's load with the elapsed running time before entering
2946 * idle. if the last scheduled task is not a CFS task, idle_enter will
2947 * be the only way to update the runnable statistic.
2949 void idle_enter_fair(struct rq *this_rq)
2954 * Update the rq's load with the elapsed idle time before a task is
2955 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2956 * be the only way to update the runnable statistic.
2958 void idle_exit_fair(struct rq *this_rq)
2962 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2964 return cfs_rq->runnable_load_avg;
2967 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2969 return cfs_rq->avg.load_avg;
2972 static int idle_balance(struct rq *this_rq);
2974 #else /* CONFIG_SMP */
2976 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2978 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
2982 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2984 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2985 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2988 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2990 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2992 static inline int idle_balance(struct rq *rq)
2997 #endif /* CONFIG_SMP */
2999 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3001 #ifdef CONFIG_SCHEDSTATS
3002 struct task_struct *tsk = NULL;
3004 if (entity_is_task(se))
3007 if (se->statistics.sleep_start) {
3008 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3013 if (unlikely(delta > se->statistics.sleep_max))
3014 se->statistics.sleep_max = delta;
3016 se->statistics.sleep_start = 0;
3017 se->statistics.sum_sleep_runtime += delta;
3020 account_scheduler_latency(tsk, delta >> 10, 1);
3021 trace_sched_stat_sleep(tsk, delta);
3024 if (se->statistics.block_start) {
3025 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3030 if (unlikely(delta > se->statistics.block_max))
3031 se->statistics.block_max = delta;
3033 se->statistics.block_start = 0;
3034 se->statistics.sum_sleep_runtime += delta;
3037 if (tsk->in_iowait) {
3038 se->statistics.iowait_sum += delta;
3039 se->statistics.iowait_count++;
3040 trace_sched_stat_iowait(tsk, delta);
3043 trace_sched_stat_blocked(tsk, delta);
3044 trace_sched_blocked_reason(tsk);
3047 * Blocking time is in units of nanosecs, so shift by
3048 * 20 to get a milliseconds-range estimation of the
3049 * amount of time that the task spent sleeping:
3051 if (unlikely(prof_on == SLEEP_PROFILING)) {
3052 profile_hits(SLEEP_PROFILING,
3053 (void *)get_wchan(tsk),
3056 account_scheduler_latency(tsk, delta >> 10, 0);
3062 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3064 #ifdef CONFIG_SCHED_DEBUG
3065 s64 d = se->vruntime - cfs_rq->min_vruntime;
3070 if (d > 3*sysctl_sched_latency)
3071 schedstat_inc(cfs_rq, nr_spread_over);
3076 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3078 u64 vruntime = cfs_rq->min_vruntime;
3081 * The 'current' period is already promised to the current tasks,
3082 * however the extra weight of the new task will slow them down a
3083 * little, place the new task so that it fits in the slot that
3084 * stays open at the end.
3086 if (initial && sched_feat(START_DEBIT))
3087 vruntime += sched_vslice(cfs_rq, se);
3089 /* sleeps up to a single latency don't count. */
3091 unsigned long thresh = sysctl_sched_latency;
3094 * Halve their sleep time's effect, to allow
3095 * for a gentler effect of sleepers:
3097 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3103 /* ensure we never gain time by being placed backwards. */
3104 se->vruntime = max_vruntime(se->vruntime, vruntime);
3107 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3110 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3113 * Update the normalized vruntime before updating min_vruntime
3114 * through calling update_curr().
3116 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3117 se->vruntime += cfs_rq->min_vruntime;
3120 * Update run-time statistics of the 'current'.
3122 update_curr(cfs_rq);
3123 enqueue_entity_load_avg(cfs_rq, se);
3124 account_entity_enqueue(cfs_rq, se);
3125 update_cfs_shares(cfs_rq);
3127 if (flags & ENQUEUE_WAKEUP) {
3128 place_entity(cfs_rq, se, 0);
3129 enqueue_sleeper(cfs_rq, se);
3132 update_stats_enqueue(cfs_rq, se);
3133 check_spread(cfs_rq, se);
3134 if (se != cfs_rq->curr)
3135 __enqueue_entity(cfs_rq, se);
3138 if (cfs_rq->nr_running == 1) {
3139 list_add_leaf_cfs_rq(cfs_rq);
3140 check_enqueue_throttle(cfs_rq);
3144 static void __clear_buddies_last(struct sched_entity *se)
3146 for_each_sched_entity(se) {
3147 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3148 if (cfs_rq->last != se)
3151 cfs_rq->last = NULL;
3155 static void __clear_buddies_next(struct sched_entity *se)
3157 for_each_sched_entity(se) {
3158 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3159 if (cfs_rq->next != se)
3162 cfs_rq->next = NULL;
3166 static void __clear_buddies_skip(struct sched_entity *se)
3168 for_each_sched_entity(se) {
3169 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3170 if (cfs_rq->skip != se)
3173 cfs_rq->skip = NULL;
3177 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3179 if (cfs_rq->last == se)
3180 __clear_buddies_last(se);
3182 if (cfs_rq->next == se)
3183 __clear_buddies_next(se);
3185 if (cfs_rq->skip == se)
3186 __clear_buddies_skip(se);
3189 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3192 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3195 * Update run-time statistics of the 'current'.
3197 update_curr(cfs_rq);
3198 dequeue_entity_load_avg(cfs_rq, se);
3200 update_stats_dequeue(cfs_rq, se);
3201 if (flags & DEQUEUE_SLEEP) {
3202 #ifdef CONFIG_SCHEDSTATS
3203 if (entity_is_task(se)) {
3204 struct task_struct *tsk = task_of(se);
3206 if (tsk->state & TASK_INTERRUPTIBLE)
3207 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3208 if (tsk->state & TASK_UNINTERRUPTIBLE)
3209 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3214 clear_buddies(cfs_rq, se);
3216 if (se != cfs_rq->curr)
3217 __dequeue_entity(cfs_rq, se);
3219 account_entity_dequeue(cfs_rq, se);
3222 * Normalize the entity after updating the min_vruntime because the
3223 * update can refer to the ->curr item and we need to reflect this
3224 * movement in our normalized position.
3226 if (!(flags & DEQUEUE_SLEEP))
3227 se->vruntime -= cfs_rq->min_vruntime;
3229 /* return excess runtime on last dequeue */
3230 return_cfs_rq_runtime(cfs_rq);
3232 update_min_vruntime(cfs_rq);
3233 update_cfs_shares(cfs_rq);
3237 * Preempt the current task with a newly woken task if needed:
3240 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3242 unsigned long ideal_runtime, delta_exec;
3243 struct sched_entity *se;
3246 ideal_runtime = sched_slice(cfs_rq, curr);
3247 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3248 if (delta_exec > ideal_runtime) {
3249 resched_curr(rq_of(cfs_rq));
3251 * The current task ran long enough, ensure it doesn't get
3252 * re-elected due to buddy favours.
3254 clear_buddies(cfs_rq, curr);
3259 * Ensure that a task that missed wakeup preemption by a
3260 * narrow margin doesn't have to wait for a full slice.
3261 * This also mitigates buddy induced latencies under load.
3263 if (delta_exec < sysctl_sched_min_granularity)
3266 se = __pick_first_entity(cfs_rq);
3267 delta = curr->vruntime - se->vruntime;
3272 if (delta > ideal_runtime)
3273 resched_curr(rq_of(cfs_rq));
3277 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3279 /* 'current' is not kept within the tree. */
3282 * Any task has to be enqueued before it get to execute on
3283 * a CPU. So account for the time it spent waiting on the
3286 update_stats_wait_end(cfs_rq, se);
3287 __dequeue_entity(cfs_rq, se);
3288 update_load_avg(se, 1);
3291 update_stats_curr_start(cfs_rq, se);
3293 #ifdef CONFIG_SCHEDSTATS
3295 * Track our maximum slice length, if the CPU's load is at
3296 * least twice that of our own weight (i.e. dont track it
3297 * when there are only lesser-weight tasks around):
3299 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3300 se->statistics.slice_max = max(se->statistics.slice_max,
3301 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3304 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3308 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3311 * Pick the next process, keeping these things in mind, in this order:
3312 * 1) keep things fair between processes/task groups
3313 * 2) pick the "next" process, since someone really wants that to run
3314 * 3) pick the "last" process, for cache locality
3315 * 4) do not run the "skip" process, if something else is available
3317 static struct sched_entity *
3318 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3320 struct sched_entity *left = __pick_first_entity(cfs_rq);
3321 struct sched_entity *se;
3324 * If curr is set we have to see if its left of the leftmost entity
3325 * still in the tree, provided there was anything in the tree at all.
3327 if (!left || (curr && entity_before(curr, left)))
3330 se = left; /* ideally we run the leftmost entity */
3333 * Avoid running the skip buddy, if running something else can
3334 * be done without getting too unfair.
3336 if (cfs_rq->skip == se) {
3337 struct sched_entity *second;
3340 second = __pick_first_entity(cfs_rq);
3342 second = __pick_next_entity(se);
3343 if (!second || (curr && entity_before(curr, second)))
3347 if (second && wakeup_preempt_entity(second, left) < 1)
3352 * Prefer last buddy, try to return the CPU to a preempted task.
3354 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3358 * Someone really wants this to run. If it's not unfair, run it.
3360 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3363 clear_buddies(cfs_rq, se);
3368 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3370 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3373 * If still on the runqueue then deactivate_task()
3374 * was not called and update_curr() has to be done:
3377 update_curr(cfs_rq);
3379 /* throttle cfs_rqs exceeding runtime */
3380 check_cfs_rq_runtime(cfs_rq);
3382 check_spread(cfs_rq, prev);
3384 update_stats_wait_start(cfs_rq, prev);
3385 /* Put 'current' back into the tree. */
3386 __enqueue_entity(cfs_rq, prev);
3387 /* in !on_rq case, update occurred at dequeue */
3388 update_load_avg(prev, 0);
3390 cfs_rq->curr = NULL;
3394 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3397 * Update run-time statistics of the 'current'.
3399 update_curr(cfs_rq);
3402 * Ensure that runnable average is periodically updated.
3404 update_load_avg(curr, 1);
3405 update_cfs_shares(cfs_rq);
3407 #ifdef CONFIG_SCHED_HRTICK
3409 * queued ticks are scheduled to match the slice, so don't bother
3410 * validating it and just reschedule.
3413 resched_curr(rq_of(cfs_rq));
3417 * don't let the period tick interfere with the hrtick preemption
3419 if (!sched_feat(DOUBLE_TICK) &&
3420 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3424 if (cfs_rq->nr_running > 1)
3425 check_preempt_tick(cfs_rq, curr);
3429 /**************************************************
3430 * CFS bandwidth control machinery
3433 #ifdef CONFIG_CFS_BANDWIDTH
3435 #ifdef HAVE_JUMP_LABEL
3436 static struct static_key __cfs_bandwidth_used;
3438 static inline bool cfs_bandwidth_used(void)
3440 return static_key_false(&__cfs_bandwidth_used);
3443 void cfs_bandwidth_usage_inc(void)
3445 static_key_slow_inc(&__cfs_bandwidth_used);
3448 void cfs_bandwidth_usage_dec(void)
3450 static_key_slow_dec(&__cfs_bandwidth_used);
3452 #else /* HAVE_JUMP_LABEL */
3453 static bool cfs_bandwidth_used(void)
3458 void cfs_bandwidth_usage_inc(void) {}
3459 void cfs_bandwidth_usage_dec(void) {}
3460 #endif /* HAVE_JUMP_LABEL */
3463 * default period for cfs group bandwidth.
3464 * default: 0.1s, units: nanoseconds
3466 static inline u64 default_cfs_period(void)
3468 return 100000000ULL;
3471 static inline u64 sched_cfs_bandwidth_slice(void)
3473 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3477 * Replenish runtime according to assigned quota and update expiration time.
3478 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3479 * additional synchronization around rq->lock.
3481 * requires cfs_b->lock
3483 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3487 if (cfs_b->quota == RUNTIME_INF)
3490 now = sched_clock_cpu(smp_processor_id());
3491 cfs_b->runtime = cfs_b->quota;
3492 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3495 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3497 return &tg->cfs_bandwidth;
3500 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3501 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3503 if (unlikely(cfs_rq->throttle_count))
3504 return cfs_rq->throttled_clock_task;
3506 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3509 /* returns 0 on failure to allocate runtime */
3510 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3512 struct task_group *tg = cfs_rq->tg;
3513 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3514 u64 amount = 0, min_amount, expires;
3516 /* note: this is a positive sum as runtime_remaining <= 0 */
3517 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3519 raw_spin_lock(&cfs_b->lock);
3520 if (cfs_b->quota == RUNTIME_INF)
3521 amount = min_amount;
3523 start_cfs_bandwidth(cfs_b);
3525 if (cfs_b->runtime > 0) {
3526 amount = min(cfs_b->runtime, min_amount);
3527 cfs_b->runtime -= amount;
3531 expires = cfs_b->runtime_expires;
3532 raw_spin_unlock(&cfs_b->lock);
3534 cfs_rq->runtime_remaining += amount;
3536 * we may have advanced our local expiration to account for allowed
3537 * spread between our sched_clock and the one on which runtime was
3540 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3541 cfs_rq->runtime_expires = expires;
3543 return cfs_rq->runtime_remaining > 0;
3547 * Note: This depends on the synchronization provided by sched_clock and the
3548 * fact that rq->clock snapshots this value.
3550 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3552 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3554 /* if the deadline is ahead of our clock, nothing to do */
3555 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3558 if (cfs_rq->runtime_remaining < 0)
3562 * If the local deadline has passed we have to consider the
3563 * possibility that our sched_clock is 'fast' and the global deadline
3564 * has not truly expired.
3566 * Fortunately we can check determine whether this the case by checking
3567 * whether the global deadline has advanced. It is valid to compare
3568 * cfs_b->runtime_expires without any locks since we only care about
3569 * exact equality, so a partial write will still work.
3572 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3573 /* extend local deadline, drift is bounded above by 2 ticks */
3574 cfs_rq->runtime_expires += TICK_NSEC;
3576 /* global deadline is ahead, expiration has passed */
3577 cfs_rq->runtime_remaining = 0;
3581 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3583 /* dock delta_exec before expiring quota (as it could span periods) */
3584 cfs_rq->runtime_remaining -= delta_exec;
3585 expire_cfs_rq_runtime(cfs_rq);
3587 if (likely(cfs_rq->runtime_remaining > 0))
3591 * if we're unable to extend our runtime we resched so that the active
3592 * hierarchy can be throttled
3594 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3595 resched_curr(rq_of(cfs_rq));
3598 static __always_inline
3599 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3601 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3604 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3607 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3609 return cfs_bandwidth_used() && cfs_rq->throttled;
3612 /* check whether cfs_rq, or any parent, is throttled */
3613 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3615 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3619 * Ensure that neither of the group entities corresponding to src_cpu or
3620 * dest_cpu are members of a throttled hierarchy when performing group
3621 * load-balance operations.
3623 static inline int throttled_lb_pair(struct task_group *tg,
3624 int src_cpu, int dest_cpu)
3626 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3628 src_cfs_rq = tg->cfs_rq[src_cpu];
3629 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3631 return throttled_hierarchy(src_cfs_rq) ||
3632 throttled_hierarchy(dest_cfs_rq);
3635 /* updated child weight may affect parent so we have to do this bottom up */
3636 static int tg_unthrottle_up(struct task_group *tg, void *data)
3638 struct rq *rq = data;
3639 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3641 cfs_rq->throttle_count--;
3643 if (!cfs_rq->throttle_count) {
3644 /* adjust cfs_rq_clock_task() */
3645 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3646 cfs_rq->throttled_clock_task;
3653 static int tg_throttle_down(struct task_group *tg, void *data)
3655 struct rq *rq = data;
3656 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3658 /* group is entering throttled state, stop time */
3659 if (!cfs_rq->throttle_count)
3660 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3661 cfs_rq->throttle_count++;
3666 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3668 struct rq *rq = rq_of(cfs_rq);
3669 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3670 struct sched_entity *se;
3671 long task_delta, dequeue = 1;
3674 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3676 /* freeze hierarchy runnable averages while throttled */
3678 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3681 task_delta = cfs_rq->h_nr_running;
3682 for_each_sched_entity(se) {
3683 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3684 /* throttled entity or throttle-on-deactivate */
3689 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3690 qcfs_rq->h_nr_running -= task_delta;
3692 if (qcfs_rq->load.weight)
3697 sub_nr_running(rq, task_delta);
3699 cfs_rq->throttled = 1;
3700 cfs_rq->throttled_clock = rq_clock(rq);
3701 raw_spin_lock(&cfs_b->lock);
3702 empty = list_empty(&cfs_b->throttled_cfs_rq);
3705 * Add to the _head_ of the list, so that an already-started
3706 * distribute_cfs_runtime will not see us
3708 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3711 * If we're the first throttled task, make sure the bandwidth
3715 start_cfs_bandwidth(cfs_b);
3717 raw_spin_unlock(&cfs_b->lock);
3720 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3722 struct rq *rq = rq_of(cfs_rq);
3723 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3724 struct sched_entity *se;
3728 se = cfs_rq->tg->se[cpu_of(rq)];
3730 cfs_rq->throttled = 0;
3732 update_rq_clock(rq);
3734 raw_spin_lock(&cfs_b->lock);
3735 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3736 list_del_rcu(&cfs_rq->throttled_list);
3737 raw_spin_unlock(&cfs_b->lock);
3739 /* update hierarchical throttle state */
3740 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3742 if (!cfs_rq->load.weight)
3745 task_delta = cfs_rq->h_nr_running;
3746 for_each_sched_entity(se) {
3750 cfs_rq = cfs_rq_of(se);
3752 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3753 cfs_rq->h_nr_running += task_delta;
3755 if (cfs_rq_throttled(cfs_rq))
3760 add_nr_running(rq, task_delta);
3762 /* determine whether we need to wake up potentially idle cpu */
3763 if (rq->curr == rq->idle && rq->cfs.nr_running)
3767 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3768 u64 remaining, u64 expires)
3770 struct cfs_rq *cfs_rq;
3772 u64 starting_runtime = remaining;
3775 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3777 struct rq *rq = rq_of(cfs_rq);
3779 raw_spin_lock(&rq->lock);
3780 if (!cfs_rq_throttled(cfs_rq))
3783 runtime = -cfs_rq->runtime_remaining + 1;
3784 if (runtime > remaining)
3785 runtime = remaining;
3786 remaining -= runtime;
3788 cfs_rq->runtime_remaining += runtime;
3789 cfs_rq->runtime_expires = expires;
3791 /* we check whether we're throttled above */
3792 if (cfs_rq->runtime_remaining > 0)
3793 unthrottle_cfs_rq(cfs_rq);
3796 raw_spin_unlock(&rq->lock);
3803 return starting_runtime - remaining;
3807 * Responsible for refilling a task_group's bandwidth and unthrottling its
3808 * cfs_rqs as appropriate. If there has been no activity within the last
3809 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3810 * used to track this state.
3812 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3814 u64 runtime, runtime_expires;
3817 /* no need to continue the timer with no bandwidth constraint */
3818 if (cfs_b->quota == RUNTIME_INF)
3819 goto out_deactivate;
3821 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3822 cfs_b->nr_periods += overrun;
3825 * idle depends on !throttled (for the case of a large deficit), and if
3826 * we're going inactive then everything else can be deferred
3828 if (cfs_b->idle && !throttled)
3829 goto out_deactivate;
3831 __refill_cfs_bandwidth_runtime(cfs_b);
3834 /* mark as potentially idle for the upcoming period */
3839 /* account preceding periods in which throttling occurred */
3840 cfs_b->nr_throttled += overrun;
3842 runtime_expires = cfs_b->runtime_expires;
3845 * This check is repeated as we are holding onto the new bandwidth while
3846 * we unthrottle. This can potentially race with an unthrottled group
3847 * trying to acquire new bandwidth from the global pool. This can result
3848 * in us over-using our runtime if it is all used during this loop, but
3849 * only by limited amounts in that extreme case.
3851 while (throttled && cfs_b->runtime > 0) {
3852 runtime = cfs_b->runtime;
3853 raw_spin_unlock(&cfs_b->lock);
3854 /* we can't nest cfs_b->lock while distributing bandwidth */
3855 runtime = distribute_cfs_runtime(cfs_b, runtime,
3857 raw_spin_lock(&cfs_b->lock);
3859 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3861 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3865 * While we are ensured activity in the period following an
3866 * unthrottle, this also covers the case in which the new bandwidth is
3867 * insufficient to cover the existing bandwidth deficit. (Forcing the
3868 * timer to remain active while there are any throttled entities.)
3878 /* a cfs_rq won't donate quota below this amount */
3879 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3880 /* minimum remaining period time to redistribute slack quota */
3881 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3882 /* how long we wait to gather additional slack before distributing */
3883 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3886 * Are we near the end of the current quota period?
3888 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3889 * hrtimer base being cleared by hrtimer_start. In the case of
3890 * migrate_hrtimers, base is never cleared, so we are fine.
3892 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3894 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3897 /* if the call-back is running a quota refresh is already occurring */
3898 if (hrtimer_callback_running(refresh_timer))
3901 /* is a quota refresh about to occur? */
3902 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3903 if (remaining < min_expire)
3909 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3911 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3913 /* if there's a quota refresh soon don't bother with slack */
3914 if (runtime_refresh_within(cfs_b, min_left))
3917 hrtimer_start(&cfs_b->slack_timer,
3918 ns_to_ktime(cfs_bandwidth_slack_period),
3922 /* we know any runtime found here is valid as update_curr() precedes return */
3923 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3925 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3926 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3928 if (slack_runtime <= 0)
3931 raw_spin_lock(&cfs_b->lock);
3932 if (cfs_b->quota != RUNTIME_INF &&
3933 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3934 cfs_b->runtime += slack_runtime;
3936 /* we are under rq->lock, defer unthrottling using a timer */
3937 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3938 !list_empty(&cfs_b->throttled_cfs_rq))
3939 start_cfs_slack_bandwidth(cfs_b);
3941 raw_spin_unlock(&cfs_b->lock);
3943 /* even if it's not valid for return we don't want to try again */
3944 cfs_rq->runtime_remaining -= slack_runtime;
3947 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3949 if (!cfs_bandwidth_used())
3952 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3955 __return_cfs_rq_runtime(cfs_rq);
3959 * This is done with a timer (instead of inline with bandwidth return) since
3960 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3962 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3964 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3967 /* confirm we're still not at a refresh boundary */
3968 raw_spin_lock(&cfs_b->lock);
3969 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3970 raw_spin_unlock(&cfs_b->lock);
3974 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3975 runtime = cfs_b->runtime;
3977 expires = cfs_b->runtime_expires;
3978 raw_spin_unlock(&cfs_b->lock);
3983 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3985 raw_spin_lock(&cfs_b->lock);
3986 if (expires == cfs_b->runtime_expires)
3987 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3988 raw_spin_unlock(&cfs_b->lock);
3992 * When a group wakes up we want to make sure that its quota is not already
3993 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3994 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3996 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3998 if (!cfs_bandwidth_used())
4001 /* Synchronize hierarchical throttle counter: */
4002 if (unlikely(!cfs_rq->throttle_uptodate)) {
4003 struct rq *rq = rq_of(cfs_rq);
4004 struct cfs_rq *pcfs_rq;
4005 struct task_group *tg;
4007 cfs_rq->throttle_uptodate = 1;
4009 /* Get closest up-to-date node, because leaves go first: */
4010 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4011 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4012 if (pcfs_rq->throttle_uptodate)
4016 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4017 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4021 /* an active group must be handled by the update_curr()->put() path */
4022 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4025 /* ensure the group is not already throttled */
4026 if (cfs_rq_throttled(cfs_rq))
4029 /* update runtime allocation */
4030 account_cfs_rq_runtime(cfs_rq, 0);
4031 if (cfs_rq->runtime_remaining <= 0)
4032 throttle_cfs_rq(cfs_rq);
4035 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4036 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4038 if (!cfs_bandwidth_used())
4041 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4045 * it's possible for a throttled entity to be forced into a running
4046 * state (e.g. set_curr_task), in this case we're finished.
4048 if (cfs_rq_throttled(cfs_rq))
4051 throttle_cfs_rq(cfs_rq);
4055 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4057 struct cfs_bandwidth *cfs_b =
4058 container_of(timer, struct cfs_bandwidth, slack_timer);
4060 do_sched_cfs_slack_timer(cfs_b);
4062 return HRTIMER_NORESTART;
4065 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4067 struct cfs_bandwidth *cfs_b =
4068 container_of(timer, struct cfs_bandwidth, period_timer);
4072 raw_spin_lock(&cfs_b->lock);
4074 overrun = hrtimer_forward_now(timer, cfs_b->period);
4078 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4081 cfs_b->period_active = 0;
4082 raw_spin_unlock(&cfs_b->lock);
4084 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4087 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4089 raw_spin_lock_init(&cfs_b->lock);
4091 cfs_b->quota = RUNTIME_INF;
4092 cfs_b->period = ns_to_ktime(default_cfs_period());
4094 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4095 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4096 cfs_b->period_timer.function = sched_cfs_period_timer;
4097 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4098 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4101 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4103 cfs_rq->runtime_enabled = 0;
4104 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4107 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4109 lockdep_assert_held(&cfs_b->lock);
4111 if (!cfs_b->period_active) {
4112 cfs_b->period_active = 1;
4113 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4114 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4118 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4120 /* init_cfs_bandwidth() was not called */
4121 if (!cfs_b->throttled_cfs_rq.next)
4124 hrtimer_cancel(&cfs_b->period_timer);
4125 hrtimer_cancel(&cfs_b->slack_timer);
4128 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4130 struct cfs_rq *cfs_rq;
4132 for_each_leaf_cfs_rq(rq, cfs_rq) {
4133 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4135 raw_spin_lock(&cfs_b->lock);
4136 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4137 raw_spin_unlock(&cfs_b->lock);
4141 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4143 struct cfs_rq *cfs_rq;
4145 for_each_leaf_cfs_rq(rq, cfs_rq) {
4146 if (!cfs_rq->runtime_enabled)
4150 * clock_task is not advancing so we just need to make sure
4151 * there's some valid quota amount
4153 cfs_rq->runtime_remaining = 1;
4155 * Offline rq is schedulable till cpu is completely disabled
4156 * in take_cpu_down(), so we prevent new cfs throttling here.
4158 cfs_rq->runtime_enabled = 0;
4160 if (cfs_rq_throttled(cfs_rq))
4161 unthrottle_cfs_rq(cfs_rq);
4165 #else /* CONFIG_CFS_BANDWIDTH */
4166 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4168 return rq_clock_task(rq_of(cfs_rq));
4171 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4172 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4173 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4174 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4176 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4181 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4186 static inline int throttled_lb_pair(struct task_group *tg,
4187 int src_cpu, int dest_cpu)
4192 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4194 #ifdef CONFIG_FAIR_GROUP_SCHED
4195 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4198 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4202 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4203 static inline void update_runtime_enabled(struct rq *rq) {}
4204 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4206 #endif /* CONFIG_CFS_BANDWIDTH */
4208 /**************************************************
4209 * CFS operations on tasks:
4212 #ifdef CONFIG_SCHED_HRTICK
4213 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4215 struct sched_entity *se = &p->se;
4216 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4218 WARN_ON(task_rq(p) != rq);
4220 if (cfs_rq->nr_running > 1) {
4221 u64 slice = sched_slice(cfs_rq, se);
4222 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4223 s64 delta = slice - ran;
4230 hrtick_start(rq, delta);
4235 * called from enqueue/dequeue and updates the hrtick when the
4236 * current task is from our class and nr_running is low enough
4239 static void hrtick_update(struct rq *rq)
4241 struct task_struct *curr = rq->curr;
4243 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4246 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4247 hrtick_start_fair(rq, curr);
4249 #else /* !CONFIG_SCHED_HRTICK */
4251 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4255 static inline void hrtick_update(struct rq *rq)
4261 static bool cpu_overutilized(int cpu);
4262 unsigned long boosted_cpu_util(int cpu);
4264 #define boosted_cpu_util(cpu) cpu_util(cpu)
4268 static void update_capacity_of(int cpu)
4270 unsigned long req_cap;
4275 /* Convert scale-invariant capacity to cpu. */
4276 req_cap = boosted_cpu_util(cpu);
4277 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4278 set_cfs_cpu_capacity(cpu, true, req_cap);
4283 * The enqueue_task method is called before nr_running is
4284 * increased. Here we update the fair scheduling stats and
4285 * then put the task into the rbtree:
4288 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4290 struct cfs_rq *cfs_rq;
4291 struct sched_entity *se = &p->se;
4293 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4294 int task_wakeup = flags & ENQUEUE_WAKEUP;
4298 * If in_iowait is set, the code below may not trigger any cpufreq
4299 * utilization updates, so do it here explicitly with the IOWAIT flag
4303 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4305 for_each_sched_entity(se) {
4308 cfs_rq = cfs_rq_of(se);
4309 enqueue_entity(cfs_rq, se, flags);
4312 * end evaluation on encountering a throttled cfs_rq
4314 * note: in the case of encountering a throttled cfs_rq we will
4315 * post the final h_nr_running increment below.
4317 if (cfs_rq_throttled(cfs_rq))
4319 cfs_rq->h_nr_running++;
4320 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4322 flags = ENQUEUE_WAKEUP;
4325 for_each_sched_entity(se) {
4326 cfs_rq = cfs_rq_of(se);
4327 cfs_rq->h_nr_running++;
4328 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4330 if (cfs_rq_throttled(cfs_rq))
4333 update_load_avg(se, 1);
4334 update_cfs_shares(cfs_rq);
4338 add_nr_running(rq, 1);
4343 * Update SchedTune accounting.
4345 * We do it before updating the CPU capacity to ensure the
4346 * boost value of the current task is accounted for in the
4347 * selection of the OPP.
4349 * We do it also in the case where we enqueue a throttled task;
4350 * we could argue that a throttled task should not boost a CPU,
4352 * a) properly implementing CPU boosting considering throttled
4353 * tasks will increase a lot the complexity of the solution
4354 * b) it's not easy to quantify the benefits introduced by
4355 * such a more complex solution.
4356 * Thus, for the time being we go for the simple solution and boost
4357 * also for throttled RQs.
4359 schedtune_enqueue_task(p, cpu_of(rq));
4362 walt_inc_cumulative_runnable_avg(rq, p);
4363 if (!task_new && !rq->rd->overutilized &&
4364 cpu_overutilized(rq->cpu)) {
4365 rq->rd->overutilized = true;
4366 trace_sched_overutilized(true);
4370 * We want to potentially trigger a freq switch
4371 * request only for tasks that are waking up; this is
4372 * because we get here also during load balancing, but
4373 * in these cases it seems wise to trigger as single
4374 * request after load balancing is done.
4376 if (task_new || task_wakeup)
4377 update_capacity_of(cpu_of(rq));
4380 #endif /* CONFIG_SMP */
4384 static void set_next_buddy(struct sched_entity *se);
4387 * The dequeue_task method is called before nr_running is
4388 * decreased. We remove the task from the rbtree and
4389 * update the fair scheduling stats:
4391 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4393 struct cfs_rq *cfs_rq;
4394 struct sched_entity *se = &p->se;
4395 int task_sleep = flags & DEQUEUE_SLEEP;
4397 for_each_sched_entity(se) {
4398 cfs_rq = cfs_rq_of(se);
4399 dequeue_entity(cfs_rq, se, flags);
4402 * end evaluation on encountering a throttled cfs_rq
4404 * note: in the case of encountering a throttled cfs_rq we will
4405 * post the final h_nr_running decrement below.
4407 if (cfs_rq_throttled(cfs_rq))
4409 cfs_rq->h_nr_running--;
4410 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4412 /* Don't dequeue parent if it has other entities besides us */
4413 if (cfs_rq->load.weight) {
4414 /* Avoid re-evaluating load for this entity: */
4415 se = parent_entity(se);
4417 * Bias pick_next to pick a task from this cfs_rq, as
4418 * p is sleeping when it is within its sched_slice.
4420 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4424 flags |= DEQUEUE_SLEEP;
4427 for_each_sched_entity(se) {
4428 cfs_rq = cfs_rq_of(se);
4429 cfs_rq->h_nr_running--;
4430 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4432 if (cfs_rq_throttled(cfs_rq))
4435 update_load_avg(se, 1);
4436 update_cfs_shares(cfs_rq);
4440 sub_nr_running(rq, 1);
4445 * Update SchedTune accounting
4447 * We do it before updating the CPU capacity to ensure the
4448 * boost value of the current task is accounted for in the
4449 * selection of the OPP.
4451 schedtune_dequeue_task(p, cpu_of(rq));
4454 walt_dec_cumulative_runnable_avg(rq, p);
4457 * We want to potentially trigger a freq switch
4458 * request only for tasks that are going to sleep;
4459 * this is because we get here also during load
4460 * balancing, but in these cases it seems wise to
4461 * trigger as single request after load balancing is
4465 if (rq->cfs.nr_running)
4466 update_capacity_of(cpu_of(rq));
4467 else if (sched_freq())
4468 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4472 #endif /* CONFIG_SMP */
4480 * per rq 'load' arrray crap; XXX kill this.
4484 * The exact cpuload at various idx values, calculated at every tick would be
4485 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4487 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4488 * on nth tick when cpu may be busy, then we have:
4489 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4490 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4492 * decay_load_missed() below does efficient calculation of
4493 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4494 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4496 * The calculation is approximated on a 128 point scale.
4497 * degrade_zero_ticks is the number of ticks after which load at any
4498 * particular idx is approximated to be zero.
4499 * degrade_factor is a precomputed table, a row for each load idx.
4500 * Each column corresponds to degradation factor for a power of two ticks,
4501 * based on 128 point scale.
4503 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4504 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4506 * With this power of 2 load factors, we can degrade the load n times
4507 * by looking at 1 bits in n and doing as many mult/shift instead of
4508 * n mult/shifts needed by the exact degradation.
4510 #define DEGRADE_SHIFT 7
4511 static const unsigned char
4512 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4513 static const unsigned char
4514 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4515 {0, 0, 0, 0, 0, 0, 0, 0},
4516 {64, 32, 8, 0, 0, 0, 0, 0},
4517 {96, 72, 40, 12, 1, 0, 0},
4518 {112, 98, 75, 43, 15, 1, 0},
4519 {120, 112, 98, 76, 45, 16, 2} };
4522 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4523 * would be when CPU is idle and so we just decay the old load without
4524 * adding any new load.
4526 static unsigned long
4527 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4531 if (!missed_updates)
4534 if (missed_updates >= degrade_zero_ticks[idx])
4538 return load >> missed_updates;
4540 while (missed_updates) {
4541 if (missed_updates % 2)
4542 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4544 missed_updates >>= 1;
4551 * Update rq->cpu_load[] statistics. This function is usually called every
4552 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4553 * every tick. We fix it up based on jiffies.
4555 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4556 unsigned long pending_updates)
4560 this_rq->nr_load_updates++;
4562 /* Update our load: */
4563 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4564 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4565 unsigned long old_load, new_load;
4567 /* scale is effectively 1 << i now, and >> i divides by scale */
4569 old_load = this_rq->cpu_load[i];
4570 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4571 new_load = this_load;
4573 * Round up the averaging division if load is increasing. This
4574 * prevents us from getting stuck on 9 if the load is 10, for
4577 if (new_load > old_load)
4578 new_load += scale - 1;
4580 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4583 sched_avg_update(this_rq);
4586 /* Used instead of source_load when we know the type == 0 */
4587 static unsigned long weighted_cpuload(const int cpu)
4589 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4592 #ifdef CONFIG_NO_HZ_COMMON
4594 * There is no sane way to deal with nohz on smp when using jiffies because the
4595 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4596 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4598 * Therefore we cannot use the delta approach from the regular tick since that
4599 * would seriously skew the load calculation. However we'll make do for those
4600 * updates happening while idle (nohz_idle_balance) or coming out of idle
4601 * (tick_nohz_idle_exit).
4603 * This means we might still be one tick off for nohz periods.
4607 * Called from nohz_idle_balance() to update the load ratings before doing the
4610 static void update_idle_cpu_load(struct rq *this_rq)
4612 unsigned long curr_jiffies = READ_ONCE(jiffies);
4613 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4614 unsigned long pending_updates;
4617 * bail if there's load or we're actually up-to-date.
4619 if (load || curr_jiffies == this_rq->last_load_update_tick)
4622 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4623 this_rq->last_load_update_tick = curr_jiffies;
4625 __update_cpu_load(this_rq, load, pending_updates);
4629 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4631 void update_cpu_load_nohz(void)
4633 struct rq *this_rq = this_rq();
4634 unsigned long curr_jiffies = READ_ONCE(jiffies);
4635 unsigned long pending_updates;
4637 if (curr_jiffies == this_rq->last_load_update_tick)
4640 raw_spin_lock(&this_rq->lock);
4641 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4642 if (pending_updates) {
4643 this_rq->last_load_update_tick = curr_jiffies;
4645 * We were idle, this means load 0, the current load might be
4646 * !0 due to remote wakeups and the sort.
4648 __update_cpu_load(this_rq, 0, pending_updates);
4650 raw_spin_unlock(&this_rq->lock);
4652 #endif /* CONFIG_NO_HZ */
4655 * Called from scheduler_tick()
4657 void update_cpu_load_active(struct rq *this_rq)
4659 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4661 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4663 this_rq->last_load_update_tick = jiffies;
4664 __update_cpu_load(this_rq, load, 1);
4668 * Return a low guess at the load of a migration-source cpu weighted
4669 * according to the scheduling class and "nice" value.
4671 * We want to under-estimate the load of migration sources, to
4672 * balance conservatively.
4674 static unsigned long source_load(int cpu, int type)
4676 struct rq *rq = cpu_rq(cpu);
4677 unsigned long total = weighted_cpuload(cpu);
4679 if (type == 0 || !sched_feat(LB_BIAS))
4682 return min(rq->cpu_load[type-1], total);
4686 * Return a high guess at the load of a migration-target cpu weighted
4687 * according to the scheduling class and "nice" value.
4689 static unsigned long target_load(int cpu, int type)
4691 struct rq *rq = cpu_rq(cpu);
4692 unsigned long total = weighted_cpuload(cpu);
4694 if (type == 0 || !sched_feat(LB_BIAS))
4697 return max(rq->cpu_load[type-1], total);
4701 static unsigned long cpu_avg_load_per_task(int cpu)
4703 struct rq *rq = cpu_rq(cpu);
4704 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4705 unsigned long load_avg = weighted_cpuload(cpu);
4708 return load_avg / nr_running;
4713 static void record_wakee(struct task_struct *p)
4716 * Rough decay (wiping) for cost saving, don't worry
4717 * about the boundary, really active task won't care
4720 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4721 current->wakee_flips >>= 1;
4722 current->wakee_flip_decay_ts = jiffies;
4725 if (current->last_wakee != p) {
4726 current->last_wakee = p;
4727 current->wakee_flips++;
4731 static void task_waking_fair(struct task_struct *p)
4733 struct sched_entity *se = &p->se;
4734 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4737 #ifndef CONFIG_64BIT
4738 u64 min_vruntime_copy;
4741 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4743 min_vruntime = cfs_rq->min_vruntime;
4744 } while (min_vruntime != min_vruntime_copy);
4746 min_vruntime = cfs_rq->min_vruntime;
4749 se->vruntime -= min_vruntime;
4753 #ifdef CONFIG_FAIR_GROUP_SCHED
4755 * effective_load() calculates the load change as seen from the root_task_group
4757 * Adding load to a group doesn't make a group heavier, but can cause movement
4758 * of group shares between cpus. Assuming the shares were perfectly aligned one
4759 * can calculate the shift in shares.
4761 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4762 * on this @cpu and results in a total addition (subtraction) of @wg to the
4763 * total group weight.
4765 * Given a runqueue weight distribution (rw_i) we can compute a shares
4766 * distribution (s_i) using:
4768 * s_i = rw_i / \Sum rw_j (1)
4770 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4771 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4772 * shares distribution (s_i):
4774 * rw_i = { 2, 4, 1, 0 }
4775 * s_i = { 2/7, 4/7, 1/7, 0 }
4777 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4778 * task used to run on and the CPU the waker is running on), we need to
4779 * compute the effect of waking a task on either CPU and, in case of a sync
4780 * wakeup, compute the effect of the current task going to sleep.
4782 * So for a change of @wl to the local @cpu with an overall group weight change
4783 * of @wl we can compute the new shares distribution (s'_i) using:
4785 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4787 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4788 * differences in waking a task to CPU 0. The additional task changes the
4789 * weight and shares distributions like:
4791 * rw'_i = { 3, 4, 1, 0 }
4792 * s'_i = { 3/8, 4/8, 1/8, 0 }
4794 * We can then compute the difference in effective weight by using:
4796 * dw_i = S * (s'_i - s_i) (3)
4798 * Where 'S' is the group weight as seen by its parent.
4800 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4801 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4802 * 4/7) times the weight of the group.
4804 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4806 struct sched_entity *se = tg->se[cpu];
4808 if (!tg->parent) /* the trivial, non-cgroup case */
4811 for_each_sched_entity(se) {
4812 struct cfs_rq *cfs_rq = se->my_q;
4813 long W, w = cfs_rq_load_avg(cfs_rq);
4818 * W = @wg + \Sum rw_j
4820 W = wg + atomic_long_read(&tg->load_avg);
4822 /* Ensure \Sum rw_j >= rw_i */
4823 W -= cfs_rq->tg_load_avg_contrib;
4832 * wl = S * s'_i; see (2)
4835 wl = (w * (long)tg->shares) / W;
4840 * Per the above, wl is the new se->load.weight value; since
4841 * those are clipped to [MIN_SHARES, ...) do so now. See
4842 * calc_cfs_shares().
4844 if (wl < MIN_SHARES)
4848 * wl = dw_i = S * (s'_i - s_i); see (3)
4850 wl -= se->avg.load_avg;
4853 * Recursively apply this logic to all parent groups to compute
4854 * the final effective load change on the root group. Since
4855 * only the @tg group gets extra weight, all parent groups can
4856 * only redistribute existing shares. @wl is the shift in shares
4857 * resulting from this level per the above.
4866 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4874 * Returns the current capacity of cpu after applying both
4875 * cpu and freq scaling.
4877 unsigned long capacity_curr_of(int cpu)
4879 return cpu_rq(cpu)->cpu_capacity_orig *
4880 arch_scale_freq_capacity(NULL, cpu)
4881 >> SCHED_CAPACITY_SHIFT;
4884 static inline bool energy_aware(void)
4886 return sched_feat(ENERGY_AWARE);
4890 struct sched_group *sg_top;
4891 struct sched_group *sg_cap;
4898 struct task_struct *task;
4913 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4914 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4915 * energy calculations. Using the scale-invariant util returned by
4916 * cpu_util() and approximating scale-invariant util by:
4918 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4920 * the normalized util can be found using the specific capacity.
4922 * capacity = capacity_orig * curr_freq/max_freq
4924 * norm_util = running_time/time ~ util/capacity
4926 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4928 int util = __cpu_util(cpu, delta);
4930 if (util >= capacity)
4931 return SCHED_CAPACITY_SCALE;
4933 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4936 static int calc_util_delta(struct energy_env *eenv, int cpu)
4938 if (cpu == eenv->src_cpu)
4939 return -eenv->util_delta;
4940 if (cpu == eenv->dst_cpu)
4941 return eenv->util_delta;
4946 unsigned long group_max_util(struct energy_env *eenv)
4949 unsigned long max_util = 0;
4951 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4952 delta = calc_util_delta(eenv, i);
4953 max_util = max(max_util, __cpu_util(i, delta));
4960 * group_norm_util() returns the approximated group util relative to it's
4961 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4962 * energy calculations. Since task executions may or may not overlap in time in
4963 * the group the true normalized util is between max(cpu_norm_util(i)) and
4964 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4965 * latter is used as the estimate as it leads to a more pessimistic energy
4966 * estimate (more busy).
4969 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4972 unsigned long util_sum = 0;
4973 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4975 for_each_cpu(i, sched_group_cpus(sg)) {
4976 delta = calc_util_delta(eenv, i);
4977 util_sum += __cpu_norm_util(i, capacity, delta);
4980 if (util_sum > SCHED_CAPACITY_SCALE)
4981 return SCHED_CAPACITY_SCALE;
4985 static int find_new_capacity(struct energy_env *eenv,
4986 const struct sched_group_energy * const sge)
4989 unsigned long util = group_max_util(eenv);
4991 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4992 if (sge->cap_states[idx].cap >= util)
4996 eenv->cap_idx = idx;
5001 static int group_idle_state(struct sched_group *sg)
5003 int i, state = INT_MAX;
5005 /* Find the shallowest idle state in the sched group. */
5006 for_each_cpu(i, sched_group_cpus(sg))
5007 state = min(state, idle_get_state_idx(cpu_rq(i)));
5009 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5016 * sched_group_energy(): Computes the absolute energy consumption of cpus
5017 * belonging to the sched_group including shared resources shared only by
5018 * members of the group. Iterates over all cpus in the hierarchy below the
5019 * sched_group starting from the bottom working it's way up before going to
5020 * the next cpu until all cpus are covered at all levels. The current
5021 * implementation is likely to gather the same util statistics multiple times.
5022 * This can probably be done in a faster but more complex way.
5023 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5025 static int sched_group_energy(struct energy_env *eenv)
5027 struct sched_domain *sd;
5028 int cpu, total_energy = 0;
5029 struct cpumask visit_cpus;
5030 struct sched_group *sg;
5032 WARN_ON(!eenv->sg_top->sge);
5034 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5036 while (!cpumask_empty(&visit_cpus)) {
5037 struct sched_group *sg_shared_cap = NULL;
5039 cpu = cpumask_first(&visit_cpus);
5042 * Is the group utilization affected by cpus outside this
5045 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5049 * We most probably raced with hotplug; returning a
5050 * wrong energy estimation is better than entering an
5056 sg_shared_cap = sd->parent->groups;
5058 for_each_domain(cpu, sd) {
5061 /* Has this sched_domain already been visited? */
5062 if (sd->child && group_first_cpu(sg) != cpu)
5066 unsigned long group_util;
5067 int sg_busy_energy, sg_idle_energy;
5068 int cap_idx, idle_idx;
5070 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5071 eenv->sg_cap = sg_shared_cap;
5075 cap_idx = find_new_capacity(eenv, sg->sge);
5077 if (sg->group_weight == 1) {
5078 /* Remove capacity of src CPU (before task move) */
5079 if (eenv->util_delta == 0 &&
5080 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5081 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5082 eenv->cap.delta -= eenv->cap.before;
5084 /* Add capacity of dst CPU (after task move) */
5085 if (eenv->util_delta != 0 &&
5086 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5087 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5088 eenv->cap.delta += eenv->cap.after;
5092 idle_idx = group_idle_state(sg);
5093 group_util = group_norm_util(eenv, sg);
5094 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5095 >> SCHED_CAPACITY_SHIFT;
5096 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5097 * sg->sge->idle_states[idle_idx].power)
5098 >> SCHED_CAPACITY_SHIFT;
5100 total_energy += sg_busy_energy + sg_idle_energy;
5103 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5105 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5108 } while (sg = sg->next, sg != sd->groups);
5111 cpumask_clear_cpu(cpu, &visit_cpus);
5115 eenv->energy = total_energy;
5119 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5121 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5125 * energy_diff(): Estimate the energy impact of changing the utilization
5126 * distribution. eenv specifies the change: utilisation amount, source, and
5127 * destination cpu. Source or destination cpu may be -1 in which case the
5128 * utilization is removed from or added to the system (e.g. task wake-up). If
5129 * both are specified, the utilization is migrated.
5131 static inline int __energy_diff(struct energy_env *eenv)
5133 struct sched_domain *sd;
5134 struct sched_group *sg;
5135 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5137 struct energy_env eenv_before = {
5139 .src_cpu = eenv->src_cpu,
5140 .dst_cpu = eenv->dst_cpu,
5141 .nrg = { 0, 0, 0, 0},
5145 if (eenv->src_cpu == eenv->dst_cpu)
5148 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5149 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5152 return 0; /* Error */
5157 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5158 eenv_before.sg_top = eenv->sg_top = sg;
5160 if (sched_group_energy(&eenv_before))
5161 return 0; /* Invalid result abort */
5162 energy_before += eenv_before.energy;
5164 /* Keep track of SRC cpu (before) capacity */
5165 eenv->cap.before = eenv_before.cap.before;
5166 eenv->cap.delta = eenv_before.cap.delta;
5168 if (sched_group_energy(eenv))
5169 return 0; /* Invalid result abort */
5170 energy_after += eenv->energy;
5172 } while (sg = sg->next, sg != sd->groups);
5174 eenv->nrg.before = energy_before;
5175 eenv->nrg.after = energy_after;
5176 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5179 trace_sched_energy_diff(eenv->task,
5180 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5181 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5182 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5183 eenv->nrg.delta, eenv->payoff);
5185 return eenv->nrg.diff;
5188 #ifdef CONFIG_SCHED_TUNE
5190 struct target_nrg schedtune_target_nrg;
5193 * System energy normalization
5194 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5195 * corresponding to the specified energy variation.
5198 normalize_energy(int energy_diff)
5201 #ifdef CONFIG_SCHED_DEBUG
5204 /* Check for boundaries */
5205 max_delta = schedtune_target_nrg.max_power;
5206 max_delta -= schedtune_target_nrg.min_power;
5207 WARN_ON(abs(energy_diff) >= max_delta);
5210 /* Do scaling using positive numbers to increase the range */
5211 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5213 /* Scale by energy magnitude */
5214 normalized_nrg <<= SCHED_LOAD_SHIFT;
5216 /* Normalize on max energy for target platform */
5217 normalized_nrg = reciprocal_divide(
5218 normalized_nrg, schedtune_target_nrg.rdiv);
5220 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5224 energy_diff(struct energy_env *eenv)
5226 int boost = schedtune_task_boost(eenv->task);
5229 /* Conpute "absolute" energy diff */
5230 __energy_diff(eenv);
5232 /* Return energy diff when boost margin is 0 */
5234 return eenv->nrg.diff;
5236 /* Compute normalized energy diff */
5237 nrg_delta = normalize_energy(eenv->nrg.diff);
5238 eenv->nrg.delta = nrg_delta;
5240 eenv->payoff = schedtune_accept_deltas(
5246 * When SchedTune is enabled, the energy_diff() function will return
5247 * the computed energy payoff value. Since the energy_diff() return
5248 * value is expected to be negative by its callers, this evaluation
5249 * function return a negative value each time the evaluation return a
5250 * positive payoff, which is the condition for the acceptance of
5251 * a scheduling decision
5253 return -eenv->payoff;
5255 #else /* CONFIG_SCHED_TUNE */
5256 #define energy_diff(eenv) __energy_diff(eenv)
5260 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5261 * A waker of many should wake a different task than the one last awakened
5262 * at a frequency roughly N times higher than one of its wakees. In order
5263 * to determine whether we should let the load spread vs consolodating to
5264 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5265 * partner, and a factor of lls_size higher frequency in the other. With
5266 * both conditions met, we can be relatively sure that the relationship is
5267 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5268 * being client/server, worker/dispatcher, interrupt source or whatever is
5269 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5271 static int wake_wide(struct task_struct *p)
5273 unsigned int master = current->wakee_flips;
5274 unsigned int slave = p->wakee_flips;
5275 int factor = this_cpu_read(sd_llc_size);
5278 swap(master, slave);
5279 if (slave < factor || master < slave * factor)
5284 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5285 int prev_cpu, int sync)
5287 s64 this_load, load;
5288 s64 this_eff_load, prev_eff_load;
5290 struct task_group *tg;
5291 unsigned long weight;
5295 this_cpu = smp_processor_id();
5296 load = source_load(prev_cpu, idx);
5297 this_load = target_load(this_cpu, idx);
5300 * If sync wakeup then subtract the (maximum possible)
5301 * effect of the currently running task from the load
5302 * of the current CPU:
5305 tg = task_group(current);
5306 weight = current->se.avg.load_avg;
5308 this_load += effective_load(tg, this_cpu, -weight, -weight);
5309 load += effective_load(tg, prev_cpu, 0, -weight);
5313 weight = p->se.avg.load_avg;
5316 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5317 * due to the sync cause above having dropped this_load to 0, we'll
5318 * always have an imbalance, but there's really nothing you can do
5319 * about that, so that's good too.
5321 * Otherwise check if either cpus are near enough in load to allow this
5322 * task to be woken on this_cpu.
5324 this_eff_load = 100;
5325 this_eff_load *= capacity_of(prev_cpu);
5327 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5328 prev_eff_load *= capacity_of(this_cpu);
5330 if (this_load > 0) {
5331 this_eff_load *= this_load +
5332 effective_load(tg, this_cpu, weight, weight);
5334 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5337 balanced = this_eff_load <= prev_eff_load;
5339 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5344 schedstat_inc(sd, ttwu_move_affine);
5345 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5350 static inline unsigned long task_util(struct task_struct *p)
5352 #ifdef CONFIG_SCHED_WALT
5353 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5354 unsigned long demand = p->ravg.demand;
5355 return (demand << 10) / walt_ravg_window;
5358 return p->se.avg.util_avg;
5361 unsigned int capacity_margin = 1280; /* ~20% margin */
5363 static inline unsigned long boosted_task_util(struct task_struct *task);
5365 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5367 unsigned long capacity = capacity_of(cpu);
5369 util += boosted_task_util(p);
5371 return (capacity * 1024) > (util * capacity_margin);
5374 static inline bool task_fits_max(struct task_struct *p, int cpu)
5376 unsigned long capacity = capacity_of(cpu);
5377 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5379 if (capacity == max_capacity)
5382 if (capacity * capacity_margin > max_capacity * 1024)
5385 return __task_fits(p, cpu, 0);
5388 static bool cpu_overutilized(int cpu)
5390 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5393 #ifdef CONFIG_SCHED_TUNE
5396 schedtune_margin(unsigned long signal, long boost)
5398 long long margin = 0;
5401 * Signal proportional compensation (SPC)
5403 * The Boost (B) value is used to compute a Margin (M) which is
5404 * proportional to the complement of the original Signal (S):
5405 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5406 * M = B * S, if B is negative
5407 * The obtained M could be used by the caller to "boost" S.
5410 margin = SCHED_LOAD_SCALE - signal;
5413 margin = -signal * boost;
5415 * Fast integer division by constant:
5416 * Constant : (C) = 100
5417 * Precision : 0.1% (P) = 0.1
5418 * Reference : C * 100 / P (R) = 100000
5421 * Shift bits : ceil(log(R,2)) (S) = 17
5422 * Mult const : round(2^S/C) (M) = 1311
5435 schedtune_cpu_margin(unsigned long util, int cpu)
5437 int boost = schedtune_cpu_boost(cpu);
5442 return schedtune_margin(util, boost);
5446 schedtune_task_margin(struct task_struct *task)
5448 int boost = schedtune_task_boost(task);
5455 util = task_util(task);
5456 margin = schedtune_margin(util, boost);
5461 #else /* CONFIG_SCHED_TUNE */
5464 schedtune_cpu_margin(unsigned long util, int cpu)
5470 schedtune_task_margin(struct task_struct *task)
5475 #endif /* CONFIG_SCHED_TUNE */
5478 boosted_cpu_util(int cpu)
5480 unsigned long util = cpu_util(cpu);
5481 long margin = schedtune_cpu_margin(util, cpu);
5483 trace_sched_boost_cpu(cpu, util, margin);
5485 return util + margin;
5488 static inline unsigned long
5489 boosted_task_util(struct task_struct *task)
5491 unsigned long util = task_util(task);
5492 long margin = schedtune_task_margin(task);
5494 trace_sched_boost_task(task, util, margin);
5496 return util + margin;
5500 * find_idlest_group finds and returns the least busy CPU group within the
5503 static struct sched_group *
5504 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5505 int this_cpu, int sd_flag)
5507 struct sched_group *idlest = NULL, *group = sd->groups;
5508 unsigned long min_load = ULONG_MAX, this_load = 0;
5509 int load_idx = sd->forkexec_idx;
5510 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5512 if (sd_flag & SD_BALANCE_WAKE)
5513 load_idx = sd->wake_idx;
5516 unsigned long load, avg_load;
5520 /* Skip over this group if it has no CPUs allowed */
5521 if (!cpumask_intersects(sched_group_cpus(group),
5522 tsk_cpus_allowed(p)))
5525 local_group = cpumask_test_cpu(this_cpu,
5526 sched_group_cpus(group));
5528 /* Tally up the load of all CPUs in the group */
5531 for_each_cpu(i, sched_group_cpus(group)) {
5532 /* Bias balancing toward cpus of our domain */
5534 load = source_load(i, load_idx);
5536 load = target_load(i, load_idx);
5541 /* Adjust by relative CPU capacity of the group */
5542 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5545 this_load = avg_load;
5546 } else if (avg_load < min_load) {
5547 min_load = avg_load;
5550 } while (group = group->next, group != sd->groups);
5552 if (!idlest || 100*this_load < imbalance*min_load)
5558 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5561 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5563 unsigned long load, min_load = ULONG_MAX;
5564 unsigned int min_exit_latency = UINT_MAX;
5565 u64 latest_idle_timestamp = 0;
5566 int least_loaded_cpu = this_cpu;
5567 int shallowest_idle_cpu = -1;
5570 /* Traverse only the allowed CPUs */
5571 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5573 struct rq *rq = cpu_rq(i);
5574 struct cpuidle_state *idle = idle_get_state(rq);
5575 if (idle && idle->exit_latency < min_exit_latency) {
5577 * We give priority to a CPU whose idle state
5578 * has the smallest exit latency irrespective
5579 * of any idle timestamp.
5581 min_exit_latency = idle->exit_latency;
5582 latest_idle_timestamp = rq->idle_stamp;
5583 shallowest_idle_cpu = i;
5584 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5585 rq->idle_stamp > latest_idle_timestamp) {
5587 * If equal or no active idle state, then
5588 * the most recently idled CPU might have
5591 latest_idle_timestamp = rq->idle_stamp;
5592 shallowest_idle_cpu = i;
5594 } else if (shallowest_idle_cpu == -1) {
5595 load = weighted_cpuload(i);
5596 if (load < min_load || (load == min_load && i == this_cpu)) {
5598 least_loaded_cpu = i;
5603 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5607 * Try and locate an idle CPU in the sched_domain.
5609 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5611 struct sched_domain *sd;
5612 struct sched_group *sg;
5614 int best_idle_cstate = -1;
5615 int best_idle_capacity = INT_MAX;
5617 if (!sysctl_sched_cstate_aware) {
5618 if (idle_cpu(target))
5622 * If the prevous cpu is cache affine and idle, don't be stupid.
5624 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5629 * Otherwise, iterate the domains and find an elegible idle cpu.
5631 sd = rcu_dereference(per_cpu(sd_llc, target));
5632 for_each_lower_domain(sd) {
5636 if (!cpumask_intersects(sched_group_cpus(sg),
5637 tsk_cpus_allowed(p)))
5640 if (sysctl_sched_cstate_aware) {
5641 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5642 struct rq *rq = cpu_rq(i);
5643 int idle_idx = idle_get_state_idx(rq);
5644 unsigned long new_usage = boosted_task_util(p);
5645 unsigned long capacity_orig = capacity_orig_of(i);
5646 if (new_usage > capacity_orig || !idle_cpu(i))
5649 if (i == target && new_usage <= capacity_curr_of(target))
5652 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5654 best_idle_cstate = idle_idx;
5655 best_idle_capacity = capacity_orig;
5659 for_each_cpu(i, sched_group_cpus(sg)) {
5660 if (i == target || !idle_cpu(i))
5664 target = cpumask_first_and(sched_group_cpus(sg),
5665 tsk_cpus_allowed(p));
5670 } while (sg != sd->groups);
5679 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5682 int target_cpu = -1;
5683 int target_util = 0;
5684 int backup_capacity = 0;
5685 int best_idle_cpu = -1;
5686 int best_idle_cstate = INT_MAX;
5687 int backup_cpu = -1;
5688 unsigned long task_util_boosted, new_util;
5690 task_util_boosted = boosted_task_util(p);
5691 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5697 * Iterate from higher cpus for boosted tasks.
5699 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5701 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5705 * p's blocked utilization is still accounted for on prev_cpu
5706 * so prev_cpu will receive a negative bias due to the double
5707 * accounting. However, the blocked utilization may be zero.
5709 new_util = cpu_util(i) + task_util_boosted;
5712 * Ensure minimum capacity to grant the required boost.
5713 * The target CPU can be already at a capacity level higher
5714 * than the one required to boost the task.
5716 if (new_util > capacity_orig_of(i))
5719 #ifdef CONFIG_SCHED_WALT
5720 if (walt_cpu_high_irqload(i))
5724 * Unconditionally favoring tasks that prefer idle cpus to
5727 if (idle_cpu(i) && prefer_idle) {
5728 if (best_idle_cpu < 0)
5733 cur_capacity = capacity_curr_of(i);
5735 idle_idx = idle_get_state_idx(rq);
5737 if (new_util < cur_capacity) {
5738 if (cpu_rq(i)->nr_running) {
5740 /* Find a target cpu with highest
5743 if (target_util == 0 ||
5744 target_util < new_util) {
5746 target_util = new_util;
5749 /* Find a target cpu with lowest
5752 if (target_util == 0 ||
5753 target_util > new_util) {
5755 target_util = new_util;
5758 } else if (!prefer_idle) {
5759 if (best_idle_cpu < 0 ||
5760 (sysctl_sched_cstate_aware &&
5761 best_idle_cstate > idle_idx)) {
5762 best_idle_cstate = idle_idx;
5766 } else if (backup_capacity == 0 ||
5767 backup_capacity > cur_capacity) {
5768 // Find a backup cpu with least capacity.
5769 backup_capacity = cur_capacity;
5774 if (prefer_idle && best_idle_cpu >= 0)
5775 target_cpu = best_idle_cpu;
5776 else if (target_cpu < 0)
5777 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5782 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5784 struct sched_domain *sd;
5785 struct sched_group *sg, *sg_target;
5786 int target_max_cap = INT_MAX;
5787 int target_cpu = task_cpu(p);
5788 unsigned long task_util_boosted, new_util;
5791 if (sysctl_sched_sync_hint_enable && sync) {
5792 int cpu = smp_processor_id();
5793 cpumask_t search_cpus;
5794 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5795 if (cpumask_test_cpu(cpu, &search_cpus))
5799 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5807 if (sysctl_sched_is_big_little) {
5810 * Find group with sufficient capacity. We only get here if no cpu is
5811 * overutilized. We may end up overutilizing a cpu by adding the task,
5812 * but that should not be any worse than select_idle_sibling().
5813 * load_balance() should sort it out later as we get above the tipping
5817 /* Assuming all cpus are the same in group */
5818 int max_cap_cpu = group_first_cpu(sg);
5821 * Assume smaller max capacity means more energy-efficient.
5822 * Ideally we should query the energy model for the right
5823 * answer but it easily ends up in an exhaustive search.
5825 if (capacity_of(max_cap_cpu) < target_max_cap &&
5826 task_fits_max(p, max_cap_cpu)) {
5828 target_max_cap = capacity_of(max_cap_cpu);
5830 } while (sg = sg->next, sg != sd->groups);
5832 task_util_boosted = boosted_task_util(p);
5833 /* Find cpu with sufficient capacity */
5834 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5836 * p's blocked utilization is still accounted for on prev_cpu
5837 * so prev_cpu will receive a negative bias due to the double
5838 * accounting. However, the blocked utilization may be zero.
5840 new_util = cpu_util(i) + task_util_boosted;
5843 * Ensure minimum capacity to grant the required boost.
5844 * The target CPU can be already at a capacity level higher
5845 * than the one required to boost the task.
5847 if (new_util > capacity_orig_of(i))
5850 if (new_util < capacity_curr_of(i)) {
5852 if (cpu_rq(i)->nr_running)
5856 /* cpu has capacity at higher OPP, keep it as fallback */
5857 if (target_cpu == task_cpu(p))
5862 * Find a cpu with sufficient capacity
5864 #ifdef CONFIG_CGROUP_SCHEDTUNE
5865 bool boosted = schedtune_task_boost(p) > 0;
5866 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5869 bool prefer_idle = 0;
5871 int tmp_target = find_best_target(p, boosted, prefer_idle);
5872 if (tmp_target >= 0) {
5873 target_cpu = tmp_target;
5874 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5879 if (target_cpu != task_cpu(p)) {
5880 struct energy_env eenv = {
5881 .util_delta = task_util(p),
5882 .src_cpu = task_cpu(p),
5883 .dst_cpu = target_cpu,
5887 /* Not enough spare capacity on previous cpu */
5888 if (cpu_overutilized(task_cpu(p)))
5891 if (energy_diff(&eenv) >= 0)
5899 * select_task_rq_fair: Select target runqueue for the waking task in domains
5900 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5901 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5903 * Balances load by selecting the idlest cpu in the idlest group, or under
5904 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5906 * Returns the target cpu number.
5908 * preempt must be disabled.
5911 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5913 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5914 int cpu = smp_processor_id();
5915 int new_cpu = prev_cpu;
5916 int want_affine = 0;
5917 int sync = wake_flags & WF_SYNC;
5919 if (sd_flag & SD_BALANCE_WAKE)
5920 want_affine = (!wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5924 for_each_domain(cpu, tmp) {
5925 if (!(tmp->flags & SD_LOAD_BALANCE))
5929 * If both cpu and prev_cpu are part of this domain,
5930 * cpu is a valid SD_WAKE_AFFINE target.
5932 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5933 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5938 if (tmp->flags & sd_flag)
5940 else if (!want_affine)
5945 sd = NULL; /* Prefer wake_affine over balance flags */
5946 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5951 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5952 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5953 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5954 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5957 struct sched_group *group;
5960 if (!(sd->flags & sd_flag)) {
5965 group = find_idlest_group(sd, p, cpu, sd_flag);
5971 new_cpu = find_idlest_cpu(group, p, cpu);
5972 if (new_cpu == -1 || new_cpu == cpu) {
5973 /* Now try balancing at a lower domain level of cpu */
5978 /* Now try balancing at a lower domain level of new_cpu */
5980 weight = sd->span_weight;
5982 for_each_domain(cpu, tmp) {
5983 if (weight <= tmp->span_weight)
5985 if (tmp->flags & sd_flag)
5988 /* while loop will break here if sd == NULL */
5996 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5997 * cfs_rq_of(p) references at time of call are still valid and identify the
5998 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5999 * other assumptions, including the state of rq->lock, should be made.
6001 static void migrate_task_rq_fair(struct task_struct *p)
6004 * We are supposed to update the task to "current" time, then its up to date
6005 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6006 * what current time is, so simply throw away the out-of-date time. This
6007 * will result in the wakee task is less decayed, but giving the wakee more
6008 * load sounds not bad.
6010 remove_entity_load_avg(&p->se);
6012 /* Tell new CPU we are migrated */
6013 p->se.avg.last_update_time = 0;
6015 /* We have migrated, no longer consider this task hot */
6016 p->se.exec_start = 0;
6019 static void task_dead_fair(struct task_struct *p)
6021 remove_entity_load_avg(&p->se);
6024 #define task_fits_max(p, cpu) true
6025 #endif /* CONFIG_SMP */
6027 static unsigned long
6028 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6030 unsigned long gran = sysctl_sched_wakeup_granularity;
6033 * Since its curr running now, convert the gran from real-time
6034 * to virtual-time in his units.
6036 * By using 'se' instead of 'curr' we penalize light tasks, so
6037 * they get preempted easier. That is, if 'se' < 'curr' then
6038 * the resulting gran will be larger, therefore penalizing the
6039 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6040 * be smaller, again penalizing the lighter task.
6042 * This is especially important for buddies when the leftmost
6043 * task is higher priority than the buddy.
6045 return calc_delta_fair(gran, se);
6049 * Should 'se' preempt 'curr'.
6063 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6065 s64 gran, vdiff = curr->vruntime - se->vruntime;
6070 gran = wakeup_gran(curr, se);
6077 static void set_last_buddy(struct sched_entity *se)
6079 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6082 for_each_sched_entity(se)
6083 cfs_rq_of(se)->last = se;
6086 static void set_next_buddy(struct sched_entity *se)
6088 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6091 for_each_sched_entity(se)
6092 cfs_rq_of(se)->next = se;
6095 static void set_skip_buddy(struct sched_entity *se)
6097 for_each_sched_entity(se)
6098 cfs_rq_of(se)->skip = se;
6102 * Preempt the current task with a newly woken task if needed:
6104 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6106 struct task_struct *curr = rq->curr;
6107 struct sched_entity *se = &curr->se, *pse = &p->se;
6108 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6109 int scale = cfs_rq->nr_running >= sched_nr_latency;
6110 int next_buddy_marked = 0;
6112 if (unlikely(se == pse))
6116 * This is possible from callers such as attach_tasks(), in which we
6117 * unconditionally check_prempt_curr() after an enqueue (which may have
6118 * lead to a throttle). This both saves work and prevents false
6119 * next-buddy nomination below.
6121 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6124 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6125 set_next_buddy(pse);
6126 next_buddy_marked = 1;
6130 * We can come here with TIF_NEED_RESCHED already set from new task
6133 * Note: this also catches the edge-case of curr being in a throttled
6134 * group (e.g. via set_curr_task), since update_curr() (in the
6135 * enqueue of curr) will have resulted in resched being set. This
6136 * prevents us from potentially nominating it as a false LAST_BUDDY
6139 if (test_tsk_need_resched(curr))
6142 /* Idle tasks are by definition preempted by non-idle tasks. */
6143 if (unlikely(curr->policy == SCHED_IDLE) &&
6144 likely(p->policy != SCHED_IDLE))
6148 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6149 * is driven by the tick):
6151 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6154 find_matching_se(&se, &pse);
6155 update_curr(cfs_rq_of(se));
6157 if (wakeup_preempt_entity(se, pse) == 1) {
6159 * Bias pick_next to pick the sched entity that is
6160 * triggering this preemption.
6162 if (!next_buddy_marked)
6163 set_next_buddy(pse);
6172 * Only set the backward buddy when the current task is still
6173 * on the rq. This can happen when a wakeup gets interleaved
6174 * with schedule on the ->pre_schedule() or idle_balance()
6175 * point, either of which can * drop the rq lock.
6177 * Also, during early boot the idle thread is in the fair class,
6178 * for obvious reasons its a bad idea to schedule back to it.
6180 if (unlikely(!se->on_rq || curr == rq->idle))
6183 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6187 static struct task_struct *
6188 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6190 struct cfs_rq *cfs_rq = &rq->cfs;
6191 struct sched_entity *se;
6192 struct task_struct *p;
6196 #ifdef CONFIG_FAIR_GROUP_SCHED
6197 if (!cfs_rq->nr_running)
6200 if (prev->sched_class != &fair_sched_class)
6204 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6205 * likely that a next task is from the same cgroup as the current.
6207 * Therefore attempt to avoid putting and setting the entire cgroup
6208 * hierarchy, only change the part that actually changes.
6212 struct sched_entity *curr = cfs_rq->curr;
6215 * Since we got here without doing put_prev_entity() we also
6216 * have to consider cfs_rq->curr. If it is still a runnable
6217 * entity, update_curr() will update its vruntime, otherwise
6218 * forget we've ever seen it.
6222 update_curr(cfs_rq);
6227 * This call to check_cfs_rq_runtime() will do the
6228 * throttle and dequeue its entity in the parent(s).
6229 * Therefore the 'simple' nr_running test will indeed
6232 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6236 se = pick_next_entity(cfs_rq, curr);
6237 cfs_rq = group_cfs_rq(se);
6243 * Since we haven't yet done put_prev_entity and if the selected task
6244 * is a different task than we started out with, try and touch the
6245 * least amount of cfs_rqs.
6248 struct sched_entity *pse = &prev->se;
6250 while (!(cfs_rq = is_same_group(se, pse))) {
6251 int se_depth = se->depth;
6252 int pse_depth = pse->depth;
6254 if (se_depth <= pse_depth) {
6255 put_prev_entity(cfs_rq_of(pse), pse);
6256 pse = parent_entity(pse);
6258 if (se_depth >= pse_depth) {
6259 set_next_entity(cfs_rq_of(se), se);
6260 se = parent_entity(se);
6264 put_prev_entity(cfs_rq, pse);
6265 set_next_entity(cfs_rq, se);
6268 if (hrtick_enabled(rq))
6269 hrtick_start_fair(rq, p);
6271 rq->misfit_task = !task_fits_max(p, rq->cpu);
6278 if (!cfs_rq->nr_running)
6281 put_prev_task(rq, prev);
6284 se = pick_next_entity(cfs_rq, NULL);
6285 set_next_entity(cfs_rq, se);
6286 cfs_rq = group_cfs_rq(se);
6291 if (hrtick_enabled(rq))
6292 hrtick_start_fair(rq, p);
6294 rq->misfit_task = !task_fits_max(p, rq->cpu);
6299 rq->misfit_task = 0;
6301 * This is OK, because current is on_cpu, which avoids it being picked
6302 * for load-balance and preemption/IRQs are still disabled avoiding
6303 * further scheduler activity on it and we're being very careful to
6304 * re-start the picking loop.
6306 lockdep_unpin_lock(&rq->lock);
6307 new_tasks = idle_balance(rq);
6308 lockdep_pin_lock(&rq->lock);
6310 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6311 * possible for any higher priority task to appear. In that case we
6312 * must re-start the pick_next_entity() loop.
6324 * Account for a descheduled task:
6326 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6328 struct sched_entity *se = &prev->se;
6329 struct cfs_rq *cfs_rq;
6331 for_each_sched_entity(se) {
6332 cfs_rq = cfs_rq_of(se);
6333 put_prev_entity(cfs_rq, se);
6338 * sched_yield() is very simple
6340 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6342 static void yield_task_fair(struct rq *rq)
6344 struct task_struct *curr = rq->curr;
6345 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6346 struct sched_entity *se = &curr->se;
6349 * Are we the only task in the tree?
6351 if (unlikely(rq->nr_running == 1))
6354 clear_buddies(cfs_rq, se);
6356 if (curr->policy != SCHED_BATCH) {
6357 update_rq_clock(rq);
6359 * Update run-time statistics of the 'current'.
6361 update_curr(cfs_rq);
6363 * Tell update_rq_clock() that we've just updated,
6364 * so we don't do microscopic update in schedule()
6365 * and double the fastpath cost.
6367 rq_clock_skip_update(rq, true);
6373 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6375 struct sched_entity *se = &p->se;
6377 /* throttled hierarchies are not runnable */
6378 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6381 /* Tell the scheduler that we'd really like pse to run next. */
6384 yield_task_fair(rq);
6390 /**************************************************
6391 * Fair scheduling class load-balancing methods.
6395 * The purpose of load-balancing is to achieve the same basic fairness the
6396 * per-cpu scheduler provides, namely provide a proportional amount of compute
6397 * time to each task. This is expressed in the following equation:
6399 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6401 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6402 * W_i,0 is defined as:
6404 * W_i,0 = \Sum_j w_i,j (2)
6406 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6407 * is derived from the nice value as per prio_to_weight[].
6409 * The weight average is an exponential decay average of the instantaneous
6412 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6414 * C_i is the compute capacity of cpu i, typically it is the
6415 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6416 * can also include other factors [XXX].
6418 * To achieve this balance we define a measure of imbalance which follows
6419 * directly from (1):
6421 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6423 * We them move tasks around to minimize the imbalance. In the continuous
6424 * function space it is obvious this converges, in the discrete case we get
6425 * a few fun cases generally called infeasible weight scenarios.
6428 * - infeasible weights;
6429 * - local vs global optima in the discrete case. ]
6434 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6435 * for all i,j solution, we create a tree of cpus that follows the hardware
6436 * topology where each level pairs two lower groups (or better). This results
6437 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6438 * tree to only the first of the previous level and we decrease the frequency
6439 * of load-balance at each level inv. proportional to the number of cpus in
6445 * \Sum { --- * --- * 2^i } = O(n) (5)
6447 * `- size of each group
6448 * | | `- number of cpus doing load-balance
6450 * `- sum over all levels
6452 * Coupled with a limit on how many tasks we can migrate every balance pass,
6453 * this makes (5) the runtime complexity of the balancer.
6455 * An important property here is that each CPU is still (indirectly) connected
6456 * to every other cpu in at most O(log n) steps:
6458 * The adjacency matrix of the resulting graph is given by:
6461 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6464 * And you'll find that:
6466 * A^(log_2 n)_i,j != 0 for all i,j (7)
6468 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6469 * The task movement gives a factor of O(m), giving a convergence complexity
6472 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6477 * In order to avoid CPUs going idle while there's still work to do, new idle
6478 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6479 * tree itself instead of relying on other CPUs to bring it work.
6481 * This adds some complexity to both (5) and (8) but it reduces the total idle
6489 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6492 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6497 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6499 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6501 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6504 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6505 * rewrite all of this once again.]
6508 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6510 enum fbq_type { regular, remote, all };
6519 #define LBF_ALL_PINNED 0x01
6520 #define LBF_NEED_BREAK 0x02
6521 #define LBF_DST_PINNED 0x04
6522 #define LBF_SOME_PINNED 0x08
6525 struct sched_domain *sd;
6533 struct cpumask *dst_grpmask;
6535 enum cpu_idle_type idle;
6537 unsigned int src_grp_nr_running;
6538 /* The set of CPUs under consideration for load-balancing */
6539 struct cpumask *cpus;
6544 unsigned int loop_break;
6545 unsigned int loop_max;
6547 enum fbq_type fbq_type;
6548 enum group_type busiest_group_type;
6549 struct list_head tasks;
6553 * Is this task likely cache-hot:
6555 static int task_hot(struct task_struct *p, struct lb_env *env)
6559 lockdep_assert_held(&env->src_rq->lock);
6561 if (p->sched_class != &fair_sched_class)
6564 if (unlikely(p->policy == SCHED_IDLE))
6568 * Buddy candidates are cache hot:
6570 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6571 (&p->se == cfs_rq_of(&p->se)->next ||
6572 &p->se == cfs_rq_of(&p->se)->last))
6575 if (sysctl_sched_migration_cost == -1)
6577 if (sysctl_sched_migration_cost == 0)
6580 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6582 return delta < (s64)sysctl_sched_migration_cost;
6585 #ifdef CONFIG_NUMA_BALANCING
6587 * Returns 1, if task migration degrades locality
6588 * Returns 0, if task migration improves locality i.e migration preferred.
6589 * Returns -1, if task migration is not affected by locality.
6591 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6593 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6594 unsigned long src_faults, dst_faults;
6595 int src_nid, dst_nid;
6597 if (!static_branch_likely(&sched_numa_balancing))
6600 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6603 src_nid = cpu_to_node(env->src_cpu);
6604 dst_nid = cpu_to_node(env->dst_cpu);
6606 if (src_nid == dst_nid)
6609 /* Migrating away from the preferred node is always bad. */
6610 if (src_nid == p->numa_preferred_nid) {
6611 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6617 /* Encourage migration to the preferred node. */
6618 if (dst_nid == p->numa_preferred_nid)
6622 src_faults = group_faults(p, src_nid);
6623 dst_faults = group_faults(p, dst_nid);
6625 src_faults = task_faults(p, src_nid);
6626 dst_faults = task_faults(p, dst_nid);
6629 return dst_faults < src_faults;
6633 static inline int migrate_degrades_locality(struct task_struct *p,
6641 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6644 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6648 lockdep_assert_held(&env->src_rq->lock);
6651 * We do not migrate tasks that are:
6652 * 1) throttled_lb_pair, or
6653 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6654 * 3) running (obviously), or
6655 * 4) are cache-hot on their current CPU.
6657 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6660 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6663 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6665 env->flags |= LBF_SOME_PINNED;
6668 * Remember if this task can be migrated to any other cpu in
6669 * our sched_group. We may want to revisit it if we couldn't
6670 * meet load balance goals by pulling other tasks on src_cpu.
6672 * Also avoid computing new_dst_cpu if we have already computed
6673 * one in current iteration.
6675 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6678 /* Prevent to re-select dst_cpu via env's cpus */
6679 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6680 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6681 env->flags |= LBF_DST_PINNED;
6682 env->new_dst_cpu = cpu;
6690 /* Record that we found atleast one task that could run on dst_cpu */
6691 env->flags &= ~LBF_ALL_PINNED;
6693 if (task_running(env->src_rq, p)) {
6694 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6699 * Aggressive migration if:
6700 * 1) destination numa is preferred
6701 * 2) task is cache cold, or
6702 * 3) too many balance attempts have failed.
6704 tsk_cache_hot = migrate_degrades_locality(p, env);
6705 if (tsk_cache_hot == -1)
6706 tsk_cache_hot = task_hot(p, env);
6708 if (tsk_cache_hot <= 0 ||
6709 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6710 if (tsk_cache_hot == 1) {
6711 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6712 schedstat_inc(p, se.statistics.nr_forced_migrations);
6717 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6722 * detach_task() -- detach the task for the migration specified in env
6724 static void detach_task(struct task_struct *p, struct lb_env *env)
6726 lockdep_assert_held(&env->src_rq->lock);
6728 deactivate_task(env->src_rq, p, 0);
6729 p->on_rq = TASK_ON_RQ_MIGRATING;
6730 double_lock_balance(env->src_rq, env->dst_rq);
6731 set_task_cpu(p, env->dst_cpu);
6732 double_unlock_balance(env->src_rq, env->dst_rq);
6736 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6737 * part of active balancing operations within "domain".
6739 * Returns a task if successful and NULL otherwise.
6741 static struct task_struct *detach_one_task(struct lb_env *env)
6743 struct task_struct *p, *n;
6745 lockdep_assert_held(&env->src_rq->lock);
6747 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6748 if (!can_migrate_task(p, env))
6751 detach_task(p, env);
6754 * Right now, this is only the second place where
6755 * lb_gained[env->idle] is updated (other is detach_tasks)
6756 * so we can safely collect stats here rather than
6757 * inside detach_tasks().
6759 schedstat_inc(env->sd, lb_gained[env->idle]);
6765 static const unsigned int sched_nr_migrate_break = 32;
6768 * detach_tasks() -- tries to detach up to imbalance weighted load from
6769 * busiest_rq, as part of a balancing operation within domain "sd".
6771 * Returns number of detached tasks if successful and 0 otherwise.
6773 static int detach_tasks(struct lb_env *env)
6775 struct list_head *tasks = &env->src_rq->cfs_tasks;
6776 struct task_struct *p;
6780 lockdep_assert_held(&env->src_rq->lock);
6782 if (env->imbalance <= 0)
6785 while (!list_empty(tasks)) {
6787 * We don't want to steal all, otherwise we may be treated likewise,
6788 * which could at worst lead to a livelock crash.
6790 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6793 p = list_first_entry(tasks, struct task_struct, se.group_node);
6796 /* We've more or less seen every task there is, call it quits */
6797 if (env->loop > env->loop_max)
6800 /* take a breather every nr_migrate tasks */
6801 if (env->loop > env->loop_break) {
6802 env->loop_break += sched_nr_migrate_break;
6803 env->flags |= LBF_NEED_BREAK;
6807 if (!can_migrate_task(p, env))
6810 load = task_h_load(p);
6812 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6815 if ((load / 2) > env->imbalance)
6818 detach_task(p, env);
6819 list_add(&p->se.group_node, &env->tasks);
6822 env->imbalance -= load;
6824 #ifdef CONFIG_PREEMPT
6826 * NEWIDLE balancing is a source of latency, so preemptible
6827 * kernels will stop after the first task is detached to minimize
6828 * the critical section.
6830 if (env->idle == CPU_NEWLY_IDLE)
6835 * We only want to steal up to the prescribed amount of
6838 if (env->imbalance <= 0)
6843 list_move_tail(&p->se.group_node, tasks);
6847 * Right now, this is one of only two places we collect this stat
6848 * so we can safely collect detach_one_task() stats here rather
6849 * than inside detach_one_task().
6851 schedstat_add(env->sd, lb_gained[env->idle], detached);
6857 * attach_task() -- attach the task detached by detach_task() to its new rq.
6859 static void attach_task(struct rq *rq, struct task_struct *p)
6861 lockdep_assert_held(&rq->lock);
6863 BUG_ON(task_rq(p) != rq);
6864 p->on_rq = TASK_ON_RQ_QUEUED;
6865 activate_task(rq, p, 0);
6866 check_preempt_curr(rq, p, 0);
6870 * attach_one_task() -- attaches the task returned from detach_one_task() to
6873 static void attach_one_task(struct rq *rq, struct task_struct *p)
6875 raw_spin_lock(&rq->lock);
6878 * We want to potentially raise target_cpu's OPP.
6880 update_capacity_of(cpu_of(rq));
6881 raw_spin_unlock(&rq->lock);
6885 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6888 static void attach_tasks(struct lb_env *env)
6890 struct list_head *tasks = &env->tasks;
6891 struct task_struct *p;
6893 raw_spin_lock(&env->dst_rq->lock);
6895 while (!list_empty(tasks)) {
6896 p = list_first_entry(tasks, struct task_struct, se.group_node);
6897 list_del_init(&p->se.group_node);
6899 attach_task(env->dst_rq, p);
6903 * We want to potentially raise env.dst_cpu's OPP.
6905 update_capacity_of(env->dst_cpu);
6907 raw_spin_unlock(&env->dst_rq->lock);
6910 #ifdef CONFIG_FAIR_GROUP_SCHED
6911 static void update_blocked_averages(int cpu)
6913 struct rq *rq = cpu_rq(cpu);
6914 struct cfs_rq *cfs_rq;
6915 unsigned long flags;
6917 raw_spin_lock_irqsave(&rq->lock, flags);
6918 update_rq_clock(rq);
6921 * Iterates the task_group tree in a bottom up fashion, see
6922 * list_add_leaf_cfs_rq() for details.
6924 for_each_leaf_cfs_rq(rq, cfs_rq) {
6925 /* throttled entities do not contribute to load */
6926 if (throttled_hierarchy(cfs_rq))
6929 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
6931 update_tg_load_avg(cfs_rq, 0);
6933 raw_spin_unlock_irqrestore(&rq->lock, flags);
6937 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6938 * This needs to be done in a top-down fashion because the load of a child
6939 * group is a fraction of its parents load.
6941 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6943 struct rq *rq = rq_of(cfs_rq);
6944 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6945 unsigned long now = jiffies;
6948 if (cfs_rq->last_h_load_update == now)
6951 cfs_rq->h_load_next = NULL;
6952 for_each_sched_entity(se) {
6953 cfs_rq = cfs_rq_of(se);
6954 cfs_rq->h_load_next = se;
6955 if (cfs_rq->last_h_load_update == now)
6960 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6961 cfs_rq->last_h_load_update = now;
6964 while ((se = cfs_rq->h_load_next) != NULL) {
6965 load = cfs_rq->h_load;
6966 load = div64_ul(load * se->avg.load_avg,
6967 cfs_rq_load_avg(cfs_rq) + 1);
6968 cfs_rq = group_cfs_rq(se);
6969 cfs_rq->h_load = load;
6970 cfs_rq->last_h_load_update = now;
6974 static unsigned long task_h_load(struct task_struct *p)
6976 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6978 update_cfs_rq_h_load(cfs_rq);
6979 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6980 cfs_rq_load_avg(cfs_rq) + 1);
6983 static inline void update_blocked_averages(int cpu)
6985 struct rq *rq = cpu_rq(cpu);
6986 struct cfs_rq *cfs_rq = &rq->cfs;
6987 unsigned long flags;
6989 raw_spin_lock_irqsave(&rq->lock, flags);
6990 update_rq_clock(rq);
6991 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6992 raw_spin_unlock_irqrestore(&rq->lock, flags);
6995 static unsigned long task_h_load(struct task_struct *p)
6997 return p->se.avg.load_avg;
7001 /********** Helpers for find_busiest_group ************************/
7004 * sg_lb_stats - stats of a sched_group required for load_balancing
7006 struct sg_lb_stats {
7007 unsigned long avg_load; /*Avg load across the CPUs of the group */
7008 unsigned long group_load; /* Total load over the CPUs of the group */
7009 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7010 unsigned long load_per_task;
7011 unsigned long group_capacity;
7012 unsigned long group_util; /* Total utilization of the group */
7013 unsigned int sum_nr_running; /* Nr tasks running in the group */
7014 unsigned int idle_cpus;
7015 unsigned int group_weight;
7016 enum group_type group_type;
7017 int group_no_capacity;
7018 int group_misfit_task; /* A cpu has a task too big for its capacity */
7019 #ifdef CONFIG_NUMA_BALANCING
7020 unsigned int nr_numa_running;
7021 unsigned int nr_preferred_running;
7026 * sd_lb_stats - Structure to store the statistics of a sched_domain
7027 * during load balancing.
7029 struct sd_lb_stats {
7030 struct sched_group *busiest; /* Busiest group in this sd */
7031 struct sched_group *local; /* Local group in this sd */
7032 unsigned long total_load; /* Total load of all groups in sd */
7033 unsigned long total_capacity; /* Total capacity of all groups in sd */
7034 unsigned long avg_load; /* Average load across all groups in sd */
7036 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7037 struct sg_lb_stats local_stat; /* Statistics of the local group */
7040 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7043 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7044 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7045 * We must however clear busiest_stat::avg_load because
7046 * update_sd_pick_busiest() reads this before assignment.
7048 *sds = (struct sd_lb_stats){
7052 .total_capacity = 0UL,
7055 .sum_nr_running = 0,
7056 .group_type = group_other,
7062 * get_sd_load_idx - Obtain the load index for a given sched domain.
7063 * @sd: The sched_domain whose load_idx is to be obtained.
7064 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7066 * Return: The load index.
7068 static inline int get_sd_load_idx(struct sched_domain *sd,
7069 enum cpu_idle_type idle)
7075 load_idx = sd->busy_idx;
7078 case CPU_NEWLY_IDLE:
7079 load_idx = sd->newidle_idx;
7082 load_idx = sd->idle_idx;
7089 static unsigned long scale_rt_capacity(int cpu)
7091 struct rq *rq = cpu_rq(cpu);
7092 u64 total, used, age_stamp, avg;
7096 * Since we're reading these variables without serialization make sure
7097 * we read them once before doing sanity checks on them.
7099 age_stamp = READ_ONCE(rq->age_stamp);
7100 avg = READ_ONCE(rq->rt_avg);
7101 delta = __rq_clock_broken(rq) - age_stamp;
7103 if (unlikely(delta < 0))
7106 total = sched_avg_period() + delta;
7108 used = div_u64(avg, total);
7111 * deadline bandwidth is defined at system level so we must
7112 * weight this bandwidth with the max capacity of the system.
7113 * As a reminder, avg_bw is 20bits width and
7114 * scale_cpu_capacity is 10 bits width
7116 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7118 if (likely(used < SCHED_CAPACITY_SCALE))
7119 return SCHED_CAPACITY_SCALE - used;
7124 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7126 raw_spin_lock_init(&mcc->lock);
7131 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7133 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7134 struct sched_group *sdg = sd->groups;
7135 struct max_cpu_capacity *mcc;
7136 unsigned long max_capacity;
7138 unsigned long flags;
7140 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7142 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7144 raw_spin_lock_irqsave(&mcc->lock, flags);
7145 max_capacity = mcc->val;
7146 max_cap_cpu = mcc->cpu;
7148 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7149 (max_capacity < capacity)) {
7150 mcc->val = capacity;
7152 #ifdef CONFIG_SCHED_DEBUG
7153 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7154 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7159 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7161 skip_unlock: __attribute__ ((unused));
7162 capacity *= scale_rt_capacity(cpu);
7163 capacity >>= SCHED_CAPACITY_SHIFT;
7168 cpu_rq(cpu)->cpu_capacity = capacity;
7169 sdg->sgc->capacity = capacity;
7170 sdg->sgc->max_capacity = capacity;
7173 void update_group_capacity(struct sched_domain *sd, int cpu)
7175 struct sched_domain *child = sd->child;
7176 struct sched_group *group, *sdg = sd->groups;
7177 unsigned long capacity, max_capacity;
7178 unsigned long interval;
7180 interval = msecs_to_jiffies(sd->balance_interval);
7181 interval = clamp(interval, 1UL, max_load_balance_interval);
7182 sdg->sgc->next_update = jiffies + interval;
7185 update_cpu_capacity(sd, cpu);
7192 if (child->flags & SD_OVERLAP) {
7194 * SD_OVERLAP domains cannot assume that child groups
7195 * span the current group.
7198 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7199 struct sched_group_capacity *sgc;
7200 struct rq *rq = cpu_rq(cpu);
7203 * build_sched_domains() -> init_sched_groups_capacity()
7204 * gets here before we've attached the domains to the
7207 * Use capacity_of(), which is set irrespective of domains
7208 * in update_cpu_capacity().
7210 * This avoids capacity from being 0 and
7211 * causing divide-by-zero issues on boot.
7213 if (unlikely(!rq->sd)) {
7214 capacity += capacity_of(cpu);
7216 sgc = rq->sd->groups->sgc;
7217 capacity += sgc->capacity;
7220 max_capacity = max(capacity, max_capacity);
7224 * !SD_OVERLAP domains can assume that child groups
7225 * span the current group.
7228 group = child->groups;
7230 struct sched_group_capacity *sgc = group->sgc;
7232 capacity += sgc->capacity;
7233 max_capacity = max(sgc->max_capacity, max_capacity);
7234 group = group->next;
7235 } while (group != child->groups);
7238 sdg->sgc->capacity = capacity;
7239 sdg->sgc->max_capacity = max_capacity;
7243 * Check whether the capacity of the rq has been noticeably reduced by side
7244 * activity. The imbalance_pct is used for the threshold.
7245 * Return true is the capacity is reduced
7248 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7250 return ((rq->cpu_capacity * sd->imbalance_pct) <
7251 (rq->cpu_capacity_orig * 100));
7255 * Group imbalance indicates (and tries to solve) the problem where balancing
7256 * groups is inadequate due to tsk_cpus_allowed() constraints.
7258 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7259 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7262 * { 0 1 2 3 } { 4 5 6 7 }
7265 * If we were to balance group-wise we'd place two tasks in the first group and
7266 * two tasks in the second group. Clearly this is undesired as it will overload
7267 * cpu 3 and leave one of the cpus in the second group unused.
7269 * The current solution to this issue is detecting the skew in the first group
7270 * by noticing the lower domain failed to reach balance and had difficulty
7271 * moving tasks due to affinity constraints.
7273 * When this is so detected; this group becomes a candidate for busiest; see
7274 * update_sd_pick_busiest(). And calculate_imbalance() and
7275 * find_busiest_group() avoid some of the usual balance conditions to allow it
7276 * to create an effective group imbalance.
7278 * This is a somewhat tricky proposition since the next run might not find the
7279 * group imbalance and decide the groups need to be balanced again. A most
7280 * subtle and fragile situation.
7283 static inline int sg_imbalanced(struct sched_group *group)
7285 return group->sgc->imbalance;
7289 * group_has_capacity returns true if the group has spare capacity that could
7290 * be used by some tasks.
7291 * We consider that a group has spare capacity if the * number of task is
7292 * smaller than the number of CPUs or if the utilization is lower than the
7293 * available capacity for CFS tasks.
7294 * For the latter, we use a threshold to stabilize the state, to take into
7295 * account the variance of the tasks' load and to return true if the available
7296 * capacity in meaningful for the load balancer.
7297 * As an example, an available capacity of 1% can appear but it doesn't make
7298 * any benefit for the load balance.
7301 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7303 if (sgs->sum_nr_running < sgs->group_weight)
7306 if ((sgs->group_capacity * 100) >
7307 (sgs->group_util * env->sd->imbalance_pct))
7314 * group_is_overloaded returns true if the group has more tasks than it can
7316 * group_is_overloaded is not equals to !group_has_capacity because a group
7317 * with the exact right number of tasks, has no more spare capacity but is not
7318 * overloaded so both group_has_capacity and group_is_overloaded return
7322 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7324 if (sgs->sum_nr_running <= sgs->group_weight)
7327 if ((sgs->group_capacity * 100) <
7328 (sgs->group_util * env->sd->imbalance_pct))
7336 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7337 * per-cpu capacity than sched_group ref.
7340 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7342 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7343 ref->sgc->max_capacity;
7347 group_type group_classify(struct sched_group *group,
7348 struct sg_lb_stats *sgs)
7350 if (sgs->group_no_capacity)
7351 return group_overloaded;
7353 if (sg_imbalanced(group))
7354 return group_imbalanced;
7356 if (sgs->group_misfit_task)
7357 return group_misfit_task;
7363 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7364 * @env: The load balancing environment.
7365 * @group: sched_group whose statistics are to be updated.
7366 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7367 * @local_group: Does group contain this_cpu.
7368 * @sgs: variable to hold the statistics for this group.
7369 * @overload: Indicate more than one runnable task for any CPU.
7370 * @overutilized: Indicate overutilization for any CPU.
7372 static inline void update_sg_lb_stats(struct lb_env *env,
7373 struct sched_group *group, int load_idx,
7374 int local_group, struct sg_lb_stats *sgs,
7375 bool *overload, bool *overutilized)
7380 memset(sgs, 0, sizeof(*sgs));
7382 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7383 struct rq *rq = cpu_rq(i);
7385 /* Bias balancing toward cpus of our domain */
7387 load = target_load(i, load_idx);
7389 load = source_load(i, load_idx);
7391 sgs->group_load += load;
7392 sgs->group_util += cpu_util(i);
7393 sgs->sum_nr_running += rq->cfs.h_nr_running;
7395 nr_running = rq->nr_running;
7399 #ifdef CONFIG_NUMA_BALANCING
7400 sgs->nr_numa_running += rq->nr_numa_running;
7401 sgs->nr_preferred_running += rq->nr_preferred_running;
7403 sgs->sum_weighted_load += weighted_cpuload(i);
7405 * No need to call idle_cpu() if nr_running is not 0
7407 if (!nr_running && idle_cpu(i))
7410 if (cpu_overutilized(i)) {
7411 *overutilized = true;
7412 if (!sgs->group_misfit_task && rq->misfit_task)
7413 sgs->group_misfit_task = capacity_of(i);
7417 /* Adjust by relative CPU capacity of the group */
7418 sgs->group_capacity = group->sgc->capacity;
7419 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7421 if (sgs->sum_nr_running)
7422 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7424 sgs->group_weight = group->group_weight;
7426 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7427 sgs->group_type = group_classify(group, sgs);
7431 * update_sd_pick_busiest - return 1 on busiest group
7432 * @env: The load balancing environment.
7433 * @sds: sched_domain statistics
7434 * @sg: sched_group candidate to be checked for being the busiest
7435 * @sgs: sched_group statistics
7437 * Determine if @sg is a busier group than the previously selected
7440 * Return: %true if @sg is a busier group than the previously selected
7441 * busiest group. %false otherwise.
7443 static bool update_sd_pick_busiest(struct lb_env *env,
7444 struct sd_lb_stats *sds,
7445 struct sched_group *sg,
7446 struct sg_lb_stats *sgs)
7448 struct sg_lb_stats *busiest = &sds->busiest_stat;
7450 if (sgs->group_type > busiest->group_type)
7453 if (sgs->group_type < busiest->group_type)
7457 * Candidate sg doesn't face any serious load-balance problems
7458 * so don't pick it if the local sg is already filled up.
7460 if (sgs->group_type == group_other &&
7461 !group_has_capacity(env, &sds->local_stat))
7464 if (sgs->avg_load <= busiest->avg_load)
7468 * Candiate sg has no more than one task per cpu and has higher
7469 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7471 if (sgs->sum_nr_running <= sgs->group_weight &&
7472 group_smaller_cpu_capacity(sds->local, sg))
7475 /* This is the busiest node in its class. */
7476 if (!(env->sd->flags & SD_ASYM_PACKING))
7480 * ASYM_PACKING needs to move all the work to the lowest
7481 * numbered CPUs in the group, therefore mark all groups
7482 * higher than ourself as busy.
7484 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7488 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7495 #ifdef CONFIG_NUMA_BALANCING
7496 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7498 if (sgs->sum_nr_running > sgs->nr_numa_running)
7500 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7505 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7507 if (rq->nr_running > rq->nr_numa_running)
7509 if (rq->nr_running > rq->nr_preferred_running)
7514 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7519 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7523 #endif /* CONFIG_NUMA_BALANCING */
7526 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7527 * @env: The load balancing environment.
7528 * @sds: variable to hold the statistics for this sched_domain.
7530 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7532 struct sched_domain *child = env->sd->child;
7533 struct sched_group *sg = env->sd->groups;
7534 struct sg_lb_stats tmp_sgs;
7535 int load_idx, prefer_sibling = 0;
7536 bool overload = false, overutilized = false;
7538 if (child && child->flags & SD_PREFER_SIBLING)
7541 load_idx = get_sd_load_idx(env->sd, env->idle);
7544 struct sg_lb_stats *sgs = &tmp_sgs;
7547 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7550 sgs = &sds->local_stat;
7552 if (env->idle != CPU_NEWLY_IDLE ||
7553 time_after_eq(jiffies, sg->sgc->next_update))
7554 update_group_capacity(env->sd, env->dst_cpu);
7557 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7558 &overload, &overutilized);
7564 * In case the child domain prefers tasks go to siblings
7565 * first, lower the sg capacity so that we'll try
7566 * and move all the excess tasks away. We lower the capacity
7567 * of a group only if the local group has the capacity to fit
7568 * these excess tasks. The extra check prevents the case where
7569 * you always pull from the heaviest group when it is already
7570 * under-utilized (possible with a large weight task outweighs
7571 * the tasks on the system).
7573 if (prefer_sibling && sds->local &&
7574 group_has_capacity(env, &sds->local_stat) &&
7575 (sgs->sum_nr_running > 1)) {
7576 sgs->group_no_capacity = 1;
7577 sgs->group_type = group_classify(sg, sgs);
7581 * Ignore task groups with misfit tasks if local group has no
7582 * capacity or if per-cpu capacity isn't higher.
7584 if (sgs->group_type == group_misfit_task &&
7585 (!group_has_capacity(env, &sds->local_stat) ||
7586 !group_smaller_cpu_capacity(sg, sds->local)))
7587 sgs->group_type = group_other;
7589 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7591 sds->busiest_stat = *sgs;
7595 /* Now, start updating sd_lb_stats */
7596 sds->total_load += sgs->group_load;
7597 sds->total_capacity += sgs->group_capacity;
7600 } while (sg != env->sd->groups);
7602 if (env->sd->flags & SD_NUMA)
7603 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7605 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7607 if (!env->sd->parent) {
7608 /* update overload indicator if we are at root domain */
7609 if (env->dst_rq->rd->overload != overload)
7610 env->dst_rq->rd->overload = overload;
7612 /* Update over-utilization (tipping point, U >= 0) indicator */
7613 if (env->dst_rq->rd->overutilized != overutilized) {
7614 env->dst_rq->rd->overutilized = overutilized;
7615 trace_sched_overutilized(overutilized);
7618 if (!env->dst_rq->rd->overutilized && overutilized) {
7619 env->dst_rq->rd->overutilized = true;
7620 trace_sched_overutilized(true);
7627 * check_asym_packing - Check to see if the group is packed into the
7630 * This is primarily intended to used at the sibling level. Some
7631 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7632 * case of POWER7, it can move to lower SMT modes only when higher
7633 * threads are idle. When in lower SMT modes, the threads will
7634 * perform better since they share less core resources. Hence when we
7635 * have idle threads, we want them to be the higher ones.
7637 * This packing function is run on idle threads. It checks to see if
7638 * the busiest CPU in this domain (core in the P7 case) has a higher
7639 * CPU number than the packing function is being run on. Here we are
7640 * assuming lower CPU number will be equivalent to lower a SMT thread
7643 * Return: 1 when packing is required and a task should be moved to
7644 * this CPU. The amount of the imbalance is returned in *imbalance.
7646 * @env: The load balancing environment.
7647 * @sds: Statistics of the sched_domain which is to be packed
7649 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7653 if (!(env->sd->flags & SD_ASYM_PACKING))
7659 busiest_cpu = group_first_cpu(sds->busiest);
7660 if (env->dst_cpu > busiest_cpu)
7663 env->imbalance = DIV_ROUND_CLOSEST(
7664 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7665 SCHED_CAPACITY_SCALE);
7671 * fix_small_imbalance - Calculate the minor imbalance that exists
7672 * amongst the groups of a sched_domain, during
7674 * @env: The load balancing environment.
7675 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7678 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7680 unsigned long tmp, capa_now = 0, capa_move = 0;
7681 unsigned int imbn = 2;
7682 unsigned long scaled_busy_load_per_task;
7683 struct sg_lb_stats *local, *busiest;
7685 local = &sds->local_stat;
7686 busiest = &sds->busiest_stat;
7688 if (!local->sum_nr_running)
7689 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7690 else if (busiest->load_per_task > local->load_per_task)
7693 scaled_busy_load_per_task =
7694 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7695 busiest->group_capacity;
7697 if (busiest->avg_load + scaled_busy_load_per_task >=
7698 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7699 env->imbalance = busiest->load_per_task;
7704 * OK, we don't have enough imbalance to justify moving tasks,
7705 * however we may be able to increase total CPU capacity used by
7709 capa_now += busiest->group_capacity *
7710 min(busiest->load_per_task, busiest->avg_load);
7711 capa_now += local->group_capacity *
7712 min(local->load_per_task, local->avg_load);
7713 capa_now /= SCHED_CAPACITY_SCALE;
7715 /* Amount of load we'd subtract */
7716 if (busiest->avg_load > scaled_busy_load_per_task) {
7717 capa_move += busiest->group_capacity *
7718 min(busiest->load_per_task,
7719 busiest->avg_load - scaled_busy_load_per_task);
7722 /* Amount of load we'd add */
7723 if (busiest->avg_load * busiest->group_capacity <
7724 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7725 tmp = (busiest->avg_load * busiest->group_capacity) /
7726 local->group_capacity;
7728 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7729 local->group_capacity;
7731 capa_move += local->group_capacity *
7732 min(local->load_per_task, local->avg_load + tmp);
7733 capa_move /= SCHED_CAPACITY_SCALE;
7735 /* Move if we gain throughput */
7736 if (capa_move > capa_now)
7737 env->imbalance = busiest->load_per_task;
7741 * calculate_imbalance - Calculate the amount of imbalance present within the
7742 * groups of a given sched_domain during load balance.
7743 * @env: load balance environment
7744 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7746 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7748 unsigned long max_pull, load_above_capacity = ~0UL;
7749 struct sg_lb_stats *local, *busiest;
7751 local = &sds->local_stat;
7752 busiest = &sds->busiest_stat;
7754 if (busiest->group_type == group_imbalanced) {
7756 * In the group_imb case we cannot rely on group-wide averages
7757 * to ensure cpu-load equilibrium, look at wider averages. XXX
7759 busiest->load_per_task =
7760 min(busiest->load_per_task, sds->avg_load);
7764 * In the presence of smp nice balancing, certain scenarios can have
7765 * max load less than avg load(as we skip the groups at or below
7766 * its cpu_capacity, while calculating max_load..)
7768 if (busiest->avg_load <= sds->avg_load ||
7769 local->avg_load >= sds->avg_load) {
7770 /* Misfitting tasks should be migrated in any case */
7771 if (busiest->group_type == group_misfit_task) {
7772 env->imbalance = busiest->group_misfit_task;
7777 * Busiest group is overloaded, local is not, use the spare
7778 * cycles to maximize throughput
7780 if (busiest->group_type == group_overloaded &&
7781 local->group_type <= group_misfit_task) {
7782 env->imbalance = busiest->load_per_task;
7787 return fix_small_imbalance(env, sds);
7791 * If there aren't any idle cpus, avoid creating some.
7793 if (busiest->group_type == group_overloaded &&
7794 local->group_type == group_overloaded) {
7795 load_above_capacity = busiest->sum_nr_running *
7797 if (load_above_capacity > busiest->group_capacity)
7798 load_above_capacity -= busiest->group_capacity;
7800 load_above_capacity = ~0UL;
7804 * We're trying to get all the cpus to the average_load, so we don't
7805 * want to push ourselves above the average load, nor do we wish to
7806 * reduce the max loaded cpu below the average load. At the same time,
7807 * we also don't want to reduce the group load below the group capacity
7808 * (so that we can implement power-savings policies etc). Thus we look
7809 * for the minimum possible imbalance.
7811 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7813 /* How much load to actually move to equalise the imbalance */
7814 env->imbalance = min(
7815 max_pull * busiest->group_capacity,
7816 (sds->avg_load - local->avg_load) * local->group_capacity
7817 ) / SCHED_CAPACITY_SCALE;
7819 /* Boost imbalance to allow misfit task to be balanced. */
7820 if (busiest->group_type == group_misfit_task)
7821 env->imbalance = max_t(long, env->imbalance,
7822 busiest->group_misfit_task);
7825 * if *imbalance is less than the average load per runnable task
7826 * there is no guarantee that any tasks will be moved so we'll have
7827 * a think about bumping its value to force at least one task to be
7830 if (env->imbalance < busiest->load_per_task)
7831 return fix_small_imbalance(env, sds);
7834 /******* find_busiest_group() helpers end here *********************/
7837 * find_busiest_group - Returns the busiest group within the sched_domain
7838 * if there is an imbalance. If there isn't an imbalance, and
7839 * the user has opted for power-savings, it returns a group whose
7840 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7841 * such a group exists.
7843 * Also calculates the amount of weighted load which should be moved
7844 * to restore balance.
7846 * @env: The load balancing environment.
7848 * Return: - The busiest group if imbalance exists.
7849 * - If no imbalance and user has opted for power-savings balance,
7850 * return the least loaded group whose CPUs can be
7851 * put to idle by rebalancing its tasks onto our group.
7853 static struct sched_group *find_busiest_group(struct lb_env *env)
7855 struct sg_lb_stats *local, *busiest;
7856 struct sd_lb_stats sds;
7858 init_sd_lb_stats(&sds);
7861 * Compute the various statistics relavent for load balancing at
7864 update_sd_lb_stats(env, &sds);
7866 if (energy_aware() && !env->dst_rq->rd->overutilized)
7869 local = &sds.local_stat;
7870 busiest = &sds.busiest_stat;
7872 /* ASYM feature bypasses nice load balance check */
7873 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7874 check_asym_packing(env, &sds))
7877 /* There is no busy sibling group to pull tasks from */
7878 if (!sds.busiest || busiest->sum_nr_running == 0)
7881 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7882 / sds.total_capacity;
7885 * If the busiest group is imbalanced the below checks don't
7886 * work because they assume all things are equal, which typically
7887 * isn't true due to cpus_allowed constraints and the like.
7889 if (busiest->group_type == group_imbalanced)
7892 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7893 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7894 busiest->group_no_capacity)
7897 /* Misfitting tasks should be dealt with regardless of the avg load */
7898 if (busiest->group_type == group_misfit_task) {
7903 * If the local group is busier than the selected busiest group
7904 * don't try and pull any tasks.
7906 if (local->avg_load >= busiest->avg_load)
7910 * Don't pull any tasks if this group is already above the domain
7913 if (local->avg_load >= sds.avg_load)
7916 if (env->idle == CPU_IDLE) {
7918 * This cpu is idle. If the busiest group is not overloaded
7919 * and there is no imbalance between this and busiest group
7920 * wrt idle cpus, it is balanced. The imbalance becomes
7921 * significant if the diff is greater than 1 otherwise we
7922 * might end up to just move the imbalance on another group
7924 if ((busiest->group_type != group_overloaded) &&
7925 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7926 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7930 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7931 * imbalance_pct to be conservative.
7933 if (100 * busiest->avg_load <=
7934 env->sd->imbalance_pct * local->avg_load)
7939 env->busiest_group_type = busiest->group_type;
7940 /* Looks like there is an imbalance. Compute it */
7941 calculate_imbalance(env, &sds);
7950 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7952 static struct rq *find_busiest_queue(struct lb_env *env,
7953 struct sched_group *group)
7955 struct rq *busiest = NULL, *rq;
7956 unsigned long busiest_load = 0, busiest_capacity = 1;
7959 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7960 unsigned long capacity, wl;
7964 rt = fbq_classify_rq(rq);
7967 * We classify groups/runqueues into three groups:
7968 * - regular: there are !numa tasks
7969 * - remote: there are numa tasks that run on the 'wrong' node
7970 * - all: there is no distinction
7972 * In order to avoid migrating ideally placed numa tasks,
7973 * ignore those when there's better options.
7975 * If we ignore the actual busiest queue to migrate another
7976 * task, the next balance pass can still reduce the busiest
7977 * queue by moving tasks around inside the node.
7979 * If we cannot move enough load due to this classification
7980 * the next pass will adjust the group classification and
7981 * allow migration of more tasks.
7983 * Both cases only affect the total convergence complexity.
7985 if (rt > env->fbq_type)
7988 capacity = capacity_of(i);
7990 wl = weighted_cpuload(i);
7993 * When comparing with imbalance, use weighted_cpuload()
7994 * which is not scaled with the cpu capacity.
7997 if (rq->nr_running == 1 && wl > env->imbalance &&
7998 !check_cpu_capacity(rq, env->sd) &&
7999 env->busiest_group_type != group_misfit_task)
8003 * For the load comparisons with the other cpu's, consider
8004 * the weighted_cpuload() scaled with the cpu capacity, so
8005 * that the load can be moved away from the cpu that is
8006 * potentially running at a lower capacity.
8008 * Thus we're looking for max(wl_i / capacity_i), crosswise
8009 * multiplication to rid ourselves of the division works out
8010 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8011 * our previous maximum.
8013 if (wl * busiest_capacity > busiest_load * capacity) {
8015 busiest_capacity = capacity;
8024 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8025 * so long as it is large enough.
8027 #define MAX_PINNED_INTERVAL 512
8029 /* Working cpumask for load_balance and load_balance_newidle. */
8030 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8032 static int need_active_balance(struct lb_env *env)
8034 struct sched_domain *sd = env->sd;
8036 if (env->idle == CPU_NEWLY_IDLE) {
8039 * ASYM_PACKING needs to force migrate tasks from busy but
8040 * higher numbered CPUs in order to pack all tasks in the
8041 * lowest numbered CPUs.
8043 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8048 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8049 * It's worth migrating the task if the src_cpu's capacity is reduced
8050 * because of other sched_class or IRQs if more capacity stays
8051 * available on dst_cpu.
8053 if ((env->idle != CPU_NOT_IDLE) &&
8054 (env->src_rq->cfs.h_nr_running == 1)) {
8055 if ((check_cpu_capacity(env->src_rq, sd)) &&
8056 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8060 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8061 env->src_rq->cfs.h_nr_running == 1 &&
8062 cpu_overutilized(env->src_cpu) &&
8063 !cpu_overutilized(env->dst_cpu)) {
8067 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8070 static int active_load_balance_cpu_stop(void *data);
8072 static int should_we_balance(struct lb_env *env)
8074 struct sched_group *sg = env->sd->groups;
8075 struct cpumask *sg_cpus, *sg_mask;
8076 int cpu, balance_cpu = -1;
8079 * In the newly idle case, we will allow all the cpu's
8080 * to do the newly idle load balance.
8082 if (env->idle == CPU_NEWLY_IDLE)
8085 sg_cpus = sched_group_cpus(sg);
8086 sg_mask = sched_group_mask(sg);
8087 /* Try to find first idle cpu */
8088 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8089 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8096 if (balance_cpu == -1)
8097 balance_cpu = group_balance_cpu(sg);
8100 * First idle cpu or the first cpu(busiest) in this sched group
8101 * is eligible for doing load balancing at this and above domains.
8103 return balance_cpu == env->dst_cpu;
8107 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8108 * tasks if there is an imbalance.
8110 static int load_balance(int this_cpu, struct rq *this_rq,
8111 struct sched_domain *sd, enum cpu_idle_type idle,
8112 int *continue_balancing)
8114 int ld_moved, cur_ld_moved, active_balance = 0;
8115 struct sched_domain *sd_parent = sd->parent;
8116 struct sched_group *group;
8118 unsigned long flags;
8119 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8121 struct lb_env env = {
8123 .dst_cpu = this_cpu,
8125 .dst_grpmask = sched_group_cpus(sd->groups),
8127 .loop_break = sched_nr_migrate_break,
8130 .tasks = LIST_HEAD_INIT(env.tasks),
8134 * For NEWLY_IDLE load_balancing, we don't need to consider
8135 * other cpus in our group
8137 if (idle == CPU_NEWLY_IDLE)
8138 env.dst_grpmask = NULL;
8140 cpumask_copy(cpus, cpu_active_mask);
8142 schedstat_inc(sd, lb_count[idle]);
8145 if (!should_we_balance(&env)) {
8146 *continue_balancing = 0;
8150 group = find_busiest_group(&env);
8152 schedstat_inc(sd, lb_nobusyg[idle]);
8156 busiest = find_busiest_queue(&env, group);
8158 schedstat_inc(sd, lb_nobusyq[idle]);
8162 BUG_ON(busiest == env.dst_rq);
8164 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8166 env.src_cpu = busiest->cpu;
8167 env.src_rq = busiest;
8170 if (busiest->nr_running > 1) {
8172 * Attempt to move tasks. If find_busiest_group has found
8173 * an imbalance but busiest->nr_running <= 1, the group is
8174 * still unbalanced. ld_moved simply stays zero, so it is
8175 * correctly treated as an imbalance.
8177 env.flags |= LBF_ALL_PINNED;
8178 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8181 raw_spin_lock_irqsave(&busiest->lock, flags);
8184 * cur_ld_moved - load moved in current iteration
8185 * ld_moved - cumulative load moved across iterations
8187 cur_ld_moved = detach_tasks(&env);
8189 * We want to potentially lower env.src_cpu's OPP.
8192 update_capacity_of(env.src_cpu);
8195 * We've detached some tasks from busiest_rq. Every
8196 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8197 * unlock busiest->lock, and we are able to be sure
8198 * that nobody can manipulate the tasks in parallel.
8199 * See task_rq_lock() family for the details.
8202 raw_spin_unlock(&busiest->lock);
8206 ld_moved += cur_ld_moved;
8209 local_irq_restore(flags);
8211 if (env.flags & LBF_NEED_BREAK) {
8212 env.flags &= ~LBF_NEED_BREAK;
8217 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8218 * us and move them to an alternate dst_cpu in our sched_group
8219 * where they can run. The upper limit on how many times we
8220 * iterate on same src_cpu is dependent on number of cpus in our
8223 * This changes load balance semantics a bit on who can move
8224 * load to a given_cpu. In addition to the given_cpu itself
8225 * (or a ilb_cpu acting on its behalf where given_cpu is
8226 * nohz-idle), we now have balance_cpu in a position to move
8227 * load to given_cpu. In rare situations, this may cause
8228 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8229 * _independently_ and at _same_ time to move some load to
8230 * given_cpu) causing exceess load to be moved to given_cpu.
8231 * This however should not happen so much in practice and
8232 * moreover subsequent load balance cycles should correct the
8233 * excess load moved.
8235 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8237 /* Prevent to re-select dst_cpu via env's cpus */
8238 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8240 env.dst_rq = cpu_rq(env.new_dst_cpu);
8241 env.dst_cpu = env.new_dst_cpu;
8242 env.flags &= ~LBF_DST_PINNED;
8244 env.loop_break = sched_nr_migrate_break;
8247 * Go back to "more_balance" rather than "redo" since we
8248 * need to continue with same src_cpu.
8254 * We failed to reach balance because of affinity.
8257 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8259 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8260 *group_imbalance = 1;
8263 /* All tasks on this runqueue were pinned by CPU affinity */
8264 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8265 cpumask_clear_cpu(cpu_of(busiest), cpus);
8266 if (!cpumask_empty(cpus)) {
8268 env.loop_break = sched_nr_migrate_break;
8271 goto out_all_pinned;
8276 schedstat_inc(sd, lb_failed[idle]);
8278 * Increment the failure counter only on periodic balance.
8279 * We do not want newidle balance, which can be very
8280 * frequent, pollute the failure counter causing
8281 * excessive cache_hot migrations and active balances.
8283 if (idle != CPU_NEWLY_IDLE)
8284 if (env.src_grp_nr_running > 1)
8285 sd->nr_balance_failed++;
8287 if (need_active_balance(&env)) {
8288 raw_spin_lock_irqsave(&busiest->lock, flags);
8290 /* don't kick the active_load_balance_cpu_stop,
8291 * if the curr task on busiest cpu can't be
8294 if (!cpumask_test_cpu(this_cpu,
8295 tsk_cpus_allowed(busiest->curr))) {
8296 raw_spin_unlock_irqrestore(&busiest->lock,
8298 env.flags |= LBF_ALL_PINNED;
8299 goto out_one_pinned;
8303 * ->active_balance synchronizes accesses to
8304 * ->active_balance_work. Once set, it's cleared
8305 * only after active load balance is finished.
8307 if (!busiest->active_balance) {
8308 busiest->active_balance = 1;
8309 busiest->push_cpu = this_cpu;
8312 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8314 if (active_balance) {
8315 stop_one_cpu_nowait(cpu_of(busiest),
8316 active_load_balance_cpu_stop, busiest,
8317 &busiest->active_balance_work);
8321 * We've kicked active balancing, reset the failure
8324 sd->nr_balance_failed = sd->cache_nice_tries+1;
8327 sd->nr_balance_failed = 0;
8329 if (likely(!active_balance)) {
8330 /* We were unbalanced, so reset the balancing interval */
8331 sd->balance_interval = sd->min_interval;
8334 * If we've begun active balancing, start to back off. This
8335 * case may not be covered by the all_pinned logic if there
8336 * is only 1 task on the busy runqueue (because we don't call
8339 if (sd->balance_interval < sd->max_interval)
8340 sd->balance_interval *= 2;
8347 * We reach balance although we may have faced some affinity
8348 * constraints. Clear the imbalance flag if it was set.
8351 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8353 if (*group_imbalance)
8354 *group_imbalance = 0;
8359 * We reach balance because all tasks are pinned at this level so
8360 * we can't migrate them. Let the imbalance flag set so parent level
8361 * can try to migrate them.
8363 schedstat_inc(sd, lb_balanced[idle]);
8365 sd->nr_balance_failed = 0;
8368 /* tune up the balancing interval */
8369 if (((env.flags & LBF_ALL_PINNED) &&
8370 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8371 (sd->balance_interval < sd->max_interval))
8372 sd->balance_interval *= 2;
8379 static inline unsigned long
8380 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8382 unsigned long interval = sd->balance_interval;
8385 interval *= sd->busy_factor;
8387 /* scale ms to jiffies */
8388 interval = msecs_to_jiffies(interval);
8389 interval = clamp(interval, 1UL, max_load_balance_interval);
8395 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8397 unsigned long interval, next;
8399 interval = get_sd_balance_interval(sd, cpu_busy);
8400 next = sd->last_balance + interval;
8402 if (time_after(*next_balance, next))
8403 *next_balance = next;
8407 * idle_balance is called by schedule() if this_cpu is about to become
8408 * idle. Attempts to pull tasks from other CPUs.
8410 static int idle_balance(struct rq *this_rq)
8412 unsigned long next_balance = jiffies + HZ;
8413 int this_cpu = this_rq->cpu;
8414 struct sched_domain *sd;
8415 int pulled_task = 0;
8417 long removed_util=0;
8419 idle_enter_fair(this_rq);
8422 * We must set idle_stamp _before_ calling idle_balance(), such that we
8423 * measure the duration of idle_balance() as idle time.
8425 this_rq->idle_stamp = rq_clock(this_rq);
8427 if (!energy_aware() &&
8428 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8429 !this_rq->rd->overload)) {
8431 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8433 update_next_balance(sd, 0, &next_balance);
8439 raw_spin_unlock(&this_rq->lock);
8442 * If removed_util_avg is !0 we most probably migrated some task away
8443 * from this_cpu. In this case we might be willing to trigger an OPP
8444 * update, but we want to do so if we don't find anybody else to pull
8445 * here (we will trigger an OPP update with the pulled task's enqueue
8448 * Record removed_util before calling update_blocked_averages, and use
8449 * it below (before returning) to see if an OPP update is required.
8451 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8452 update_blocked_averages(this_cpu);
8454 for_each_domain(this_cpu, sd) {
8455 int continue_balancing = 1;
8456 u64 t0, domain_cost;
8458 if (!(sd->flags & SD_LOAD_BALANCE))
8461 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8462 update_next_balance(sd, 0, &next_balance);
8466 if (sd->flags & SD_BALANCE_NEWIDLE) {
8467 t0 = sched_clock_cpu(this_cpu);
8469 pulled_task = load_balance(this_cpu, this_rq,
8471 &continue_balancing);
8473 domain_cost = sched_clock_cpu(this_cpu) - t0;
8474 if (domain_cost > sd->max_newidle_lb_cost)
8475 sd->max_newidle_lb_cost = domain_cost;
8477 curr_cost += domain_cost;
8480 update_next_balance(sd, 0, &next_balance);
8483 * Stop searching for tasks to pull if there are
8484 * now runnable tasks on this rq.
8486 if (pulled_task || this_rq->nr_running > 0)
8491 raw_spin_lock(&this_rq->lock);
8493 if (curr_cost > this_rq->max_idle_balance_cost)
8494 this_rq->max_idle_balance_cost = curr_cost;
8497 * While browsing the domains, we released the rq lock, a task could
8498 * have been enqueued in the meantime. Since we're not going idle,
8499 * pretend we pulled a task.
8501 if (this_rq->cfs.h_nr_running && !pulled_task)
8505 /* Move the next balance forward */
8506 if (time_after(this_rq->next_balance, next_balance))
8507 this_rq->next_balance = next_balance;
8509 /* Is there a task of a high priority class? */
8510 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8514 idle_exit_fair(this_rq);
8515 this_rq->idle_stamp = 0;
8516 } else if (removed_util) {
8518 * No task pulled and someone has been migrated away.
8519 * Good case to trigger an OPP update.
8521 update_capacity_of(this_cpu);
8528 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8529 * running tasks off the busiest CPU onto idle CPUs. It requires at
8530 * least 1 task to be running on each physical CPU where possible, and
8531 * avoids physical / logical imbalances.
8533 static int active_load_balance_cpu_stop(void *data)
8535 struct rq *busiest_rq = data;
8536 int busiest_cpu = cpu_of(busiest_rq);
8537 int target_cpu = busiest_rq->push_cpu;
8538 struct rq *target_rq = cpu_rq(target_cpu);
8539 struct sched_domain *sd;
8540 struct task_struct *p = NULL;
8542 raw_spin_lock_irq(&busiest_rq->lock);
8544 /* make sure the requested cpu hasn't gone down in the meantime */
8545 if (unlikely(busiest_cpu != smp_processor_id() ||
8546 !busiest_rq->active_balance))
8549 /* Is there any task to move? */
8550 if (busiest_rq->nr_running <= 1)
8554 * This condition is "impossible", if it occurs
8555 * we need to fix it. Originally reported by
8556 * Bjorn Helgaas on a 128-cpu setup.
8558 BUG_ON(busiest_rq == target_rq);
8560 /* Search for an sd spanning us and the target CPU. */
8562 for_each_domain(target_cpu, sd) {
8563 if ((sd->flags & SD_LOAD_BALANCE) &&
8564 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8569 struct lb_env env = {
8571 .dst_cpu = target_cpu,
8572 .dst_rq = target_rq,
8573 .src_cpu = busiest_rq->cpu,
8574 .src_rq = busiest_rq,
8578 schedstat_inc(sd, alb_count);
8580 p = detach_one_task(&env);
8582 schedstat_inc(sd, alb_pushed);
8584 * We want to potentially lower env.src_cpu's OPP.
8586 update_capacity_of(env.src_cpu);
8589 schedstat_inc(sd, alb_failed);
8593 busiest_rq->active_balance = 0;
8594 raw_spin_unlock(&busiest_rq->lock);
8597 attach_one_task(target_rq, p);
8604 static inline int on_null_domain(struct rq *rq)
8606 return unlikely(!rcu_dereference_sched(rq->sd));
8609 #ifdef CONFIG_NO_HZ_COMMON
8611 * idle load balancing details
8612 * - When one of the busy CPUs notice that there may be an idle rebalancing
8613 * needed, they will kick the idle load balancer, which then does idle
8614 * load balancing for all the idle CPUs.
8617 cpumask_var_t idle_cpus_mask;
8619 unsigned long next_balance; /* in jiffy units */
8620 } nohz ____cacheline_aligned;
8622 static inline int find_new_ilb(void)
8624 int ilb = cpumask_first(nohz.idle_cpus_mask);
8626 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8633 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8634 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8635 * CPU (if there is one).
8637 static void nohz_balancer_kick(void)
8641 nohz.next_balance++;
8643 ilb_cpu = find_new_ilb();
8645 if (ilb_cpu >= nr_cpu_ids)
8648 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8651 * Use smp_send_reschedule() instead of resched_cpu().
8652 * This way we generate a sched IPI on the target cpu which
8653 * is idle. And the softirq performing nohz idle load balance
8654 * will be run before returning from the IPI.
8656 smp_send_reschedule(ilb_cpu);
8660 static inline void nohz_balance_exit_idle(int cpu)
8662 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8664 * Completely isolated CPUs don't ever set, so we must test.
8666 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8667 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8668 atomic_dec(&nohz.nr_cpus);
8670 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8674 static inline void set_cpu_sd_state_busy(void)
8676 struct sched_domain *sd;
8677 int cpu = smp_processor_id();
8680 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8682 if (!sd || !sd->nohz_idle)
8686 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8691 void set_cpu_sd_state_idle(void)
8693 struct sched_domain *sd;
8694 int cpu = smp_processor_id();
8697 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8699 if (!sd || sd->nohz_idle)
8703 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8709 * This routine will record that the cpu is going idle with tick stopped.
8710 * This info will be used in performing idle load balancing in the future.
8712 void nohz_balance_enter_idle(int cpu)
8715 * If this cpu is going down, then nothing needs to be done.
8717 if (!cpu_active(cpu))
8720 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8724 * If we're a completely isolated CPU, we don't play.
8726 if (on_null_domain(cpu_rq(cpu)))
8729 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8730 atomic_inc(&nohz.nr_cpus);
8731 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8734 static int sched_ilb_notifier(struct notifier_block *nfb,
8735 unsigned long action, void *hcpu)
8737 switch (action & ~CPU_TASKS_FROZEN) {
8739 nohz_balance_exit_idle(smp_processor_id());
8747 static DEFINE_SPINLOCK(balancing);
8750 * Scale the max load_balance interval with the number of CPUs in the system.
8751 * This trades load-balance latency on larger machines for less cross talk.
8753 void update_max_interval(void)
8755 max_load_balance_interval = HZ*num_online_cpus()/10;
8759 * It checks each scheduling domain to see if it is due to be balanced,
8760 * and initiates a balancing operation if so.
8762 * Balancing parameters are set up in init_sched_domains.
8764 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8766 int continue_balancing = 1;
8768 unsigned long interval;
8769 struct sched_domain *sd;
8770 /* Earliest time when we have to do rebalance again */
8771 unsigned long next_balance = jiffies + 60*HZ;
8772 int update_next_balance = 0;
8773 int need_serialize, need_decay = 0;
8776 update_blocked_averages(cpu);
8779 for_each_domain(cpu, sd) {
8781 * Decay the newidle max times here because this is a regular
8782 * visit to all the domains. Decay ~1% per second.
8784 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8785 sd->max_newidle_lb_cost =
8786 (sd->max_newidle_lb_cost * 253) / 256;
8787 sd->next_decay_max_lb_cost = jiffies + HZ;
8790 max_cost += sd->max_newidle_lb_cost;
8792 if (!(sd->flags & SD_LOAD_BALANCE))
8796 * Stop the load balance at this level. There is another
8797 * CPU in our sched group which is doing load balancing more
8800 if (!continue_balancing) {
8806 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8808 need_serialize = sd->flags & SD_SERIALIZE;
8809 if (need_serialize) {
8810 if (!spin_trylock(&balancing))
8814 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8815 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8817 * The LBF_DST_PINNED logic could have changed
8818 * env->dst_cpu, so we can't know our idle
8819 * state even if we migrated tasks. Update it.
8821 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8823 sd->last_balance = jiffies;
8824 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8827 spin_unlock(&balancing);
8829 if (time_after(next_balance, sd->last_balance + interval)) {
8830 next_balance = sd->last_balance + interval;
8831 update_next_balance = 1;
8836 * Ensure the rq-wide value also decays but keep it at a
8837 * reasonable floor to avoid funnies with rq->avg_idle.
8839 rq->max_idle_balance_cost =
8840 max((u64)sysctl_sched_migration_cost, max_cost);
8845 * next_balance will be updated only when there is a need.
8846 * When the cpu is attached to null domain for ex, it will not be
8849 if (likely(update_next_balance)) {
8850 rq->next_balance = next_balance;
8852 #ifdef CONFIG_NO_HZ_COMMON
8854 * If this CPU has been elected to perform the nohz idle
8855 * balance. Other idle CPUs have already rebalanced with
8856 * nohz_idle_balance() and nohz.next_balance has been
8857 * updated accordingly. This CPU is now running the idle load
8858 * balance for itself and we need to update the
8859 * nohz.next_balance accordingly.
8861 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8862 nohz.next_balance = rq->next_balance;
8867 #ifdef CONFIG_NO_HZ_COMMON
8869 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8870 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8872 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8874 int this_cpu = this_rq->cpu;
8877 /* Earliest time when we have to do rebalance again */
8878 unsigned long next_balance = jiffies + 60*HZ;
8879 int update_next_balance = 0;
8881 if (idle != CPU_IDLE ||
8882 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8885 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8886 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8890 * If this cpu gets work to do, stop the load balancing
8891 * work being done for other cpus. Next load
8892 * balancing owner will pick it up.
8897 rq = cpu_rq(balance_cpu);
8900 * If time for next balance is due,
8903 if (time_after_eq(jiffies, rq->next_balance)) {
8904 raw_spin_lock_irq(&rq->lock);
8905 update_rq_clock(rq);
8906 update_idle_cpu_load(rq);
8907 raw_spin_unlock_irq(&rq->lock);
8908 rebalance_domains(rq, CPU_IDLE);
8911 if (time_after(next_balance, rq->next_balance)) {
8912 next_balance = rq->next_balance;
8913 update_next_balance = 1;
8918 * next_balance will be updated only when there is a need.
8919 * When the CPU is attached to null domain for ex, it will not be
8922 if (likely(update_next_balance))
8923 nohz.next_balance = next_balance;
8925 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8929 * Current heuristic for kicking the idle load balancer in the presence
8930 * of an idle cpu in the system.
8931 * - This rq has more than one task.
8932 * - This rq has at least one CFS task and the capacity of the CPU is
8933 * significantly reduced because of RT tasks or IRQs.
8934 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8935 * multiple busy cpu.
8936 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8937 * domain span are idle.
8939 static inline bool nohz_kick_needed(struct rq *rq)
8941 unsigned long now = jiffies;
8942 struct sched_domain *sd;
8943 struct sched_group_capacity *sgc;
8944 int nr_busy, cpu = rq->cpu;
8947 if (unlikely(rq->idle_balance))
8951 * We may be recently in ticked or tickless idle mode. At the first
8952 * busy tick after returning from idle, we will update the busy stats.
8954 set_cpu_sd_state_busy();
8955 nohz_balance_exit_idle(cpu);
8958 * None are in tickless mode and hence no need for NOHZ idle load
8961 if (likely(!atomic_read(&nohz.nr_cpus)))
8964 if (time_before(now, nohz.next_balance))
8967 if (rq->nr_running >= 2 &&
8968 (!energy_aware() || cpu_overutilized(cpu)))
8972 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8973 if (sd && !energy_aware()) {
8974 sgc = sd->groups->sgc;
8975 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8984 sd = rcu_dereference(rq->sd);
8986 if ((rq->cfs.h_nr_running >= 1) &&
8987 check_cpu_capacity(rq, sd)) {
8993 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8994 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8995 sched_domain_span(sd)) < cpu)) {
9005 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9009 * run_rebalance_domains is triggered when needed from the scheduler tick.
9010 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9012 static void run_rebalance_domains(struct softirq_action *h)
9014 struct rq *this_rq = this_rq();
9015 enum cpu_idle_type idle = this_rq->idle_balance ?
9016 CPU_IDLE : CPU_NOT_IDLE;
9019 * If this cpu has a pending nohz_balance_kick, then do the
9020 * balancing on behalf of the other idle cpus whose ticks are
9021 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9022 * give the idle cpus a chance to load balance. Else we may
9023 * load balance only within the local sched_domain hierarchy
9024 * and abort nohz_idle_balance altogether if we pull some load.
9026 nohz_idle_balance(this_rq, idle);
9027 rebalance_domains(this_rq, idle);
9031 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9033 void trigger_load_balance(struct rq *rq)
9035 /* Don't need to rebalance while attached to NULL domain */
9036 if (unlikely(on_null_domain(rq)))
9039 if (time_after_eq(jiffies, rq->next_balance))
9040 raise_softirq(SCHED_SOFTIRQ);
9041 #ifdef CONFIG_NO_HZ_COMMON
9042 if (nohz_kick_needed(rq))
9043 nohz_balancer_kick();
9047 static void rq_online_fair(struct rq *rq)
9051 update_runtime_enabled(rq);
9054 static void rq_offline_fair(struct rq *rq)
9058 /* Ensure any throttled groups are reachable by pick_next_task */
9059 unthrottle_offline_cfs_rqs(rq);
9062 #endif /* CONFIG_SMP */
9065 * scheduler tick hitting a task of our scheduling class:
9067 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9069 struct cfs_rq *cfs_rq;
9070 struct sched_entity *se = &curr->se;
9072 for_each_sched_entity(se) {
9073 cfs_rq = cfs_rq_of(se);
9074 entity_tick(cfs_rq, se, queued);
9077 if (static_branch_unlikely(&sched_numa_balancing))
9078 task_tick_numa(rq, curr);
9081 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9082 rq->rd->overutilized = true;
9083 trace_sched_overutilized(true);
9086 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9092 * called on fork with the child task as argument from the parent's context
9093 * - child not yet on the tasklist
9094 * - preemption disabled
9096 static void task_fork_fair(struct task_struct *p)
9098 struct cfs_rq *cfs_rq;
9099 struct sched_entity *se = &p->se, *curr;
9100 int this_cpu = smp_processor_id();
9101 struct rq *rq = this_rq();
9102 unsigned long flags;
9104 raw_spin_lock_irqsave(&rq->lock, flags);
9106 update_rq_clock(rq);
9108 cfs_rq = task_cfs_rq(current);
9109 curr = cfs_rq->curr;
9112 * Not only the cpu but also the task_group of the parent might have
9113 * been changed after parent->se.parent,cfs_rq were copied to
9114 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9115 * of child point to valid ones.
9118 __set_task_cpu(p, this_cpu);
9121 update_curr(cfs_rq);
9124 se->vruntime = curr->vruntime;
9125 place_entity(cfs_rq, se, 1);
9127 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9129 * Upon rescheduling, sched_class::put_prev_task() will place
9130 * 'current' within the tree based on its new key value.
9132 swap(curr->vruntime, se->vruntime);
9136 se->vruntime -= cfs_rq->min_vruntime;
9138 raw_spin_unlock_irqrestore(&rq->lock, flags);
9142 * Priority of the task has changed. Check to see if we preempt
9146 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9148 if (!task_on_rq_queued(p))
9152 * Reschedule if we are currently running on this runqueue and
9153 * our priority decreased, or if we are not currently running on
9154 * this runqueue and our priority is higher than the current's
9156 if (rq->curr == p) {
9157 if (p->prio > oldprio)
9160 check_preempt_curr(rq, p, 0);
9163 static inline bool vruntime_normalized(struct task_struct *p)
9165 struct sched_entity *se = &p->se;
9168 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9169 * the dequeue_entity(.flags=0) will already have normalized the
9176 * When !on_rq, vruntime of the task has usually NOT been normalized.
9177 * But there are some cases where it has already been normalized:
9179 * - A forked child which is waiting for being woken up by
9180 * wake_up_new_task().
9181 * - A task which has been woken up by try_to_wake_up() and
9182 * waiting for actually being woken up by sched_ttwu_pending().
9184 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9190 static void detach_task_cfs_rq(struct task_struct *p)
9192 struct sched_entity *se = &p->se;
9193 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9195 if (!vruntime_normalized(p)) {
9197 * Fix up our vruntime so that the current sleep doesn't
9198 * cause 'unlimited' sleep bonus.
9200 place_entity(cfs_rq, se, 0);
9201 se->vruntime -= cfs_rq->min_vruntime;
9204 /* Catch up with the cfs_rq and remove our load when we leave */
9205 detach_entity_load_avg(cfs_rq, se);
9208 static void attach_task_cfs_rq(struct task_struct *p)
9210 struct sched_entity *se = &p->se;
9211 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9213 #ifdef CONFIG_FAIR_GROUP_SCHED
9215 * Since the real-depth could have been changed (only FAIR
9216 * class maintain depth value), reset depth properly.
9218 se->depth = se->parent ? se->parent->depth + 1 : 0;
9221 /* Synchronize task with its cfs_rq */
9222 attach_entity_load_avg(cfs_rq, se);
9224 if (!vruntime_normalized(p))
9225 se->vruntime += cfs_rq->min_vruntime;
9228 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9230 detach_task_cfs_rq(p);
9233 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9235 attach_task_cfs_rq(p);
9237 if (task_on_rq_queued(p)) {
9239 * We were most likely switched from sched_rt, so
9240 * kick off the schedule if running, otherwise just see
9241 * if we can still preempt the current task.
9246 check_preempt_curr(rq, p, 0);
9250 /* Account for a task changing its policy or group.
9252 * This routine is mostly called to set cfs_rq->curr field when a task
9253 * migrates between groups/classes.
9255 static void set_curr_task_fair(struct rq *rq)
9257 struct sched_entity *se = &rq->curr->se;
9259 for_each_sched_entity(se) {
9260 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9262 set_next_entity(cfs_rq, se);
9263 /* ensure bandwidth has been allocated on our new cfs_rq */
9264 account_cfs_rq_runtime(cfs_rq, 0);
9268 void init_cfs_rq(struct cfs_rq *cfs_rq)
9270 cfs_rq->tasks_timeline = RB_ROOT;
9271 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9272 #ifndef CONFIG_64BIT
9273 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9276 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9277 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9281 #ifdef CONFIG_FAIR_GROUP_SCHED
9282 static void task_move_group_fair(struct task_struct *p)
9284 detach_task_cfs_rq(p);
9285 set_task_rq(p, task_cpu(p));
9288 /* Tell se's cfs_rq has been changed -- migrated */
9289 p->se.avg.last_update_time = 0;
9291 attach_task_cfs_rq(p);
9294 void free_fair_sched_group(struct task_group *tg)
9298 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9300 for_each_possible_cpu(i) {
9302 kfree(tg->cfs_rq[i]);
9305 remove_entity_load_avg(tg->se[i]);
9314 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9316 struct cfs_rq *cfs_rq;
9317 struct sched_entity *se;
9320 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9323 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9327 tg->shares = NICE_0_LOAD;
9329 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9331 for_each_possible_cpu(i) {
9332 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9333 GFP_KERNEL, cpu_to_node(i));
9337 se = kzalloc_node(sizeof(struct sched_entity),
9338 GFP_KERNEL, cpu_to_node(i));
9342 init_cfs_rq(cfs_rq);
9343 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9344 init_entity_runnable_average(se);
9355 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9357 struct rq *rq = cpu_rq(cpu);
9358 unsigned long flags;
9361 * Only empty task groups can be destroyed; so we can speculatively
9362 * check on_list without danger of it being re-added.
9364 if (!tg->cfs_rq[cpu]->on_list)
9367 raw_spin_lock_irqsave(&rq->lock, flags);
9368 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9369 raw_spin_unlock_irqrestore(&rq->lock, flags);
9372 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9373 struct sched_entity *se, int cpu,
9374 struct sched_entity *parent)
9376 struct rq *rq = cpu_rq(cpu);
9380 init_cfs_rq_runtime(cfs_rq);
9382 tg->cfs_rq[cpu] = cfs_rq;
9385 /* se could be NULL for root_task_group */
9390 se->cfs_rq = &rq->cfs;
9393 se->cfs_rq = parent->my_q;
9394 se->depth = parent->depth + 1;
9398 /* guarantee group entities always have weight */
9399 update_load_set(&se->load, NICE_0_LOAD);
9400 se->parent = parent;
9403 static DEFINE_MUTEX(shares_mutex);
9405 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9408 unsigned long flags;
9411 * We can't change the weight of the root cgroup.
9416 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9418 mutex_lock(&shares_mutex);
9419 if (tg->shares == shares)
9422 tg->shares = shares;
9423 for_each_possible_cpu(i) {
9424 struct rq *rq = cpu_rq(i);
9425 struct sched_entity *se;
9428 /* Propagate contribution to hierarchy */
9429 raw_spin_lock_irqsave(&rq->lock, flags);
9431 /* Possible calls to update_curr() need rq clock */
9432 update_rq_clock(rq);
9433 for_each_sched_entity(se)
9434 update_cfs_shares(group_cfs_rq(se));
9435 raw_spin_unlock_irqrestore(&rq->lock, flags);
9439 mutex_unlock(&shares_mutex);
9442 #else /* CONFIG_FAIR_GROUP_SCHED */
9444 void free_fair_sched_group(struct task_group *tg) { }
9446 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9451 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9453 #endif /* CONFIG_FAIR_GROUP_SCHED */
9456 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9458 struct sched_entity *se = &task->se;
9459 unsigned int rr_interval = 0;
9462 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9465 if (rq->cfs.load.weight)
9466 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9472 * All the scheduling class methods:
9474 const struct sched_class fair_sched_class = {
9475 .next = &idle_sched_class,
9476 .enqueue_task = enqueue_task_fair,
9477 .dequeue_task = dequeue_task_fair,
9478 .yield_task = yield_task_fair,
9479 .yield_to_task = yield_to_task_fair,
9481 .check_preempt_curr = check_preempt_wakeup,
9483 .pick_next_task = pick_next_task_fair,
9484 .put_prev_task = put_prev_task_fair,
9487 .select_task_rq = select_task_rq_fair,
9488 .migrate_task_rq = migrate_task_rq_fair,
9490 .rq_online = rq_online_fair,
9491 .rq_offline = rq_offline_fair,
9493 .task_waking = task_waking_fair,
9494 .task_dead = task_dead_fair,
9495 .set_cpus_allowed = set_cpus_allowed_common,
9498 .set_curr_task = set_curr_task_fair,
9499 .task_tick = task_tick_fair,
9500 .task_fork = task_fork_fair,
9502 .prio_changed = prio_changed_fair,
9503 .switched_from = switched_from_fair,
9504 .switched_to = switched_to_fair,
9506 .get_rr_interval = get_rr_interval_fair,
9508 .update_curr = update_curr_fair,
9510 #ifdef CONFIG_FAIR_GROUP_SCHED
9511 .task_move_group = task_move_group_fair,
9515 #ifdef CONFIG_SCHED_DEBUG
9516 void print_cfs_stats(struct seq_file *m, int cpu)
9518 struct cfs_rq *cfs_rq;
9521 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9522 print_cfs_rq(m, cpu, cfs_rq);
9526 #ifdef CONFIG_NUMA_BALANCING
9527 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9530 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9532 for_each_online_node(node) {
9533 if (p->numa_faults) {
9534 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9535 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9537 if (p->numa_group) {
9538 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9539 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9541 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9544 #endif /* CONFIG_NUMA_BALANCING */
9545 #endif /* CONFIG_SCHED_DEBUG */
9547 __init void init_sched_fair_class(void)
9550 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9552 #ifdef CONFIG_NO_HZ_COMMON
9553 nohz.next_balance = jiffies;
9554 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9555 cpu_notifier(sched_ilb_notifier, 0);