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 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->dst_cpu);
1411 task_numa_assign(env, cur, imp);
1415 * The dst_rq->curr isn't assigned. The protection for task_struct is
1418 if (cur && !assigned)
1419 put_task_struct(cur);
1422 static void task_numa_find_cpu(struct task_numa_env *env,
1423 long taskimp, long groupimp)
1427 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1428 /* Skip this CPU if the source task cannot migrate */
1429 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1433 task_numa_compare(env, taskimp, groupimp);
1437 /* Only move tasks to a NUMA node less busy than the current node. */
1438 static bool numa_has_capacity(struct task_numa_env *env)
1440 struct numa_stats *src = &env->src_stats;
1441 struct numa_stats *dst = &env->dst_stats;
1443 if (src->has_free_capacity && !dst->has_free_capacity)
1447 * Only consider a task move if the source has a higher load
1448 * than the destination, corrected for CPU capacity on each node.
1450 * src->load dst->load
1451 * --------------------- vs ---------------------
1452 * src->compute_capacity dst->compute_capacity
1454 if (src->load * dst->compute_capacity * env->imbalance_pct >
1456 dst->load * src->compute_capacity * 100)
1462 static int task_numa_migrate(struct task_struct *p)
1464 struct task_numa_env env = {
1467 .src_cpu = task_cpu(p),
1468 .src_nid = task_node(p),
1470 .imbalance_pct = 112,
1476 struct sched_domain *sd;
1477 unsigned long taskweight, groupweight;
1479 long taskimp, groupimp;
1482 * Pick the lowest SD_NUMA domain, as that would have the smallest
1483 * imbalance and would be the first to start moving tasks about.
1485 * And we want to avoid any moving of tasks about, as that would create
1486 * random movement of tasks -- counter the numa conditions we're trying
1490 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1492 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1496 * Cpusets can break the scheduler domain tree into smaller
1497 * balance domains, some of which do not cross NUMA boundaries.
1498 * Tasks that are "trapped" in such domains cannot be migrated
1499 * elsewhere, so there is no point in (re)trying.
1501 if (unlikely(!sd)) {
1502 p->numa_preferred_nid = task_node(p);
1506 env.dst_nid = p->numa_preferred_nid;
1507 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1508 taskweight = task_weight(p, env.src_nid, dist);
1509 groupweight = group_weight(p, env.src_nid, dist);
1510 update_numa_stats(&env.src_stats, env.src_nid);
1511 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1512 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1513 update_numa_stats(&env.dst_stats, env.dst_nid);
1515 /* Try to find a spot on the preferred nid. */
1516 if (numa_has_capacity(&env))
1517 task_numa_find_cpu(&env, taskimp, groupimp);
1520 * Look at other nodes in these cases:
1521 * - there is no space available on the preferred_nid
1522 * - the task is part of a numa_group that is interleaved across
1523 * multiple NUMA nodes; in order to better consolidate the group,
1524 * we need to check other locations.
1526 if (env.best_cpu == -1 || (p->numa_group &&
1527 nodes_weight(p->numa_group->active_nodes) > 1)) {
1528 for_each_online_node(nid) {
1529 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1532 dist = node_distance(env.src_nid, env.dst_nid);
1533 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1535 taskweight = task_weight(p, env.src_nid, dist);
1536 groupweight = group_weight(p, env.src_nid, dist);
1539 /* Only consider nodes where both task and groups benefit */
1540 taskimp = task_weight(p, nid, dist) - taskweight;
1541 groupimp = group_weight(p, nid, dist) - groupweight;
1542 if (taskimp < 0 && groupimp < 0)
1547 update_numa_stats(&env.dst_stats, env.dst_nid);
1548 if (numa_has_capacity(&env))
1549 task_numa_find_cpu(&env, taskimp, groupimp);
1554 * If the task is part of a workload that spans multiple NUMA nodes,
1555 * and is migrating into one of the workload's active nodes, remember
1556 * this node as the task's preferred numa node, so the workload can
1558 * A task that migrated to a second choice node will be better off
1559 * trying for a better one later. Do not set the preferred node here.
1561 if (p->numa_group) {
1562 if (env.best_cpu == -1)
1567 if (node_isset(nid, p->numa_group->active_nodes))
1568 sched_setnuma(p, env.dst_nid);
1571 /* No better CPU than the current one was found. */
1572 if (env.best_cpu == -1)
1576 * Reset the scan period if the task is being rescheduled on an
1577 * alternative node to recheck if the tasks is now properly placed.
1579 p->numa_scan_period = task_scan_min(p);
1581 if (env.best_task == NULL) {
1582 ret = migrate_task_to(p, env.best_cpu);
1584 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1588 ret = migrate_swap(p, env.best_task);
1590 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1591 put_task_struct(env.best_task);
1595 /* Attempt to migrate a task to a CPU on the preferred node. */
1596 static void numa_migrate_preferred(struct task_struct *p)
1598 unsigned long interval = HZ;
1600 /* This task has no NUMA fault statistics yet */
1601 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1604 /* Periodically retry migrating the task to the preferred node */
1605 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1606 p->numa_migrate_retry = jiffies + interval;
1608 /* Success if task is already running on preferred CPU */
1609 if (task_node(p) == p->numa_preferred_nid)
1612 /* Otherwise, try migrate to a CPU on the preferred node */
1613 task_numa_migrate(p);
1617 * Find the nodes on which the workload is actively running. We do this by
1618 * tracking the nodes from which NUMA hinting faults are triggered. This can
1619 * be different from the set of nodes where the workload's memory is currently
1622 * The bitmask is used to make smarter decisions on when to do NUMA page
1623 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1624 * are added when they cause over 6/16 of the maximum number of faults, but
1625 * only removed when they drop below 3/16.
1627 static void update_numa_active_node_mask(struct numa_group *numa_group)
1629 unsigned long faults, max_faults = 0;
1632 for_each_online_node(nid) {
1633 faults = group_faults_cpu(numa_group, nid);
1634 if (faults > max_faults)
1635 max_faults = faults;
1638 for_each_online_node(nid) {
1639 faults = group_faults_cpu(numa_group, nid);
1640 if (!node_isset(nid, numa_group->active_nodes)) {
1641 if (faults > max_faults * 6 / 16)
1642 node_set(nid, numa_group->active_nodes);
1643 } else if (faults < max_faults * 3 / 16)
1644 node_clear(nid, numa_group->active_nodes);
1649 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1650 * increments. The more local the fault statistics are, the higher the scan
1651 * period will be for the next scan window. If local/(local+remote) ratio is
1652 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1653 * the scan period will decrease. Aim for 70% local accesses.
1655 #define NUMA_PERIOD_SLOTS 10
1656 #define NUMA_PERIOD_THRESHOLD 7
1659 * Increase the scan period (slow down scanning) if the majority of
1660 * our memory is already on our local node, or if the majority of
1661 * the page accesses are shared with other processes.
1662 * Otherwise, decrease the scan period.
1664 static void update_task_scan_period(struct task_struct *p,
1665 unsigned long shared, unsigned long private)
1667 unsigned int period_slot;
1671 unsigned long remote = p->numa_faults_locality[0];
1672 unsigned long local = p->numa_faults_locality[1];
1675 * If there were no record hinting faults then either the task is
1676 * completely idle or all activity is areas that are not of interest
1677 * to automatic numa balancing. Related to that, if there were failed
1678 * migration then it implies we are migrating too quickly or the local
1679 * node is overloaded. In either case, scan slower
1681 if (local + shared == 0 || p->numa_faults_locality[2]) {
1682 p->numa_scan_period = min(p->numa_scan_period_max,
1683 p->numa_scan_period << 1);
1685 p->mm->numa_next_scan = jiffies +
1686 msecs_to_jiffies(p->numa_scan_period);
1692 * Prepare to scale scan period relative to the current period.
1693 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1694 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1695 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1697 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1698 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1699 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1700 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1703 diff = slot * period_slot;
1705 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1708 * Scale scan rate increases based on sharing. There is an
1709 * inverse relationship between the degree of sharing and
1710 * the adjustment made to the scanning period. Broadly
1711 * speaking the intent is that there is little point
1712 * scanning faster if shared accesses dominate as it may
1713 * simply bounce migrations uselessly
1715 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1716 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1719 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1720 task_scan_min(p), task_scan_max(p));
1721 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1725 * Get the fraction of time the task has been running since the last
1726 * NUMA placement cycle. The scheduler keeps similar statistics, but
1727 * decays those on a 32ms period, which is orders of magnitude off
1728 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1729 * stats only if the task is so new there are no NUMA statistics yet.
1731 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1733 u64 runtime, delta, now;
1734 /* Use the start of this time slice to avoid calculations. */
1735 now = p->se.exec_start;
1736 runtime = p->se.sum_exec_runtime;
1738 if (p->last_task_numa_placement) {
1739 delta = runtime - p->last_sum_exec_runtime;
1740 *period = now - p->last_task_numa_placement;
1742 delta = p->se.avg.load_sum / p->se.load.weight;
1743 *period = LOAD_AVG_MAX;
1746 p->last_sum_exec_runtime = runtime;
1747 p->last_task_numa_placement = now;
1753 * Determine the preferred nid for a task in a numa_group. This needs to
1754 * be done in a way that produces consistent results with group_weight,
1755 * otherwise workloads might not converge.
1757 static int preferred_group_nid(struct task_struct *p, int nid)
1762 /* Direct connections between all NUMA nodes. */
1763 if (sched_numa_topology_type == NUMA_DIRECT)
1767 * On a system with glueless mesh NUMA topology, group_weight
1768 * scores nodes according to the number of NUMA hinting faults on
1769 * both the node itself, and on nearby nodes.
1771 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1772 unsigned long score, max_score = 0;
1773 int node, max_node = nid;
1775 dist = sched_max_numa_distance;
1777 for_each_online_node(node) {
1778 score = group_weight(p, node, dist);
1779 if (score > max_score) {
1788 * Finding the preferred nid in a system with NUMA backplane
1789 * interconnect topology is more involved. The goal is to locate
1790 * tasks from numa_groups near each other in the system, and
1791 * untangle workloads from different sides of the system. This requires
1792 * searching down the hierarchy of node groups, recursively searching
1793 * inside the highest scoring group of nodes. The nodemask tricks
1794 * keep the complexity of the search down.
1796 nodes = node_online_map;
1797 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1798 unsigned long max_faults = 0;
1799 nodemask_t max_group = NODE_MASK_NONE;
1802 /* Are there nodes at this distance from each other? */
1803 if (!find_numa_distance(dist))
1806 for_each_node_mask(a, nodes) {
1807 unsigned long faults = 0;
1808 nodemask_t this_group;
1809 nodes_clear(this_group);
1811 /* Sum group's NUMA faults; includes a==b case. */
1812 for_each_node_mask(b, nodes) {
1813 if (node_distance(a, b) < dist) {
1814 faults += group_faults(p, b);
1815 node_set(b, this_group);
1816 node_clear(b, nodes);
1820 /* Remember the top group. */
1821 if (faults > max_faults) {
1822 max_faults = faults;
1823 max_group = this_group;
1825 * subtle: at the smallest distance there is
1826 * just one node left in each "group", the
1827 * winner is the preferred nid.
1832 /* Next round, evaluate the nodes within max_group. */
1840 static void task_numa_placement(struct task_struct *p)
1842 int seq, nid, max_nid = -1, max_group_nid = -1;
1843 unsigned long max_faults = 0, max_group_faults = 0;
1844 unsigned long fault_types[2] = { 0, 0 };
1845 unsigned long total_faults;
1846 u64 runtime, period;
1847 spinlock_t *group_lock = NULL;
1850 * The p->mm->numa_scan_seq field gets updated without
1851 * exclusive access. Use READ_ONCE() here to ensure
1852 * that the field is read in a single access:
1854 seq = READ_ONCE(p->mm->numa_scan_seq);
1855 if (p->numa_scan_seq == seq)
1857 p->numa_scan_seq = seq;
1858 p->numa_scan_period_max = task_scan_max(p);
1860 total_faults = p->numa_faults_locality[0] +
1861 p->numa_faults_locality[1];
1862 runtime = numa_get_avg_runtime(p, &period);
1864 /* If the task is part of a group prevent parallel updates to group stats */
1865 if (p->numa_group) {
1866 group_lock = &p->numa_group->lock;
1867 spin_lock_irq(group_lock);
1870 /* Find the node with the highest number of faults */
1871 for_each_online_node(nid) {
1872 /* Keep track of the offsets in numa_faults array */
1873 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1874 unsigned long faults = 0, group_faults = 0;
1877 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1878 long diff, f_diff, f_weight;
1880 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1881 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1882 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1883 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1885 /* Decay existing window, copy faults since last scan */
1886 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1887 fault_types[priv] += p->numa_faults[membuf_idx];
1888 p->numa_faults[membuf_idx] = 0;
1891 * Normalize the faults_from, so all tasks in a group
1892 * count according to CPU use, instead of by the raw
1893 * number of faults. Tasks with little runtime have
1894 * little over-all impact on throughput, and thus their
1895 * faults are less important.
1897 f_weight = div64_u64(runtime << 16, period + 1);
1898 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1900 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1901 p->numa_faults[cpubuf_idx] = 0;
1903 p->numa_faults[mem_idx] += diff;
1904 p->numa_faults[cpu_idx] += f_diff;
1905 faults += p->numa_faults[mem_idx];
1906 p->total_numa_faults += diff;
1907 if (p->numa_group) {
1909 * safe because we can only change our own group
1911 * mem_idx represents the offset for a given
1912 * nid and priv in a specific region because it
1913 * is at the beginning of the numa_faults array.
1915 p->numa_group->faults[mem_idx] += diff;
1916 p->numa_group->faults_cpu[mem_idx] += f_diff;
1917 p->numa_group->total_faults += diff;
1918 group_faults += p->numa_group->faults[mem_idx];
1922 if (faults > max_faults) {
1923 max_faults = faults;
1927 if (group_faults > max_group_faults) {
1928 max_group_faults = group_faults;
1929 max_group_nid = nid;
1933 update_task_scan_period(p, fault_types[0], fault_types[1]);
1935 if (p->numa_group) {
1936 update_numa_active_node_mask(p->numa_group);
1937 spin_unlock_irq(group_lock);
1938 max_nid = preferred_group_nid(p, max_group_nid);
1942 /* Set the new preferred node */
1943 if (max_nid != p->numa_preferred_nid)
1944 sched_setnuma(p, max_nid);
1946 if (task_node(p) != p->numa_preferred_nid)
1947 numa_migrate_preferred(p);
1951 static inline int get_numa_group(struct numa_group *grp)
1953 return atomic_inc_not_zero(&grp->refcount);
1956 static inline void put_numa_group(struct numa_group *grp)
1958 if (atomic_dec_and_test(&grp->refcount))
1959 kfree_rcu(grp, rcu);
1962 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1965 struct numa_group *grp, *my_grp;
1966 struct task_struct *tsk;
1968 int cpu = cpupid_to_cpu(cpupid);
1971 if (unlikely(!p->numa_group)) {
1972 unsigned int size = sizeof(struct numa_group) +
1973 4*nr_node_ids*sizeof(unsigned long);
1975 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1979 atomic_set(&grp->refcount, 1);
1980 spin_lock_init(&grp->lock);
1982 /* Second half of the array tracks nids where faults happen */
1983 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1986 node_set(task_node(current), grp->active_nodes);
1988 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1989 grp->faults[i] = p->numa_faults[i];
1991 grp->total_faults = p->total_numa_faults;
1994 rcu_assign_pointer(p->numa_group, grp);
1998 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2000 if (!cpupid_match_pid(tsk, cpupid))
2003 grp = rcu_dereference(tsk->numa_group);
2007 my_grp = p->numa_group;
2012 * Only join the other group if its bigger; if we're the bigger group,
2013 * the other task will join us.
2015 if (my_grp->nr_tasks > grp->nr_tasks)
2019 * Tie-break on the grp address.
2021 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2024 /* Always join threads in the same process. */
2025 if (tsk->mm == current->mm)
2028 /* Simple filter to avoid false positives due to PID collisions */
2029 if (flags & TNF_SHARED)
2032 /* Update priv based on whether false sharing was detected */
2035 if (join && !get_numa_group(grp))
2043 BUG_ON(irqs_disabled());
2044 double_lock_irq(&my_grp->lock, &grp->lock);
2046 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2047 my_grp->faults[i] -= p->numa_faults[i];
2048 grp->faults[i] += p->numa_faults[i];
2050 my_grp->total_faults -= p->total_numa_faults;
2051 grp->total_faults += p->total_numa_faults;
2056 spin_unlock(&my_grp->lock);
2057 spin_unlock_irq(&grp->lock);
2059 rcu_assign_pointer(p->numa_group, grp);
2061 put_numa_group(my_grp);
2069 void task_numa_free(struct task_struct *p)
2071 struct numa_group *grp = p->numa_group;
2072 void *numa_faults = p->numa_faults;
2073 unsigned long flags;
2077 spin_lock_irqsave(&grp->lock, flags);
2078 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2079 grp->faults[i] -= p->numa_faults[i];
2080 grp->total_faults -= p->total_numa_faults;
2083 spin_unlock_irqrestore(&grp->lock, flags);
2084 RCU_INIT_POINTER(p->numa_group, NULL);
2085 put_numa_group(grp);
2088 p->numa_faults = NULL;
2093 * Got a PROT_NONE fault for a page on @node.
2095 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2097 struct task_struct *p = current;
2098 bool migrated = flags & TNF_MIGRATED;
2099 int cpu_node = task_node(current);
2100 int local = !!(flags & TNF_FAULT_LOCAL);
2103 if (!static_branch_likely(&sched_numa_balancing))
2106 /* for example, ksmd faulting in a user's mm */
2110 /* Allocate buffer to track faults on a per-node basis */
2111 if (unlikely(!p->numa_faults)) {
2112 int size = sizeof(*p->numa_faults) *
2113 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2115 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2116 if (!p->numa_faults)
2119 p->total_numa_faults = 0;
2120 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2124 * First accesses are treated as private, otherwise consider accesses
2125 * to be private if the accessing pid has not changed
2127 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2130 priv = cpupid_match_pid(p, last_cpupid);
2131 if (!priv && !(flags & TNF_NO_GROUP))
2132 task_numa_group(p, last_cpupid, flags, &priv);
2136 * If a workload spans multiple NUMA nodes, a shared fault that
2137 * occurs wholly within the set of nodes that the workload is
2138 * actively using should be counted as local. This allows the
2139 * scan rate to slow down when a workload has settled down.
2141 if (!priv && !local && p->numa_group &&
2142 node_isset(cpu_node, p->numa_group->active_nodes) &&
2143 node_isset(mem_node, p->numa_group->active_nodes))
2146 task_numa_placement(p);
2149 * Retry task to preferred node migration periodically, in case it
2150 * case it previously failed, or the scheduler moved us.
2152 if (time_after(jiffies, p->numa_migrate_retry))
2153 numa_migrate_preferred(p);
2156 p->numa_pages_migrated += pages;
2157 if (flags & TNF_MIGRATE_FAIL)
2158 p->numa_faults_locality[2] += pages;
2160 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2161 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2162 p->numa_faults_locality[local] += pages;
2165 static void reset_ptenuma_scan(struct task_struct *p)
2168 * We only did a read acquisition of the mmap sem, so
2169 * p->mm->numa_scan_seq is written to without exclusive access
2170 * and the update is not guaranteed to be atomic. That's not
2171 * much of an issue though, since this is just used for
2172 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2173 * expensive, to avoid any form of compiler optimizations:
2175 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2176 p->mm->numa_scan_offset = 0;
2180 * The expensive part of numa migration is done from task_work context.
2181 * Triggered from task_tick_numa().
2183 void task_numa_work(struct callback_head *work)
2185 unsigned long migrate, next_scan, now = jiffies;
2186 struct task_struct *p = current;
2187 struct mm_struct *mm = p->mm;
2188 struct vm_area_struct *vma;
2189 unsigned long start, end;
2190 unsigned long nr_pte_updates = 0;
2191 long pages, virtpages;
2193 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2195 work->next = work; /* protect against double add */
2197 * Who cares about NUMA placement when they're dying.
2199 * NOTE: make sure not to dereference p->mm before this check,
2200 * exit_task_work() happens _after_ exit_mm() so we could be called
2201 * without p->mm even though we still had it when we enqueued this
2204 if (p->flags & PF_EXITING)
2207 if (!mm->numa_next_scan) {
2208 mm->numa_next_scan = now +
2209 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2213 * Enforce maximal scan/migration frequency..
2215 migrate = mm->numa_next_scan;
2216 if (time_before(now, migrate))
2219 if (p->numa_scan_period == 0) {
2220 p->numa_scan_period_max = task_scan_max(p);
2221 p->numa_scan_period = task_scan_min(p);
2224 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2225 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2229 * Delay this task enough that another task of this mm will likely win
2230 * the next time around.
2232 p->node_stamp += 2 * TICK_NSEC;
2234 start = mm->numa_scan_offset;
2235 pages = sysctl_numa_balancing_scan_size;
2236 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2237 virtpages = pages * 8; /* Scan up to this much virtual space */
2242 down_read(&mm->mmap_sem);
2243 vma = find_vma(mm, start);
2245 reset_ptenuma_scan(p);
2249 for (; vma; vma = vma->vm_next) {
2250 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2251 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2256 * Shared library pages mapped by multiple processes are not
2257 * migrated as it is expected they are cache replicated. Avoid
2258 * hinting faults in read-only file-backed mappings or the vdso
2259 * as migrating the pages will be of marginal benefit.
2262 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2266 * Skip inaccessible VMAs to avoid any confusion between
2267 * PROT_NONE and NUMA hinting ptes
2269 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2273 start = max(start, vma->vm_start);
2274 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2275 end = min(end, vma->vm_end);
2276 nr_pte_updates = change_prot_numa(vma, start, end);
2279 * Try to scan sysctl_numa_balancing_size worth of
2280 * hpages that have at least one present PTE that
2281 * is not already pte-numa. If the VMA contains
2282 * areas that are unused or already full of prot_numa
2283 * PTEs, scan up to virtpages, to skip through those
2287 pages -= (end - start) >> PAGE_SHIFT;
2288 virtpages -= (end - start) >> PAGE_SHIFT;
2291 if (pages <= 0 || virtpages <= 0)
2295 } while (end != vma->vm_end);
2300 * It is possible to reach the end of the VMA list but the last few
2301 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2302 * would find the !migratable VMA on the next scan but not reset the
2303 * scanner to the start so check it now.
2306 mm->numa_scan_offset = start;
2308 reset_ptenuma_scan(p);
2309 up_read(&mm->mmap_sem);
2313 * Drive the periodic memory faults..
2315 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2317 struct callback_head *work = &curr->numa_work;
2321 * We don't care about NUMA placement if we don't have memory.
2323 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2327 * Using runtime rather than walltime has the dual advantage that
2328 * we (mostly) drive the selection from busy threads and that the
2329 * task needs to have done some actual work before we bother with
2332 now = curr->se.sum_exec_runtime;
2333 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2335 if (now > curr->node_stamp + period) {
2336 if (!curr->node_stamp)
2337 curr->numa_scan_period = task_scan_min(curr);
2338 curr->node_stamp += period;
2340 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2341 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2342 task_work_add(curr, work, true);
2347 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2351 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2355 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2358 #endif /* CONFIG_NUMA_BALANCING */
2361 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2363 update_load_add(&cfs_rq->load, se->load.weight);
2364 if (!parent_entity(se))
2365 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2367 if (entity_is_task(se)) {
2368 struct rq *rq = rq_of(cfs_rq);
2370 account_numa_enqueue(rq, task_of(se));
2371 list_add(&se->group_node, &rq->cfs_tasks);
2374 cfs_rq->nr_running++;
2378 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2380 update_load_sub(&cfs_rq->load, se->load.weight);
2381 if (!parent_entity(se))
2382 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2383 if (entity_is_task(se)) {
2384 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2385 list_del_init(&se->group_node);
2387 cfs_rq->nr_running--;
2390 #ifdef CONFIG_FAIR_GROUP_SCHED
2392 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2397 * Use this CPU's real-time load instead of the last load contribution
2398 * as the updating of the contribution is delayed, and we will use the
2399 * the real-time load to calc the share. See update_tg_load_avg().
2401 tg_weight = atomic_long_read(&tg->load_avg);
2402 tg_weight -= cfs_rq->tg_load_avg_contrib;
2403 tg_weight += cfs_rq->load.weight;
2408 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2410 long tg_weight, load, shares;
2412 tg_weight = calc_tg_weight(tg, cfs_rq);
2413 load = cfs_rq->load.weight;
2415 shares = (tg->shares * load);
2417 shares /= tg_weight;
2419 if (shares < MIN_SHARES)
2420 shares = MIN_SHARES;
2421 if (shares > tg->shares)
2422 shares = tg->shares;
2426 # else /* CONFIG_SMP */
2427 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2431 # endif /* CONFIG_SMP */
2432 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2433 unsigned long weight)
2436 /* commit outstanding execution time */
2437 if (cfs_rq->curr == se)
2438 update_curr(cfs_rq);
2439 account_entity_dequeue(cfs_rq, se);
2442 update_load_set(&se->load, weight);
2445 account_entity_enqueue(cfs_rq, se);
2448 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2450 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2452 struct task_group *tg;
2453 struct sched_entity *se;
2457 se = tg->se[cpu_of(rq_of(cfs_rq))];
2458 if (!se || throttled_hierarchy(cfs_rq))
2461 if (likely(se->load.weight == tg->shares))
2464 shares = calc_cfs_shares(cfs_rq, tg);
2466 reweight_entity(cfs_rq_of(se), se, shares);
2468 #else /* CONFIG_FAIR_GROUP_SCHED */
2469 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2472 #endif /* CONFIG_FAIR_GROUP_SCHED */
2475 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2476 static const u32 runnable_avg_yN_inv[] = {
2477 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2478 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2479 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2480 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2481 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2482 0x85aac367, 0x82cd8698,
2486 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2487 * over-estimates when re-combining.
2489 static const u32 runnable_avg_yN_sum[] = {
2490 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2491 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2492 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2497 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2499 static __always_inline u64 decay_load(u64 val, u64 n)
2501 unsigned int local_n;
2505 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2508 /* after bounds checking we can collapse to 32-bit */
2512 * As y^PERIOD = 1/2, we can combine
2513 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2514 * With a look-up table which covers y^n (n<PERIOD)
2516 * To achieve constant time decay_load.
2518 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2519 val >>= local_n / LOAD_AVG_PERIOD;
2520 local_n %= LOAD_AVG_PERIOD;
2523 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2528 * For updates fully spanning n periods, the contribution to runnable
2529 * average will be: \Sum 1024*y^n
2531 * We can compute this reasonably efficiently by combining:
2532 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2534 static u32 __compute_runnable_contrib(u64 n)
2538 if (likely(n <= LOAD_AVG_PERIOD))
2539 return runnable_avg_yN_sum[n];
2540 else if (unlikely(n >= LOAD_AVG_MAX_N))
2541 return LOAD_AVG_MAX;
2543 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2545 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2546 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2548 n -= LOAD_AVG_PERIOD;
2549 } while (n > LOAD_AVG_PERIOD);
2551 contrib = decay_load(contrib, n);
2552 return contrib + runnable_avg_yN_sum[n];
2555 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2556 #error "load tracking assumes 2^10 as unit"
2559 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2562 * We can represent the historical contribution to runnable average as the
2563 * coefficients of a geometric series. To do this we sub-divide our runnable
2564 * history into segments of approximately 1ms (1024us); label the segment that
2565 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2567 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2569 * (now) (~1ms ago) (~2ms ago)
2571 * Let u_i denote the fraction of p_i that the entity was runnable.
2573 * We then designate the fractions u_i as our co-efficients, yielding the
2574 * following representation of historical load:
2575 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2577 * We choose y based on the with of a reasonably scheduling period, fixing:
2580 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2581 * approximately half as much as the contribution to load within the last ms
2584 * When a period "rolls over" and we have new u_0`, multiplying the previous
2585 * sum again by y is sufficient to update:
2586 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2587 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2589 static __always_inline int
2590 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2591 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2593 u64 delta, scaled_delta, periods;
2595 unsigned int delta_w, scaled_delta_w, decayed = 0;
2596 unsigned long scale_freq, scale_cpu;
2598 delta = now - sa->last_update_time;
2600 * This should only happen when time goes backwards, which it
2601 * unfortunately does during sched clock init when we swap over to TSC.
2603 if ((s64)delta < 0) {
2604 sa->last_update_time = now;
2609 * Use 1024ns as the unit of measurement since it's a reasonable
2610 * approximation of 1us and fast to compute.
2615 sa->last_update_time = now;
2617 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2618 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2619 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2621 /* delta_w is the amount already accumulated against our next period */
2622 delta_w = sa->period_contrib;
2623 if (delta + delta_w >= 1024) {
2626 /* how much left for next period will start over, we don't know yet */
2627 sa->period_contrib = 0;
2630 * Now that we know we're crossing a period boundary, figure
2631 * out how much from delta we need to complete the current
2632 * period and accrue it.
2634 delta_w = 1024 - delta_w;
2635 scaled_delta_w = cap_scale(delta_w, scale_freq);
2637 sa->load_sum += weight * scaled_delta_w;
2639 cfs_rq->runnable_load_sum +=
2640 weight * scaled_delta_w;
2644 sa->util_sum += scaled_delta_w * scale_cpu;
2648 /* Figure out how many additional periods this update spans */
2649 periods = delta / 1024;
2652 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2654 cfs_rq->runnable_load_sum =
2655 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2657 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2659 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2660 contrib = __compute_runnable_contrib(periods);
2661 contrib = cap_scale(contrib, scale_freq);
2663 sa->load_sum += weight * contrib;
2665 cfs_rq->runnable_load_sum += weight * contrib;
2668 sa->util_sum += contrib * scale_cpu;
2671 /* Remainder of delta accrued against u_0` */
2672 scaled_delta = cap_scale(delta, scale_freq);
2674 sa->load_sum += weight * scaled_delta;
2676 cfs_rq->runnable_load_sum += weight * scaled_delta;
2679 sa->util_sum += scaled_delta * scale_cpu;
2681 sa->period_contrib += delta;
2684 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2686 cfs_rq->runnable_load_avg =
2687 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2689 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2695 #ifdef CONFIG_FAIR_GROUP_SCHED
2697 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2698 * and effective_load (which is not done because it is too costly).
2700 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2702 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2704 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2705 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2706 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2710 #else /* CONFIG_FAIR_GROUP_SCHED */
2711 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2712 #endif /* CONFIG_FAIR_GROUP_SCHED */
2714 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2716 if (&this_rq()->cfs == cfs_rq) {
2718 * There are a few boundary cases this might miss but it should
2719 * get called often enough that that should (hopefully) not be
2720 * a real problem -- added to that it only calls on the local
2721 * CPU, so if we enqueue remotely we'll miss an update, but
2722 * the next tick/schedule should update.
2724 * It will not get called when we go idle, because the idle
2725 * thread is a different class (!fair), nor will the utilization
2726 * number include things like RT tasks.
2728 * As is, the util number is not freq-invariant (we'd have to
2729 * implement arch_scale_freq_capacity() for that).
2733 cpufreq_update_util(rq_of(cfs_rq), 0);
2737 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2740 * Unsigned subtract and clamp on underflow.
2742 * Explicitly do a load-store to ensure the intermediate value never hits
2743 * memory. This allows lockless observations without ever seeing the negative
2746 #define sub_positive(_ptr, _val) do { \
2747 typeof(_ptr) ptr = (_ptr); \
2748 typeof(*ptr) val = (_val); \
2749 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2753 WRITE_ONCE(*ptr, res); \
2756 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2757 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq,
2760 struct sched_avg *sa = &cfs_rq->avg;
2761 int decayed, removed = 0, removed_util = 0;
2763 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2764 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2765 sub_positive(&sa->load_avg, r);
2766 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2770 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2771 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2772 sub_positive(&sa->util_avg, r);
2773 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2777 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2778 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2780 #ifndef CONFIG_64BIT
2782 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2785 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2786 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2787 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2789 if (update_freq && (decayed || removed_util))
2790 cfs_rq_util_change(cfs_rq);
2792 return decayed || removed;
2795 /* Update task and its cfs_rq load average */
2796 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2798 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2799 u64 now = cfs_rq_clock_task(cfs_rq);
2800 int cpu = cpu_of(rq_of(cfs_rq));
2803 * Track task load average for carrying it to new CPU after migrated, and
2804 * track group sched_entity load average for task_h_load calc in migration
2806 __update_load_avg(now, cpu, &se->avg,
2807 se->on_rq * scale_load_down(se->load.weight),
2808 cfs_rq->curr == se, NULL);
2810 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2811 update_tg_load_avg(cfs_rq, 0);
2813 if (entity_is_task(se))
2814 trace_sched_load_avg_task(task_of(se), &se->avg);
2817 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2819 if (!sched_feat(ATTACH_AGE_LOAD))
2823 * If we got migrated (either between CPUs or between cgroups) we'll
2824 * have aged the average right before clearing @last_update_time.
2826 if (se->avg.last_update_time) {
2827 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2828 &se->avg, 0, 0, NULL);
2831 * XXX: we could have just aged the entire load away if we've been
2832 * absent from the fair class for too long.
2837 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2838 cfs_rq->avg.load_avg += se->avg.load_avg;
2839 cfs_rq->avg.load_sum += se->avg.load_sum;
2840 cfs_rq->avg.util_avg += se->avg.util_avg;
2841 cfs_rq->avg.util_sum += se->avg.util_sum;
2843 cfs_rq_util_change(cfs_rq);
2846 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2848 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2849 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2850 cfs_rq->curr == se, NULL);
2852 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2853 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2854 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2855 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2857 cfs_rq_util_change(cfs_rq);
2860 /* Add the load generated by se into cfs_rq's load average */
2862 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2864 struct sched_avg *sa = &se->avg;
2865 u64 now = cfs_rq_clock_task(cfs_rq);
2866 int migrated, decayed;
2868 migrated = !sa->last_update_time;
2870 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2871 se->on_rq * scale_load_down(se->load.weight),
2872 cfs_rq->curr == se, NULL);
2875 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
2877 cfs_rq->runnable_load_avg += sa->load_avg;
2878 cfs_rq->runnable_load_sum += sa->load_sum;
2881 attach_entity_load_avg(cfs_rq, se);
2883 if (decayed || migrated)
2884 update_tg_load_avg(cfs_rq, 0);
2887 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2889 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2891 update_load_avg(se, 1);
2893 cfs_rq->runnable_load_avg =
2894 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2895 cfs_rq->runnable_load_sum =
2896 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2899 #ifndef CONFIG_64BIT
2900 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2902 u64 last_update_time_copy;
2903 u64 last_update_time;
2906 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2908 last_update_time = cfs_rq->avg.last_update_time;
2909 } while (last_update_time != last_update_time_copy);
2911 return last_update_time;
2914 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2916 return cfs_rq->avg.last_update_time;
2921 * Task first catches up with cfs_rq, and then subtract
2922 * itself from the cfs_rq (task must be off the queue now).
2924 void remove_entity_load_avg(struct sched_entity *se)
2926 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2927 u64 last_update_time;
2930 * Newly created task or never used group entity should not be removed
2931 * from its (source) cfs_rq
2933 if (se->avg.last_update_time == 0)
2936 last_update_time = cfs_rq_last_update_time(cfs_rq);
2938 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2939 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2940 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2944 * Update the rq's load with the elapsed running time before entering
2945 * idle. if the last scheduled task is not a CFS task, idle_enter will
2946 * be the only way to update the runnable statistic.
2948 void idle_enter_fair(struct rq *this_rq)
2953 * Update the rq's load with the elapsed idle time before a task is
2954 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2955 * be the only way to update the runnable statistic.
2957 void idle_exit_fair(struct rq *this_rq)
2961 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2963 return cfs_rq->runnable_load_avg;
2966 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2968 return cfs_rq->avg.load_avg;
2971 static int idle_balance(struct rq *this_rq);
2973 #else /* CONFIG_SMP */
2975 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2977 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
2981 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2983 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2984 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2987 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2989 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2991 static inline int idle_balance(struct rq *rq)
2996 #endif /* CONFIG_SMP */
2998 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3000 #ifdef CONFIG_SCHEDSTATS
3001 struct task_struct *tsk = NULL;
3003 if (entity_is_task(se))
3006 if (se->statistics.sleep_start) {
3007 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3012 if (unlikely(delta > se->statistics.sleep_max))
3013 se->statistics.sleep_max = delta;
3015 se->statistics.sleep_start = 0;
3016 se->statistics.sum_sleep_runtime += delta;
3019 account_scheduler_latency(tsk, delta >> 10, 1);
3020 trace_sched_stat_sleep(tsk, delta);
3023 if (se->statistics.block_start) {
3024 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3029 if (unlikely(delta > se->statistics.block_max))
3030 se->statistics.block_max = delta;
3032 se->statistics.block_start = 0;
3033 se->statistics.sum_sleep_runtime += delta;
3036 if (tsk->in_iowait) {
3037 se->statistics.iowait_sum += delta;
3038 se->statistics.iowait_count++;
3039 trace_sched_stat_iowait(tsk, delta);
3042 trace_sched_stat_blocked(tsk, delta);
3043 trace_sched_blocked_reason(tsk);
3046 * Blocking time is in units of nanosecs, so shift by
3047 * 20 to get a milliseconds-range estimation of the
3048 * amount of time that the task spent sleeping:
3050 if (unlikely(prof_on == SLEEP_PROFILING)) {
3051 profile_hits(SLEEP_PROFILING,
3052 (void *)get_wchan(tsk),
3055 account_scheduler_latency(tsk, delta >> 10, 0);
3061 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3063 #ifdef CONFIG_SCHED_DEBUG
3064 s64 d = se->vruntime - cfs_rq->min_vruntime;
3069 if (d > 3*sysctl_sched_latency)
3070 schedstat_inc(cfs_rq, nr_spread_over);
3075 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3077 u64 vruntime = cfs_rq->min_vruntime;
3080 * The 'current' period is already promised to the current tasks,
3081 * however the extra weight of the new task will slow them down a
3082 * little, place the new task so that it fits in the slot that
3083 * stays open at the end.
3085 if (initial && sched_feat(START_DEBIT))
3086 vruntime += sched_vslice(cfs_rq, se);
3088 /* sleeps up to a single latency don't count. */
3090 unsigned long thresh = sysctl_sched_latency;
3093 * Halve their sleep time's effect, to allow
3094 * for a gentler effect of sleepers:
3096 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3102 /* ensure we never gain time by being placed backwards. */
3103 se->vruntime = max_vruntime(se->vruntime, vruntime);
3106 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3109 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3112 * Update the normalized vruntime before updating min_vruntime
3113 * through calling update_curr().
3115 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3116 se->vruntime += cfs_rq->min_vruntime;
3119 * Update run-time statistics of the 'current'.
3121 update_curr(cfs_rq);
3122 enqueue_entity_load_avg(cfs_rq, se);
3123 account_entity_enqueue(cfs_rq, se);
3124 update_cfs_shares(cfs_rq);
3126 if (flags & ENQUEUE_WAKEUP) {
3127 place_entity(cfs_rq, se, 0);
3128 enqueue_sleeper(cfs_rq, se);
3131 update_stats_enqueue(cfs_rq, se);
3132 check_spread(cfs_rq, se);
3133 if (se != cfs_rq->curr)
3134 __enqueue_entity(cfs_rq, se);
3137 if (cfs_rq->nr_running == 1) {
3138 list_add_leaf_cfs_rq(cfs_rq);
3139 check_enqueue_throttle(cfs_rq);
3143 static void __clear_buddies_last(struct sched_entity *se)
3145 for_each_sched_entity(se) {
3146 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3147 if (cfs_rq->last != se)
3150 cfs_rq->last = NULL;
3154 static void __clear_buddies_next(struct sched_entity *se)
3156 for_each_sched_entity(se) {
3157 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3158 if (cfs_rq->next != se)
3161 cfs_rq->next = NULL;
3165 static void __clear_buddies_skip(struct sched_entity *se)
3167 for_each_sched_entity(se) {
3168 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3169 if (cfs_rq->skip != se)
3172 cfs_rq->skip = NULL;
3176 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3178 if (cfs_rq->last == se)
3179 __clear_buddies_last(se);
3181 if (cfs_rq->next == se)
3182 __clear_buddies_next(se);
3184 if (cfs_rq->skip == se)
3185 __clear_buddies_skip(se);
3188 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3191 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3194 * Update run-time statistics of the 'current'.
3196 update_curr(cfs_rq);
3197 dequeue_entity_load_avg(cfs_rq, se);
3199 update_stats_dequeue(cfs_rq, se);
3200 if (flags & DEQUEUE_SLEEP) {
3201 #ifdef CONFIG_SCHEDSTATS
3202 if (entity_is_task(se)) {
3203 struct task_struct *tsk = task_of(se);
3205 if (tsk->state & TASK_INTERRUPTIBLE)
3206 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3207 if (tsk->state & TASK_UNINTERRUPTIBLE)
3208 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3213 clear_buddies(cfs_rq, se);
3215 if (se != cfs_rq->curr)
3216 __dequeue_entity(cfs_rq, se);
3218 account_entity_dequeue(cfs_rq, se);
3221 * Normalize the entity after updating the min_vruntime because the
3222 * update can refer to the ->curr item and we need to reflect this
3223 * movement in our normalized position.
3225 if (!(flags & DEQUEUE_SLEEP))
3226 se->vruntime -= cfs_rq->min_vruntime;
3228 /* return excess runtime on last dequeue */
3229 return_cfs_rq_runtime(cfs_rq);
3231 update_min_vruntime(cfs_rq);
3232 update_cfs_shares(cfs_rq);
3236 * Preempt the current task with a newly woken task if needed:
3239 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3241 unsigned long ideal_runtime, delta_exec;
3242 struct sched_entity *se;
3245 ideal_runtime = sched_slice(cfs_rq, curr);
3246 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3247 if (delta_exec > ideal_runtime) {
3248 resched_curr(rq_of(cfs_rq));
3250 * The current task ran long enough, ensure it doesn't get
3251 * re-elected due to buddy favours.
3253 clear_buddies(cfs_rq, curr);
3258 * Ensure that a task that missed wakeup preemption by a
3259 * narrow margin doesn't have to wait for a full slice.
3260 * This also mitigates buddy induced latencies under load.
3262 if (delta_exec < sysctl_sched_min_granularity)
3265 se = __pick_first_entity(cfs_rq);
3266 delta = curr->vruntime - se->vruntime;
3271 if (delta > ideal_runtime)
3272 resched_curr(rq_of(cfs_rq));
3276 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3278 /* 'current' is not kept within the tree. */
3281 * Any task has to be enqueued before it get to execute on
3282 * a CPU. So account for the time it spent waiting on the
3285 update_stats_wait_end(cfs_rq, se);
3286 __dequeue_entity(cfs_rq, se);
3287 update_load_avg(se, 1);
3290 update_stats_curr_start(cfs_rq, se);
3292 #ifdef CONFIG_SCHEDSTATS
3294 * Track our maximum slice length, if the CPU's load is at
3295 * least twice that of our own weight (i.e. dont track it
3296 * when there are only lesser-weight tasks around):
3298 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3299 se->statistics.slice_max = max(se->statistics.slice_max,
3300 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3303 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3307 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3310 * Pick the next process, keeping these things in mind, in this order:
3311 * 1) keep things fair between processes/task groups
3312 * 2) pick the "next" process, since someone really wants that to run
3313 * 3) pick the "last" process, for cache locality
3314 * 4) do not run the "skip" process, if something else is available
3316 static struct sched_entity *
3317 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3319 struct sched_entity *left = __pick_first_entity(cfs_rq);
3320 struct sched_entity *se;
3323 * If curr is set we have to see if its left of the leftmost entity
3324 * still in the tree, provided there was anything in the tree at all.
3326 if (!left || (curr && entity_before(curr, left)))
3329 se = left; /* ideally we run the leftmost entity */
3332 * Avoid running the skip buddy, if running something else can
3333 * be done without getting too unfair.
3335 if (cfs_rq->skip == se) {
3336 struct sched_entity *second;
3339 second = __pick_first_entity(cfs_rq);
3341 second = __pick_next_entity(se);
3342 if (!second || (curr && entity_before(curr, second)))
3346 if (second && wakeup_preempt_entity(second, left) < 1)
3351 * Prefer last buddy, try to return the CPU to a preempted task.
3353 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3357 * Someone really wants this to run. If it's not unfair, run it.
3359 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3362 clear_buddies(cfs_rq, se);
3367 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3369 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3372 * If still on the runqueue then deactivate_task()
3373 * was not called and update_curr() has to be done:
3376 update_curr(cfs_rq);
3378 /* throttle cfs_rqs exceeding runtime */
3379 check_cfs_rq_runtime(cfs_rq);
3381 check_spread(cfs_rq, prev);
3383 update_stats_wait_start(cfs_rq, prev);
3384 /* Put 'current' back into the tree. */
3385 __enqueue_entity(cfs_rq, prev);
3386 /* in !on_rq case, update occurred at dequeue */
3387 update_load_avg(prev, 0);
3389 cfs_rq->curr = NULL;
3393 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3396 * Update run-time statistics of the 'current'.
3398 update_curr(cfs_rq);
3401 * Ensure that runnable average is periodically updated.
3403 update_load_avg(curr, 1);
3404 update_cfs_shares(cfs_rq);
3406 #ifdef CONFIG_SCHED_HRTICK
3408 * queued ticks are scheduled to match the slice, so don't bother
3409 * validating it and just reschedule.
3412 resched_curr(rq_of(cfs_rq));
3416 * don't let the period tick interfere with the hrtick preemption
3418 if (!sched_feat(DOUBLE_TICK) &&
3419 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3423 if (cfs_rq->nr_running > 1)
3424 check_preempt_tick(cfs_rq, curr);
3428 /**************************************************
3429 * CFS bandwidth control machinery
3432 #ifdef CONFIG_CFS_BANDWIDTH
3434 #ifdef HAVE_JUMP_LABEL
3435 static struct static_key __cfs_bandwidth_used;
3437 static inline bool cfs_bandwidth_used(void)
3439 return static_key_false(&__cfs_bandwidth_used);
3442 void cfs_bandwidth_usage_inc(void)
3444 static_key_slow_inc(&__cfs_bandwidth_used);
3447 void cfs_bandwidth_usage_dec(void)
3449 static_key_slow_dec(&__cfs_bandwidth_used);
3451 #else /* HAVE_JUMP_LABEL */
3452 static bool cfs_bandwidth_used(void)
3457 void cfs_bandwidth_usage_inc(void) {}
3458 void cfs_bandwidth_usage_dec(void) {}
3459 #endif /* HAVE_JUMP_LABEL */
3462 * default period for cfs group bandwidth.
3463 * default: 0.1s, units: nanoseconds
3465 static inline u64 default_cfs_period(void)
3467 return 100000000ULL;
3470 static inline u64 sched_cfs_bandwidth_slice(void)
3472 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3476 * Replenish runtime according to assigned quota and update expiration time.
3477 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3478 * additional synchronization around rq->lock.
3480 * requires cfs_b->lock
3482 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3486 if (cfs_b->quota == RUNTIME_INF)
3489 now = sched_clock_cpu(smp_processor_id());
3490 cfs_b->runtime = cfs_b->quota;
3491 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3494 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3496 return &tg->cfs_bandwidth;
3499 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3500 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3502 if (unlikely(cfs_rq->throttle_count))
3503 return cfs_rq->throttled_clock_task;
3505 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3508 /* returns 0 on failure to allocate runtime */
3509 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3511 struct task_group *tg = cfs_rq->tg;
3512 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3513 u64 amount = 0, min_amount, expires;
3515 /* note: this is a positive sum as runtime_remaining <= 0 */
3516 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3518 raw_spin_lock(&cfs_b->lock);
3519 if (cfs_b->quota == RUNTIME_INF)
3520 amount = min_amount;
3522 start_cfs_bandwidth(cfs_b);
3524 if (cfs_b->runtime > 0) {
3525 amount = min(cfs_b->runtime, min_amount);
3526 cfs_b->runtime -= amount;
3530 expires = cfs_b->runtime_expires;
3531 raw_spin_unlock(&cfs_b->lock);
3533 cfs_rq->runtime_remaining += amount;
3535 * we may have advanced our local expiration to account for allowed
3536 * spread between our sched_clock and the one on which runtime was
3539 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3540 cfs_rq->runtime_expires = expires;
3542 return cfs_rq->runtime_remaining > 0;
3546 * Note: This depends on the synchronization provided by sched_clock and the
3547 * fact that rq->clock snapshots this value.
3549 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3551 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3553 /* if the deadline is ahead of our clock, nothing to do */
3554 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3557 if (cfs_rq->runtime_remaining < 0)
3561 * If the local deadline has passed we have to consider the
3562 * possibility that our sched_clock is 'fast' and the global deadline
3563 * has not truly expired.
3565 * Fortunately we can check determine whether this the case by checking
3566 * whether the global deadline has advanced. It is valid to compare
3567 * cfs_b->runtime_expires without any locks since we only care about
3568 * exact equality, so a partial write will still work.
3571 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3572 /* extend local deadline, drift is bounded above by 2 ticks */
3573 cfs_rq->runtime_expires += TICK_NSEC;
3575 /* global deadline is ahead, expiration has passed */
3576 cfs_rq->runtime_remaining = 0;
3580 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3582 /* dock delta_exec before expiring quota (as it could span periods) */
3583 cfs_rq->runtime_remaining -= delta_exec;
3584 expire_cfs_rq_runtime(cfs_rq);
3586 if (likely(cfs_rq->runtime_remaining > 0))
3590 * if we're unable to extend our runtime we resched so that the active
3591 * hierarchy can be throttled
3593 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3594 resched_curr(rq_of(cfs_rq));
3597 static __always_inline
3598 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3600 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3603 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3606 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3608 return cfs_bandwidth_used() && cfs_rq->throttled;
3611 /* check whether cfs_rq, or any parent, is throttled */
3612 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3614 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3618 * Ensure that neither of the group entities corresponding to src_cpu or
3619 * dest_cpu are members of a throttled hierarchy when performing group
3620 * load-balance operations.
3622 static inline int throttled_lb_pair(struct task_group *tg,
3623 int src_cpu, int dest_cpu)
3625 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3627 src_cfs_rq = tg->cfs_rq[src_cpu];
3628 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3630 return throttled_hierarchy(src_cfs_rq) ||
3631 throttled_hierarchy(dest_cfs_rq);
3634 /* updated child weight may affect parent so we have to do this bottom up */
3635 static int tg_unthrottle_up(struct task_group *tg, void *data)
3637 struct rq *rq = data;
3638 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3640 cfs_rq->throttle_count--;
3642 if (!cfs_rq->throttle_count) {
3643 /* adjust cfs_rq_clock_task() */
3644 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3645 cfs_rq->throttled_clock_task;
3652 static int tg_throttle_down(struct task_group *tg, void *data)
3654 struct rq *rq = data;
3655 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3657 /* group is entering throttled state, stop time */
3658 if (!cfs_rq->throttle_count)
3659 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3660 cfs_rq->throttle_count++;
3665 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3667 struct rq *rq = rq_of(cfs_rq);
3668 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3669 struct sched_entity *se;
3670 long task_delta, dequeue = 1;
3673 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3675 /* freeze hierarchy runnable averages while throttled */
3677 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3680 task_delta = cfs_rq->h_nr_running;
3681 for_each_sched_entity(se) {
3682 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3683 /* throttled entity or throttle-on-deactivate */
3688 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3689 qcfs_rq->h_nr_running -= task_delta;
3691 if (qcfs_rq->load.weight)
3696 sub_nr_running(rq, task_delta);
3698 cfs_rq->throttled = 1;
3699 cfs_rq->throttled_clock = rq_clock(rq);
3700 raw_spin_lock(&cfs_b->lock);
3701 empty = list_empty(&cfs_b->throttled_cfs_rq);
3704 * Add to the _head_ of the list, so that an already-started
3705 * distribute_cfs_runtime will not see us
3707 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3710 * If we're the first throttled task, make sure the bandwidth
3714 start_cfs_bandwidth(cfs_b);
3716 raw_spin_unlock(&cfs_b->lock);
3719 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3721 struct rq *rq = rq_of(cfs_rq);
3722 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3723 struct sched_entity *se;
3727 se = cfs_rq->tg->se[cpu_of(rq)];
3729 cfs_rq->throttled = 0;
3731 update_rq_clock(rq);
3733 raw_spin_lock(&cfs_b->lock);
3734 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3735 list_del_rcu(&cfs_rq->throttled_list);
3736 raw_spin_unlock(&cfs_b->lock);
3738 /* update hierarchical throttle state */
3739 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3741 if (!cfs_rq->load.weight)
3744 task_delta = cfs_rq->h_nr_running;
3745 for_each_sched_entity(se) {
3749 cfs_rq = cfs_rq_of(se);
3751 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3752 cfs_rq->h_nr_running += task_delta;
3754 if (cfs_rq_throttled(cfs_rq))
3759 add_nr_running(rq, task_delta);
3761 /* determine whether we need to wake up potentially idle cpu */
3762 if (rq->curr == rq->idle && rq->cfs.nr_running)
3766 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3767 u64 remaining, u64 expires)
3769 struct cfs_rq *cfs_rq;
3771 u64 starting_runtime = remaining;
3774 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3776 struct rq *rq = rq_of(cfs_rq);
3778 raw_spin_lock(&rq->lock);
3779 if (!cfs_rq_throttled(cfs_rq))
3782 runtime = -cfs_rq->runtime_remaining + 1;
3783 if (runtime > remaining)
3784 runtime = remaining;
3785 remaining -= runtime;
3787 cfs_rq->runtime_remaining += runtime;
3788 cfs_rq->runtime_expires = expires;
3790 /* we check whether we're throttled above */
3791 if (cfs_rq->runtime_remaining > 0)
3792 unthrottle_cfs_rq(cfs_rq);
3795 raw_spin_unlock(&rq->lock);
3802 return starting_runtime - remaining;
3806 * Responsible for refilling a task_group's bandwidth and unthrottling its
3807 * cfs_rqs as appropriate. If there has been no activity within the last
3808 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3809 * used to track this state.
3811 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3813 u64 runtime, runtime_expires;
3816 /* no need to continue the timer with no bandwidth constraint */
3817 if (cfs_b->quota == RUNTIME_INF)
3818 goto out_deactivate;
3820 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3821 cfs_b->nr_periods += overrun;
3824 * idle depends on !throttled (for the case of a large deficit), and if
3825 * we're going inactive then everything else can be deferred
3827 if (cfs_b->idle && !throttled)
3828 goto out_deactivate;
3830 __refill_cfs_bandwidth_runtime(cfs_b);
3833 /* mark as potentially idle for the upcoming period */
3838 /* account preceding periods in which throttling occurred */
3839 cfs_b->nr_throttled += overrun;
3841 runtime_expires = cfs_b->runtime_expires;
3844 * This check is repeated as we are holding onto the new bandwidth while
3845 * we unthrottle. This can potentially race with an unthrottled group
3846 * trying to acquire new bandwidth from the global pool. This can result
3847 * in us over-using our runtime if it is all used during this loop, but
3848 * only by limited amounts in that extreme case.
3850 while (throttled && cfs_b->runtime > 0) {
3851 runtime = cfs_b->runtime;
3852 raw_spin_unlock(&cfs_b->lock);
3853 /* we can't nest cfs_b->lock while distributing bandwidth */
3854 runtime = distribute_cfs_runtime(cfs_b, runtime,
3856 raw_spin_lock(&cfs_b->lock);
3858 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3860 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3864 * While we are ensured activity in the period following an
3865 * unthrottle, this also covers the case in which the new bandwidth is
3866 * insufficient to cover the existing bandwidth deficit. (Forcing the
3867 * timer to remain active while there are any throttled entities.)
3877 /* a cfs_rq won't donate quota below this amount */
3878 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3879 /* minimum remaining period time to redistribute slack quota */
3880 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3881 /* how long we wait to gather additional slack before distributing */
3882 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3885 * Are we near the end of the current quota period?
3887 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3888 * hrtimer base being cleared by hrtimer_start. In the case of
3889 * migrate_hrtimers, base is never cleared, so we are fine.
3891 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3893 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3896 /* if the call-back is running a quota refresh is already occurring */
3897 if (hrtimer_callback_running(refresh_timer))
3900 /* is a quota refresh about to occur? */
3901 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3902 if (remaining < min_expire)
3908 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3910 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3912 /* if there's a quota refresh soon don't bother with slack */
3913 if (runtime_refresh_within(cfs_b, min_left))
3916 hrtimer_start(&cfs_b->slack_timer,
3917 ns_to_ktime(cfs_bandwidth_slack_period),
3921 /* we know any runtime found here is valid as update_curr() precedes return */
3922 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3924 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3925 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3927 if (slack_runtime <= 0)
3930 raw_spin_lock(&cfs_b->lock);
3931 if (cfs_b->quota != RUNTIME_INF &&
3932 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3933 cfs_b->runtime += slack_runtime;
3935 /* we are under rq->lock, defer unthrottling using a timer */
3936 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3937 !list_empty(&cfs_b->throttled_cfs_rq))
3938 start_cfs_slack_bandwidth(cfs_b);
3940 raw_spin_unlock(&cfs_b->lock);
3942 /* even if it's not valid for return we don't want to try again */
3943 cfs_rq->runtime_remaining -= slack_runtime;
3946 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3948 if (!cfs_bandwidth_used())
3951 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3954 __return_cfs_rq_runtime(cfs_rq);
3958 * This is done with a timer (instead of inline with bandwidth return) since
3959 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3961 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3963 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3966 /* confirm we're still not at a refresh boundary */
3967 raw_spin_lock(&cfs_b->lock);
3968 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3969 raw_spin_unlock(&cfs_b->lock);
3973 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3974 runtime = cfs_b->runtime;
3976 expires = cfs_b->runtime_expires;
3977 raw_spin_unlock(&cfs_b->lock);
3982 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3984 raw_spin_lock(&cfs_b->lock);
3985 if (expires == cfs_b->runtime_expires)
3986 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3987 raw_spin_unlock(&cfs_b->lock);
3991 * When a group wakes up we want to make sure that its quota is not already
3992 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3993 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3995 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3997 if (!cfs_bandwidth_used())
4000 /* Synchronize hierarchical throttle counter: */
4001 if (unlikely(!cfs_rq->throttle_uptodate)) {
4002 struct rq *rq = rq_of(cfs_rq);
4003 struct cfs_rq *pcfs_rq;
4004 struct task_group *tg;
4006 cfs_rq->throttle_uptodate = 1;
4008 /* Get closest up-to-date node, because leaves go first: */
4009 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4010 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4011 if (pcfs_rq->throttle_uptodate)
4015 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4016 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4020 /* an active group must be handled by the update_curr()->put() path */
4021 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4024 /* ensure the group is not already throttled */
4025 if (cfs_rq_throttled(cfs_rq))
4028 /* update runtime allocation */
4029 account_cfs_rq_runtime(cfs_rq, 0);
4030 if (cfs_rq->runtime_remaining <= 0)
4031 throttle_cfs_rq(cfs_rq);
4034 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4035 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4037 if (!cfs_bandwidth_used())
4040 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4044 * it's possible for a throttled entity to be forced into a running
4045 * state (e.g. set_curr_task), in this case we're finished.
4047 if (cfs_rq_throttled(cfs_rq))
4050 throttle_cfs_rq(cfs_rq);
4054 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4056 struct cfs_bandwidth *cfs_b =
4057 container_of(timer, struct cfs_bandwidth, slack_timer);
4059 do_sched_cfs_slack_timer(cfs_b);
4061 return HRTIMER_NORESTART;
4064 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4066 struct cfs_bandwidth *cfs_b =
4067 container_of(timer, struct cfs_bandwidth, period_timer);
4071 raw_spin_lock(&cfs_b->lock);
4073 overrun = hrtimer_forward_now(timer, cfs_b->period);
4077 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4080 cfs_b->period_active = 0;
4081 raw_spin_unlock(&cfs_b->lock);
4083 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4086 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4088 raw_spin_lock_init(&cfs_b->lock);
4090 cfs_b->quota = RUNTIME_INF;
4091 cfs_b->period = ns_to_ktime(default_cfs_period());
4093 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4094 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4095 cfs_b->period_timer.function = sched_cfs_period_timer;
4096 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4097 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4100 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4102 cfs_rq->runtime_enabled = 0;
4103 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4106 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4108 lockdep_assert_held(&cfs_b->lock);
4110 if (!cfs_b->period_active) {
4111 cfs_b->period_active = 1;
4112 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4113 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4117 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4119 /* init_cfs_bandwidth() was not called */
4120 if (!cfs_b->throttled_cfs_rq.next)
4123 hrtimer_cancel(&cfs_b->period_timer);
4124 hrtimer_cancel(&cfs_b->slack_timer);
4127 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4129 struct cfs_rq *cfs_rq;
4131 for_each_leaf_cfs_rq(rq, cfs_rq) {
4132 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4134 raw_spin_lock(&cfs_b->lock);
4135 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4136 raw_spin_unlock(&cfs_b->lock);
4140 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4142 struct cfs_rq *cfs_rq;
4144 for_each_leaf_cfs_rq(rq, cfs_rq) {
4145 if (!cfs_rq->runtime_enabled)
4149 * clock_task is not advancing so we just need to make sure
4150 * there's some valid quota amount
4152 cfs_rq->runtime_remaining = 1;
4154 * Offline rq is schedulable till cpu is completely disabled
4155 * in take_cpu_down(), so we prevent new cfs throttling here.
4157 cfs_rq->runtime_enabled = 0;
4159 if (cfs_rq_throttled(cfs_rq))
4160 unthrottle_cfs_rq(cfs_rq);
4164 #else /* CONFIG_CFS_BANDWIDTH */
4165 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4167 return rq_clock_task(rq_of(cfs_rq));
4170 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4171 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4172 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4173 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4175 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4180 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4185 static inline int throttled_lb_pair(struct task_group *tg,
4186 int src_cpu, int dest_cpu)
4191 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4193 #ifdef CONFIG_FAIR_GROUP_SCHED
4194 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4197 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4201 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4202 static inline void update_runtime_enabled(struct rq *rq) {}
4203 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4205 #endif /* CONFIG_CFS_BANDWIDTH */
4207 /**************************************************
4208 * CFS operations on tasks:
4211 #ifdef CONFIG_SCHED_HRTICK
4212 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4214 struct sched_entity *se = &p->se;
4215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4217 WARN_ON(task_rq(p) != rq);
4219 if (cfs_rq->nr_running > 1) {
4220 u64 slice = sched_slice(cfs_rq, se);
4221 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4222 s64 delta = slice - ran;
4229 hrtick_start(rq, delta);
4234 * called from enqueue/dequeue and updates the hrtick when the
4235 * current task is from our class and nr_running is low enough
4238 static void hrtick_update(struct rq *rq)
4240 struct task_struct *curr = rq->curr;
4242 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4245 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4246 hrtick_start_fair(rq, curr);
4248 #else /* !CONFIG_SCHED_HRTICK */
4250 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4254 static inline void hrtick_update(struct rq *rq)
4260 static bool cpu_overutilized(int cpu);
4261 unsigned long boosted_cpu_util(int cpu);
4263 #define boosted_cpu_util(cpu) cpu_util(cpu)
4267 static void update_capacity_of(int cpu)
4269 unsigned long req_cap;
4274 /* Convert scale-invariant capacity to cpu. */
4275 req_cap = boosted_cpu_util(cpu);
4276 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4277 set_cfs_cpu_capacity(cpu, true, req_cap);
4282 * The enqueue_task method is called before nr_running is
4283 * increased. Here we update the fair scheduling stats and
4284 * then put the task into the rbtree:
4287 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4289 struct cfs_rq *cfs_rq;
4290 struct sched_entity *se = &p->se;
4292 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4293 int task_wakeup = flags & ENQUEUE_WAKEUP;
4297 * If in_iowait is set, the code below may not trigger any cpufreq
4298 * utilization updates, so do it here explicitly with the IOWAIT flag
4302 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4304 for_each_sched_entity(se) {
4307 cfs_rq = cfs_rq_of(se);
4308 enqueue_entity(cfs_rq, se, flags);
4311 * end evaluation on encountering a throttled cfs_rq
4313 * note: in the case of encountering a throttled cfs_rq we will
4314 * post the final h_nr_running increment below.
4316 if (cfs_rq_throttled(cfs_rq))
4318 cfs_rq->h_nr_running++;
4319 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4321 flags = ENQUEUE_WAKEUP;
4324 for_each_sched_entity(se) {
4325 cfs_rq = cfs_rq_of(se);
4326 cfs_rq->h_nr_running++;
4327 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4329 if (cfs_rq_throttled(cfs_rq))
4332 update_load_avg(se, 1);
4333 update_cfs_shares(cfs_rq);
4337 add_nr_running(rq, 1);
4342 * Update SchedTune accounting.
4344 * We do it before updating the CPU capacity to ensure the
4345 * boost value of the current task is accounted for in the
4346 * selection of the OPP.
4348 * We do it also in the case where we enqueue a throttled task;
4349 * we could argue that a throttled task should not boost a CPU,
4351 * a) properly implementing CPU boosting considering throttled
4352 * tasks will increase a lot the complexity of the solution
4353 * b) it's not easy to quantify the benefits introduced by
4354 * such a more complex solution.
4355 * Thus, for the time being we go for the simple solution and boost
4356 * also for throttled RQs.
4358 schedtune_enqueue_task(p, cpu_of(rq));
4361 walt_inc_cumulative_runnable_avg(rq, p);
4362 if (!task_new && !rq->rd->overutilized &&
4363 cpu_overutilized(rq->cpu)) {
4364 rq->rd->overutilized = true;
4365 trace_sched_overutilized(true);
4369 * We want to potentially trigger a freq switch
4370 * request only for tasks that are waking up; this is
4371 * because we get here also during load balancing, but
4372 * in these cases it seems wise to trigger as single
4373 * request after load balancing is done.
4375 if (task_new || task_wakeup)
4376 update_capacity_of(cpu_of(rq));
4379 #endif /* CONFIG_SMP */
4383 static void set_next_buddy(struct sched_entity *se);
4386 * The dequeue_task method is called before nr_running is
4387 * decreased. We remove the task from the rbtree and
4388 * update the fair scheduling stats:
4390 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4392 struct cfs_rq *cfs_rq;
4393 struct sched_entity *se = &p->se;
4394 int task_sleep = flags & DEQUEUE_SLEEP;
4396 for_each_sched_entity(se) {
4397 cfs_rq = cfs_rq_of(se);
4398 dequeue_entity(cfs_rq, se, flags);
4401 * end evaluation on encountering a throttled cfs_rq
4403 * note: in the case of encountering a throttled cfs_rq we will
4404 * post the final h_nr_running decrement below.
4406 if (cfs_rq_throttled(cfs_rq))
4408 cfs_rq->h_nr_running--;
4409 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4411 /* Don't dequeue parent if it has other entities besides us */
4412 if (cfs_rq->load.weight) {
4413 /* Avoid re-evaluating load for this entity: */
4414 se = parent_entity(se);
4416 * Bias pick_next to pick a task from this cfs_rq, as
4417 * p is sleeping when it is within its sched_slice.
4419 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4423 flags |= DEQUEUE_SLEEP;
4426 for_each_sched_entity(se) {
4427 cfs_rq = cfs_rq_of(se);
4428 cfs_rq->h_nr_running--;
4429 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4431 if (cfs_rq_throttled(cfs_rq))
4434 update_load_avg(se, 1);
4435 update_cfs_shares(cfs_rq);
4439 sub_nr_running(rq, 1);
4444 * Update SchedTune accounting
4446 * We do it before updating the CPU capacity to ensure the
4447 * boost value of the current task is accounted for in the
4448 * selection of the OPP.
4450 schedtune_dequeue_task(p, cpu_of(rq));
4453 walt_dec_cumulative_runnable_avg(rq, p);
4456 * We want to potentially trigger a freq switch
4457 * request only for tasks that are going to sleep;
4458 * this is because we get here also during load
4459 * balancing, but in these cases it seems wise to
4460 * trigger as single request after load balancing is
4464 if (rq->cfs.nr_running)
4465 update_capacity_of(cpu_of(rq));
4466 else if (sched_freq())
4467 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4471 #endif /* CONFIG_SMP */
4479 * per rq 'load' arrray crap; XXX kill this.
4483 * The exact cpuload at various idx values, calculated at every tick would be
4484 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4486 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4487 * on nth tick when cpu may be busy, then we have:
4488 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4489 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4491 * decay_load_missed() below does efficient calculation of
4492 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4493 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4495 * The calculation is approximated on a 128 point scale.
4496 * degrade_zero_ticks is the number of ticks after which load at any
4497 * particular idx is approximated to be zero.
4498 * degrade_factor is a precomputed table, a row for each load idx.
4499 * Each column corresponds to degradation factor for a power of two ticks,
4500 * based on 128 point scale.
4502 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4503 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4505 * With this power of 2 load factors, we can degrade the load n times
4506 * by looking at 1 bits in n and doing as many mult/shift instead of
4507 * n mult/shifts needed by the exact degradation.
4509 #define DEGRADE_SHIFT 7
4510 static const unsigned char
4511 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4512 static const unsigned char
4513 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4514 {0, 0, 0, 0, 0, 0, 0, 0},
4515 {64, 32, 8, 0, 0, 0, 0, 0},
4516 {96, 72, 40, 12, 1, 0, 0},
4517 {112, 98, 75, 43, 15, 1, 0},
4518 {120, 112, 98, 76, 45, 16, 2} };
4521 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4522 * would be when CPU is idle and so we just decay the old load without
4523 * adding any new load.
4525 static unsigned long
4526 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4530 if (!missed_updates)
4533 if (missed_updates >= degrade_zero_ticks[idx])
4537 return load >> missed_updates;
4539 while (missed_updates) {
4540 if (missed_updates % 2)
4541 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4543 missed_updates >>= 1;
4550 * Update rq->cpu_load[] statistics. This function is usually called every
4551 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4552 * every tick. We fix it up based on jiffies.
4554 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4555 unsigned long pending_updates)
4559 this_rq->nr_load_updates++;
4561 /* Update our load: */
4562 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4563 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4564 unsigned long old_load, new_load;
4566 /* scale is effectively 1 << i now, and >> i divides by scale */
4568 old_load = this_rq->cpu_load[i];
4569 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4570 new_load = this_load;
4572 * Round up the averaging division if load is increasing. This
4573 * prevents us from getting stuck on 9 if the load is 10, for
4576 if (new_load > old_load)
4577 new_load += scale - 1;
4579 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4582 sched_avg_update(this_rq);
4585 /* Used instead of source_load when we know the type == 0 */
4586 static unsigned long weighted_cpuload(const int cpu)
4588 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4591 #ifdef CONFIG_NO_HZ_COMMON
4593 * There is no sane way to deal with nohz on smp when using jiffies because the
4594 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4595 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4597 * Therefore we cannot use the delta approach from the regular tick since that
4598 * would seriously skew the load calculation. However we'll make do for those
4599 * updates happening while idle (nohz_idle_balance) or coming out of idle
4600 * (tick_nohz_idle_exit).
4602 * This means we might still be one tick off for nohz periods.
4606 * Called from nohz_idle_balance() to update the load ratings before doing the
4609 static void update_idle_cpu_load(struct rq *this_rq)
4611 unsigned long curr_jiffies = READ_ONCE(jiffies);
4612 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4613 unsigned long pending_updates;
4616 * bail if there's load or we're actually up-to-date.
4618 if (load || curr_jiffies == this_rq->last_load_update_tick)
4621 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4622 this_rq->last_load_update_tick = curr_jiffies;
4624 __update_cpu_load(this_rq, load, pending_updates);
4628 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4630 void update_cpu_load_nohz(void)
4632 struct rq *this_rq = this_rq();
4633 unsigned long curr_jiffies = READ_ONCE(jiffies);
4634 unsigned long pending_updates;
4636 if (curr_jiffies == this_rq->last_load_update_tick)
4639 raw_spin_lock(&this_rq->lock);
4640 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4641 if (pending_updates) {
4642 this_rq->last_load_update_tick = curr_jiffies;
4644 * We were idle, this means load 0, the current load might be
4645 * !0 due to remote wakeups and the sort.
4647 __update_cpu_load(this_rq, 0, pending_updates);
4649 raw_spin_unlock(&this_rq->lock);
4651 #endif /* CONFIG_NO_HZ */
4654 * Called from scheduler_tick()
4656 void update_cpu_load_active(struct rq *this_rq)
4658 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4660 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4662 this_rq->last_load_update_tick = jiffies;
4663 __update_cpu_load(this_rq, load, 1);
4667 * Return a low guess at the load of a migration-source cpu weighted
4668 * according to the scheduling class and "nice" value.
4670 * We want to under-estimate the load of migration sources, to
4671 * balance conservatively.
4673 static unsigned long source_load(int cpu, int type)
4675 struct rq *rq = cpu_rq(cpu);
4676 unsigned long total = weighted_cpuload(cpu);
4678 if (type == 0 || !sched_feat(LB_BIAS))
4681 return min(rq->cpu_load[type-1], total);
4685 * Return a high guess at the load of a migration-target cpu weighted
4686 * according to the scheduling class and "nice" value.
4688 static unsigned long target_load(int cpu, int type)
4690 struct rq *rq = cpu_rq(cpu);
4691 unsigned long total = weighted_cpuload(cpu);
4693 if (type == 0 || !sched_feat(LB_BIAS))
4696 return max(rq->cpu_load[type-1], total);
4700 static unsigned long cpu_avg_load_per_task(int cpu)
4702 struct rq *rq = cpu_rq(cpu);
4703 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4704 unsigned long load_avg = weighted_cpuload(cpu);
4707 return load_avg / nr_running;
4712 static void record_wakee(struct task_struct *p)
4715 * Rough decay (wiping) for cost saving, don't worry
4716 * about the boundary, really active task won't care
4719 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4720 current->wakee_flips >>= 1;
4721 current->wakee_flip_decay_ts = jiffies;
4724 if (current->last_wakee != p) {
4725 current->last_wakee = p;
4726 current->wakee_flips++;
4730 static void task_waking_fair(struct task_struct *p)
4732 struct sched_entity *se = &p->se;
4733 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4736 #ifndef CONFIG_64BIT
4737 u64 min_vruntime_copy;
4740 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4742 min_vruntime = cfs_rq->min_vruntime;
4743 } while (min_vruntime != min_vruntime_copy);
4745 min_vruntime = cfs_rq->min_vruntime;
4748 se->vruntime -= min_vruntime;
4752 #ifdef CONFIG_FAIR_GROUP_SCHED
4754 * effective_load() calculates the load change as seen from the root_task_group
4756 * Adding load to a group doesn't make a group heavier, but can cause movement
4757 * of group shares between cpus. Assuming the shares were perfectly aligned one
4758 * can calculate the shift in shares.
4760 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4761 * on this @cpu and results in a total addition (subtraction) of @wg to the
4762 * total group weight.
4764 * Given a runqueue weight distribution (rw_i) we can compute a shares
4765 * distribution (s_i) using:
4767 * s_i = rw_i / \Sum rw_j (1)
4769 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4770 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4771 * shares distribution (s_i):
4773 * rw_i = { 2, 4, 1, 0 }
4774 * s_i = { 2/7, 4/7, 1/7, 0 }
4776 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4777 * task used to run on and the CPU the waker is running on), we need to
4778 * compute the effect of waking a task on either CPU and, in case of a sync
4779 * wakeup, compute the effect of the current task going to sleep.
4781 * So for a change of @wl to the local @cpu with an overall group weight change
4782 * of @wl we can compute the new shares distribution (s'_i) using:
4784 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4786 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4787 * differences in waking a task to CPU 0. The additional task changes the
4788 * weight and shares distributions like:
4790 * rw'_i = { 3, 4, 1, 0 }
4791 * s'_i = { 3/8, 4/8, 1/8, 0 }
4793 * We can then compute the difference in effective weight by using:
4795 * dw_i = S * (s'_i - s_i) (3)
4797 * Where 'S' is the group weight as seen by its parent.
4799 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4800 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4801 * 4/7) times the weight of the group.
4803 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4805 struct sched_entity *se = tg->se[cpu];
4807 if (!tg->parent) /* the trivial, non-cgroup case */
4810 for_each_sched_entity(se) {
4811 struct cfs_rq *cfs_rq = se->my_q;
4812 long W, w = cfs_rq_load_avg(cfs_rq);
4817 * W = @wg + \Sum rw_j
4819 W = wg + atomic_long_read(&tg->load_avg);
4821 /* Ensure \Sum rw_j >= rw_i */
4822 W -= cfs_rq->tg_load_avg_contrib;
4831 * wl = S * s'_i; see (2)
4834 wl = (w * (long)tg->shares) / W;
4839 * Per the above, wl is the new se->load.weight value; since
4840 * those are clipped to [MIN_SHARES, ...) do so now. See
4841 * calc_cfs_shares().
4843 if (wl < MIN_SHARES)
4847 * wl = dw_i = S * (s'_i - s_i); see (3)
4849 wl -= se->avg.load_avg;
4852 * Recursively apply this logic to all parent groups to compute
4853 * the final effective load change on the root group. Since
4854 * only the @tg group gets extra weight, all parent groups can
4855 * only redistribute existing shares. @wl is the shift in shares
4856 * resulting from this level per the above.
4865 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4873 * Returns the current capacity of cpu after applying both
4874 * cpu and freq scaling.
4876 unsigned long capacity_curr_of(int cpu)
4878 return cpu_rq(cpu)->cpu_capacity_orig *
4879 arch_scale_freq_capacity(NULL, cpu)
4880 >> SCHED_CAPACITY_SHIFT;
4883 static inline bool energy_aware(void)
4885 return sched_feat(ENERGY_AWARE);
4889 struct sched_group *sg_top;
4890 struct sched_group *sg_cap;
4897 struct task_struct *task;
4912 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4913 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4914 * energy calculations. Using the scale-invariant util returned by
4915 * cpu_util() and approximating scale-invariant util by:
4917 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4919 * the normalized util can be found using the specific capacity.
4921 * capacity = capacity_orig * curr_freq/max_freq
4923 * norm_util = running_time/time ~ util/capacity
4925 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4927 int util = __cpu_util(cpu, delta);
4929 if (util >= capacity)
4930 return SCHED_CAPACITY_SCALE;
4932 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4935 static int calc_util_delta(struct energy_env *eenv, int cpu)
4937 if (cpu == eenv->src_cpu)
4938 return -eenv->util_delta;
4939 if (cpu == eenv->dst_cpu)
4940 return eenv->util_delta;
4945 unsigned long group_max_util(struct energy_env *eenv)
4948 unsigned long max_util = 0;
4950 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4951 delta = calc_util_delta(eenv, i);
4952 max_util = max(max_util, __cpu_util(i, delta));
4959 * group_norm_util() returns the approximated group util relative to it's
4960 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4961 * energy calculations. Since task executions may or may not overlap in time in
4962 * the group the true normalized util is between max(cpu_norm_util(i)) and
4963 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4964 * latter is used as the estimate as it leads to a more pessimistic energy
4965 * estimate (more busy).
4968 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4971 unsigned long util_sum = 0;
4972 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4974 for_each_cpu(i, sched_group_cpus(sg)) {
4975 delta = calc_util_delta(eenv, i);
4976 util_sum += __cpu_norm_util(i, capacity, delta);
4979 if (util_sum > SCHED_CAPACITY_SCALE)
4980 return SCHED_CAPACITY_SCALE;
4984 static int find_new_capacity(struct energy_env *eenv,
4985 const struct sched_group_energy * const sge)
4988 unsigned long util = group_max_util(eenv);
4990 for (idx = 0; idx < sge->nr_cap_states; idx++) {
4991 if (sge->cap_states[idx].cap >= util)
4995 eenv->cap_idx = idx;
5000 static int group_idle_state(struct sched_group *sg)
5002 int i, state = INT_MAX;
5004 /* Find the shallowest idle state in the sched group. */
5005 for_each_cpu(i, sched_group_cpus(sg))
5006 state = min(state, idle_get_state_idx(cpu_rq(i)));
5008 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5015 * sched_group_energy(): Computes the absolute energy consumption of cpus
5016 * belonging to the sched_group including shared resources shared only by
5017 * members of the group. Iterates over all cpus in the hierarchy below the
5018 * sched_group starting from the bottom working it's way up before going to
5019 * the next cpu until all cpus are covered at all levels. The current
5020 * implementation is likely to gather the same util statistics multiple times.
5021 * This can probably be done in a faster but more complex way.
5022 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5024 static int sched_group_energy(struct energy_env *eenv)
5026 struct sched_domain *sd;
5027 int cpu, total_energy = 0;
5028 struct cpumask visit_cpus;
5029 struct sched_group *sg;
5031 WARN_ON(!eenv->sg_top->sge);
5033 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5035 while (!cpumask_empty(&visit_cpus)) {
5036 struct sched_group *sg_shared_cap = NULL;
5038 cpu = cpumask_first(&visit_cpus);
5041 * Is the group utilization affected by cpus outside this
5044 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5048 * We most probably raced with hotplug; returning a
5049 * wrong energy estimation is better than entering an
5055 sg_shared_cap = sd->parent->groups;
5057 for_each_domain(cpu, sd) {
5060 /* Has this sched_domain already been visited? */
5061 if (sd->child && group_first_cpu(sg) != cpu)
5065 unsigned long group_util;
5066 int sg_busy_energy, sg_idle_energy;
5067 int cap_idx, idle_idx;
5069 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5070 eenv->sg_cap = sg_shared_cap;
5074 cap_idx = find_new_capacity(eenv, sg->sge);
5076 if (sg->group_weight == 1) {
5077 /* Remove capacity of src CPU (before task move) */
5078 if (eenv->util_delta == 0 &&
5079 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5080 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5081 eenv->cap.delta -= eenv->cap.before;
5083 /* Add capacity of dst CPU (after task move) */
5084 if (eenv->util_delta != 0 &&
5085 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5086 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5087 eenv->cap.delta += eenv->cap.after;
5091 idle_idx = group_idle_state(sg);
5092 group_util = group_norm_util(eenv, sg);
5093 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5094 >> SCHED_CAPACITY_SHIFT;
5095 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5096 * sg->sge->idle_states[idle_idx].power)
5097 >> SCHED_CAPACITY_SHIFT;
5099 total_energy += sg_busy_energy + sg_idle_energy;
5102 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5104 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5107 } while (sg = sg->next, sg != sd->groups);
5110 cpumask_clear_cpu(cpu, &visit_cpus);
5114 eenv->energy = total_energy;
5118 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5120 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5124 * energy_diff(): Estimate the energy impact of changing the utilization
5125 * distribution. eenv specifies the change: utilisation amount, source, and
5126 * destination cpu. Source or destination cpu may be -1 in which case the
5127 * utilization is removed from or added to the system (e.g. task wake-up). If
5128 * both are specified, the utilization is migrated.
5130 static inline int __energy_diff(struct energy_env *eenv)
5132 struct sched_domain *sd;
5133 struct sched_group *sg;
5134 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5136 struct energy_env eenv_before = {
5138 .src_cpu = eenv->src_cpu,
5139 .dst_cpu = eenv->dst_cpu,
5140 .nrg = { 0, 0, 0, 0},
5144 if (eenv->src_cpu == eenv->dst_cpu)
5147 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5148 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5151 return 0; /* Error */
5156 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5157 eenv_before.sg_top = eenv->sg_top = sg;
5159 if (sched_group_energy(&eenv_before))
5160 return 0; /* Invalid result abort */
5161 energy_before += eenv_before.energy;
5163 /* Keep track of SRC cpu (before) capacity */
5164 eenv->cap.before = eenv_before.cap.before;
5165 eenv->cap.delta = eenv_before.cap.delta;
5167 if (sched_group_energy(eenv))
5168 return 0; /* Invalid result abort */
5169 energy_after += eenv->energy;
5171 } while (sg = sg->next, sg != sd->groups);
5173 eenv->nrg.before = energy_before;
5174 eenv->nrg.after = energy_after;
5175 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5178 trace_sched_energy_diff(eenv->task,
5179 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5180 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5181 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5182 eenv->nrg.delta, eenv->payoff);
5184 return eenv->nrg.diff;
5187 #ifdef CONFIG_SCHED_TUNE
5189 struct target_nrg schedtune_target_nrg;
5192 * System energy normalization
5193 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5194 * corresponding to the specified energy variation.
5197 normalize_energy(int energy_diff)
5200 #ifdef CONFIG_SCHED_DEBUG
5203 /* Check for boundaries */
5204 max_delta = schedtune_target_nrg.max_power;
5205 max_delta -= schedtune_target_nrg.min_power;
5206 WARN_ON(abs(energy_diff) >= max_delta);
5209 /* Do scaling using positive numbers to increase the range */
5210 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5212 /* Scale by energy magnitude */
5213 normalized_nrg <<= SCHED_LOAD_SHIFT;
5215 /* Normalize on max energy for target platform */
5216 normalized_nrg = reciprocal_divide(
5217 normalized_nrg, schedtune_target_nrg.rdiv);
5219 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5223 energy_diff(struct energy_env *eenv)
5225 int boost = schedtune_task_boost(eenv->task);
5228 /* Conpute "absolute" energy diff */
5229 __energy_diff(eenv);
5231 /* Return energy diff when boost margin is 0 */
5233 return eenv->nrg.diff;
5235 /* Compute normalized energy diff */
5236 nrg_delta = normalize_energy(eenv->nrg.diff);
5237 eenv->nrg.delta = nrg_delta;
5239 eenv->payoff = schedtune_accept_deltas(
5245 * When SchedTune is enabled, the energy_diff() function will return
5246 * the computed energy payoff value. Since the energy_diff() return
5247 * value is expected to be negative by its callers, this evaluation
5248 * function return a negative value each time the evaluation return a
5249 * positive payoff, which is the condition for the acceptance of
5250 * a scheduling decision
5252 return -eenv->payoff;
5254 #else /* CONFIG_SCHED_TUNE */
5255 #define energy_diff(eenv) __energy_diff(eenv)
5259 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5260 * A waker of many should wake a different task than the one last awakened
5261 * at a frequency roughly N times higher than one of its wakees. In order
5262 * to determine whether we should let the load spread vs consolodating to
5263 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5264 * partner, and a factor of lls_size higher frequency in the other. With
5265 * both conditions met, we can be relatively sure that the relationship is
5266 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5267 * being client/server, worker/dispatcher, interrupt source or whatever is
5268 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5270 static int wake_wide(struct task_struct *p)
5272 unsigned int master = current->wakee_flips;
5273 unsigned int slave = p->wakee_flips;
5274 int factor = this_cpu_read(sd_llc_size);
5277 swap(master, slave);
5278 if (slave < factor || master < slave * factor)
5283 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5285 s64 this_load, load;
5286 s64 this_eff_load, prev_eff_load;
5287 int idx, this_cpu, prev_cpu;
5288 struct task_group *tg;
5289 unsigned long weight;
5293 this_cpu = smp_processor_id();
5294 prev_cpu = task_cpu(p);
5295 load = source_load(prev_cpu, idx);
5296 this_load = target_load(this_cpu, idx);
5299 * If sync wakeup then subtract the (maximum possible)
5300 * effect of the currently running task from the load
5301 * of the current CPU:
5304 tg = task_group(current);
5305 weight = current->se.avg.load_avg;
5307 this_load += effective_load(tg, this_cpu, -weight, -weight);
5308 load += effective_load(tg, prev_cpu, 0, -weight);
5312 weight = p->se.avg.load_avg;
5315 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5316 * due to the sync cause above having dropped this_load to 0, we'll
5317 * always have an imbalance, but there's really nothing you can do
5318 * about that, so that's good too.
5320 * Otherwise check if either cpus are near enough in load to allow this
5321 * task to be woken on this_cpu.
5323 this_eff_load = 100;
5324 this_eff_load *= capacity_of(prev_cpu);
5326 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5327 prev_eff_load *= capacity_of(this_cpu);
5329 if (this_load > 0) {
5330 this_eff_load *= this_load +
5331 effective_load(tg, this_cpu, weight, weight);
5333 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5336 balanced = this_eff_load <= prev_eff_load;
5338 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5343 schedstat_inc(sd, ttwu_move_affine);
5344 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5349 static inline unsigned long task_util(struct task_struct *p)
5351 #ifdef CONFIG_SCHED_WALT
5352 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5353 unsigned long demand = p->ravg.demand;
5354 return (demand << 10) / walt_ravg_window;
5357 return p->se.avg.util_avg;
5360 unsigned int capacity_margin = 1280; /* ~20% margin */
5362 static inline unsigned long boosted_task_util(struct task_struct *task);
5364 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5366 unsigned long capacity = capacity_of(cpu);
5368 util += boosted_task_util(p);
5370 return (capacity * 1024) > (util * capacity_margin);
5373 static inline bool task_fits_max(struct task_struct *p, int cpu)
5375 unsigned long capacity = capacity_of(cpu);
5376 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5378 if (capacity == max_capacity)
5381 if (capacity * capacity_margin > max_capacity * 1024)
5384 return __task_fits(p, cpu, 0);
5387 static inline bool task_fits_spare(struct task_struct *p, int cpu)
5389 return __task_fits(p, cpu, cpu_util(cpu));
5392 static bool cpu_overutilized(int cpu)
5394 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5397 #ifdef CONFIG_SCHED_TUNE
5400 schedtune_margin(unsigned long signal, long boost)
5402 long long margin = 0;
5405 * Signal proportional compensation (SPC)
5407 * The Boost (B) value is used to compute a Margin (M) which is
5408 * proportional to the complement of the original Signal (S):
5409 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5410 * M = B * S, if B is negative
5411 * The obtained M could be used by the caller to "boost" S.
5414 margin = SCHED_LOAD_SCALE - signal;
5417 margin = -signal * boost;
5419 * Fast integer division by constant:
5420 * Constant : (C) = 100
5421 * Precision : 0.1% (P) = 0.1
5422 * Reference : C * 100 / P (R) = 100000
5425 * Shift bits : ceil(log(R,2)) (S) = 17
5426 * Mult const : round(2^S/C) (M) = 1311
5439 schedtune_cpu_margin(unsigned long util, int cpu)
5441 int boost = schedtune_cpu_boost(cpu);
5446 return schedtune_margin(util, boost);
5450 schedtune_task_margin(struct task_struct *task)
5452 int boost = schedtune_task_boost(task);
5459 util = task_util(task);
5460 margin = schedtune_margin(util, boost);
5465 #else /* CONFIG_SCHED_TUNE */
5468 schedtune_cpu_margin(unsigned long util, int cpu)
5474 schedtune_task_margin(struct task_struct *task)
5479 #endif /* CONFIG_SCHED_TUNE */
5482 boosted_cpu_util(int cpu)
5484 unsigned long util = cpu_util(cpu);
5485 long margin = schedtune_cpu_margin(util, cpu);
5487 trace_sched_boost_cpu(cpu, util, margin);
5489 return util + margin;
5492 static inline unsigned long
5493 boosted_task_util(struct task_struct *task)
5495 unsigned long util = task_util(task);
5496 long margin = schedtune_task_margin(task);
5498 trace_sched_boost_task(task, util, margin);
5500 return util + margin;
5504 * find_idlest_group finds and returns the least busy CPU group within the
5507 static struct sched_group *
5508 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5509 int this_cpu, int sd_flag)
5511 struct sched_group *idlest = NULL, *group = sd->groups;
5512 struct sched_group *fit_group = NULL, *spare_group = NULL;
5513 unsigned long min_load = ULONG_MAX, this_load = 0;
5514 unsigned long fit_capacity = ULONG_MAX;
5515 unsigned long max_spare_capacity = capacity_margin - SCHED_LOAD_SCALE;
5516 int load_idx = sd->forkexec_idx;
5517 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5519 if (sd_flag & SD_BALANCE_WAKE)
5520 load_idx = sd->wake_idx;
5523 unsigned long load, avg_load, spare_capacity;
5527 /* Skip over this group if it has no CPUs allowed */
5528 if (!cpumask_intersects(sched_group_cpus(group),
5529 tsk_cpus_allowed(p)))
5532 local_group = cpumask_test_cpu(this_cpu,
5533 sched_group_cpus(group));
5535 /* Tally up the load of all CPUs in the group */
5538 for_each_cpu(i, sched_group_cpus(group)) {
5539 /* Bias balancing toward cpus of our domain */
5541 load = source_load(i, load_idx);
5543 load = target_load(i, load_idx);
5548 * Look for most energy-efficient group that can fit
5549 * that can fit the task.
5551 if (capacity_of(i) < fit_capacity && task_fits_spare(p, i)) {
5552 fit_capacity = capacity_of(i);
5557 * Look for group which has most spare capacity on a
5560 spare_capacity = capacity_of(i) - cpu_util(i);
5561 if (spare_capacity > max_spare_capacity) {
5562 max_spare_capacity = spare_capacity;
5563 spare_group = group;
5567 /* Adjust by relative CPU capacity of the group */
5568 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5571 this_load = avg_load;
5572 } else if (avg_load < min_load) {
5573 min_load = avg_load;
5576 } while (group = group->next, group != sd->groups);
5584 if (!idlest || 100*this_load < imbalance*min_load)
5590 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5593 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5595 unsigned long load, min_load = ULONG_MAX;
5596 unsigned int min_exit_latency = UINT_MAX;
5597 u64 latest_idle_timestamp = 0;
5598 int least_loaded_cpu = this_cpu;
5599 int shallowest_idle_cpu = -1;
5602 /* Traverse only the allowed CPUs */
5603 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5604 if (task_fits_spare(p, i)) {
5605 struct rq *rq = cpu_rq(i);
5606 struct cpuidle_state *idle = idle_get_state(rq);
5607 if (idle && idle->exit_latency < min_exit_latency) {
5609 * We give priority to a CPU whose idle state
5610 * has the smallest exit latency irrespective
5611 * of any idle timestamp.
5613 min_exit_latency = idle->exit_latency;
5614 latest_idle_timestamp = rq->idle_stamp;
5615 shallowest_idle_cpu = i;
5616 } else if (idle_cpu(i) &&
5617 (!idle || idle->exit_latency == min_exit_latency) &&
5618 rq->idle_stamp > latest_idle_timestamp) {
5620 * If equal or no active idle state, then
5621 * the most recently idled CPU might have
5624 latest_idle_timestamp = rq->idle_stamp;
5625 shallowest_idle_cpu = i;
5626 } else if (shallowest_idle_cpu == -1) {
5628 * If we haven't found an idle CPU yet
5629 * pick a non-idle one that can fit the task as
5632 shallowest_idle_cpu = i;
5634 } else if (shallowest_idle_cpu == -1) {
5635 load = weighted_cpuload(i);
5636 if (load < min_load || (load == min_load && i == this_cpu)) {
5638 least_loaded_cpu = i;
5643 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5647 * Try and locate an idle CPU in the sched_domain.
5649 static int select_idle_sibling(struct task_struct *p, int target)
5651 struct sched_domain *sd;
5652 struct sched_group *sg;
5653 int i = task_cpu(p);
5655 int best_idle_cstate = -1;
5656 int best_idle_capacity = INT_MAX;
5658 if (!sysctl_sched_cstate_aware) {
5659 if (idle_cpu(target))
5663 * If the prevous cpu is cache affine and idle, don't be stupid.
5665 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5670 * Otherwise, iterate the domains and find an elegible idle cpu.
5672 sd = rcu_dereference(per_cpu(sd_llc, target));
5673 for_each_lower_domain(sd) {
5676 if (!cpumask_intersects(sched_group_cpus(sg),
5677 tsk_cpus_allowed(p)))
5680 if (sysctl_sched_cstate_aware) {
5681 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5682 struct rq *rq = cpu_rq(i);
5683 int idle_idx = idle_get_state_idx(rq);
5684 unsigned long new_usage = boosted_task_util(p);
5685 unsigned long capacity_orig = capacity_orig_of(i);
5686 if (new_usage > capacity_orig || !idle_cpu(i))
5689 if (i == target && new_usage <= capacity_curr_of(target))
5692 if (best_idle < 0 || (idle_idx < best_idle_cstate && capacity_orig <= best_idle_capacity)) {
5694 best_idle_cstate = idle_idx;
5695 best_idle_capacity = capacity_orig;
5699 for_each_cpu(i, sched_group_cpus(sg)) {
5700 if (i == target || !idle_cpu(i))
5704 target = cpumask_first_and(sched_group_cpus(sg),
5705 tsk_cpus_allowed(p));
5710 } while (sg != sd->groups);
5719 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5722 int target_cpu = -1;
5723 int target_util = 0;
5724 int backup_capacity = 0;
5725 int best_idle_cpu = -1;
5726 int best_idle_cstate = INT_MAX;
5727 int backup_cpu = -1;
5728 unsigned long task_util_boosted, new_util;
5730 task_util_boosted = boosted_task_util(p);
5731 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5737 * Iterate from higher cpus for boosted tasks.
5739 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5741 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5745 * p's blocked utilization is still accounted for on prev_cpu
5746 * so prev_cpu will receive a negative bias due to the double
5747 * accounting. However, the blocked utilization may be zero.
5749 new_util = cpu_util(i) + task_util_boosted;
5752 * Ensure minimum capacity to grant the required boost.
5753 * The target CPU can be already at a capacity level higher
5754 * than the one required to boost the task.
5756 if (new_util > capacity_orig_of(i))
5759 #ifdef CONFIG_SCHED_WALT
5760 if (walt_cpu_high_irqload(i))
5764 * Unconditionally favoring tasks that prefer idle cpus to
5767 if (idle_cpu(i) && prefer_idle) {
5768 if (best_idle_cpu < 0)
5773 cur_capacity = capacity_curr_of(i);
5775 idle_idx = idle_get_state_idx(rq);
5777 if (new_util < cur_capacity) {
5778 if (cpu_rq(i)->nr_running) {
5780 /* Find a target cpu with highest
5783 if (target_util == 0 ||
5784 target_util < new_util) {
5786 target_util = new_util;
5789 /* Find a target cpu with lowest
5792 if (target_util == 0 ||
5793 target_util > new_util) {
5795 target_util = new_util;
5798 } else if (!prefer_idle) {
5799 if (best_idle_cpu < 0 ||
5800 (sysctl_sched_cstate_aware &&
5801 best_idle_cstate > idle_idx)) {
5802 best_idle_cstate = idle_idx;
5806 } else if (backup_capacity == 0 ||
5807 backup_capacity > cur_capacity) {
5808 // Find a backup cpu with least capacity.
5809 backup_capacity = cur_capacity;
5814 if (prefer_idle && best_idle_cpu >= 0)
5815 target_cpu = best_idle_cpu;
5816 else if (target_cpu < 0)
5817 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5822 static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync)
5824 struct sched_domain *sd;
5825 struct sched_group *sg, *sg_target;
5826 int target_max_cap = INT_MAX;
5827 int target_cpu = task_cpu(p);
5828 unsigned long task_util_boosted, new_util;
5831 if (sysctl_sched_sync_hint_enable && sync) {
5832 int cpu = smp_processor_id();
5833 cpumask_t search_cpus;
5834 cpumask_and(&search_cpus, tsk_cpus_allowed(p), cpu_online_mask);
5835 if (cpumask_test_cpu(cpu, &search_cpus))
5839 sd = rcu_dereference(per_cpu(sd_ea, task_cpu(p)));
5847 if (sysctl_sched_is_big_little) {
5850 * Find group with sufficient capacity. We only get here if no cpu is
5851 * overutilized. We may end up overutilizing a cpu by adding the task,
5852 * but that should not be any worse than select_idle_sibling().
5853 * load_balance() should sort it out later as we get above the tipping
5857 /* Assuming all cpus are the same in group */
5858 int max_cap_cpu = group_first_cpu(sg);
5861 * Assume smaller max capacity means more energy-efficient.
5862 * Ideally we should query the energy model for the right
5863 * answer but it easily ends up in an exhaustive search.
5865 if (capacity_of(max_cap_cpu) < target_max_cap &&
5866 task_fits_max(p, max_cap_cpu)) {
5868 target_max_cap = capacity_of(max_cap_cpu);
5870 } while (sg = sg->next, sg != sd->groups);
5872 task_util_boosted = boosted_task_util(p);
5873 /* Find cpu with sufficient capacity */
5874 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg_target)) {
5876 * p's blocked utilization is still accounted for on prev_cpu
5877 * so prev_cpu will receive a negative bias due to the double
5878 * accounting. However, the blocked utilization may be zero.
5880 new_util = cpu_util(i) + task_util_boosted;
5883 * Ensure minimum capacity to grant the required boost.
5884 * The target CPU can be already at a capacity level higher
5885 * than the one required to boost the task.
5887 if (new_util > capacity_orig_of(i))
5890 if (new_util < capacity_curr_of(i)) {
5892 if (cpu_rq(i)->nr_running)
5896 /* cpu has capacity at higher OPP, keep it as fallback */
5897 if (target_cpu == task_cpu(p))
5902 * Find a cpu with sufficient capacity
5904 #ifdef CONFIG_CGROUP_SCHEDTUNE
5905 bool boosted = schedtune_task_boost(p) > 0;
5906 bool prefer_idle = schedtune_prefer_idle(p) > 0;
5909 bool prefer_idle = 0;
5911 int tmp_target = find_best_target(p, boosted, prefer_idle);
5912 if (tmp_target >= 0) {
5913 target_cpu = tmp_target;
5914 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5919 if (target_cpu != task_cpu(p)) {
5920 struct energy_env eenv = {
5921 .util_delta = task_util(p),
5922 .src_cpu = task_cpu(p),
5923 .dst_cpu = target_cpu,
5927 /* Not enough spare capacity on previous cpu */
5928 if (cpu_overutilized(task_cpu(p)))
5931 if (energy_diff(&eenv) >= 0)
5939 * select_task_rq_fair: Select target runqueue for the waking task in domains
5940 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5941 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5943 * Balances load by selecting the idlest cpu in the idlest group, or under
5944 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5946 * Returns the target cpu number.
5948 * preempt must be disabled.
5951 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5953 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5954 int cpu = smp_processor_id();
5955 int new_cpu = prev_cpu;
5956 int want_affine = 0;
5957 int sync = wake_flags & WF_SYNC;
5959 if (sd_flag & SD_BALANCE_WAKE)
5960 want_affine = (!wake_wide(p) && task_fits_max(p, cpu) &&
5961 cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) ||
5965 for_each_domain(cpu, tmp) {
5966 if (!(tmp->flags & SD_LOAD_BALANCE))
5970 * If both cpu and prev_cpu are part of this domain,
5971 * cpu is a valid SD_WAKE_AFFINE target.
5973 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5974 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5979 if (tmp->flags & sd_flag)
5981 else if (!want_affine)
5986 sd = NULL; /* Prefer wake_affine over balance flags */
5987 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5992 if (energy_aware() && !cpu_rq(cpu)->rd->overutilized)
5993 new_cpu = energy_aware_wake_cpu(p, prev_cpu, sync);
5994 else if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5995 new_cpu = select_idle_sibling(p, new_cpu);
5998 struct sched_group *group;
6001 if (!(sd->flags & sd_flag)) {
6006 group = find_idlest_group(sd, p, cpu, sd_flag);
6012 new_cpu = find_idlest_cpu(group, p, cpu);
6013 if (new_cpu == -1 || new_cpu == cpu) {
6014 /* Now try balancing at a lower domain level of cpu */
6019 /* Now try balancing at a lower domain level of new_cpu */
6021 weight = sd->span_weight;
6023 for_each_domain(cpu, tmp) {
6024 if (weight <= tmp->span_weight)
6026 if (tmp->flags & sd_flag)
6029 /* while loop will break here if sd == NULL */
6037 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6038 * cfs_rq_of(p) references at time of call are still valid and identify the
6039 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6040 * other assumptions, including the state of rq->lock, should be made.
6042 static void migrate_task_rq_fair(struct task_struct *p)
6045 * We are supposed to update the task to "current" time, then its up to date
6046 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6047 * what current time is, so simply throw away the out-of-date time. This
6048 * will result in the wakee task is less decayed, but giving the wakee more
6049 * load sounds not bad.
6051 remove_entity_load_avg(&p->se);
6053 /* Tell new CPU we are migrated */
6054 p->se.avg.last_update_time = 0;
6056 /* We have migrated, no longer consider this task hot */
6057 p->se.exec_start = 0;
6060 static void task_dead_fair(struct task_struct *p)
6062 remove_entity_load_avg(&p->se);
6065 #define task_fits_max(p, cpu) true
6066 #endif /* CONFIG_SMP */
6068 static unsigned long
6069 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6071 unsigned long gran = sysctl_sched_wakeup_granularity;
6074 * Since its curr running now, convert the gran from real-time
6075 * to virtual-time in his units.
6077 * By using 'se' instead of 'curr' we penalize light tasks, so
6078 * they get preempted easier. That is, if 'se' < 'curr' then
6079 * the resulting gran will be larger, therefore penalizing the
6080 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6081 * be smaller, again penalizing the lighter task.
6083 * This is especially important for buddies when the leftmost
6084 * task is higher priority than the buddy.
6086 return calc_delta_fair(gran, se);
6090 * Should 'se' preempt 'curr'.
6104 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6106 s64 gran, vdiff = curr->vruntime - se->vruntime;
6111 gran = wakeup_gran(curr, se);
6118 static void set_last_buddy(struct sched_entity *se)
6120 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6123 for_each_sched_entity(se)
6124 cfs_rq_of(se)->last = se;
6127 static void set_next_buddy(struct sched_entity *se)
6129 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6132 for_each_sched_entity(se)
6133 cfs_rq_of(se)->next = se;
6136 static void set_skip_buddy(struct sched_entity *se)
6138 for_each_sched_entity(se)
6139 cfs_rq_of(se)->skip = se;
6143 * Preempt the current task with a newly woken task if needed:
6145 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6147 struct task_struct *curr = rq->curr;
6148 struct sched_entity *se = &curr->se, *pse = &p->se;
6149 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6150 int scale = cfs_rq->nr_running >= sched_nr_latency;
6151 int next_buddy_marked = 0;
6153 if (unlikely(se == pse))
6157 * This is possible from callers such as attach_tasks(), in which we
6158 * unconditionally check_prempt_curr() after an enqueue (which may have
6159 * lead to a throttle). This both saves work and prevents false
6160 * next-buddy nomination below.
6162 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6165 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6166 set_next_buddy(pse);
6167 next_buddy_marked = 1;
6171 * We can come here with TIF_NEED_RESCHED already set from new task
6174 * Note: this also catches the edge-case of curr being in a throttled
6175 * group (e.g. via set_curr_task), since update_curr() (in the
6176 * enqueue of curr) will have resulted in resched being set. This
6177 * prevents us from potentially nominating it as a false LAST_BUDDY
6180 if (test_tsk_need_resched(curr))
6183 /* Idle tasks are by definition preempted by non-idle tasks. */
6184 if (unlikely(curr->policy == SCHED_IDLE) &&
6185 likely(p->policy != SCHED_IDLE))
6189 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6190 * is driven by the tick):
6192 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6195 find_matching_se(&se, &pse);
6196 update_curr(cfs_rq_of(se));
6198 if (wakeup_preempt_entity(se, pse) == 1) {
6200 * Bias pick_next to pick the sched entity that is
6201 * triggering this preemption.
6203 if (!next_buddy_marked)
6204 set_next_buddy(pse);
6213 * Only set the backward buddy when the current task is still
6214 * on the rq. This can happen when a wakeup gets interleaved
6215 * with schedule on the ->pre_schedule() or idle_balance()
6216 * point, either of which can * drop the rq lock.
6218 * Also, during early boot the idle thread is in the fair class,
6219 * for obvious reasons its a bad idea to schedule back to it.
6221 if (unlikely(!se->on_rq || curr == rq->idle))
6224 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6228 static struct task_struct *
6229 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6231 struct cfs_rq *cfs_rq = &rq->cfs;
6232 struct sched_entity *se;
6233 struct task_struct *p;
6237 #ifdef CONFIG_FAIR_GROUP_SCHED
6238 if (!cfs_rq->nr_running)
6241 if (prev->sched_class != &fair_sched_class)
6245 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6246 * likely that a next task is from the same cgroup as the current.
6248 * Therefore attempt to avoid putting and setting the entire cgroup
6249 * hierarchy, only change the part that actually changes.
6253 struct sched_entity *curr = cfs_rq->curr;
6256 * Since we got here without doing put_prev_entity() we also
6257 * have to consider cfs_rq->curr. If it is still a runnable
6258 * entity, update_curr() will update its vruntime, otherwise
6259 * forget we've ever seen it.
6263 update_curr(cfs_rq);
6268 * This call to check_cfs_rq_runtime() will do the
6269 * throttle and dequeue its entity in the parent(s).
6270 * Therefore the 'simple' nr_running test will indeed
6273 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6277 se = pick_next_entity(cfs_rq, curr);
6278 cfs_rq = group_cfs_rq(se);
6284 * Since we haven't yet done put_prev_entity and if the selected task
6285 * is a different task than we started out with, try and touch the
6286 * least amount of cfs_rqs.
6289 struct sched_entity *pse = &prev->se;
6291 while (!(cfs_rq = is_same_group(se, pse))) {
6292 int se_depth = se->depth;
6293 int pse_depth = pse->depth;
6295 if (se_depth <= pse_depth) {
6296 put_prev_entity(cfs_rq_of(pse), pse);
6297 pse = parent_entity(pse);
6299 if (se_depth >= pse_depth) {
6300 set_next_entity(cfs_rq_of(se), se);
6301 se = parent_entity(se);
6305 put_prev_entity(cfs_rq, pse);
6306 set_next_entity(cfs_rq, se);
6309 if (hrtick_enabled(rq))
6310 hrtick_start_fair(rq, p);
6312 rq->misfit_task = !task_fits_max(p, rq->cpu);
6319 if (!cfs_rq->nr_running)
6322 put_prev_task(rq, prev);
6325 se = pick_next_entity(cfs_rq, NULL);
6326 set_next_entity(cfs_rq, se);
6327 cfs_rq = group_cfs_rq(se);
6332 if (hrtick_enabled(rq))
6333 hrtick_start_fair(rq, p);
6335 rq->misfit_task = !task_fits_max(p, rq->cpu);
6340 rq->misfit_task = 0;
6342 * This is OK, because current is on_cpu, which avoids it being picked
6343 * for load-balance and preemption/IRQs are still disabled avoiding
6344 * further scheduler activity on it and we're being very careful to
6345 * re-start the picking loop.
6347 lockdep_unpin_lock(&rq->lock);
6348 new_tasks = idle_balance(rq);
6349 lockdep_pin_lock(&rq->lock);
6351 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6352 * possible for any higher priority task to appear. In that case we
6353 * must re-start the pick_next_entity() loop.
6365 * Account for a descheduled task:
6367 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6369 struct sched_entity *se = &prev->se;
6370 struct cfs_rq *cfs_rq;
6372 for_each_sched_entity(se) {
6373 cfs_rq = cfs_rq_of(se);
6374 put_prev_entity(cfs_rq, se);
6379 * sched_yield() is very simple
6381 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6383 static void yield_task_fair(struct rq *rq)
6385 struct task_struct *curr = rq->curr;
6386 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6387 struct sched_entity *se = &curr->se;
6390 * Are we the only task in the tree?
6392 if (unlikely(rq->nr_running == 1))
6395 clear_buddies(cfs_rq, se);
6397 if (curr->policy != SCHED_BATCH) {
6398 update_rq_clock(rq);
6400 * Update run-time statistics of the 'current'.
6402 update_curr(cfs_rq);
6404 * Tell update_rq_clock() that we've just updated,
6405 * so we don't do microscopic update in schedule()
6406 * and double the fastpath cost.
6408 rq_clock_skip_update(rq, true);
6414 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6416 struct sched_entity *se = &p->se;
6418 /* throttled hierarchies are not runnable */
6419 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6422 /* Tell the scheduler that we'd really like pse to run next. */
6425 yield_task_fair(rq);
6431 /**************************************************
6432 * Fair scheduling class load-balancing methods.
6436 * The purpose of load-balancing is to achieve the same basic fairness the
6437 * per-cpu scheduler provides, namely provide a proportional amount of compute
6438 * time to each task. This is expressed in the following equation:
6440 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6442 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6443 * W_i,0 is defined as:
6445 * W_i,0 = \Sum_j w_i,j (2)
6447 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6448 * is derived from the nice value as per prio_to_weight[].
6450 * The weight average is an exponential decay average of the instantaneous
6453 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6455 * C_i is the compute capacity of cpu i, typically it is the
6456 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6457 * can also include other factors [XXX].
6459 * To achieve this balance we define a measure of imbalance which follows
6460 * directly from (1):
6462 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6464 * We them move tasks around to minimize the imbalance. In the continuous
6465 * function space it is obvious this converges, in the discrete case we get
6466 * a few fun cases generally called infeasible weight scenarios.
6469 * - infeasible weights;
6470 * - local vs global optima in the discrete case. ]
6475 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6476 * for all i,j solution, we create a tree of cpus that follows the hardware
6477 * topology where each level pairs two lower groups (or better). This results
6478 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6479 * tree to only the first of the previous level and we decrease the frequency
6480 * of load-balance at each level inv. proportional to the number of cpus in
6486 * \Sum { --- * --- * 2^i } = O(n) (5)
6488 * `- size of each group
6489 * | | `- number of cpus doing load-balance
6491 * `- sum over all levels
6493 * Coupled with a limit on how many tasks we can migrate every balance pass,
6494 * this makes (5) the runtime complexity of the balancer.
6496 * An important property here is that each CPU is still (indirectly) connected
6497 * to every other cpu in at most O(log n) steps:
6499 * The adjacency matrix of the resulting graph is given by:
6502 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6505 * And you'll find that:
6507 * A^(log_2 n)_i,j != 0 for all i,j (7)
6509 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6510 * The task movement gives a factor of O(m), giving a convergence complexity
6513 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6518 * In order to avoid CPUs going idle while there's still work to do, new idle
6519 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6520 * tree itself instead of relying on other CPUs to bring it work.
6522 * This adds some complexity to both (5) and (8) but it reduces the total idle
6530 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6533 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6538 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6540 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6542 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6545 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6546 * rewrite all of this once again.]
6549 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6551 enum fbq_type { regular, remote, all };
6560 #define LBF_ALL_PINNED 0x01
6561 #define LBF_NEED_BREAK 0x02
6562 #define LBF_DST_PINNED 0x04
6563 #define LBF_SOME_PINNED 0x08
6566 struct sched_domain *sd;
6574 struct cpumask *dst_grpmask;
6576 enum cpu_idle_type idle;
6578 unsigned int src_grp_nr_running;
6579 /* The set of CPUs under consideration for load-balancing */
6580 struct cpumask *cpus;
6585 unsigned int loop_break;
6586 unsigned int loop_max;
6588 enum fbq_type fbq_type;
6589 enum group_type busiest_group_type;
6590 struct list_head tasks;
6594 * Is this task likely cache-hot:
6596 static int task_hot(struct task_struct *p, struct lb_env *env)
6600 lockdep_assert_held(&env->src_rq->lock);
6602 if (p->sched_class != &fair_sched_class)
6605 if (unlikely(p->policy == SCHED_IDLE))
6609 * Buddy candidates are cache hot:
6611 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6612 (&p->se == cfs_rq_of(&p->se)->next ||
6613 &p->se == cfs_rq_of(&p->se)->last))
6616 if (sysctl_sched_migration_cost == -1)
6618 if (sysctl_sched_migration_cost == 0)
6621 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6623 return delta < (s64)sysctl_sched_migration_cost;
6626 #ifdef CONFIG_NUMA_BALANCING
6628 * Returns 1, if task migration degrades locality
6629 * Returns 0, if task migration improves locality i.e migration preferred.
6630 * Returns -1, if task migration is not affected by locality.
6632 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6634 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6635 unsigned long src_faults, dst_faults;
6636 int src_nid, dst_nid;
6638 if (!static_branch_likely(&sched_numa_balancing))
6641 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6644 src_nid = cpu_to_node(env->src_cpu);
6645 dst_nid = cpu_to_node(env->dst_cpu);
6647 if (src_nid == dst_nid)
6650 /* Migrating away from the preferred node is always bad. */
6651 if (src_nid == p->numa_preferred_nid) {
6652 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6658 /* Encourage migration to the preferred node. */
6659 if (dst_nid == p->numa_preferred_nid)
6663 src_faults = group_faults(p, src_nid);
6664 dst_faults = group_faults(p, dst_nid);
6666 src_faults = task_faults(p, src_nid);
6667 dst_faults = task_faults(p, dst_nid);
6670 return dst_faults < src_faults;
6674 static inline int migrate_degrades_locality(struct task_struct *p,
6682 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6685 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6689 lockdep_assert_held(&env->src_rq->lock);
6692 * We do not migrate tasks that are:
6693 * 1) throttled_lb_pair, or
6694 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6695 * 3) running (obviously), or
6696 * 4) are cache-hot on their current CPU.
6698 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6701 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6704 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6706 env->flags |= LBF_SOME_PINNED;
6709 * Remember if this task can be migrated to any other cpu in
6710 * our sched_group. We may want to revisit it if we couldn't
6711 * meet load balance goals by pulling other tasks on src_cpu.
6713 * Also avoid computing new_dst_cpu if we have already computed
6714 * one in current iteration.
6716 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6719 /* Prevent to re-select dst_cpu via env's cpus */
6720 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6721 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6722 env->flags |= LBF_DST_PINNED;
6723 env->new_dst_cpu = cpu;
6731 /* Record that we found atleast one task that could run on dst_cpu */
6732 env->flags &= ~LBF_ALL_PINNED;
6734 if (task_running(env->src_rq, p)) {
6735 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6740 * Aggressive migration if:
6741 * 1) destination numa is preferred
6742 * 2) task is cache cold, or
6743 * 3) too many balance attempts have failed.
6745 tsk_cache_hot = migrate_degrades_locality(p, env);
6746 if (tsk_cache_hot == -1)
6747 tsk_cache_hot = task_hot(p, env);
6749 if (tsk_cache_hot <= 0 ||
6750 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6751 if (tsk_cache_hot == 1) {
6752 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6753 schedstat_inc(p, se.statistics.nr_forced_migrations);
6758 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6763 * detach_task() -- detach the task for the migration specified in env
6765 static void detach_task(struct task_struct *p, struct lb_env *env)
6767 lockdep_assert_held(&env->src_rq->lock);
6769 deactivate_task(env->src_rq, p, 0);
6770 p->on_rq = TASK_ON_RQ_MIGRATING;
6771 double_lock_balance(env->src_rq, env->dst_rq);
6772 set_task_cpu(p, env->dst_cpu);
6773 double_unlock_balance(env->src_rq, env->dst_rq);
6777 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6778 * part of active balancing operations within "domain".
6780 * Returns a task if successful and NULL otherwise.
6782 static struct task_struct *detach_one_task(struct lb_env *env)
6784 struct task_struct *p, *n;
6786 lockdep_assert_held(&env->src_rq->lock);
6788 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6789 if (!can_migrate_task(p, env))
6792 detach_task(p, env);
6795 * Right now, this is only the second place where
6796 * lb_gained[env->idle] is updated (other is detach_tasks)
6797 * so we can safely collect stats here rather than
6798 * inside detach_tasks().
6800 schedstat_inc(env->sd, lb_gained[env->idle]);
6806 static const unsigned int sched_nr_migrate_break = 32;
6809 * detach_tasks() -- tries to detach up to imbalance weighted load from
6810 * busiest_rq, as part of a balancing operation within domain "sd".
6812 * Returns number of detached tasks if successful and 0 otherwise.
6814 static int detach_tasks(struct lb_env *env)
6816 struct list_head *tasks = &env->src_rq->cfs_tasks;
6817 struct task_struct *p;
6821 lockdep_assert_held(&env->src_rq->lock);
6823 if (env->imbalance <= 0)
6826 while (!list_empty(tasks)) {
6828 * We don't want to steal all, otherwise we may be treated likewise,
6829 * which could at worst lead to a livelock crash.
6831 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6834 p = list_first_entry(tasks, struct task_struct, se.group_node);
6837 /* We've more or less seen every task there is, call it quits */
6838 if (env->loop > env->loop_max)
6841 /* take a breather every nr_migrate tasks */
6842 if (env->loop > env->loop_break) {
6843 env->loop_break += sched_nr_migrate_break;
6844 env->flags |= LBF_NEED_BREAK;
6848 if (!can_migrate_task(p, env))
6851 load = task_h_load(p);
6853 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6856 if ((load / 2) > env->imbalance)
6859 detach_task(p, env);
6860 list_add(&p->se.group_node, &env->tasks);
6863 env->imbalance -= load;
6865 #ifdef CONFIG_PREEMPT
6867 * NEWIDLE balancing is a source of latency, so preemptible
6868 * kernels will stop after the first task is detached to minimize
6869 * the critical section.
6871 if (env->idle == CPU_NEWLY_IDLE)
6876 * We only want to steal up to the prescribed amount of
6879 if (env->imbalance <= 0)
6884 list_move_tail(&p->se.group_node, tasks);
6888 * Right now, this is one of only two places we collect this stat
6889 * so we can safely collect detach_one_task() stats here rather
6890 * than inside detach_one_task().
6892 schedstat_add(env->sd, lb_gained[env->idle], detached);
6898 * attach_task() -- attach the task detached by detach_task() to its new rq.
6900 static void attach_task(struct rq *rq, struct task_struct *p)
6902 lockdep_assert_held(&rq->lock);
6904 BUG_ON(task_rq(p) != rq);
6905 p->on_rq = TASK_ON_RQ_QUEUED;
6906 activate_task(rq, p, 0);
6907 check_preempt_curr(rq, p, 0);
6911 * attach_one_task() -- attaches the task returned from detach_one_task() to
6914 static void attach_one_task(struct rq *rq, struct task_struct *p)
6916 raw_spin_lock(&rq->lock);
6919 * We want to potentially raise target_cpu's OPP.
6921 update_capacity_of(cpu_of(rq));
6922 raw_spin_unlock(&rq->lock);
6926 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6929 static void attach_tasks(struct lb_env *env)
6931 struct list_head *tasks = &env->tasks;
6932 struct task_struct *p;
6934 raw_spin_lock(&env->dst_rq->lock);
6936 while (!list_empty(tasks)) {
6937 p = list_first_entry(tasks, struct task_struct, se.group_node);
6938 list_del_init(&p->se.group_node);
6940 attach_task(env->dst_rq, p);
6944 * We want to potentially raise env.dst_cpu's OPP.
6946 update_capacity_of(env->dst_cpu);
6948 raw_spin_unlock(&env->dst_rq->lock);
6951 #ifdef CONFIG_FAIR_GROUP_SCHED
6952 static void update_blocked_averages(int cpu)
6954 struct rq *rq = cpu_rq(cpu);
6955 struct cfs_rq *cfs_rq;
6956 unsigned long flags;
6958 raw_spin_lock_irqsave(&rq->lock, flags);
6959 update_rq_clock(rq);
6962 * Iterates the task_group tree in a bottom up fashion, see
6963 * list_add_leaf_cfs_rq() for details.
6965 for_each_leaf_cfs_rq(rq, cfs_rq) {
6966 /* throttled entities do not contribute to load */
6967 if (throttled_hierarchy(cfs_rq))
6970 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
6972 update_tg_load_avg(cfs_rq, 0);
6974 raw_spin_unlock_irqrestore(&rq->lock, flags);
6978 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6979 * This needs to be done in a top-down fashion because the load of a child
6980 * group is a fraction of its parents load.
6982 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6984 struct rq *rq = rq_of(cfs_rq);
6985 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6986 unsigned long now = jiffies;
6989 if (cfs_rq->last_h_load_update == now)
6992 cfs_rq->h_load_next = NULL;
6993 for_each_sched_entity(se) {
6994 cfs_rq = cfs_rq_of(se);
6995 cfs_rq->h_load_next = se;
6996 if (cfs_rq->last_h_load_update == now)
7001 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7002 cfs_rq->last_h_load_update = now;
7005 while ((se = cfs_rq->h_load_next) != NULL) {
7006 load = cfs_rq->h_load;
7007 load = div64_ul(load * se->avg.load_avg,
7008 cfs_rq_load_avg(cfs_rq) + 1);
7009 cfs_rq = group_cfs_rq(se);
7010 cfs_rq->h_load = load;
7011 cfs_rq->last_h_load_update = now;
7015 static unsigned long task_h_load(struct task_struct *p)
7017 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7019 update_cfs_rq_h_load(cfs_rq);
7020 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7021 cfs_rq_load_avg(cfs_rq) + 1);
7024 static inline void update_blocked_averages(int cpu)
7026 struct rq *rq = cpu_rq(cpu);
7027 struct cfs_rq *cfs_rq = &rq->cfs;
7028 unsigned long flags;
7030 raw_spin_lock_irqsave(&rq->lock, flags);
7031 update_rq_clock(rq);
7032 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7033 raw_spin_unlock_irqrestore(&rq->lock, flags);
7036 static unsigned long task_h_load(struct task_struct *p)
7038 return p->se.avg.load_avg;
7042 /********** Helpers for find_busiest_group ************************/
7045 * sg_lb_stats - stats of a sched_group required for load_balancing
7047 struct sg_lb_stats {
7048 unsigned long avg_load; /*Avg load across the CPUs of the group */
7049 unsigned long group_load; /* Total load over the CPUs of the group */
7050 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7051 unsigned long load_per_task;
7052 unsigned long group_capacity;
7053 unsigned long group_util; /* Total utilization of the group */
7054 unsigned int sum_nr_running; /* Nr tasks running in the group */
7055 unsigned int idle_cpus;
7056 unsigned int group_weight;
7057 enum group_type group_type;
7058 int group_no_capacity;
7059 int group_misfit_task; /* A cpu has a task too big for its capacity */
7060 #ifdef CONFIG_NUMA_BALANCING
7061 unsigned int nr_numa_running;
7062 unsigned int nr_preferred_running;
7067 * sd_lb_stats - Structure to store the statistics of a sched_domain
7068 * during load balancing.
7070 struct sd_lb_stats {
7071 struct sched_group *busiest; /* Busiest group in this sd */
7072 struct sched_group *local; /* Local group in this sd */
7073 unsigned long total_load; /* Total load of all groups in sd */
7074 unsigned long total_capacity; /* Total capacity of all groups in sd */
7075 unsigned long avg_load; /* Average load across all groups in sd */
7077 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7078 struct sg_lb_stats local_stat; /* Statistics of the local group */
7081 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7084 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7085 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7086 * We must however clear busiest_stat::avg_load because
7087 * update_sd_pick_busiest() reads this before assignment.
7089 *sds = (struct sd_lb_stats){
7093 .total_capacity = 0UL,
7096 .sum_nr_running = 0,
7097 .group_type = group_other,
7103 * get_sd_load_idx - Obtain the load index for a given sched domain.
7104 * @sd: The sched_domain whose load_idx is to be obtained.
7105 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7107 * Return: The load index.
7109 static inline int get_sd_load_idx(struct sched_domain *sd,
7110 enum cpu_idle_type idle)
7116 load_idx = sd->busy_idx;
7119 case CPU_NEWLY_IDLE:
7120 load_idx = sd->newidle_idx;
7123 load_idx = sd->idle_idx;
7130 static unsigned long scale_rt_capacity(int cpu)
7132 struct rq *rq = cpu_rq(cpu);
7133 u64 total, used, age_stamp, avg;
7137 * Since we're reading these variables without serialization make sure
7138 * we read them once before doing sanity checks on them.
7140 age_stamp = READ_ONCE(rq->age_stamp);
7141 avg = READ_ONCE(rq->rt_avg);
7142 delta = __rq_clock_broken(rq) - age_stamp;
7144 if (unlikely(delta < 0))
7147 total = sched_avg_period() + delta;
7149 used = div_u64(avg, total);
7152 * deadline bandwidth is defined at system level so we must
7153 * weight this bandwidth with the max capacity of the system.
7154 * As a reminder, avg_bw is 20bits width and
7155 * scale_cpu_capacity is 10 bits width
7157 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7159 if (likely(used < SCHED_CAPACITY_SCALE))
7160 return SCHED_CAPACITY_SCALE - used;
7165 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7167 raw_spin_lock_init(&mcc->lock);
7172 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7174 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7175 struct sched_group *sdg = sd->groups;
7176 struct max_cpu_capacity *mcc;
7177 unsigned long max_capacity;
7179 unsigned long flags;
7181 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7183 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7185 raw_spin_lock_irqsave(&mcc->lock, flags);
7186 max_capacity = mcc->val;
7187 max_cap_cpu = mcc->cpu;
7189 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7190 (max_capacity < capacity)) {
7191 mcc->val = capacity;
7193 #ifdef CONFIG_SCHED_DEBUG
7194 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7195 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7200 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7202 skip_unlock: __attribute__ ((unused));
7203 capacity *= scale_rt_capacity(cpu);
7204 capacity >>= SCHED_CAPACITY_SHIFT;
7209 cpu_rq(cpu)->cpu_capacity = capacity;
7210 sdg->sgc->capacity = capacity;
7211 sdg->sgc->max_capacity = capacity;
7214 void update_group_capacity(struct sched_domain *sd, int cpu)
7216 struct sched_domain *child = sd->child;
7217 struct sched_group *group, *sdg = sd->groups;
7218 unsigned long capacity, max_capacity;
7219 unsigned long interval;
7221 interval = msecs_to_jiffies(sd->balance_interval);
7222 interval = clamp(interval, 1UL, max_load_balance_interval);
7223 sdg->sgc->next_update = jiffies + interval;
7226 update_cpu_capacity(sd, cpu);
7233 if (child->flags & SD_OVERLAP) {
7235 * SD_OVERLAP domains cannot assume that child groups
7236 * span the current group.
7239 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7240 struct sched_group_capacity *sgc;
7241 struct rq *rq = cpu_rq(cpu);
7244 * build_sched_domains() -> init_sched_groups_capacity()
7245 * gets here before we've attached the domains to the
7248 * Use capacity_of(), which is set irrespective of domains
7249 * in update_cpu_capacity().
7251 * This avoids capacity from being 0 and
7252 * causing divide-by-zero issues on boot.
7254 if (unlikely(!rq->sd)) {
7255 capacity += capacity_of(cpu);
7257 sgc = rq->sd->groups->sgc;
7258 capacity += sgc->capacity;
7261 max_capacity = max(capacity, max_capacity);
7265 * !SD_OVERLAP domains can assume that child groups
7266 * span the current group.
7269 group = child->groups;
7271 struct sched_group_capacity *sgc = group->sgc;
7273 capacity += sgc->capacity;
7274 max_capacity = max(sgc->max_capacity, max_capacity);
7275 group = group->next;
7276 } while (group != child->groups);
7279 sdg->sgc->capacity = capacity;
7280 sdg->sgc->max_capacity = max_capacity;
7284 * Check whether the capacity of the rq has been noticeably reduced by side
7285 * activity. The imbalance_pct is used for the threshold.
7286 * Return true is the capacity is reduced
7289 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7291 return ((rq->cpu_capacity * sd->imbalance_pct) <
7292 (rq->cpu_capacity_orig * 100));
7296 * Group imbalance indicates (and tries to solve) the problem where balancing
7297 * groups is inadequate due to tsk_cpus_allowed() constraints.
7299 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7300 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7303 * { 0 1 2 3 } { 4 5 6 7 }
7306 * If we were to balance group-wise we'd place two tasks in the first group and
7307 * two tasks in the second group. Clearly this is undesired as it will overload
7308 * cpu 3 and leave one of the cpus in the second group unused.
7310 * The current solution to this issue is detecting the skew in the first group
7311 * by noticing the lower domain failed to reach balance and had difficulty
7312 * moving tasks due to affinity constraints.
7314 * When this is so detected; this group becomes a candidate for busiest; see
7315 * update_sd_pick_busiest(). And calculate_imbalance() and
7316 * find_busiest_group() avoid some of the usual balance conditions to allow it
7317 * to create an effective group imbalance.
7319 * This is a somewhat tricky proposition since the next run might not find the
7320 * group imbalance and decide the groups need to be balanced again. A most
7321 * subtle and fragile situation.
7324 static inline int sg_imbalanced(struct sched_group *group)
7326 return group->sgc->imbalance;
7330 * group_has_capacity returns true if the group has spare capacity that could
7331 * be used by some tasks.
7332 * We consider that a group has spare capacity if the * number of task is
7333 * smaller than the number of CPUs or if the utilization is lower than the
7334 * available capacity for CFS tasks.
7335 * For the latter, we use a threshold to stabilize the state, to take into
7336 * account the variance of the tasks' load and to return true if the available
7337 * capacity in meaningful for the load balancer.
7338 * As an example, an available capacity of 1% can appear but it doesn't make
7339 * any benefit for the load balance.
7342 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7344 if (sgs->sum_nr_running < sgs->group_weight)
7347 if ((sgs->group_capacity * 100) >
7348 (sgs->group_util * env->sd->imbalance_pct))
7355 * group_is_overloaded returns true if the group has more tasks than it can
7357 * group_is_overloaded is not equals to !group_has_capacity because a group
7358 * with the exact right number of tasks, has no more spare capacity but is not
7359 * overloaded so both group_has_capacity and group_is_overloaded return
7363 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7365 if (sgs->sum_nr_running <= sgs->group_weight)
7368 if ((sgs->group_capacity * 100) <
7369 (sgs->group_util * env->sd->imbalance_pct))
7377 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7378 * per-cpu capacity than sched_group ref.
7381 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7383 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7384 ref->sgc->max_capacity;
7388 group_type group_classify(struct sched_group *group,
7389 struct sg_lb_stats *sgs)
7391 if (sgs->group_no_capacity)
7392 return group_overloaded;
7394 if (sg_imbalanced(group))
7395 return group_imbalanced;
7397 if (sgs->group_misfit_task)
7398 return group_misfit_task;
7404 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7405 * @env: The load balancing environment.
7406 * @group: sched_group whose statistics are to be updated.
7407 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7408 * @local_group: Does group contain this_cpu.
7409 * @sgs: variable to hold the statistics for this group.
7410 * @overload: Indicate more than one runnable task for any CPU.
7411 * @overutilized: Indicate overutilization for any CPU.
7413 static inline void update_sg_lb_stats(struct lb_env *env,
7414 struct sched_group *group, int load_idx,
7415 int local_group, struct sg_lb_stats *sgs,
7416 bool *overload, bool *overutilized)
7421 memset(sgs, 0, sizeof(*sgs));
7423 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7424 struct rq *rq = cpu_rq(i);
7426 /* Bias balancing toward cpus of our domain */
7428 load = target_load(i, load_idx);
7430 load = source_load(i, load_idx);
7432 sgs->group_load += load;
7433 sgs->group_util += cpu_util(i);
7434 sgs->sum_nr_running += rq->cfs.h_nr_running;
7436 nr_running = rq->nr_running;
7440 #ifdef CONFIG_NUMA_BALANCING
7441 sgs->nr_numa_running += rq->nr_numa_running;
7442 sgs->nr_preferred_running += rq->nr_preferred_running;
7444 sgs->sum_weighted_load += weighted_cpuload(i);
7446 * No need to call idle_cpu() if nr_running is not 0
7448 if (!nr_running && idle_cpu(i))
7451 if (cpu_overutilized(i)) {
7452 *overutilized = true;
7453 if (!sgs->group_misfit_task && rq->misfit_task)
7454 sgs->group_misfit_task = capacity_of(i);
7458 /* Adjust by relative CPU capacity of the group */
7459 sgs->group_capacity = group->sgc->capacity;
7460 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7462 if (sgs->sum_nr_running)
7463 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7465 sgs->group_weight = group->group_weight;
7467 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7468 sgs->group_type = group_classify(group, sgs);
7472 * update_sd_pick_busiest - return 1 on busiest group
7473 * @env: The load balancing environment.
7474 * @sds: sched_domain statistics
7475 * @sg: sched_group candidate to be checked for being the busiest
7476 * @sgs: sched_group statistics
7478 * Determine if @sg is a busier group than the previously selected
7481 * Return: %true if @sg is a busier group than the previously selected
7482 * busiest group. %false otherwise.
7484 static bool update_sd_pick_busiest(struct lb_env *env,
7485 struct sd_lb_stats *sds,
7486 struct sched_group *sg,
7487 struct sg_lb_stats *sgs)
7489 struct sg_lb_stats *busiest = &sds->busiest_stat;
7491 if (sgs->group_type > busiest->group_type)
7494 if (sgs->group_type < busiest->group_type)
7498 * Candidate sg doesn't face any serious load-balance problems
7499 * so don't pick it if the local sg is already filled up.
7501 if (sgs->group_type == group_other &&
7502 !group_has_capacity(env, &sds->local_stat))
7505 if (sgs->avg_load <= busiest->avg_load)
7509 * Candiate sg has no more than one task per cpu and has higher
7510 * per-cpu capacity. No reason to pull tasks to less capable cpus.
7512 if (sgs->sum_nr_running <= sgs->group_weight &&
7513 group_smaller_cpu_capacity(sds->local, sg))
7516 /* This is the busiest node in its class. */
7517 if (!(env->sd->flags & SD_ASYM_PACKING))
7521 * ASYM_PACKING needs to move all the work to the lowest
7522 * numbered CPUs in the group, therefore mark all groups
7523 * higher than ourself as busy.
7525 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7529 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7536 #ifdef CONFIG_NUMA_BALANCING
7537 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7539 if (sgs->sum_nr_running > sgs->nr_numa_running)
7541 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7546 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7548 if (rq->nr_running > rq->nr_numa_running)
7550 if (rq->nr_running > rq->nr_preferred_running)
7555 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7560 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7564 #endif /* CONFIG_NUMA_BALANCING */
7567 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7568 * @env: The load balancing environment.
7569 * @sds: variable to hold the statistics for this sched_domain.
7571 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7573 struct sched_domain *child = env->sd->child;
7574 struct sched_group *sg = env->sd->groups;
7575 struct sg_lb_stats tmp_sgs;
7576 int load_idx, prefer_sibling = 0;
7577 bool overload = false, overutilized = false;
7579 if (child && child->flags & SD_PREFER_SIBLING)
7582 load_idx = get_sd_load_idx(env->sd, env->idle);
7585 struct sg_lb_stats *sgs = &tmp_sgs;
7588 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7591 sgs = &sds->local_stat;
7593 if (env->idle != CPU_NEWLY_IDLE ||
7594 time_after_eq(jiffies, sg->sgc->next_update))
7595 update_group_capacity(env->sd, env->dst_cpu);
7598 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7599 &overload, &overutilized);
7605 * In case the child domain prefers tasks go to siblings
7606 * first, lower the sg capacity so that we'll try
7607 * and move all the excess tasks away. We lower the capacity
7608 * of a group only if the local group has the capacity to fit
7609 * these excess tasks. The extra check prevents the case where
7610 * you always pull from the heaviest group when it is already
7611 * under-utilized (possible with a large weight task outweighs
7612 * the tasks on the system).
7614 if (prefer_sibling && sds->local &&
7615 group_has_capacity(env, &sds->local_stat) &&
7616 (sgs->sum_nr_running > 1)) {
7617 sgs->group_no_capacity = 1;
7618 sgs->group_type = group_classify(sg, sgs);
7622 * Ignore task groups with misfit tasks if local group has no
7623 * capacity or if per-cpu capacity isn't higher.
7625 if (sgs->group_type == group_misfit_task &&
7626 (!group_has_capacity(env, &sds->local_stat) ||
7627 !group_smaller_cpu_capacity(sg, sds->local)))
7628 sgs->group_type = group_other;
7630 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7632 sds->busiest_stat = *sgs;
7636 /* Now, start updating sd_lb_stats */
7637 sds->total_load += sgs->group_load;
7638 sds->total_capacity += sgs->group_capacity;
7641 } while (sg != env->sd->groups);
7643 if (env->sd->flags & SD_NUMA)
7644 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7646 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7648 if (!env->sd->parent) {
7649 /* update overload indicator if we are at root domain */
7650 if (env->dst_rq->rd->overload != overload)
7651 env->dst_rq->rd->overload = overload;
7653 /* Update over-utilization (tipping point, U >= 0) indicator */
7654 if (env->dst_rq->rd->overutilized != overutilized) {
7655 env->dst_rq->rd->overutilized = overutilized;
7656 trace_sched_overutilized(overutilized);
7659 if (!env->dst_rq->rd->overutilized && overutilized) {
7660 env->dst_rq->rd->overutilized = true;
7661 trace_sched_overutilized(true);
7668 * check_asym_packing - Check to see if the group is packed into the
7671 * This is primarily intended to used at the sibling level. Some
7672 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7673 * case of POWER7, it can move to lower SMT modes only when higher
7674 * threads are idle. When in lower SMT modes, the threads will
7675 * perform better since they share less core resources. Hence when we
7676 * have idle threads, we want them to be the higher ones.
7678 * This packing function is run on idle threads. It checks to see if
7679 * the busiest CPU in this domain (core in the P7 case) has a higher
7680 * CPU number than the packing function is being run on. Here we are
7681 * assuming lower CPU number will be equivalent to lower a SMT thread
7684 * Return: 1 when packing is required and a task should be moved to
7685 * this CPU. The amount of the imbalance is returned in *imbalance.
7687 * @env: The load balancing environment.
7688 * @sds: Statistics of the sched_domain which is to be packed
7690 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7694 if (!(env->sd->flags & SD_ASYM_PACKING))
7700 busiest_cpu = group_first_cpu(sds->busiest);
7701 if (env->dst_cpu > busiest_cpu)
7704 env->imbalance = DIV_ROUND_CLOSEST(
7705 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7706 SCHED_CAPACITY_SCALE);
7712 * fix_small_imbalance - Calculate the minor imbalance that exists
7713 * amongst the groups of a sched_domain, during
7715 * @env: The load balancing environment.
7716 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7719 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7721 unsigned long tmp, capa_now = 0, capa_move = 0;
7722 unsigned int imbn = 2;
7723 unsigned long scaled_busy_load_per_task;
7724 struct sg_lb_stats *local, *busiest;
7726 local = &sds->local_stat;
7727 busiest = &sds->busiest_stat;
7729 if (!local->sum_nr_running)
7730 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7731 else if (busiest->load_per_task > local->load_per_task)
7734 scaled_busy_load_per_task =
7735 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7736 busiest->group_capacity;
7738 if (busiest->avg_load + scaled_busy_load_per_task >=
7739 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7740 env->imbalance = busiest->load_per_task;
7745 * OK, we don't have enough imbalance to justify moving tasks,
7746 * however we may be able to increase total CPU capacity used by
7750 capa_now += busiest->group_capacity *
7751 min(busiest->load_per_task, busiest->avg_load);
7752 capa_now += local->group_capacity *
7753 min(local->load_per_task, local->avg_load);
7754 capa_now /= SCHED_CAPACITY_SCALE;
7756 /* Amount of load we'd subtract */
7757 if (busiest->avg_load > scaled_busy_load_per_task) {
7758 capa_move += busiest->group_capacity *
7759 min(busiest->load_per_task,
7760 busiest->avg_load - scaled_busy_load_per_task);
7763 /* Amount of load we'd add */
7764 if (busiest->avg_load * busiest->group_capacity <
7765 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7766 tmp = (busiest->avg_load * busiest->group_capacity) /
7767 local->group_capacity;
7769 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7770 local->group_capacity;
7772 capa_move += local->group_capacity *
7773 min(local->load_per_task, local->avg_load + tmp);
7774 capa_move /= SCHED_CAPACITY_SCALE;
7776 /* Move if we gain throughput */
7777 if (capa_move > capa_now)
7778 env->imbalance = busiest->load_per_task;
7782 * calculate_imbalance - Calculate the amount of imbalance present within the
7783 * groups of a given sched_domain during load balance.
7784 * @env: load balance environment
7785 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7787 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7789 unsigned long max_pull, load_above_capacity = ~0UL;
7790 struct sg_lb_stats *local, *busiest;
7792 local = &sds->local_stat;
7793 busiest = &sds->busiest_stat;
7795 if (busiest->group_type == group_imbalanced) {
7797 * In the group_imb case we cannot rely on group-wide averages
7798 * to ensure cpu-load equilibrium, look at wider averages. XXX
7800 busiest->load_per_task =
7801 min(busiest->load_per_task, sds->avg_load);
7805 * In the presence of smp nice balancing, certain scenarios can have
7806 * max load less than avg load(as we skip the groups at or below
7807 * its cpu_capacity, while calculating max_load..)
7809 if (busiest->avg_load <= sds->avg_load ||
7810 local->avg_load >= sds->avg_load) {
7811 /* Misfitting tasks should be migrated in any case */
7812 if (busiest->group_type == group_misfit_task) {
7813 env->imbalance = busiest->group_misfit_task;
7818 * Busiest group is overloaded, local is not, use the spare
7819 * cycles to maximize throughput
7821 if (busiest->group_type == group_overloaded &&
7822 local->group_type <= group_misfit_task) {
7823 env->imbalance = busiest->load_per_task;
7828 return fix_small_imbalance(env, sds);
7832 * If there aren't any idle cpus, avoid creating some.
7834 if (busiest->group_type == group_overloaded &&
7835 local->group_type == group_overloaded) {
7836 load_above_capacity = busiest->sum_nr_running *
7838 if (load_above_capacity > busiest->group_capacity)
7839 load_above_capacity -= busiest->group_capacity;
7841 load_above_capacity = ~0UL;
7845 * We're trying to get all the cpus to the average_load, so we don't
7846 * want to push ourselves above the average load, nor do we wish to
7847 * reduce the max loaded cpu below the average load. At the same time,
7848 * we also don't want to reduce the group load below the group capacity
7849 * (so that we can implement power-savings policies etc). Thus we look
7850 * for the minimum possible imbalance.
7852 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7854 /* How much load to actually move to equalise the imbalance */
7855 env->imbalance = min(
7856 max_pull * busiest->group_capacity,
7857 (sds->avg_load - local->avg_load) * local->group_capacity
7858 ) / SCHED_CAPACITY_SCALE;
7860 /* Boost imbalance to allow misfit task to be balanced. */
7861 if (busiest->group_type == group_misfit_task)
7862 env->imbalance = max_t(long, env->imbalance,
7863 busiest->group_misfit_task);
7866 * if *imbalance is less than the average load per runnable task
7867 * there is no guarantee that any tasks will be moved so we'll have
7868 * a think about bumping its value to force at least one task to be
7871 if (env->imbalance < busiest->load_per_task)
7872 return fix_small_imbalance(env, sds);
7875 /******* find_busiest_group() helpers end here *********************/
7878 * find_busiest_group - Returns the busiest group within the sched_domain
7879 * if there is an imbalance. If there isn't an imbalance, and
7880 * the user has opted for power-savings, it returns a group whose
7881 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7882 * such a group exists.
7884 * Also calculates the amount of weighted load which should be moved
7885 * to restore balance.
7887 * @env: The load balancing environment.
7889 * Return: - The busiest group if imbalance exists.
7890 * - If no imbalance and user has opted for power-savings balance,
7891 * return the least loaded group whose CPUs can be
7892 * put to idle by rebalancing its tasks onto our group.
7894 static struct sched_group *find_busiest_group(struct lb_env *env)
7896 struct sg_lb_stats *local, *busiest;
7897 struct sd_lb_stats sds;
7899 init_sd_lb_stats(&sds);
7902 * Compute the various statistics relavent for load balancing at
7905 update_sd_lb_stats(env, &sds);
7907 if (energy_aware() && !env->dst_rq->rd->overutilized)
7910 local = &sds.local_stat;
7911 busiest = &sds.busiest_stat;
7913 /* ASYM feature bypasses nice load balance check */
7914 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7915 check_asym_packing(env, &sds))
7918 /* There is no busy sibling group to pull tasks from */
7919 if (!sds.busiest || busiest->sum_nr_running == 0)
7922 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7923 / sds.total_capacity;
7926 * If the busiest group is imbalanced the below checks don't
7927 * work because they assume all things are equal, which typically
7928 * isn't true due to cpus_allowed constraints and the like.
7930 if (busiest->group_type == group_imbalanced)
7933 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7934 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7935 busiest->group_no_capacity)
7938 /* Misfitting tasks should be dealt with regardless of the avg load */
7939 if (busiest->group_type == group_misfit_task) {
7944 * If the local group is busier than the selected busiest group
7945 * don't try and pull any tasks.
7947 if (local->avg_load >= busiest->avg_load)
7951 * Don't pull any tasks if this group is already above the domain
7954 if (local->avg_load >= sds.avg_load)
7957 if (env->idle == CPU_IDLE) {
7959 * This cpu is idle. If the busiest group is not overloaded
7960 * and there is no imbalance between this and busiest group
7961 * wrt idle cpus, it is balanced. The imbalance becomes
7962 * significant if the diff is greater than 1 otherwise we
7963 * might end up to just move the imbalance on another group
7965 if ((busiest->group_type != group_overloaded) &&
7966 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7967 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7971 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7972 * imbalance_pct to be conservative.
7974 if (100 * busiest->avg_load <=
7975 env->sd->imbalance_pct * local->avg_load)
7980 env->busiest_group_type = busiest->group_type;
7981 /* Looks like there is an imbalance. Compute it */
7982 calculate_imbalance(env, &sds);
7991 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7993 static struct rq *find_busiest_queue(struct lb_env *env,
7994 struct sched_group *group)
7996 struct rq *busiest = NULL, *rq;
7997 unsigned long busiest_load = 0, busiest_capacity = 1;
8000 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8001 unsigned long capacity, wl;
8005 rt = fbq_classify_rq(rq);
8008 * We classify groups/runqueues into three groups:
8009 * - regular: there are !numa tasks
8010 * - remote: there are numa tasks that run on the 'wrong' node
8011 * - all: there is no distinction
8013 * In order to avoid migrating ideally placed numa tasks,
8014 * ignore those when there's better options.
8016 * If we ignore the actual busiest queue to migrate another
8017 * task, the next balance pass can still reduce the busiest
8018 * queue by moving tasks around inside the node.
8020 * If we cannot move enough load due to this classification
8021 * the next pass will adjust the group classification and
8022 * allow migration of more tasks.
8024 * Both cases only affect the total convergence complexity.
8026 if (rt > env->fbq_type)
8029 capacity = capacity_of(i);
8031 wl = weighted_cpuload(i);
8034 * When comparing with imbalance, use weighted_cpuload()
8035 * which is not scaled with the cpu capacity.
8038 if (rq->nr_running == 1 && wl > env->imbalance &&
8039 !check_cpu_capacity(rq, env->sd) &&
8040 env->busiest_group_type != group_misfit_task)
8044 * For the load comparisons with the other cpu's, consider
8045 * the weighted_cpuload() scaled with the cpu capacity, so
8046 * that the load can be moved away from the cpu that is
8047 * potentially running at a lower capacity.
8049 * Thus we're looking for max(wl_i / capacity_i), crosswise
8050 * multiplication to rid ourselves of the division works out
8051 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8052 * our previous maximum.
8054 if (wl * busiest_capacity > busiest_load * capacity) {
8056 busiest_capacity = capacity;
8065 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8066 * so long as it is large enough.
8068 #define MAX_PINNED_INTERVAL 512
8070 /* Working cpumask for load_balance and load_balance_newidle. */
8071 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8073 static int need_active_balance(struct lb_env *env)
8075 struct sched_domain *sd = env->sd;
8077 if (env->idle == CPU_NEWLY_IDLE) {
8080 * ASYM_PACKING needs to force migrate tasks from busy but
8081 * higher numbered CPUs in order to pack all tasks in the
8082 * lowest numbered CPUs.
8084 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8089 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8090 * It's worth migrating the task if the src_cpu's capacity is reduced
8091 * because of other sched_class or IRQs if more capacity stays
8092 * available on dst_cpu.
8094 if ((env->idle != CPU_NOT_IDLE) &&
8095 (env->src_rq->cfs.h_nr_running == 1)) {
8096 if ((check_cpu_capacity(env->src_rq, sd)) &&
8097 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8101 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8102 env->src_rq->cfs.h_nr_running == 1 &&
8103 cpu_overutilized(env->src_cpu) &&
8104 !cpu_overutilized(env->dst_cpu)) {
8108 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8111 static int active_load_balance_cpu_stop(void *data);
8113 static int should_we_balance(struct lb_env *env)
8115 struct sched_group *sg = env->sd->groups;
8116 struct cpumask *sg_cpus, *sg_mask;
8117 int cpu, balance_cpu = -1;
8120 * In the newly idle case, we will allow all the cpu's
8121 * to do the newly idle load balance.
8123 if (env->idle == CPU_NEWLY_IDLE)
8126 sg_cpus = sched_group_cpus(sg);
8127 sg_mask = sched_group_mask(sg);
8128 /* Try to find first idle cpu */
8129 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8130 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8137 if (balance_cpu == -1)
8138 balance_cpu = group_balance_cpu(sg);
8141 * First idle cpu or the first cpu(busiest) in this sched group
8142 * is eligible for doing load balancing at this and above domains.
8144 return balance_cpu == env->dst_cpu;
8148 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8149 * tasks if there is an imbalance.
8151 static int load_balance(int this_cpu, struct rq *this_rq,
8152 struct sched_domain *sd, enum cpu_idle_type idle,
8153 int *continue_balancing)
8155 int ld_moved, cur_ld_moved, active_balance = 0;
8156 struct sched_domain *sd_parent = sd->parent;
8157 struct sched_group *group;
8159 unsigned long flags;
8160 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8162 struct lb_env env = {
8164 .dst_cpu = this_cpu,
8166 .dst_grpmask = sched_group_cpus(sd->groups),
8168 .loop_break = sched_nr_migrate_break,
8171 .tasks = LIST_HEAD_INIT(env.tasks),
8175 * For NEWLY_IDLE load_balancing, we don't need to consider
8176 * other cpus in our group
8178 if (idle == CPU_NEWLY_IDLE)
8179 env.dst_grpmask = NULL;
8181 cpumask_copy(cpus, cpu_active_mask);
8183 schedstat_inc(sd, lb_count[idle]);
8186 if (!should_we_balance(&env)) {
8187 *continue_balancing = 0;
8191 group = find_busiest_group(&env);
8193 schedstat_inc(sd, lb_nobusyg[idle]);
8197 busiest = find_busiest_queue(&env, group);
8199 schedstat_inc(sd, lb_nobusyq[idle]);
8203 BUG_ON(busiest == env.dst_rq);
8205 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8207 env.src_cpu = busiest->cpu;
8208 env.src_rq = busiest;
8211 if (busiest->nr_running > 1) {
8213 * Attempt to move tasks. If find_busiest_group has found
8214 * an imbalance but busiest->nr_running <= 1, the group is
8215 * still unbalanced. ld_moved simply stays zero, so it is
8216 * correctly treated as an imbalance.
8218 env.flags |= LBF_ALL_PINNED;
8219 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8222 raw_spin_lock_irqsave(&busiest->lock, flags);
8225 * cur_ld_moved - load moved in current iteration
8226 * ld_moved - cumulative load moved across iterations
8228 cur_ld_moved = detach_tasks(&env);
8230 * We want to potentially lower env.src_cpu's OPP.
8233 update_capacity_of(env.src_cpu);
8236 * We've detached some tasks from busiest_rq. Every
8237 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8238 * unlock busiest->lock, and we are able to be sure
8239 * that nobody can manipulate the tasks in parallel.
8240 * See task_rq_lock() family for the details.
8243 raw_spin_unlock(&busiest->lock);
8247 ld_moved += cur_ld_moved;
8250 local_irq_restore(flags);
8252 if (env.flags & LBF_NEED_BREAK) {
8253 env.flags &= ~LBF_NEED_BREAK;
8258 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8259 * us and move them to an alternate dst_cpu in our sched_group
8260 * where they can run. The upper limit on how many times we
8261 * iterate on same src_cpu is dependent on number of cpus in our
8264 * This changes load balance semantics a bit on who can move
8265 * load to a given_cpu. In addition to the given_cpu itself
8266 * (or a ilb_cpu acting on its behalf where given_cpu is
8267 * nohz-idle), we now have balance_cpu in a position to move
8268 * load to given_cpu. In rare situations, this may cause
8269 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8270 * _independently_ and at _same_ time to move some load to
8271 * given_cpu) causing exceess load to be moved to given_cpu.
8272 * This however should not happen so much in practice and
8273 * moreover subsequent load balance cycles should correct the
8274 * excess load moved.
8276 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8278 /* Prevent to re-select dst_cpu via env's cpus */
8279 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8281 env.dst_rq = cpu_rq(env.new_dst_cpu);
8282 env.dst_cpu = env.new_dst_cpu;
8283 env.flags &= ~LBF_DST_PINNED;
8285 env.loop_break = sched_nr_migrate_break;
8288 * Go back to "more_balance" rather than "redo" since we
8289 * need to continue with same src_cpu.
8295 * We failed to reach balance because of affinity.
8298 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8300 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8301 *group_imbalance = 1;
8304 /* All tasks on this runqueue were pinned by CPU affinity */
8305 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8306 cpumask_clear_cpu(cpu_of(busiest), cpus);
8307 if (!cpumask_empty(cpus)) {
8309 env.loop_break = sched_nr_migrate_break;
8312 goto out_all_pinned;
8317 schedstat_inc(sd, lb_failed[idle]);
8319 * Increment the failure counter only on periodic balance.
8320 * We do not want newidle balance, which can be very
8321 * frequent, pollute the failure counter causing
8322 * excessive cache_hot migrations and active balances.
8324 if (idle != CPU_NEWLY_IDLE)
8325 if (env.src_grp_nr_running > 1)
8326 sd->nr_balance_failed++;
8328 if (need_active_balance(&env)) {
8329 raw_spin_lock_irqsave(&busiest->lock, flags);
8331 /* don't kick the active_load_balance_cpu_stop,
8332 * if the curr task on busiest cpu can't be
8335 if (!cpumask_test_cpu(this_cpu,
8336 tsk_cpus_allowed(busiest->curr))) {
8337 raw_spin_unlock_irqrestore(&busiest->lock,
8339 env.flags |= LBF_ALL_PINNED;
8340 goto out_one_pinned;
8344 * ->active_balance synchronizes accesses to
8345 * ->active_balance_work. Once set, it's cleared
8346 * only after active load balance is finished.
8348 if (!busiest->active_balance) {
8349 busiest->active_balance = 1;
8350 busiest->push_cpu = this_cpu;
8353 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8355 if (active_balance) {
8356 stop_one_cpu_nowait(cpu_of(busiest),
8357 active_load_balance_cpu_stop, busiest,
8358 &busiest->active_balance_work);
8362 * We've kicked active balancing, reset the failure
8365 sd->nr_balance_failed = sd->cache_nice_tries+1;
8368 sd->nr_balance_failed = 0;
8370 if (likely(!active_balance)) {
8371 /* We were unbalanced, so reset the balancing interval */
8372 sd->balance_interval = sd->min_interval;
8375 * If we've begun active balancing, start to back off. This
8376 * case may not be covered by the all_pinned logic if there
8377 * is only 1 task on the busy runqueue (because we don't call
8380 if (sd->balance_interval < sd->max_interval)
8381 sd->balance_interval *= 2;
8388 * We reach balance although we may have faced some affinity
8389 * constraints. Clear the imbalance flag if it was set.
8392 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8394 if (*group_imbalance)
8395 *group_imbalance = 0;
8400 * We reach balance because all tasks are pinned at this level so
8401 * we can't migrate them. Let the imbalance flag set so parent level
8402 * can try to migrate them.
8404 schedstat_inc(sd, lb_balanced[idle]);
8406 sd->nr_balance_failed = 0;
8409 /* tune up the balancing interval */
8410 if (((env.flags & LBF_ALL_PINNED) &&
8411 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8412 (sd->balance_interval < sd->max_interval))
8413 sd->balance_interval *= 2;
8420 static inline unsigned long
8421 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8423 unsigned long interval = sd->balance_interval;
8426 interval *= sd->busy_factor;
8428 /* scale ms to jiffies */
8429 interval = msecs_to_jiffies(interval);
8430 interval = clamp(interval, 1UL, max_load_balance_interval);
8436 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8438 unsigned long interval, next;
8440 interval = get_sd_balance_interval(sd, cpu_busy);
8441 next = sd->last_balance + interval;
8443 if (time_after(*next_balance, next))
8444 *next_balance = next;
8448 * idle_balance is called by schedule() if this_cpu is about to become
8449 * idle. Attempts to pull tasks from other CPUs.
8451 static int idle_balance(struct rq *this_rq)
8453 unsigned long next_balance = jiffies + HZ;
8454 int this_cpu = this_rq->cpu;
8455 struct sched_domain *sd;
8456 int pulled_task = 0;
8458 long removed_util=0;
8460 idle_enter_fair(this_rq);
8463 * We must set idle_stamp _before_ calling idle_balance(), such that we
8464 * measure the duration of idle_balance() as idle time.
8466 this_rq->idle_stamp = rq_clock(this_rq);
8468 if (!energy_aware() &&
8469 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8470 !this_rq->rd->overload)) {
8472 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8474 update_next_balance(sd, 0, &next_balance);
8480 raw_spin_unlock(&this_rq->lock);
8483 * If removed_util_avg is !0 we most probably migrated some task away
8484 * from this_cpu. In this case we might be willing to trigger an OPP
8485 * update, but we want to do so if we don't find anybody else to pull
8486 * here (we will trigger an OPP update with the pulled task's enqueue
8489 * Record removed_util before calling update_blocked_averages, and use
8490 * it below (before returning) to see if an OPP update is required.
8492 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8493 update_blocked_averages(this_cpu);
8495 for_each_domain(this_cpu, sd) {
8496 int continue_balancing = 1;
8497 u64 t0, domain_cost;
8499 if (!(sd->flags & SD_LOAD_BALANCE))
8502 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8503 update_next_balance(sd, 0, &next_balance);
8507 if (sd->flags & SD_BALANCE_NEWIDLE) {
8508 t0 = sched_clock_cpu(this_cpu);
8510 pulled_task = load_balance(this_cpu, this_rq,
8512 &continue_balancing);
8514 domain_cost = sched_clock_cpu(this_cpu) - t0;
8515 if (domain_cost > sd->max_newidle_lb_cost)
8516 sd->max_newidle_lb_cost = domain_cost;
8518 curr_cost += domain_cost;
8521 update_next_balance(sd, 0, &next_balance);
8524 * Stop searching for tasks to pull if there are
8525 * now runnable tasks on this rq.
8527 if (pulled_task || this_rq->nr_running > 0)
8532 raw_spin_lock(&this_rq->lock);
8534 if (curr_cost > this_rq->max_idle_balance_cost)
8535 this_rq->max_idle_balance_cost = curr_cost;
8538 * While browsing the domains, we released the rq lock, a task could
8539 * have been enqueued in the meantime. Since we're not going idle,
8540 * pretend we pulled a task.
8542 if (this_rq->cfs.h_nr_running && !pulled_task)
8546 /* Move the next balance forward */
8547 if (time_after(this_rq->next_balance, next_balance))
8548 this_rq->next_balance = next_balance;
8550 /* Is there a task of a high priority class? */
8551 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8555 idle_exit_fair(this_rq);
8556 this_rq->idle_stamp = 0;
8557 } else if (removed_util) {
8559 * No task pulled and someone has been migrated away.
8560 * Good case to trigger an OPP update.
8562 update_capacity_of(this_cpu);
8569 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8570 * running tasks off the busiest CPU onto idle CPUs. It requires at
8571 * least 1 task to be running on each physical CPU where possible, and
8572 * avoids physical / logical imbalances.
8574 static int active_load_balance_cpu_stop(void *data)
8576 struct rq *busiest_rq = data;
8577 int busiest_cpu = cpu_of(busiest_rq);
8578 int target_cpu = busiest_rq->push_cpu;
8579 struct rq *target_rq = cpu_rq(target_cpu);
8580 struct sched_domain *sd;
8581 struct task_struct *p = NULL;
8583 raw_spin_lock_irq(&busiest_rq->lock);
8585 /* make sure the requested cpu hasn't gone down in the meantime */
8586 if (unlikely(busiest_cpu != smp_processor_id() ||
8587 !busiest_rq->active_balance))
8590 /* Is there any task to move? */
8591 if (busiest_rq->nr_running <= 1)
8595 * This condition is "impossible", if it occurs
8596 * we need to fix it. Originally reported by
8597 * Bjorn Helgaas on a 128-cpu setup.
8599 BUG_ON(busiest_rq == target_rq);
8601 /* Search for an sd spanning us and the target CPU. */
8603 for_each_domain(target_cpu, sd) {
8604 if ((sd->flags & SD_LOAD_BALANCE) &&
8605 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8610 struct lb_env env = {
8612 .dst_cpu = target_cpu,
8613 .dst_rq = target_rq,
8614 .src_cpu = busiest_rq->cpu,
8615 .src_rq = busiest_rq,
8619 schedstat_inc(sd, alb_count);
8621 p = detach_one_task(&env);
8623 schedstat_inc(sd, alb_pushed);
8625 * We want to potentially lower env.src_cpu's OPP.
8627 update_capacity_of(env.src_cpu);
8630 schedstat_inc(sd, alb_failed);
8634 busiest_rq->active_balance = 0;
8635 raw_spin_unlock(&busiest_rq->lock);
8638 attach_one_task(target_rq, p);
8645 static inline int on_null_domain(struct rq *rq)
8647 return unlikely(!rcu_dereference_sched(rq->sd));
8650 #ifdef CONFIG_NO_HZ_COMMON
8652 * idle load balancing details
8653 * - When one of the busy CPUs notice that there may be an idle rebalancing
8654 * needed, they will kick the idle load balancer, which then does idle
8655 * load balancing for all the idle CPUs.
8658 cpumask_var_t idle_cpus_mask;
8660 unsigned long next_balance; /* in jiffy units */
8661 } nohz ____cacheline_aligned;
8663 static inline int find_new_ilb(void)
8665 int ilb = cpumask_first(nohz.idle_cpus_mask);
8667 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8674 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8675 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8676 * CPU (if there is one).
8678 static void nohz_balancer_kick(void)
8682 nohz.next_balance++;
8684 ilb_cpu = find_new_ilb();
8686 if (ilb_cpu >= nr_cpu_ids)
8689 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8692 * Use smp_send_reschedule() instead of resched_cpu().
8693 * This way we generate a sched IPI on the target cpu which
8694 * is idle. And the softirq performing nohz idle load balance
8695 * will be run before returning from the IPI.
8697 smp_send_reschedule(ilb_cpu);
8701 static inline void nohz_balance_exit_idle(int cpu)
8703 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8705 * Completely isolated CPUs don't ever set, so we must test.
8707 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8708 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8709 atomic_dec(&nohz.nr_cpus);
8711 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8715 static inline void set_cpu_sd_state_busy(void)
8717 struct sched_domain *sd;
8718 int cpu = smp_processor_id();
8721 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8723 if (!sd || !sd->nohz_idle)
8727 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8732 void set_cpu_sd_state_idle(void)
8734 struct sched_domain *sd;
8735 int cpu = smp_processor_id();
8738 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8740 if (!sd || sd->nohz_idle)
8744 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8750 * This routine will record that the cpu is going idle with tick stopped.
8751 * This info will be used in performing idle load balancing in the future.
8753 void nohz_balance_enter_idle(int cpu)
8756 * If this cpu is going down, then nothing needs to be done.
8758 if (!cpu_active(cpu))
8761 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8765 * If we're a completely isolated CPU, we don't play.
8767 if (on_null_domain(cpu_rq(cpu)))
8770 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8771 atomic_inc(&nohz.nr_cpus);
8772 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8775 static int sched_ilb_notifier(struct notifier_block *nfb,
8776 unsigned long action, void *hcpu)
8778 switch (action & ~CPU_TASKS_FROZEN) {
8780 nohz_balance_exit_idle(smp_processor_id());
8788 static DEFINE_SPINLOCK(balancing);
8791 * Scale the max load_balance interval with the number of CPUs in the system.
8792 * This trades load-balance latency on larger machines for less cross talk.
8794 void update_max_interval(void)
8796 max_load_balance_interval = HZ*num_online_cpus()/10;
8800 * It checks each scheduling domain to see if it is due to be balanced,
8801 * and initiates a balancing operation if so.
8803 * Balancing parameters are set up in init_sched_domains.
8805 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8807 int continue_balancing = 1;
8809 unsigned long interval;
8810 struct sched_domain *sd;
8811 /* Earliest time when we have to do rebalance again */
8812 unsigned long next_balance = jiffies + 60*HZ;
8813 int update_next_balance = 0;
8814 int need_serialize, need_decay = 0;
8817 update_blocked_averages(cpu);
8820 for_each_domain(cpu, sd) {
8822 * Decay the newidle max times here because this is a regular
8823 * visit to all the domains. Decay ~1% per second.
8825 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8826 sd->max_newidle_lb_cost =
8827 (sd->max_newidle_lb_cost * 253) / 256;
8828 sd->next_decay_max_lb_cost = jiffies + HZ;
8831 max_cost += sd->max_newidle_lb_cost;
8833 if (!(sd->flags & SD_LOAD_BALANCE))
8837 * Stop the load balance at this level. There is another
8838 * CPU in our sched group which is doing load balancing more
8841 if (!continue_balancing) {
8847 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8849 need_serialize = sd->flags & SD_SERIALIZE;
8850 if (need_serialize) {
8851 if (!spin_trylock(&balancing))
8855 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8856 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8858 * The LBF_DST_PINNED logic could have changed
8859 * env->dst_cpu, so we can't know our idle
8860 * state even if we migrated tasks. Update it.
8862 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8864 sd->last_balance = jiffies;
8865 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8868 spin_unlock(&balancing);
8870 if (time_after(next_balance, sd->last_balance + interval)) {
8871 next_balance = sd->last_balance + interval;
8872 update_next_balance = 1;
8877 * Ensure the rq-wide value also decays but keep it at a
8878 * reasonable floor to avoid funnies with rq->avg_idle.
8880 rq->max_idle_balance_cost =
8881 max((u64)sysctl_sched_migration_cost, max_cost);
8886 * next_balance will be updated only when there is a need.
8887 * When the cpu is attached to null domain for ex, it will not be
8890 if (likely(update_next_balance)) {
8891 rq->next_balance = next_balance;
8893 #ifdef CONFIG_NO_HZ_COMMON
8895 * If this CPU has been elected to perform the nohz idle
8896 * balance. Other idle CPUs have already rebalanced with
8897 * nohz_idle_balance() and nohz.next_balance has been
8898 * updated accordingly. This CPU is now running the idle load
8899 * balance for itself and we need to update the
8900 * nohz.next_balance accordingly.
8902 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8903 nohz.next_balance = rq->next_balance;
8908 #ifdef CONFIG_NO_HZ_COMMON
8910 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8911 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8913 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8915 int this_cpu = this_rq->cpu;
8918 /* Earliest time when we have to do rebalance again */
8919 unsigned long next_balance = jiffies + 60*HZ;
8920 int update_next_balance = 0;
8922 if (idle != CPU_IDLE ||
8923 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8926 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8927 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8931 * If this cpu gets work to do, stop the load balancing
8932 * work being done for other cpus. Next load
8933 * balancing owner will pick it up.
8938 rq = cpu_rq(balance_cpu);
8941 * If time for next balance is due,
8944 if (time_after_eq(jiffies, rq->next_balance)) {
8945 raw_spin_lock_irq(&rq->lock);
8946 update_rq_clock(rq);
8947 update_idle_cpu_load(rq);
8948 raw_spin_unlock_irq(&rq->lock);
8949 rebalance_domains(rq, CPU_IDLE);
8952 if (time_after(next_balance, rq->next_balance)) {
8953 next_balance = rq->next_balance;
8954 update_next_balance = 1;
8959 * next_balance will be updated only when there is a need.
8960 * When the CPU is attached to null domain for ex, it will not be
8963 if (likely(update_next_balance))
8964 nohz.next_balance = next_balance;
8966 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8970 * Current heuristic for kicking the idle load balancer in the presence
8971 * of an idle cpu in the system.
8972 * - This rq has more than one task.
8973 * - This rq has at least one CFS task and the capacity of the CPU is
8974 * significantly reduced because of RT tasks or IRQs.
8975 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8976 * multiple busy cpu.
8977 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8978 * domain span are idle.
8980 static inline bool nohz_kick_needed(struct rq *rq)
8982 unsigned long now = jiffies;
8983 struct sched_domain *sd;
8984 struct sched_group_capacity *sgc;
8985 int nr_busy, cpu = rq->cpu;
8988 if (unlikely(rq->idle_balance))
8992 * We may be recently in ticked or tickless idle mode. At the first
8993 * busy tick after returning from idle, we will update the busy stats.
8995 set_cpu_sd_state_busy();
8996 nohz_balance_exit_idle(cpu);
8999 * None are in tickless mode and hence no need for NOHZ idle load
9002 if (likely(!atomic_read(&nohz.nr_cpus)))
9005 if (time_before(now, nohz.next_balance))
9008 if (rq->nr_running >= 2 &&
9009 (!energy_aware() || cpu_overutilized(cpu)))
9013 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9014 if (sd && !energy_aware()) {
9015 sgc = sd->groups->sgc;
9016 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9025 sd = rcu_dereference(rq->sd);
9027 if ((rq->cfs.h_nr_running >= 1) &&
9028 check_cpu_capacity(rq, sd)) {
9034 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9035 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9036 sched_domain_span(sd)) < cpu)) {
9046 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9050 * run_rebalance_domains is triggered when needed from the scheduler tick.
9051 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9053 static void run_rebalance_domains(struct softirq_action *h)
9055 struct rq *this_rq = this_rq();
9056 enum cpu_idle_type idle = this_rq->idle_balance ?
9057 CPU_IDLE : CPU_NOT_IDLE;
9060 * If this cpu has a pending nohz_balance_kick, then do the
9061 * balancing on behalf of the other idle cpus whose ticks are
9062 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9063 * give the idle cpus a chance to load balance. Else we may
9064 * load balance only within the local sched_domain hierarchy
9065 * and abort nohz_idle_balance altogether if we pull some load.
9067 nohz_idle_balance(this_rq, idle);
9068 rebalance_domains(this_rq, idle);
9072 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9074 void trigger_load_balance(struct rq *rq)
9076 /* Don't need to rebalance while attached to NULL domain */
9077 if (unlikely(on_null_domain(rq)))
9080 if (time_after_eq(jiffies, rq->next_balance))
9081 raise_softirq(SCHED_SOFTIRQ);
9082 #ifdef CONFIG_NO_HZ_COMMON
9083 if (nohz_kick_needed(rq))
9084 nohz_balancer_kick();
9088 static void rq_online_fair(struct rq *rq)
9092 update_runtime_enabled(rq);
9095 static void rq_offline_fair(struct rq *rq)
9099 /* Ensure any throttled groups are reachable by pick_next_task */
9100 unthrottle_offline_cfs_rqs(rq);
9103 #endif /* CONFIG_SMP */
9106 * scheduler tick hitting a task of our scheduling class:
9108 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9110 struct cfs_rq *cfs_rq;
9111 struct sched_entity *se = &curr->se;
9113 for_each_sched_entity(se) {
9114 cfs_rq = cfs_rq_of(se);
9115 entity_tick(cfs_rq, se, queued);
9118 if (static_branch_unlikely(&sched_numa_balancing))
9119 task_tick_numa(rq, curr);
9122 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9123 rq->rd->overutilized = true;
9124 trace_sched_overutilized(true);
9127 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9133 * called on fork with the child task as argument from the parent's context
9134 * - child not yet on the tasklist
9135 * - preemption disabled
9137 static void task_fork_fair(struct task_struct *p)
9139 struct cfs_rq *cfs_rq;
9140 struct sched_entity *se = &p->se, *curr;
9141 int this_cpu = smp_processor_id();
9142 struct rq *rq = this_rq();
9143 unsigned long flags;
9145 raw_spin_lock_irqsave(&rq->lock, flags);
9147 update_rq_clock(rq);
9149 cfs_rq = task_cfs_rq(current);
9150 curr = cfs_rq->curr;
9153 * Not only the cpu but also the task_group of the parent might have
9154 * been changed after parent->se.parent,cfs_rq were copied to
9155 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9156 * of child point to valid ones.
9159 __set_task_cpu(p, this_cpu);
9162 update_curr(cfs_rq);
9165 se->vruntime = curr->vruntime;
9166 place_entity(cfs_rq, se, 1);
9168 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9170 * Upon rescheduling, sched_class::put_prev_task() will place
9171 * 'current' within the tree based on its new key value.
9173 swap(curr->vruntime, se->vruntime);
9177 se->vruntime -= cfs_rq->min_vruntime;
9179 raw_spin_unlock_irqrestore(&rq->lock, flags);
9183 * Priority of the task has changed. Check to see if we preempt
9187 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9189 if (!task_on_rq_queued(p))
9193 * Reschedule if we are currently running on this runqueue and
9194 * our priority decreased, or if we are not currently running on
9195 * this runqueue and our priority is higher than the current's
9197 if (rq->curr == p) {
9198 if (p->prio > oldprio)
9201 check_preempt_curr(rq, p, 0);
9204 static inline bool vruntime_normalized(struct task_struct *p)
9206 struct sched_entity *se = &p->se;
9209 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9210 * the dequeue_entity(.flags=0) will already have normalized the
9217 * When !on_rq, vruntime of the task has usually NOT been normalized.
9218 * But there are some cases where it has already been normalized:
9220 * - A forked child which is waiting for being woken up by
9221 * wake_up_new_task().
9222 * - A task which has been woken up by try_to_wake_up() and
9223 * waiting for actually being woken up by sched_ttwu_pending().
9225 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9231 static void detach_task_cfs_rq(struct task_struct *p)
9233 struct sched_entity *se = &p->se;
9234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9236 if (!vruntime_normalized(p)) {
9238 * Fix up our vruntime so that the current sleep doesn't
9239 * cause 'unlimited' sleep bonus.
9241 place_entity(cfs_rq, se, 0);
9242 se->vruntime -= cfs_rq->min_vruntime;
9245 /* Catch up with the cfs_rq and remove our load when we leave */
9246 detach_entity_load_avg(cfs_rq, se);
9249 static void attach_task_cfs_rq(struct task_struct *p)
9251 struct sched_entity *se = &p->se;
9252 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9254 #ifdef CONFIG_FAIR_GROUP_SCHED
9256 * Since the real-depth could have been changed (only FAIR
9257 * class maintain depth value), reset depth properly.
9259 se->depth = se->parent ? se->parent->depth + 1 : 0;
9262 /* Synchronize task with its cfs_rq */
9263 attach_entity_load_avg(cfs_rq, se);
9265 if (!vruntime_normalized(p))
9266 se->vruntime += cfs_rq->min_vruntime;
9269 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9271 detach_task_cfs_rq(p);
9274 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9276 attach_task_cfs_rq(p);
9278 if (task_on_rq_queued(p)) {
9280 * We were most likely switched from sched_rt, so
9281 * kick off the schedule if running, otherwise just see
9282 * if we can still preempt the current task.
9287 check_preempt_curr(rq, p, 0);
9291 /* Account for a task changing its policy or group.
9293 * This routine is mostly called to set cfs_rq->curr field when a task
9294 * migrates between groups/classes.
9296 static void set_curr_task_fair(struct rq *rq)
9298 struct sched_entity *se = &rq->curr->se;
9300 for_each_sched_entity(se) {
9301 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9303 set_next_entity(cfs_rq, se);
9304 /* ensure bandwidth has been allocated on our new cfs_rq */
9305 account_cfs_rq_runtime(cfs_rq, 0);
9309 void init_cfs_rq(struct cfs_rq *cfs_rq)
9311 cfs_rq->tasks_timeline = RB_ROOT;
9312 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9313 #ifndef CONFIG_64BIT
9314 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9317 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9318 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9322 #ifdef CONFIG_FAIR_GROUP_SCHED
9323 static void task_move_group_fair(struct task_struct *p)
9325 detach_task_cfs_rq(p);
9326 set_task_rq(p, task_cpu(p));
9329 /* Tell se's cfs_rq has been changed -- migrated */
9330 p->se.avg.last_update_time = 0;
9332 attach_task_cfs_rq(p);
9335 void free_fair_sched_group(struct task_group *tg)
9339 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9341 for_each_possible_cpu(i) {
9343 kfree(tg->cfs_rq[i]);
9346 remove_entity_load_avg(tg->se[i]);
9355 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9357 struct cfs_rq *cfs_rq;
9358 struct sched_entity *se;
9361 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9364 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9368 tg->shares = NICE_0_LOAD;
9370 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9372 for_each_possible_cpu(i) {
9373 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9374 GFP_KERNEL, cpu_to_node(i));
9378 se = kzalloc_node(sizeof(struct sched_entity),
9379 GFP_KERNEL, cpu_to_node(i));
9383 init_cfs_rq(cfs_rq);
9384 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9385 init_entity_runnable_average(se);
9396 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9398 struct rq *rq = cpu_rq(cpu);
9399 unsigned long flags;
9402 * Only empty task groups can be destroyed; so we can speculatively
9403 * check on_list without danger of it being re-added.
9405 if (!tg->cfs_rq[cpu]->on_list)
9408 raw_spin_lock_irqsave(&rq->lock, flags);
9409 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9410 raw_spin_unlock_irqrestore(&rq->lock, flags);
9413 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9414 struct sched_entity *se, int cpu,
9415 struct sched_entity *parent)
9417 struct rq *rq = cpu_rq(cpu);
9421 init_cfs_rq_runtime(cfs_rq);
9423 tg->cfs_rq[cpu] = cfs_rq;
9426 /* se could be NULL for root_task_group */
9431 se->cfs_rq = &rq->cfs;
9434 se->cfs_rq = parent->my_q;
9435 se->depth = parent->depth + 1;
9439 /* guarantee group entities always have weight */
9440 update_load_set(&se->load, NICE_0_LOAD);
9441 se->parent = parent;
9444 static DEFINE_MUTEX(shares_mutex);
9446 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9449 unsigned long flags;
9452 * We can't change the weight of the root cgroup.
9457 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9459 mutex_lock(&shares_mutex);
9460 if (tg->shares == shares)
9463 tg->shares = shares;
9464 for_each_possible_cpu(i) {
9465 struct rq *rq = cpu_rq(i);
9466 struct sched_entity *se;
9469 /* Propagate contribution to hierarchy */
9470 raw_spin_lock_irqsave(&rq->lock, flags);
9472 /* Possible calls to update_curr() need rq clock */
9473 update_rq_clock(rq);
9474 for_each_sched_entity(se)
9475 update_cfs_shares(group_cfs_rq(se));
9476 raw_spin_unlock_irqrestore(&rq->lock, flags);
9480 mutex_unlock(&shares_mutex);
9483 #else /* CONFIG_FAIR_GROUP_SCHED */
9485 void free_fair_sched_group(struct task_group *tg) { }
9487 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9492 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9494 #endif /* CONFIG_FAIR_GROUP_SCHED */
9497 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9499 struct sched_entity *se = &task->se;
9500 unsigned int rr_interval = 0;
9503 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9506 if (rq->cfs.load.weight)
9507 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9513 * All the scheduling class methods:
9515 const struct sched_class fair_sched_class = {
9516 .next = &idle_sched_class,
9517 .enqueue_task = enqueue_task_fair,
9518 .dequeue_task = dequeue_task_fair,
9519 .yield_task = yield_task_fair,
9520 .yield_to_task = yield_to_task_fair,
9522 .check_preempt_curr = check_preempt_wakeup,
9524 .pick_next_task = pick_next_task_fair,
9525 .put_prev_task = put_prev_task_fair,
9528 .select_task_rq = select_task_rq_fair,
9529 .migrate_task_rq = migrate_task_rq_fair,
9531 .rq_online = rq_online_fair,
9532 .rq_offline = rq_offline_fair,
9534 .task_waking = task_waking_fair,
9535 .task_dead = task_dead_fair,
9536 .set_cpus_allowed = set_cpus_allowed_common,
9539 .set_curr_task = set_curr_task_fair,
9540 .task_tick = task_tick_fair,
9541 .task_fork = task_fork_fair,
9543 .prio_changed = prio_changed_fair,
9544 .switched_from = switched_from_fair,
9545 .switched_to = switched_to_fair,
9547 .get_rr_interval = get_rr_interval_fair,
9549 .update_curr = update_curr_fair,
9551 #ifdef CONFIG_FAIR_GROUP_SCHED
9552 .task_move_group = task_move_group_fair,
9556 #ifdef CONFIG_SCHED_DEBUG
9557 void print_cfs_stats(struct seq_file *m, int cpu)
9559 struct cfs_rq *cfs_rq;
9562 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9563 print_cfs_rq(m, cpu, cfs_rq);
9567 #ifdef CONFIG_NUMA_BALANCING
9568 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9571 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9573 for_each_online_node(node) {
9574 if (p->numa_faults) {
9575 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9576 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9578 if (p->numa_group) {
9579 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9580 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9582 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9585 #endif /* CONFIG_NUMA_BALANCING */
9586 #endif /* CONFIG_SCHED_DEBUG */
9588 __init void init_sched_fair_class(void)
9591 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9593 #ifdef CONFIG_NO_HZ_COMMON
9594 nohz.next_balance = jiffies;
9595 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9596 cpu_notifier(sched_ilb_notifier, 0);