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_sync_hint_enable = 1;
57 unsigned int sysctl_sched_initial_task_util = 0;
58 unsigned int sysctl_sched_cstate_aware = 1;
60 #ifdef CONFIG_SCHED_WALT
61 unsigned int sysctl_sched_use_walt_cpu_util = 1;
62 unsigned int sysctl_sched_use_walt_task_util = 1;
63 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
67 * The initial- and re-scaling of tunables is configurable
68 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
72 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
73 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75 enum sched_tunable_scaling sysctl_sched_tunable_scaling
76 = SCHED_TUNABLESCALING_LOG;
79 * Minimal preemption granularity for CPU-bound tasks:
80 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_min_granularity = 750000ULL;
83 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
86 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
88 static unsigned int sched_nr_latency = 8;
91 * After fork, child runs first. If set to 0 (default) then
92 * parent will (try to) run first.
94 unsigned int sysctl_sched_child_runs_first __read_mostly;
97 * SCHED_OTHER wake-up granularity.
98 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
100 * This option delays the preemption effects of decoupled workloads
101 * and reduces their over-scheduling. Synchronous workloads will still
102 * have immediate wakeup/sleep latencies.
104 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
105 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
107 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
110 * The exponential sliding window over which load is averaged for shares
114 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
116 #ifdef CONFIG_CFS_BANDWIDTH
118 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
119 * each time a cfs_rq requests quota.
121 * Note: in the case that the slice exceeds the runtime remaining (either due
122 * to consumption or the quota being specified to be smaller than the slice)
123 * we will always only issue the remaining available time.
125 * default: 5 msec, units: microseconds
127 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
131 * The margin used when comparing utilization with CPU capacity:
132 * util * margin < capacity * 1024
134 unsigned int capacity_margin = 1280; /* ~20% */
136 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
142 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
148 static inline void update_load_set(struct load_weight *lw, unsigned long w)
155 * Increase the granularity value when there are more CPUs,
156 * because with more CPUs the 'effective latency' as visible
157 * to users decreases. But the relationship is not linear,
158 * so pick a second-best guess by going with the log2 of the
161 * This idea comes from the SD scheduler of Con Kolivas:
163 static unsigned int get_update_sysctl_factor(void)
165 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
168 switch (sysctl_sched_tunable_scaling) {
169 case SCHED_TUNABLESCALING_NONE:
172 case SCHED_TUNABLESCALING_LINEAR:
175 case SCHED_TUNABLESCALING_LOG:
177 factor = 1 + ilog2(cpus);
184 static void update_sysctl(void)
186 unsigned int factor = get_update_sysctl_factor();
188 #define SET_SYSCTL(name) \
189 (sysctl_##name = (factor) * normalized_sysctl_##name)
190 SET_SYSCTL(sched_min_granularity);
191 SET_SYSCTL(sched_latency);
192 SET_SYSCTL(sched_wakeup_granularity);
196 void sched_init_granularity(void)
201 #define WMULT_CONST (~0U)
202 #define WMULT_SHIFT 32
204 static void __update_inv_weight(struct load_weight *lw)
208 if (likely(lw->inv_weight))
211 w = scale_load_down(lw->weight);
213 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
215 else if (unlikely(!w))
216 lw->inv_weight = WMULT_CONST;
218 lw->inv_weight = WMULT_CONST / w;
222 * delta_exec * weight / lw.weight
224 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
226 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
227 * we're guaranteed shift stays positive because inv_weight is guaranteed to
228 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
230 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
231 * weight/lw.weight <= 1, and therefore our shift will also be positive.
233 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
235 u64 fact = scale_load_down(weight);
236 int shift = WMULT_SHIFT;
238 __update_inv_weight(lw);
240 if (unlikely(fact >> 32)) {
247 /* hint to use a 32x32->64 mul */
248 fact = (u64)(u32)fact * lw->inv_weight;
255 return mul_u64_u32_shr(delta_exec, fact, shift);
259 const struct sched_class fair_sched_class;
261 /**************************************************************
262 * CFS operations on generic schedulable entities:
265 #ifdef CONFIG_FAIR_GROUP_SCHED
267 /* cpu runqueue to which this cfs_rq is attached */
268 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
273 /* An entity is a task if it doesn't "own" a runqueue */
274 #define entity_is_task(se) (!se->my_q)
276 static inline struct task_struct *task_of(struct sched_entity *se)
278 #ifdef CONFIG_SCHED_DEBUG
279 WARN_ON_ONCE(!entity_is_task(se));
281 return container_of(se, struct task_struct, se);
284 /* Walk up scheduling entities hierarchy */
285 #define for_each_sched_entity(se) \
286 for (; se; se = se->parent)
288 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
293 /* runqueue on which this entity is (to be) queued */
294 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
299 /* runqueue "owned" by this group */
300 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
305 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
307 if (!cfs_rq->on_list) {
309 * Ensure we either appear before our parent (if already
310 * enqueued) or force our parent to appear after us when it is
311 * enqueued. The fact that we always enqueue bottom-up
312 * reduces this to two cases.
314 if (cfs_rq->tg->parent &&
315 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
316 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
317 &rq_of(cfs_rq)->leaf_cfs_rq_list);
319 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
320 &rq_of(cfs_rq)->leaf_cfs_rq_list);
327 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
329 if (cfs_rq->on_list) {
330 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
335 /* Iterate thr' all leaf cfs_rq's on a runqueue */
336 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
337 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
339 /* Do the two (enqueued) entities belong to the same group ? */
340 static inline struct cfs_rq *
341 is_same_group(struct sched_entity *se, struct sched_entity *pse)
343 if (se->cfs_rq == pse->cfs_rq)
349 static inline struct sched_entity *parent_entity(struct sched_entity *se)
355 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
357 int se_depth, pse_depth;
360 * preemption test can be made between sibling entities who are in the
361 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
362 * both tasks until we find their ancestors who are siblings of common
366 /* First walk up until both entities are at same depth */
367 se_depth = (*se)->depth;
368 pse_depth = (*pse)->depth;
370 while (se_depth > pse_depth) {
372 *se = parent_entity(*se);
375 while (pse_depth > se_depth) {
377 *pse = parent_entity(*pse);
380 while (!is_same_group(*se, *pse)) {
381 *se = parent_entity(*se);
382 *pse = parent_entity(*pse);
386 #else /* !CONFIG_FAIR_GROUP_SCHED */
388 static inline struct task_struct *task_of(struct sched_entity *se)
390 return container_of(se, struct task_struct, se);
393 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
395 return container_of(cfs_rq, struct rq, cfs);
398 #define entity_is_task(se) 1
400 #define for_each_sched_entity(se) \
401 for (; se; se = NULL)
403 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
405 return &task_rq(p)->cfs;
408 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
410 struct task_struct *p = task_of(se);
411 struct rq *rq = task_rq(p);
416 /* runqueue "owned" by this group */
417 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
422 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
426 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
430 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
431 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 unsigned int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
617 if (unlikely(se->load.weight != NICE_0_LOAD))
618 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
624 * The idea is to set a period in which each task runs once.
626 * When there are too many tasks (sched_nr_latency) we have to stretch
627 * this period because otherwise the slices get too small.
629 * p = (nr <= nl) ? l : l*nr/nl
631 static u64 __sched_period(unsigned long nr_running)
633 if (unlikely(nr_running > sched_nr_latency))
634 return nr_running * sysctl_sched_min_granularity;
636 return sysctl_sched_latency;
640 * We calculate the wall-time slice from the period by taking a part
641 * proportional to the weight.
645 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
649 for_each_sched_entity(se) {
650 struct load_weight *load;
651 struct load_weight lw;
653 cfs_rq = cfs_rq_of(se);
654 load = &cfs_rq->load;
656 if (unlikely(!se->on_rq)) {
659 update_load_add(&lw, se->load.weight);
662 slice = __calc_delta(slice, se->load.weight, load);
668 * We calculate the vruntime slice of a to-be-inserted task.
672 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
674 return calc_delta_fair(sched_slice(cfs_rq, se), se);
678 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
679 static unsigned long task_h_load(struct task_struct *p);
682 * We choose a half-life close to 1 scheduling period.
683 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
684 * dependent on this value.
686 #define LOAD_AVG_PERIOD 32
687 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
688 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
690 /* Give new sched_entity start runnable values to heavy its load in infant time */
691 void init_entity_runnable_average(struct sched_entity *se)
693 struct sched_avg *sa = &se->avg;
695 sa->last_update_time = 0;
697 * sched_avg's period_contrib should be strictly less then 1024, so
698 * we give it 1023 to make sure it is almost a period (1024us), and
699 * will definitely be update (after enqueue).
701 sa->period_contrib = 1023;
702 sa->load_avg = scale_load_down(se->load.weight);
703 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
704 sa->util_avg = sched_freq() ?
705 sysctl_sched_initial_task_util :
706 scale_load_down(SCHED_LOAD_SCALE);
707 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
708 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
712 void init_entity_runnable_average(struct sched_entity *se)
718 * Update the current task's runtime statistics.
720 static void update_curr(struct cfs_rq *cfs_rq)
722 struct sched_entity *curr = cfs_rq->curr;
723 u64 now = rq_clock_task(rq_of(cfs_rq));
729 delta_exec = now - curr->exec_start;
730 if (unlikely((s64)delta_exec <= 0))
733 curr->exec_start = now;
735 schedstat_set(curr->statistics.exec_max,
736 max(delta_exec, curr->statistics.exec_max));
738 curr->sum_exec_runtime += delta_exec;
739 schedstat_add(cfs_rq, exec_clock, delta_exec);
741 curr->vruntime += calc_delta_fair(delta_exec, curr);
742 update_min_vruntime(cfs_rq);
744 if (entity_is_task(curr)) {
745 struct task_struct *curtask = task_of(curr);
747 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
748 cpuacct_charge(curtask, delta_exec);
749 account_group_exec_runtime(curtask, delta_exec);
752 account_cfs_rq_runtime(cfs_rq, delta_exec);
755 static void update_curr_fair(struct rq *rq)
757 update_curr(cfs_rq_of(&rq->curr->se));
761 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
763 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
767 * Task is being enqueued - update stats:
769 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
772 * Are we enqueueing a waiting task? (for current tasks
773 * a dequeue/enqueue event is a NOP)
775 if (se != cfs_rq->curr)
776 update_stats_wait_start(cfs_rq, se);
780 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
782 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
783 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
784 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
785 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
787 #ifdef CONFIG_SCHEDSTATS
788 if (entity_is_task(se)) {
789 trace_sched_stat_wait(task_of(se),
790 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
793 schedstat_set(se->statistics.wait_start, 0);
797 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
800 * Mark the end of the wait period if dequeueing a
803 if (se != cfs_rq->curr)
804 update_stats_wait_end(cfs_rq, se);
808 * We are picking a new current task - update its stats:
811 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
814 * We are starting a new run period:
816 se->exec_start = rq_clock_task(rq_of(cfs_rq));
819 /**************************************************
820 * Scheduling class queueing methods:
823 #ifdef CONFIG_NUMA_BALANCING
825 * Approximate time to scan a full NUMA task in ms. The task scan period is
826 * calculated based on the tasks virtual memory size and
827 * numa_balancing_scan_size.
829 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
830 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
832 /* Portion of address space to scan in MB */
833 unsigned int sysctl_numa_balancing_scan_size = 256;
835 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
836 unsigned int sysctl_numa_balancing_scan_delay = 1000;
838 static unsigned int task_nr_scan_windows(struct task_struct *p)
840 unsigned long rss = 0;
841 unsigned long nr_scan_pages;
844 * Calculations based on RSS as non-present and empty pages are skipped
845 * by the PTE scanner and NUMA hinting faults should be trapped based
848 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
849 rss = get_mm_rss(p->mm);
853 rss = round_up(rss, nr_scan_pages);
854 return rss / nr_scan_pages;
857 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
858 #define MAX_SCAN_WINDOW 2560
860 static unsigned int task_scan_min(struct task_struct *p)
862 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
863 unsigned int scan, floor;
864 unsigned int windows = 1;
866 if (scan_size < MAX_SCAN_WINDOW)
867 windows = MAX_SCAN_WINDOW / scan_size;
868 floor = 1000 / windows;
870 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
871 return max_t(unsigned int, floor, scan);
874 static unsigned int task_scan_max(struct task_struct *p)
876 unsigned int smin = task_scan_min(p);
879 /* Watch for min being lower than max due to floor calculations */
880 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
881 return max(smin, smax);
884 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
886 rq->nr_numa_running += (p->numa_preferred_nid != -1);
887 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
890 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
892 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
893 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
899 spinlock_t lock; /* nr_tasks, tasks */
904 nodemask_t active_nodes;
905 unsigned long total_faults;
907 * Faults_cpu is used to decide whether memory should move
908 * towards the CPU. As a consequence, these stats are weighted
909 * more by CPU use than by memory faults.
911 unsigned long *faults_cpu;
912 unsigned long faults[0];
915 /* Shared or private faults. */
916 #define NR_NUMA_HINT_FAULT_TYPES 2
918 /* Memory and CPU locality */
919 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
921 /* Averaged statistics, and temporary buffers. */
922 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
924 pid_t task_numa_group_id(struct task_struct *p)
926 return p->numa_group ? p->numa_group->gid : 0;
930 * The averaged statistics, shared & private, memory & cpu,
931 * occupy the first half of the array. The second half of the
932 * array is for current counters, which are averaged into the
933 * first set by task_numa_placement.
935 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
937 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
940 static inline unsigned long task_faults(struct task_struct *p, int nid)
945 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
946 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
949 static inline unsigned long group_faults(struct task_struct *p, int nid)
954 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
955 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
958 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
960 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
961 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
964 /* Handle placement on systems where not all nodes are directly connected. */
965 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
966 int maxdist, bool task)
968 unsigned long score = 0;
972 * All nodes are directly connected, and the same distance
973 * from each other. No need for fancy placement algorithms.
975 if (sched_numa_topology_type == NUMA_DIRECT)
979 * This code is called for each node, introducing N^2 complexity,
980 * which should be ok given the number of nodes rarely exceeds 8.
982 for_each_online_node(node) {
983 unsigned long faults;
984 int dist = node_distance(nid, node);
987 * The furthest away nodes in the system are not interesting
988 * for placement; nid was already counted.
990 if (dist == sched_max_numa_distance || node == nid)
994 * On systems with a backplane NUMA topology, compare groups
995 * of nodes, and move tasks towards the group with the most
996 * memory accesses. When comparing two nodes at distance
997 * "hoplimit", only nodes closer by than "hoplimit" are part
998 * of each group. Skip other nodes.
1000 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1004 /* Add up the faults from nearby nodes. */
1006 faults = task_faults(p, node);
1008 faults = group_faults(p, node);
1011 * On systems with a glueless mesh NUMA topology, there are
1012 * no fixed "groups of nodes". Instead, nodes that are not
1013 * directly connected bounce traffic through intermediate
1014 * nodes; a numa_group can occupy any set of nodes.
1015 * The further away a node is, the less the faults count.
1016 * This seems to result in good task placement.
1018 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1019 faults *= (sched_max_numa_distance - dist);
1020 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1030 * These return the fraction of accesses done by a particular task, or
1031 * task group, on a particular numa node. The group weight is given a
1032 * larger multiplier, in order to group tasks together that are almost
1033 * evenly spread out between numa nodes.
1035 static inline unsigned long task_weight(struct task_struct *p, int nid,
1038 unsigned long faults, total_faults;
1040 if (!p->numa_faults)
1043 total_faults = p->total_numa_faults;
1048 faults = task_faults(p, nid);
1049 faults += score_nearby_nodes(p, nid, dist, true);
1051 return 1000 * faults / total_faults;
1054 static inline unsigned long group_weight(struct task_struct *p, int nid,
1057 unsigned long faults, total_faults;
1062 total_faults = p->numa_group->total_faults;
1067 faults = group_faults(p, nid);
1068 faults += score_nearby_nodes(p, nid, dist, false);
1070 return 1000 * faults / total_faults;
1073 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1074 int src_nid, int dst_cpu)
1076 struct numa_group *ng = p->numa_group;
1077 int dst_nid = cpu_to_node(dst_cpu);
1078 int last_cpupid, this_cpupid;
1080 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1083 * Multi-stage node selection is used in conjunction with a periodic
1084 * migration fault to build a temporal task<->page relation. By using
1085 * a two-stage filter we remove short/unlikely relations.
1087 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1088 * a task's usage of a particular page (n_p) per total usage of this
1089 * page (n_t) (in a given time-span) to a probability.
1091 * Our periodic faults will sample this probability and getting the
1092 * same result twice in a row, given these samples are fully
1093 * independent, is then given by P(n)^2, provided our sample period
1094 * is sufficiently short compared to the usage pattern.
1096 * This quadric squishes small probabilities, making it less likely we
1097 * act on an unlikely task<->page relation.
1099 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1100 if (!cpupid_pid_unset(last_cpupid) &&
1101 cpupid_to_nid(last_cpupid) != dst_nid)
1104 /* Always allow migrate on private faults */
1105 if (cpupid_match_pid(p, last_cpupid))
1108 /* A shared fault, but p->numa_group has not been set up yet. */
1113 * Do not migrate if the destination is not a node that
1114 * is actively used by this numa group.
1116 if (!node_isset(dst_nid, ng->active_nodes))
1120 * Source is a node that is not actively used by this
1121 * numa group, while the destination is. Migrate.
1123 if (!node_isset(src_nid, ng->active_nodes))
1127 * Both source and destination are nodes in active
1128 * use by this numa group. Maximize memory bandwidth
1129 * by migrating from more heavily used groups, to less
1130 * heavily used ones, spreading the load around.
1131 * Use a 1/4 hysteresis to avoid spurious page movement.
1133 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1136 static unsigned long weighted_cpuload(const int cpu);
1137 static unsigned long source_load(int cpu, int type);
1138 static unsigned long target_load(int cpu, int type);
1139 static unsigned long capacity_of(int cpu);
1140 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1142 /* Cached statistics for all CPUs within a node */
1144 unsigned long nr_running;
1147 /* Total compute capacity of CPUs on a node */
1148 unsigned long compute_capacity;
1150 /* Approximate capacity in terms of runnable tasks on a node */
1151 unsigned long task_capacity;
1152 int has_free_capacity;
1156 * XXX borrowed from update_sg_lb_stats
1158 static void update_numa_stats(struct numa_stats *ns, int nid)
1160 int smt, cpu, cpus = 0;
1161 unsigned long capacity;
1163 memset(ns, 0, sizeof(*ns));
1164 for_each_cpu(cpu, cpumask_of_node(nid)) {
1165 struct rq *rq = cpu_rq(cpu);
1167 ns->nr_running += rq->nr_running;
1168 ns->load += weighted_cpuload(cpu);
1169 ns->compute_capacity += capacity_of(cpu);
1175 * If we raced with hotplug and there are no CPUs left in our mask
1176 * the @ns structure is NULL'ed and task_numa_compare() will
1177 * not find this node attractive.
1179 * We'll either bail at !has_free_capacity, or we'll detect a huge
1180 * imbalance and bail there.
1185 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1186 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1187 capacity = cpus / smt; /* cores */
1189 ns->task_capacity = min_t(unsigned, capacity,
1190 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1191 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1194 struct task_numa_env {
1195 struct task_struct *p;
1197 int src_cpu, src_nid;
1198 int dst_cpu, dst_nid;
1200 struct numa_stats src_stats, dst_stats;
1205 struct task_struct *best_task;
1210 static void task_numa_assign(struct task_numa_env *env,
1211 struct task_struct *p, long imp)
1214 put_task_struct(env->best_task);
1217 env->best_imp = imp;
1218 env->best_cpu = env->dst_cpu;
1221 static bool load_too_imbalanced(long src_load, long dst_load,
1222 struct task_numa_env *env)
1225 long orig_src_load, orig_dst_load;
1226 long src_capacity, dst_capacity;
1229 * The load is corrected for the CPU capacity available on each node.
1232 * ------------ vs ---------
1233 * src_capacity dst_capacity
1235 src_capacity = env->src_stats.compute_capacity;
1236 dst_capacity = env->dst_stats.compute_capacity;
1238 /* We care about the slope of the imbalance, not the direction. */
1239 if (dst_load < src_load)
1240 swap(dst_load, src_load);
1242 /* Is the difference below the threshold? */
1243 imb = dst_load * src_capacity * 100 -
1244 src_load * dst_capacity * env->imbalance_pct;
1249 * The imbalance is above the allowed threshold.
1250 * Compare it with the old imbalance.
1252 orig_src_load = env->src_stats.load;
1253 orig_dst_load = env->dst_stats.load;
1255 if (orig_dst_load < orig_src_load)
1256 swap(orig_dst_load, orig_src_load);
1258 old_imb = orig_dst_load * src_capacity * 100 -
1259 orig_src_load * dst_capacity * env->imbalance_pct;
1261 /* Would this change make things worse? */
1262 return (imb > old_imb);
1266 * This checks if the overall compute and NUMA accesses of the system would
1267 * be improved if the source tasks was migrated to the target dst_cpu taking
1268 * into account that it might be best if task running on the dst_cpu should
1269 * be exchanged with the source task
1271 static void task_numa_compare(struct task_numa_env *env,
1272 long taskimp, long groupimp)
1274 struct rq *src_rq = cpu_rq(env->src_cpu);
1275 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1276 struct task_struct *cur;
1277 long src_load, dst_load;
1279 long imp = env->p->numa_group ? groupimp : taskimp;
1281 int dist = env->dist;
1282 bool assigned = false;
1286 raw_spin_lock_irq(&dst_rq->lock);
1289 * No need to move the exiting task or idle task.
1291 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1295 * The task_struct must be protected here to protect the
1296 * p->numa_faults access in the task_weight since the
1297 * numa_faults could already be freed in the following path:
1298 * finish_task_switch()
1299 * --> put_task_struct()
1300 * --> __put_task_struct()
1301 * --> task_numa_free()
1303 get_task_struct(cur);
1306 raw_spin_unlock_irq(&dst_rq->lock);
1309 * Because we have preemption enabled we can get migrated around and
1310 * end try selecting ourselves (current == env->p) as a swap candidate.
1316 * "imp" is the fault differential for the source task between the
1317 * source and destination node. Calculate the total differential for
1318 * the source task and potential destination task. The more negative
1319 * the value is, the more rmeote accesses that would be expected to
1320 * be incurred if the tasks were swapped.
1323 /* Skip this swap candidate if cannot move to the source cpu */
1324 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1328 * If dst and source tasks are in the same NUMA group, or not
1329 * in any group then look only at task weights.
1331 if (cur->numa_group == env->p->numa_group) {
1332 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1333 task_weight(cur, env->dst_nid, dist);
1335 * Add some hysteresis to prevent swapping the
1336 * tasks within a group over tiny differences.
1338 if (cur->numa_group)
1342 * Compare the group weights. If a task is all by
1343 * itself (not part of a group), use the task weight
1346 if (cur->numa_group)
1347 imp += group_weight(cur, env->src_nid, dist) -
1348 group_weight(cur, env->dst_nid, dist);
1350 imp += task_weight(cur, env->src_nid, dist) -
1351 task_weight(cur, env->dst_nid, dist);
1355 if (imp <= env->best_imp && moveimp <= env->best_imp)
1359 /* Is there capacity at our destination? */
1360 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1361 !env->dst_stats.has_free_capacity)
1367 /* Balance doesn't matter much if we're running a task per cpu */
1368 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1369 dst_rq->nr_running == 1)
1373 * In the overloaded case, try and keep the load balanced.
1376 load = task_h_load(env->p);
1377 dst_load = env->dst_stats.load + load;
1378 src_load = env->src_stats.load - load;
1380 if (moveimp > imp && moveimp > env->best_imp) {
1382 * If the improvement from just moving env->p direction is
1383 * better than swapping tasks around, check if a move is
1384 * possible. Store a slightly smaller score than moveimp,
1385 * so an actually idle CPU will win.
1387 if (!load_too_imbalanced(src_load, dst_load, env)) {
1389 put_task_struct(cur);
1395 if (imp <= env->best_imp)
1399 load = task_h_load(cur);
1404 if (load_too_imbalanced(src_load, dst_load, env))
1408 * One idle CPU per node is evaluated for a task numa move.
1409 * Call select_idle_sibling to maybe find a better one.
1412 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1417 task_numa_assign(env, cur, imp);
1421 * The dst_rq->curr isn't assigned. The protection for task_struct is
1424 if (cur && !assigned)
1425 put_task_struct(cur);
1428 static void task_numa_find_cpu(struct task_numa_env *env,
1429 long taskimp, long groupimp)
1433 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1434 /* Skip this CPU if the source task cannot migrate */
1435 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1439 task_numa_compare(env, taskimp, groupimp);
1443 /* Only move tasks to a NUMA node less busy than the current node. */
1444 static bool numa_has_capacity(struct task_numa_env *env)
1446 struct numa_stats *src = &env->src_stats;
1447 struct numa_stats *dst = &env->dst_stats;
1449 if (src->has_free_capacity && !dst->has_free_capacity)
1453 * Only consider a task move if the source has a higher load
1454 * than the destination, corrected for CPU capacity on each node.
1456 * src->load dst->load
1457 * --------------------- vs ---------------------
1458 * src->compute_capacity dst->compute_capacity
1460 if (src->load * dst->compute_capacity * env->imbalance_pct >
1462 dst->load * src->compute_capacity * 100)
1468 static int task_numa_migrate(struct task_struct *p)
1470 struct task_numa_env env = {
1473 .src_cpu = task_cpu(p),
1474 .src_nid = task_node(p),
1476 .imbalance_pct = 112,
1482 struct sched_domain *sd;
1483 unsigned long taskweight, groupweight;
1485 long taskimp, groupimp;
1488 * Pick the lowest SD_NUMA domain, as that would have the smallest
1489 * imbalance and would be the first to start moving tasks about.
1491 * And we want to avoid any moving of tasks about, as that would create
1492 * random movement of tasks -- counter the numa conditions we're trying
1496 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1498 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1502 * Cpusets can break the scheduler domain tree into smaller
1503 * balance domains, some of which do not cross NUMA boundaries.
1504 * Tasks that are "trapped" in such domains cannot be migrated
1505 * elsewhere, so there is no point in (re)trying.
1507 if (unlikely(!sd)) {
1508 p->numa_preferred_nid = task_node(p);
1512 env.dst_nid = p->numa_preferred_nid;
1513 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1514 taskweight = task_weight(p, env.src_nid, dist);
1515 groupweight = group_weight(p, env.src_nid, dist);
1516 update_numa_stats(&env.src_stats, env.src_nid);
1517 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1518 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1519 update_numa_stats(&env.dst_stats, env.dst_nid);
1521 /* Try to find a spot on the preferred nid. */
1522 if (numa_has_capacity(&env))
1523 task_numa_find_cpu(&env, taskimp, groupimp);
1526 * Look at other nodes in these cases:
1527 * - there is no space available on the preferred_nid
1528 * - the task is part of a numa_group that is interleaved across
1529 * multiple NUMA nodes; in order to better consolidate the group,
1530 * we need to check other locations.
1532 if (env.best_cpu == -1 || (p->numa_group &&
1533 nodes_weight(p->numa_group->active_nodes) > 1)) {
1534 for_each_online_node(nid) {
1535 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1538 dist = node_distance(env.src_nid, env.dst_nid);
1539 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1541 taskweight = task_weight(p, env.src_nid, dist);
1542 groupweight = group_weight(p, env.src_nid, dist);
1545 /* Only consider nodes where both task and groups benefit */
1546 taskimp = task_weight(p, nid, dist) - taskweight;
1547 groupimp = group_weight(p, nid, dist) - groupweight;
1548 if (taskimp < 0 && groupimp < 0)
1553 update_numa_stats(&env.dst_stats, env.dst_nid);
1554 if (numa_has_capacity(&env))
1555 task_numa_find_cpu(&env, taskimp, groupimp);
1560 * If the task is part of a workload that spans multiple NUMA nodes,
1561 * and is migrating into one of the workload's active nodes, remember
1562 * this node as the task's preferred numa node, so the workload can
1564 * A task that migrated to a second choice node will be better off
1565 * trying for a better one later. Do not set the preferred node here.
1567 if (p->numa_group) {
1568 if (env.best_cpu == -1)
1573 if (node_isset(nid, p->numa_group->active_nodes))
1574 sched_setnuma(p, env.dst_nid);
1577 /* No better CPU than the current one was found. */
1578 if (env.best_cpu == -1)
1582 * Reset the scan period if the task is being rescheduled on an
1583 * alternative node to recheck if the tasks is now properly placed.
1585 p->numa_scan_period = task_scan_min(p);
1587 if (env.best_task == NULL) {
1588 ret = migrate_task_to(p, env.best_cpu);
1590 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1594 ret = migrate_swap(p, env.best_task);
1596 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1597 put_task_struct(env.best_task);
1601 /* Attempt to migrate a task to a CPU on the preferred node. */
1602 static void numa_migrate_preferred(struct task_struct *p)
1604 unsigned long interval = HZ;
1606 /* This task has no NUMA fault statistics yet */
1607 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1610 /* Periodically retry migrating the task to the preferred node */
1611 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1612 p->numa_migrate_retry = jiffies + interval;
1614 /* Success if task is already running on preferred CPU */
1615 if (task_node(p) == p->numa_preferred_nid)
1618 /* Otherwise, try migrate to a CPU on the preferred node */
1619 task_numa_migrate(p);
1623 * Find the nodes on which the workload is actively running. We do this by
1624 * tracking the nodes from which NUMA hinting faults are triggered. This can
1625 * be different from the set of nodes where the workload's memory is currently
1628 * The bitmask is used to make smarter decisions on when to do NUMA page
1629 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1630 * are added when they cause over 6/16 of the maximum number of faults, but
1631 * only removed when they drop below 3/16.
1633 static void update_numa_active_node_mask(struct numa_group *numa_group)
1635 unsigned long faults, max_faults = 0;
1638 for_each_online_node(nid) {
1639 faults = group_faults_cpu(numa_group, nid);
1640 if (faults > max_faults)
1641 max_faults = faults;
1644 for_each_online_node(nid) {
1645 faults = group_faults_cpu(numa_group, nid);
1646 if (!node_isset(nid, numa_group->active_nodes)) {
1647 if (faults > max_faults * 6 / 16)
1648 node_set(nid, numa_group->active_nodes);
1649 } else if (faults < max_faults * 3 / 16)
1650 node_clear(nid, numa_group->active_nodes);
1655 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1656 * increments. The more local the fault statistics are, the higher the scan
1657 * period will be for the next scan window. If local/(local+remote) ratio is
1658 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1659 * the scan period will decrease. Aim for 70% local accesses.
1661 #define NUMA_PERIOD_SLOTS 10
1662 #define NUMA_PERIOD_THRESHOLD 7
1665 * Increase the scan period (slow down scanning) if the majority of
1666 * our memory is already on our local node, or if the majority of
1667 * the page accesses are shared with other processes.
1668 * Otherwise, decrease the scan period.
1670 static void update_task_scan_period(struct task_struct *p,
1671 unsigned long shared, unsigned long private)
1673 unsigned int period_slot;
1677 unsigned long remote = p->numa_faults_locality[0];
1678 unsigned long local = p->numa_faults_locality[1];
1681 * If there were no record hinting faults then either the task is
1682 * completely idle or all activity is areas that are not of interest
1683 * to automatic numa balancing. Related to that, if there were failed
1684 * migration then it implies we are migrating too quickly or the local
1685 * node is overloaded. In either case, scan slower
1687 if (local + shared == 0 || p->numa_faults_locality[2]) {
1688 p->numa_scan_period = min(p->numa_scan_period_max,
1689 p->numa_scan_period << 1);
1691 p->mm->numa_next_scan = jiffies +
1692 msecs_to_jiffies(p->numa_scan_period);
1698 * Prepare to scale scan period relative to the current period.
1699 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1700 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1701 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1703 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1704 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1705 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1706 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1709 diff = slot * period_slot;
1711 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1714 * Scale scan rate increases based on sharing. There is an
1715 * inverse relationship between the degree of sharing and
1716 * the adjustment made to the scanning period. Broadly
1717 * speaking the intent is that there is little point
1718 * scanning faster if shared accesses dominate as it may
1719 * simply bounce migrations uselessly
1721 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1722 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1725 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1726 task_scan_min(p), task_scan_max(p));
1727 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1731 * Get the fraction of time the task has been running since the last
1732 * NUMA placement cycle. The scheduler keeps similar statistics, but
1733 * decays those on a 32ms period, which is orders of magnitude off
1734 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1735 * stats only if the task is so new there are no NUMA statistics yet.
1737 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1739 u64 runtime, delta, now;
1740 /* Use the start of this time slice to avoid calculations. */
1741 now = p->se.exec_start;
1742 runtime = p->se.sum_exec_runtime;
1744 if (p->last_task_numa_placement) {
1745 delta = runtime - p->last_sum_exec_runtime;
1746 *period = now - p->last_task_numa_placement;
1748 delta = p->se.avg.load_sum / p->se.load.weight;
1749 *period = LOAD_AVG_MAX;
1752 p->last_sum_exec_runtime = runtime;
1753 p->last_task_numa_placement = now;
1759 * Determine the preferred nid for a task in a numa_group. This needs to
1760 * be done in a way that produces consistent results with group_weight,
1761 * otherwise workloads might not converge.
1763 static int preferred_group_nid(struct task_struct *p, int nid)
1768 /* Direct connections between all NUMA nodes. */
1769 if (sched_numa_topology_type == NUMA_DIRECT)
1773 * On a system with glueless mesh NUMA topology, group_weight
1774 * scores nodes according to the number of NUMA hinting faults on
1775 * both the node itself, and on nearby nodes.
1777 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1778 unsigned long score, max_score = 0;
1779 int node, max_node = nid;
1781 dist = sched_max_numa_distance;
1783 for_each_online_node(node) {
1784 score = group_weight(p, node, dist);
1785 if (score > max_score) {
1794 * Finding the preferred nid in a system with NUMA backplane
1795 * interconnect topology is more involved. The goal is to locate
1796 * tasks from numa_groups near each other in the system, and
1797 * untangle workloads from different sides of the system. This requires
1798 * searching down the hierarchy of node groups, recursively searching
1799 * inside the highest scoring group of nodes. The nodemask tricks
1800 * keep the complexity of the search down.
1802 nodes = node_online_map;
1803 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1804 unsigned long max_faults = 0;
1805 nodemask_t max_group = NODE_MASK_NONE;
1808 /* Are there nodes at this distance from each other? */
1809 if (!find_numa_distance(dist))
1812 for_each_node_mask(a, nodes) {
1813 unsigned long faults = 0;
1814 nodemask_t this_group;
1815 nodes_clear(this_group);
1817 /* Sum group's NUMA faults; includes a==b case. */
1818 for_each_node_mask(b, nodes) {
1819 if (node_distance(a, b) < dist) {
1820 faults += group_faults(p, b);
1821 node_set(b, this_group);
1822 node_clear(b, nodes);
1826 /* Remember the top group. */
1827 if (faults > max_faults) {
1828 max_faults = faults;
1829 max_group = this_group;
1831 * subtle: at the smallest distance there is
1832 * just one node left in each "group", the
1833 * winner is the preferred nid.
1838 /* Next round, evaluate the nodes within max_group. */
1846 static void task_numa_placement(struct task_struct *p)
1848 int seq, nid, max_nid = -1, max_group_nid = -1;
1849 unsigned long max_faults = 0, max_group_faults = 0;
1850 unsigned long fault_types[2] = { 0, 0 };
1851 unsigned long total_faults;
1852 u64 runtime, period;
1853 spinlock_t *group_lock = NULL;
1856 * The p->mm->numa_scan_seq field gets updated without
1857 * exclusive access. Use READ_ONCE() here to ensure
1858 * that the field is read in a single access:
1860 seq = READ_ONCE(p->mm->numa_scan_seq);
1861 if (p->numa_scan_seq == seq)
1863 p->numa_scan_seq = seq;
1864 p->numa_scan_period_max = task_scan_max(p);
1866 total_faults = p->numa_faults_locality[0] +
1867 p->numa_faults_locality[1];
1868 runtime = numa_get_avg_runtime(p, &period);
1870 /* If the task is part of a group prevent parallel updates to group stats */
1871 if (p->numa_group) {
1872 group_lock = &p->numa_group->lock;
1873 spin_lock_irq(group_lock);
1876 /* Find the node with the highest number of faults */
1877 for_each_online_node(nid) {
1878 /* Keep track of the offsets in numa_faults array */
1879 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1880 unsigned long faults = 0, group_faults = 0;
1883 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1884 long diff, f_diff, f_weight;
1886 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1887 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1888 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1889 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1891 /* Decay existing window, copy faults since last scan */
1892 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1893 fault_types[priv] += p->numa_faults[membuf_idx];
1894 p->numa_faults[membuf_idx] = 0;
1897 * Normalize the faults_from, so all tasks in a group
1898 * count according to CPU use, instead of by the raw
1899 * number of faults. Tasks with little runtime have
1900 * little over-all impact on throughput, and thus their
1901 * faults are less important.
1903 f_weight = div64_u64(runtime << 16, period + 1);
1904 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1906 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1907 p->numa_faults[cpubuf_idx] = 0;
1909 p->numa_faults[mem_idx] += diff;
1910 p->numa_faults[cpu_idx] += f_diff;
1911 faults += p->numa_faults[mem_idx];
1912 p->total_numa_faults += diff;
1913 if (p->numa_group) {
1915 * safe because we can only change our own group
1917 * mem_idx represents the offset for a given
1918 * nid and priv in a specific region because it
1919 * is at the beginning of the numa_faults array.
1921 p->numa_group->faults[mem_idx] += diff;
1922 p->numa_group->faults_cpu[mem_idx] += f_diff;
1923 p->numa_group->total_faults += diff;
1924 group_faults += p->numa_group->faults[mem_idx];
1928 if (faults > max_faults) {
1929 max_faults = faults;
1933 if (group_faults > max_group_faults) {
1934 max_group_faults = group_faults;
1935 max_group_nid = nid;
1939 update_task_scan_period(p, fault_types[0], fault_types[1]);
1941 if (p->numa_group) {
1942 update_numa_active_node_mask(p->numa_group);
1943 spin_unlock_irq(group_lock);
1944 max_nid = preferred_group_nid(p, max_group_nid);
1948 /* Set the new preferred node */
1949 if (max_nid != p->numa_preferred_nid)
1950 sched_setnuma(p, max_nid);
1952 if (task_node(p) != p->numa_preferred_nid)
1953 numa_migrate_preferred(p);
1957 static inline int get_numa_group(struct numa_group *grp)
1959 return atomic_inc_not_zero(&grp->refcount);
1962 static inline void put_numa_group(struct numa_group *grp)
1964 if (atomic_dec_and_test(&grp->refcount))
1965 kfree_rcu(grp, rcu);
1968 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1971 struct numa_group *grp, *my_grp;
1972 struct task_struct *tsk;
1974 int cpu = cpupid_to_cpu(cpupid);
1977 if (unlikely(!p->numa_group)) {
1978 unsigned int size = sizeof(struct numa_group) +
1979 4*nr_node_ids*sizeof(unsigned long);
1981 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1985 atomic_set(&grp->refcount, 1);
1986 spin_lock_init(&grp->lock);
1988 /* Second half of the array tracks nids where faults happen */
1989 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1992 node_set(task_node(current), grp->active_nodes);
1994 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1995 grp->faults[i] = p->numa_faults[i];
1997 grp->total_faults = p->total_numa_faults;
2000 rcu_assign_pointer(p->numa_group, grp);
2004 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2006 if (!cpupid_match_pid(tsk, cpupid))
2009 grp = rcu_dereference(tsk->numa_group);
2013 my_grp = p->numa_group;
2018 * Only join the other group if its bigger; if we're the bigger group,
2019 * the other task will join us.
2021 if (my_grp->nr_tasks > grp->nr_tasks)
2025 * Tie-break on the grp address.
2027 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2030 /* Always join threads in the same process. */
2031 if (tsk->mm == current->mm)
2034 /* Simple filter to avoid false positives due to PID collisions */
2035 if (flags & TNF_SHARED)
2038 /* Update priv based on whether false sharing was detected */
2041 if (join && !get_numa_group(grp))
2049 BUG_ON(irqs_disabled());
2050 double_lock_irq(&my_grp->lock, &grp->lock);
2052 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2053 my_grp->faults[i] -= p->numa_faults[i];
2054 grp->faults[i] += p->numa_faults[i];
2056 my_grp->total_faults -= p->total_numa_faults;
2057 grp->total_faults += p->total_numa_faults;
2062 spin_unlock(&my_grp->lock);
2063 spin_unlock_irq(&grp->lock);
2065 rcu_assign_pointer(p->numa_group, grp);
2067 put_numa_group(my_grp);
2075 void task_numa_free(struct task_struct *p)
2077 struct numa_group *grp = p->numa_group;
2078 void *numa_faults = p->numa_faults;
2079 unsigned long flags;
2083 spin_lock_irqsave(&grp->lock, flags);
2084 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2085 grp->faults[i] -= p->numa_faults[i];
2086 grp->total_faults -= p->total_numa_faults;
2089 spin_unlock_irqrestore(&grp->lock, flags);
2090 RCU_INIT_POINTER(p->numa_group, NULL);
2091 put_numa_group(grp);
2094 p->numa_faults = NULL;
2099 * Got a PROT_NONE fault for a page on @node.
2101 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2103 struct task_struct *p = current;
2104 bool migrated = flags & TNF_MIGRATED;
2105 int cpu_node = task_node(current);
2106 int local = !!(flags & TNF_FAULT_LOCAL);
2109 if (!static_branch_likely(&sched_numa_balancing))
2112 /* for example, ksmd faulting in a user's mm */
2116 /* Allocate buffer to track faults on a per-node basis */
2117 if (unlikely(!p->numa_faults)) {
2118 int size = sizeof(*p->numa_faults) *
2119 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2121 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2122 if (!p->numa_faults)
2125 p->total_numa_faults = 0;
2126 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2130 * First accesses are treated as private, otherwise consider accesses
2131 * to be private if the accessing pid has not changed
2133 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2136 priv = cpupid_match_pid(p, last_cpupid);
2137 if (!priv && !(flags & TNF_NO_GROUP))
2138 task_numa_group(p, last_cpupid, flags, &priv);
2142 * If a workload spans multiple NUMA nodes, a shared fault that
2143 * occurs wholly within the set of nodes that the workload is
2144 * actively using should be counted as local. This allows the
2145 * scan rate to slow down when a workload has settled down.
2147 if (!priv && !local && p->numa_group &&
2148 node_isset(cpu_node, p->numa_group->active_nodes) &&
2149 node_isset(mem_node, p->numa_group->active_nodes))
2152 task_numa_placement(p);
2155 * Retry task to preferred node migration periodically, in case it
2156 * case it previously failed, or the scheduler moved us.
2158 if (time_after(jiffies, p->numa_migrate_retry))
2159 numa_migrate_preferred(p);
2162 p->numa_pages_migrated += pages;
2163 if (flags & TNF_MIGRATE_FAIL)
2164 p->numa_faults_locality[2] += pages;
2166 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2167 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2168 p->numa_faults_locality[local] += pages;
2171 static void reset_ptenuma_scan(struct task_struct *p)
2174 * We only did a read acquisition of the mmap sem, so
2175 * p->mm->numa_scan_seq is written to without exclusive access
2176 * and the update is not guaranteed to be atomic. That's not
2177 * much of an issue though, since this is just used for
2178 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2179 * expensive, to avoid any form of compiler optimizations:
2181 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2182 p->mm->numa_scan_offset = 0;
2186 * The expensive part of numa migration is done from task_work context.
2187 * Triggered from task_tick_numa().
2189 void task_numa_work(struct callback_head *work)
2191 unsigned long migrate, next_scan, now = jiffies;
2192 struct task_struct *p = current;
2193 struct mm_struct *mm = p->mm;
2194 struct vm_area_struct *vma;
2195 unsigned long start, end;
2196 unsigned long nr_pte_updates = 0;
2197 long pages, virtpages;
2199 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2201 work->next = work; /* protect against double add */
2203 * Who cares about NUMA placement when they're dying.
2205 * NOTE: make sure not to dereference p->mm before this check,
2206 * exit_task_work() happens _after_ exit_mm() so we could be called
2207 * without p->mm even though we still had it when we enqueued this
2210 if (p->flags & PF_EXITING)
2213 if (!mm->numa_next_scan) {
2214 mm->numa_next_scan = now +
2215 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2219 * Enforce maximal scan/migration frequency..
2221 migrate = mm->numa_next_scan;
2222 if (time_before(now, migrate))
2225 if (p->numa_scan_period == 0) {
2226 p->numa_scan_period_max = task_scan_max(p);
2227 p->numa_scan_period = task_scan_min(p);
2230 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2231 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2235 * Delay this task enough that another task of this mm will likely win
2236 * the next time around.
2238 p->node_stamp += 2 * TICK_NSEC;
2240 start = mm->numa_scan_offset;
2241 pages = sysctl_numa_balancing_scan_size;
2242 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2243 virtpages = pages * 8; /* Scan up to this much virtual space */
2248 down_read(&mm->mmap_sem);
2249 vma = find_vma(mm, start);
2251 reset_ptenuma_scan(p);
2255 for (; vma; vma = vma->vm_next) {
2256 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2257 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2262 * Shared library pages mapped by multiple processes are not
2263 * migrated as it is expected they are cache replicated. Avoid
2264 * hinting faults in read-only file-backed mappings or the vdso
2265 * as migrating the pages will be of marginal benefit.
2268 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2272 * Skip inaccessible VMAs to avoid any confusion between
2273 * PROT_NONE and NUMA hinting ptes
2275 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2279 start = max(start, vma->vm_start);
2280 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2281 end = min(end, vma->vm_end);
2282 nr_pte_updates = change_prot_numa(vma, start, end);
2285 * Try to scan sysctl_numa_balancing_size worth of
2286 * hpages that have at least one present PTE that
2287 * is not already pte-numa. If the VMA contains
2288 * areas that are unused or already full of prot_numa
2289 * PTEs, scan up to virtpages, to skip through those
2293 pages -= (end - start) >> PAGE_SHIFT;
2294 virtpages -= (end - start) >> PAGE_SHIFT;
2297 if (pages <= 0 || virtpages <= 0)
2301 } while (end != vma->vm_end);
2306 * It is possible to reach the end of the VMA list but the last few
2307 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2308 * would find the !migratable VMA on the next scan but not reset the
2309 * scanner to the start so check it now.
2312 mm->numa_scan_offset = start;
2314 reset_ptenuma_scan(p);
2315 up_read(&mm->mmap_sem);
2319 * Drive the periodic memory faults..
2321 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2323 struct callback_head *work = &curr->numa_work;
2327 * We don't care about NUMA placement if we don't have memory.
2329 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2333 * Using runtime rather than walltime has the dual advantage that
2334 * we (mostly) drive the selection from busy threads and that the
2335 * task needs to have done some actual work before we bother with
2338 now = curr->se.sum_exec_runtime;
2339 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2341 if (now > curr->node_stamp + period) {
2342 if (!curr->node_stamp)
2343 curr->numa_scan_period = task_scan_min(curr);
2344 curr->node_stamp += period;
2346 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2347 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2348 task_work_add(curr, work, true);
2353 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2357 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2361 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2364 #endif /* CONFIG_NUMA_BALANCING */
2367 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2369 update_load_add(&cfs_rq->load, se->load.weight);
2370 if (!parent_entity(se))
2371 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2373 if (entity_is_task(se)) {
2374 struct rq *rq = rq_of(cfs_rq);
2376 account_numa_enqueue(rq, task_of(se));
2377 list_add(&se->group_node, &rq->cfs_tasks);
2380 cfs_rq->nr_running++;
2384 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2386 update_load_sub(&cfs_rq->load, se->load.weight);
2387 if (!parent_entity(se))
2388 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2389 if (entity_is_task(se)) {
2390 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2391 list_del_init(&se->group_node);
2393 cfs_rq->nr_running--;
2396 #ifdef CONFIG_FAIR_GROUP_SCHED
2398 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2403 * Use this CPU's real-time load instead of the last load contribution
2404 * as the updating of the contribution is delayed, and we will use the
2405 * the real-time load to calc the share. See update_tg_load_avg().
2407 tg_weight = atomic_long_read(&tg->load_avg);
2408 tg_weight -= cfs_rq->tg_load_avg_contrib;
2409 tg_weight += cfs_rq->load.weight;
2414 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2416 long tg_weight, load, shares;
2418 tg_weight = calc_tg_weight(tg, cfs_rq);
2419 load = cfs_rq->load.weight;
2421 shares = (tg->shares * load);
2423 shares /= tg_weight;
2425 if (shares < MIN_SHARES)
2426 shares = MIN_SHARES;
2427 if (shares > tg->shares)
2428 shares = tg->shares;
2432 # else /* CONFIG_SMP */
2433 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2437 # endif /* CONFIG_SMP */
2438 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2439 unsigned long weight)
2442 /* commit outstanding execution time */
2443 if (cfs_rq->curr == se)
2444 update_curr(cfs_rq);
2445 account_entity_dequeue(cfs_rq, se);
2448 update_load_set(&se->load, weight);
2451 account_entity_enqueue(cfs_rq, se);
2454 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2456 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2458 struct task_group *tg;
2459 struct sched_entity *se;
2463 se = tg->se[cpu_of(rq_of(cfs_rq))];
2464 if (!se || throttled_hierarchy(cfs_rq))
2467 if (likely(se->load.weight == tg->shares))
2470 shares = calc_cfs_shares(cfs_rq, tg);
2472 reweight_entity(cfs_rq_of(se), se, shares);
2474 #else /* CONFIG_FAIR_GROUP_SCHED */
2475 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2478 #endif /* CONFIG_FAIR_GROUP_SCHED */
2481 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2482 static const u32 runnable_avg_yN_inv[] = {
2483 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2484 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2485 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2486 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2487 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2488 0x85aac367, 0x82cd8698,
2492 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2493 * over-estimates when re-combining.
2495 static const u32 runnable_avg_yN_sum[] = {
2496 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2497 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2498 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2503 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2505 static __always_inline u64 decay_load(u64 val, u64 n)
2507 unsigned int local_n;
2511 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2514 /* after bounds checking we can collapse to 32-bit */
2518 * As y^PERIOD = 1/2, we can combine
2519 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2520 * With a look-up table which covers y^n (n<PERIOD)
2522 * To achieve constant time decay_load.
2524 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2525 val >>= local_n / LOAD_AVG_PERIOD;
2526 local_n %= LOAD_AVG_PERIOD;
2529 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2534 * For updates fully spanning n periods, the contribution to runnable
2535 * average will be: \Sum 1024*y^n
2537 * We can compute this reasonably efficiently by combining:
2538 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2540 static u32 __compute_runnable_contrib(u64 n)
2544 if (likely(n <= LOAD_AVG_PERIOD))
2545 return runnable_avg_yN_sum[n];
2546 else if (unlikely(n >= LOAD_AVG_MAX_N))
2547 return LOAD_AVG_MAX;
2549 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2551 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2552 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2554 n -= LOAD_AVG_PERIOD;
2555 } while (n > LOAD_AVG_PERIOD);
2557 contrib = decay_load(contrib, n);
2558 return contrib + runnable_avg_yN_sum[n];
2561 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2562 #error "load tracking assumes 2^10 as unit"
2565 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2568 * We can represent the historical contribution to runnable average as the
2569 * coefficients of a geometric series. To do this we sub-divide our runnable
2570 * history into segments of approximately 1ms (1024us); label the segment that
2571 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2573 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2575 * (now) (~1ms ago) (~2ms ago)
2577 * Let u_i denote the fraction of p_i that the entity was runnable.
2579 * We then designate the fractions u_i as our co-efficients, yielding the
2580 * following representation of historical load:
2581 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2583 * We choose y based on the with of a reasonably scheduling period, fixing:
2586 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2587 * approximately half as much as the contribution to load within the last ms
2590 * When a period "rolls over" and we have new u_0`, multiplying the previous
2591 * sum again by y is sufficient to update:
2592 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2593 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2595 static __always_inline int
2596 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2597 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2599 u64 delta, scaled_delta, periods;
2601 unsigned int delta_w, scaled_delta_w, decayed = 0;
2602 unsigned long scale_freq, scale_cpu;
2604 delta = now - sa->last_update_time;
2606 * This should only happen when time goes backwards, which it
2607 * unfortunately does during sched clock init when we swap over to TSC.
2609 if ((s64)delta < 0) {
2610 sa->last_update_time = now;
2615 * Use 1024ns as the unit of measurement since it's a reasonable
2616 * approximation of 1us and fast to compute.
2621 sa->last_update_time = now;
2623 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2624 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2625 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2627 /* delta_w is the amount already accumulated against our next period */
2628 delta_w = sa->period_contrib;
2629 if (delta + delta_w >= 1024) {
2632 /* how much left for next period will start over, we don't know yet */
2633 sa->period_contrib = 0;
2636 * Now that we know we're crossing a period boundary, figure
2637 * out how much from delta we need to complete the current
2638 * period and accrue it.
2640 delta_w = 1024 - delta_w;
2641 scaled_delta_w = cap_scale(delta_w, scale_freq);
2643 sa->load_sum += weight * scaled_delta_w;
2645 cfs_rq->runnable_load_sum +=
2646 weight * scaled_delta_w;
2650 sa->util_sum += scaled_delta_w * scale_cpu;
2654 /* Figure out how many additional periods this update spans */
2655 periods = delta / 1024;
2658 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2660 cfs_rq->runnable_load_sum =
2661 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2663 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2665 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2666 contrib = __compute_runnable_contrib(periods);
2667 contrib = cap_scale(contrib, scale_freq);
2669 sa->load_sum += weight * contrib;
2671 cfs_rq->runnable_load_sum += weight * contrib;
2674 sa->util_sum += contrib * scale_cpu;
2677 /* Remainder of delta accrued against u_0` */
2678 scaled_delta = cap_scale(delta, scale_freq);
2680 sa->load_sum += weight * scaled_delta;
2682 cfs_rq->runnable_load_sum += weight * scaled_delta;
2685 sa->util_sum += scaled_delta * scale_cpu;
2687 sa->period_contrib += delta;
2690 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2692 cfs_rq->runnable_load_avg =
2693 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2695 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2701 #ifdef CONFIG_FAIR_GROUP_SCHED
2703 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2704 * and effective_load (which is not done because it is too costly).
2706 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2708 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2710 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2711 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2712 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2716 #else /* CONFIG_FAIR_GROUP_SCHED */
2717 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2718 #endif /* CONFIG_FAIR_GROUP_SCHED */
2720 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2722 if (&this_rq()->cfs == cfs_rq) {
2724 * There are a few boundary cases this might miss but it should
2725 * get called often enough that that should (hopefully) not be
2726 * a real problem -- added to that it only calls on the local
2727 * CPU, so if we enqueue remotely we'll miss an update, but
2728 * the next tick/schedule should update.
2730 * It will not get called when we go idle, because the idle
2731 * thread is a different class (!fair), nor will the utilization
2732 * number include things like RT tasks.
2734 * As is, the util number is not freq-invariant (we'd have to
2735 * implement arch_scale_freq_capacity() for that).
2739 cpufreq_update_util(rq_of(cfs_rq), 0);
2743 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2746 * Unsigned subtract and clamp on underflow.
2748 * Explicitly do a load-store to ensure the intermediate value never hits
2749 * memory. This allows lockless observations without ever seeing the negative
2752 #define sub_positive(_ptr, _val) do { \
2753 typeof(_ptr) ptr = (_ptr); \
2754 typeof(*ptr) val = (_val); \
2755 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2759 WRITE_ONCE(*ptr, res); \
2762 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2763 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq,
2766 struct sched_avg *sa = &cfs_rq->avg;
2767 int decayed, removed = 0, removed_util = 0;
2769 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2770 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2771 sub_positive(&sa->load_avg, r);
2772 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2776 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2777 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2778 sub_positive(&sa->util_avg, r);
2779 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2783 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2784 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2786 #ifndef CONFIG_64BIT
2788 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2791 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2792 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2793 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2795 if (update_freq && (decayed || removed_util))
2796 cfs_rq_util_change(cfs_rq);
2798 return decayed || removed;
2801 /* Update task and its cfs_rq load average */
2802 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2804 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2805 u64 now = cfs_rq_clock_task(cfs_rq);
2806 int cpu = cpu_of(rq_of(cfs_rq));
2809 * Track task load average for carrying it to new CPU after migrated, and
2810 * track group sched_entity load average for task_h_load calc in migration
2812 __update_load_avg(now, cpu, &se->avg,
2813 se->on_rq * scale_load_down(se->load.weight),
2814 cfs_rq->curr == se, NULL);
2816 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2817 update_tg_load_avg(cfs_rq, 0);
2819 if (entity_is_task(se))
2820 trace_sched_load_avg_task(task_of(se), &se->avg);
2823 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2825 if (!sched_feat(ATTACH_AGE_LOAD))
2829 * If we got migrated (either between CPUs or between cgroups) we'll
2830 * have aged the average right before clearing @last_update_time.
2832 if (se->avg.last_update_time) {
2833 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2834 &se->avg, 0, 0, NULL);
2837 * XXX: we could have just aged the entire load away if we've been
2838 * absent from the fair class for too long.
2843 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2844 cfs_rq->avg.load_avg += se->avg.load_avg;
2845 cfs_rq->avg.load_sum += se->avg.load_sum;
2846 cfs_rq->avg.util_avg += se->avg.util_avg;
2847 cfs_rq->avg.util_sum += se->avg.util_sum;
2849 cfs_rq_util_change(cfs_rq);
2852 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2854 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2855 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2856 cfs_rq->curr == se, NULL);
2858 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2859 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2860 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2861 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2863 cfs_rq_util_change(cfs_rq);
2866 /* Add the load generated by se into cfs_rq's load average */
2868 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2870 struct sched_avg *sa = &se->avg;
2871 u64 now = cfs_rq_clock_task(cfs_rq);
2872 int migrated, decayed;
2874 migrated = !sa->last_update_time;
2876 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2877 se->on_rq * scale_load_down(se->load.weight),
2878 cfs_rq->curr == se, NULL);
2881 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
2883 cfs_rq->runnable_load_avg += sa->load_avg;
2884 cfs_rq->runnable_load_sum += sa->load_sum;
2887 attach_entity_load_avg(cfs_rq, se);
2889 if (decayed || migrated)
2890 update_tg_load_avg(cfs_rq, 0);
2893 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2895 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2897 update_load_avg(se, 1);
2899 cfs_rq->runnable_load_avg =
2900 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2901 cfs_rq->runnable_load_sum =
2902 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2905 #ifndef CONFIG_64BIT
2906 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2908 u64 last_update_time_copy;
2909 u64 last_update_time;
2912 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2914 last_update_time = cfs_rq->avg.last_update_time;
2915 } while (last_update_time != last_update_time_copy);
2917 return last_update_time;
2920 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2922 return cfs_rq->avg.last_update_time;
2927 * Synchronize entity load avg of dequeued entity without locking
2930 void sync_entity_load_avg(struct sched_entity *se)
2932 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2933 u64 last_update_time;
2935 last_update_time = cfs_rq_last_update_time(cfs_rq);
2936 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2940 * Task first catches up with cfs_rq, and then subtract
2941 * itself from the cfs_rq (task must be off the queue now).
2943 void remove_entity_load_avg(struct sched_entity *se)
2945 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2948 * Newly created task or never used group entity should not be removed
2949 * from its (source) cfs_rq
2951 if (se->avg.last_update_time == 0)
2954 sync_entity_load_avg(se);
2955 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2956 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2960 * Update the rq's load with the elapsed running time before entering
2961 * idle. if the last scheduled task is not a CFS task, idle_enter will
2962 * be the only way to update the runnable statistic.
2964 void idle_enter_fair(struct rq *this_rq)
2969 * Update the rq's load with the elapsed idle time before a task is
2970 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2971 * be the only way to update the runnable statistic.
2973 void idle_exit_fair(struct rq *this_rq)
2977 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2979 return cfs_rq->runnable_load_avg;
2982 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2984 return cfs_rq->avg.load_avg;
2987 static int idle_balance(struct rq *this_rq);
2989 #else /* CONFIG_SMP */
2991 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2993 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
2997 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2999 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3000 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3003 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3005 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3007 static inline int idle_balance(struct rq *rq)
3012 #endif /* CONFIG_SMP */
3014 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3016 #ifdef CONFIG_SCHEDSTATS
3017 struct task_struct *tsk = NULL;
3019 if (entity_is_task(se))
3022 if (se->statistics.sleep_start) {
3023 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3028 if (unlikely(delta > se->statistics.sleep_max))
3029 se->statistics.sleep_max = delta;
3031 se->statistics.sleep_start = 0;
3032 se->statistics.sum_sleep_runtime += delta;
3035 account_scheduler_latency(tsk, delta >> 10, 1);
3036 trace_sched_stat_sleep(tsk, delta);
3039 if (se->statistics.block_start) {
3040 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3045 if (unlikely(delta > se->statistics.block_max))
3046 se->statistics.block_max = delta;
3048 se->statistics.block_start = 0;
3049 se->statistics.sum_sleep_runtime += delta;
3052 if (tsk->in_iowait) {
3053 se->statistics.iowait_sum += delta;
3054 se->statistics.iowait_count++;
3055 trace_sched_stat_iowait(tsk, delta);
3058 trace_sched_stat_blocked(tsk, delta);
3059 trace_sched_blocked_reason(tsk);
3062 * Blocking time is in units of nanosecs, so shift by
3063 * 20 to get a milliseconds-range estimation of the
3064 * amount of time that the task spent sleeping:
3066 if (unlikely(prof_on == SLEEP_PROFILING)) {
3067 profile_hits(SLEEP_PROFILING,
3068 (void *)get_wchan(tsk),
3071 account_scheduler_latency(tsk, delta >> 10, 0);
3077 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3079 #ifdef CONFIG_SCHED_DEBUG
3080 s64 d = se->vruntime - cfs_rq->min_vruntime;
3085 if (d > 3*sysctl_sched_latency)
3086 schedstat_inc(cfs_rq, nr_spread_over);
3091 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3093 u64 vruntime = cfs_rq->min_vruntime;
3096 * The 'current' period is already promised to the current tasks,
3097 * however the extra weight of the new task will slow them down a
3098 * little, place the new task so that it fits in the slot that
3099 * stays open at the end.
3101 if (initial && sched_feat(START_DEBIT))
3102 vruntime += sched_vslice(cfs_rq, se);
3104 /* sleeps up to a single latency don't count. */
3106 unsigned long thresh = sysctl_sched_latency;
3109 * Halve their sleep time's effect, to allow
3110 * for a gentler effect of sleepers:
3112 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3118 /* ensure we never gain time by being placed backwards. */
3119 se->vruntime = max_vruntime(se->vruntime, vruntime);
3122 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3125 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3128 * Update the normalized vruntime before updating min_vruntime
3129 * through calling update_curr().
3131 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3132 se->vruntime += cfs_rq->min_vruntime;
3135 * Update run-time statistics of the 'current'.
3137 update_curr(cfs_rq);
3138 enqueue_entity_load_avg(cfs_rq, se);
3139 account_entity_enqueue(cfs_rq, se);
3140 update_cfs_shares(cfs_rq);
3142 if (flags & ENQUEUE_WAKEUP) {
3143 place_entity(cfs_rq, se, 0);
3144 enqueue_sleeper(cfs_rq, se);
3147 update_stats_enqueue(cfs_rq, se);
3148 check_spread(cfs_rq, se);
3149 if (se != cfs_rq->curr)
3150 __enqueue_entity(cfs_rq, se);
3153 if (cfs_rq->nr_running == 1) {
3154 list_add_leaf_cfs_rq(cfs_rq);
3155 check_enqueue_throttle(cfs_rq);
3159 static void __clear_buddies_last(struct sched_entity *se)
3161 for_each_sched_entity(se) {
3162 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3163 if (cfs_rq->last != se)
3166 cfs_rq->last = NULL;
3170 static void __clear_buddies_next(struct sched_entity *se)
3172 for_each_sched_entity(se) {
3173 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3174 if (cfs_rq->next != se)
3177 cfs_rq->next = NULL;
3181 static void __clear_buddies_skip(struct sched_entity *se)
3183 for_each_sched_entity(se) {
3184 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3185 if (cfs_rq->skip != se)
3188 cfs_rq->skip = NULL;
3192 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3194 if (cfs_rq->last == se)
3195 __clear_buddies_last(se);
3197 if (cfs_rq->next == se)
3198 __clear_buddies_next(se);
3200 if (cfs_rq->skip == se)
3201 __clear_buddies_skip(se);
3204 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3207 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3210 * Update run-time statistics of the 'current'.
3212 update_curr(cfs_rq);
3213 dequeue_entity_load_avg(cfs_rq, se);
3215 update_stats_dequeue(cfs_rq, se);
3216 if (flags & DEQUEUE_SLEEP) {
3217 #ifdef CONFIG_SCHEDSTATS
3218 if (entity_is_task(se)) {
3219 struct task_struct *tsk = task_of(se);
3221 if (tsk->state & TASK_INTERRUPTIBLE)
3222 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3223 if (tsk->state & TASK_UNINTERRUPTIBLE)
3224 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3229 clear_buddies(cfs_rq, se);
3231 if (se != cfs_rq->curr)
3232 __dequeue_entity(cfs_rq, se);
3234 account_entity_dequeue(cfs_rq, se);
3237 * Normalize the entity after updating the min_vruntime because the
3238 * update can refer to the ->curr item and we need to reflect this
3239 * movement in our normalized position.
3241 if (!(flags & DEQUEUE_SLEEP))
3242 se->vruntime -= cfs_rq->min_vruntime;
3244 /* return excess runtime on last dequeue */
3245 return_cfs_rq_runtime(cfs_rq);
3247 update_min_vruntime(cfs_rq);
3248 update_cfs_shares(cfs_rq);
3252 * Preempt the current task with a newly woken task if needed:
3255 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3257 unsigned long ideal_runtime, delta_exec;
3258 struct sched_entity *se;
3261 ideal_runtime = sched_slice(cfs_rq, curr);
3262 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3263 if (delta_exec > ideal_runtime) {
3264 resched_curr(rq_of(cfs_rq));
3266 * The current task ran long enough, ensure it doesn't get
3267 * re-elected due to buddy favours.
3269 clear_buddies(cfs_rq, curr);
3274 * Ensure that a task that missed wakeup preemption by a
3275 * narrow margin doesn't have to wait for a full slice.
3276 * This also mitigates buddy induced latencies under load.
3278 if (delta_exec < sysctl_sched_min_granularity)
3281 se = __pick_first_entity(cfs_rq);
3282 delta = curr->vruntime - se->vruntime;
3287 if (delta > ideal_runtime)
3288 resched_curr(rq_of(cfs_rq));
3292 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3294 /* 'current' is not kept within the tree. */
3297 * Any task has to be enqueued before it get to execute on
3298 * a CPU. So account for the time it spent waiting on the
3301 update_stats_wait_end(cfs_rq, se);
3302 __dequeue_entity(cfs_rq, se);
3303 update_load_avg(se, 1);
3306 update_stats_curr_start(cfs_rq, se);
3308 #ifdef CONFIG_SCHEDSTATS
3310 * Track our maximum slice length, if the CPU's load is at
3311 * least twice that of our own weight (i.e. dont track it
3312 * when there are only lesser-weight tasks around):
3314 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3315 se->statistics.slice_max = max(se->statistics.slice_max,
3316 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3319 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3323 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3326 * Pick the next process, keeping these things in mind, in this order:
3327 * 1) keep things fair between processes/task groups
3328 * 2) pick the "next" process, since someone really wants that to run
3329 * 3) pick the "last" process, for cache locality
3330 * 4) do not run the "skip" process, if something else is available
3332 static struct sched_entity *
3333 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3335 struct sched_entity *left = __pick_first_entity(cfs_rq);
3336 struct sched_entity *se;
3339 * If curr is set we have to see if its left of the leftmost entity
3340 * still in the tree, provided there was anything in the tree at all.
3342 if (!left || (curr && entity_before(curr, left)))
3345 se = left; /* ideally we run the leftmost entity */
3348 * Avoid running the skip buddy, if running something else can
3349 * be done without getting too unfair.
3351 if (cfs_rq->skip == se) {
3352 struct sched_entity *second;
3355 second = __pick_first_entity(cfs_rq);
3357 second = __pick_next_entity(se);
3358 if (!second || (curr && entity_before(curr, second)))
3362 if (second && wakeup_preempt_entity(second, left) < 1)
3367 * Prefer last buddy, try to return the CPU to a preempted task.
3369 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3373 * Someone really wants this to run. If it's not unfair, run it.
3375 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3378 clear_buddies(cfs_rq, se);
3383 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3385 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3388 * If still on the runqueue then deactivate_task()
3389 * was not called and update_curr() has to be done:
3392 update_curr(cfs_rq);
3394 /* throttle cfs_rqs exceeding runtime */
3395 check_cfs_rq_runtime(cfs_rq);
3397 check_spread(cfs_rq, prev);
3399 update_stats_wait_start(cfs_rq, prev);
3400 /* Put 'current' back into the tree. */
3401 __enqueue_entity(cfs_rq, prev);
3402 /* in !on_rq case, update occurred at dequeue */
3403 update_load_avg(prev, 0);
3405 cfs_rq->curr = NULL;
3409 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3412 * Update run-time statistics of the 'current'.
3414 update_curr(cfs_rq);
3417 * Ensure that runnable average is periodically updated.
3419 update_load_avg(curr, 1);
3420 update_cfs_shares(cfs_rq);
3422 #ifdef CONFIG_SCHED_HRTICK
3424 * queued ticks are scheduled to match the slice, so don't bother
3425 * validating it and just reschedule.
3428 resched_curr(rq_of(cfs_rq));
3432 * don't let the period tick interfere with the hrtick preemption
3434 if (!sched_feat(DOUBLE_TICK) &&
3435 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3439 if (cfs_rq->nr_running > 1)
3440 check_preempt_tick(cfs_rq, curr);
3444 /**************************************************
3445 * CFS bandwidth control machinery
3448 #ifdef CONFIG_CFS_BANDWIDTH
3450 #ifdef HAVE_JUMP_LABEL
3451 static struct static_key __cfs_bandwidth_used;
3453 static inline bool cfs_bandwidth_used(void)
3455 return static_key_false(&__cfs_bandwidth_used);
3458 void cfs_bandwidth_usage_inc(void)
3460 static_key_slow_inc(&__cfs_bandwidth_used);
3463 void cfs_bandwidth_usage_dec(void)
3465 static_key_slow_dec(&__cfs_bandwidth_used);
3467 #else /* HAVE_JUMP_LABEL */
3468 static bool cfs_bandwidth_used(void)
3473 void cfs_bandwidth_usage_inc(void) {}
3474 void cfs_bandwidth_usage_dec(void) {}
3475 #endif /* HAVE_JUMP_LABEL */
3478 * default period for cfs group bandwidth.
3479 * default: 0.1s, units: nanoseconds
3481 static inline u64 default_cfs_period(void)
3483 return 100000000ULL;
3486 static inline u64 sched_cfs_bandwidth_slice(void)
3488 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3492 * Replenish runtime according to assigned quota and update expiration time.
3493 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3494 * additional synchronization around rq->lock.
3496 * requires cfs_b->lock
3498 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3502 if (cfs_b->quota == RUNTIME_INF)
3505 now = sched_clock_cpu(smp_processor_id());
3506 cfs_b->runtime = cfs_b->quota;
3507 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3510 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3512 return &tg->cfs_bandwidth;
3515 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3516 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3518 if (unlikely(cfs_rq->throttle_count))
3519 return cfs_rq->throttled_clock_task;
3521 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3524 /* returns 0 on failure to allocate runtime */
3525 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3527 struct task_group *tg = cfs_rq->tg;
3528 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3529 u64 amount = 0, min_amount, expires;
3531 /* note: this is a positive sum as runtime_remaining <= 0 */
3532 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3534 raw_spin_lock(&cfs_b->lock);
3535 if (cfs_b->quota == RUNTIME_INF)
3536 amount = min_amount;
3538 start_cfs_bandwidth(cfs_b);
3540 if (cfs_b->runtime > 0) {
3541 amount = min(cfs_b->runtime, min_amount);
3542 cfs_b->runtime -= amount;
3546 expires = cfs_b->runtime_expires;
3547 raw_spin_unlock(&cfs_b->lock);
3549 cfs_rq->runtime_remaining += amount;
3551 * we may have advanced our local expiration to account for allowed
3552 * spread between our sched_clock and the one on which runtime was
3555 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3556 cfs_rq->runtime_expires = expires;
3558 return cfs_rq->runtime_remaining > 0;
3562 * Note: This depends on the synchronization provided by sched_clock and the
3563 * fact that rq->clock snapshots this value.
3565 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3567 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3569 /* if the deadline is ahead of our clock, nothing to do */
3570 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3573 if (cfs_rq->runtime_remaining < 0)
3577 * If the local deadline has passed we have to consider the
3578 * possibility that our sched_clock is 'fast' and the global deadline
3579 * has not truly expired.
3581 * Fortunately we can check determine whether this the case by checking
3582 * whether the global deadline has advanced. It is valid to compare
3583 * cfs_b->runtime_expires without any locks since we only care about
3584 * exact equality, so a partial write will still work.
3587 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3588 /* extend local deadline, drift is bounded above by 2 ticks */
3589 cfs_rq->runtime_expires += TICK_NSEC;
3591 /* global deadline is ahead, expiration has passed */
3592 cfs_rq->runtime_remaining = 0;
3596 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3598 /* dock delta_exec before expiring quota (as it could span periods) */
3599 cfs_rq->runtime_remaining -= delta_exec;
3600 expire_cfs_rq_runtime(cfs_rq);
3602 if (likely(cfs_rq->runtime_remaining > 0))
3606 * if we're unable to extend our runtime we resched so that the active
3607 * hierarchy can be throttled
3609 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3610 resched_curr(rq_of(cfs_rq));
3613 static __always_inline
3614 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3616 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3619 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3622 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3624 return cfs_bandwidth_used() && cfs_rq->throttled;
3627 /* check whether cfs_rq, or any parent, is throttled */
3628 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3630 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3634 * Ensure that neither of the group entities corresponding to src_cpu or
3635 * dest_cpu are members of a throttled hierarchy when performing group
3636 * load-balance operations.
3638 static inline int throttled_lb_pair(struct task_group *tg,
3639 int src_cpu, int dest_cpu)
3641 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3643 src_cfs_rq = tg->cfs_rq[src_cpu];
3644 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3646 return throttled_hierarchy(src_cfs_rq) ||
3647 throttled_hierarchy(dest_cfs_rq);
3650 /* updated child weight may affect parent so we have to do this bottom up */
3651 static int tg_unthrottle_up(struct task_group *tg, void *data)
3653 struct rq *rq = data;
3654 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3656 cfs_rq->throttle_count--;
3658 if (!cfs_rq->throttle_count) {
3659 /* adjust cfs_rq_clock_task() */
3660 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3661 cfs_rq->throttled_clock_task;
3668 static int tg_throttle_down(struct task_group *tg, void *data)
3670 struct rq *rq = data;
3671 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3673 /* group is entering throttled state, stop time */
3674 if (!cfs_rq->throttle_count)
3675 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3676 cfs_rq->throttle_count++;
3681 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3683 struct rq *rq = rq_of(cfs_rq);
3684 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3685 struct sched_entity *se;
3686 long task_delta, dequeue = 1;
3689 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3691 /* freeze hierarchy runnable averages while throttled */
3693 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3696 task_delta = cfs_rq->h_nr_running;
3697 for_each_sched_entity(se) {
3698 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3699 /* throttled entity or throttle-on-deactivate */
3704 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3705 qcfs_rq->h_nr_running -= task_delta;
3707 if (qcfs_rq->load.weight)
3712 sub_nr_running(rq, task_delta);
3714 cfs_rq->throttled = 1;
3715 cfs_rq->throttled_clock = rq_clock(rq);
3716 raw_spin_lock(&cfs_b->lock);
3717 empty = list_empty(&cfs_b->throttled_cfs_rq);
3720 * Add to the _head_ of the list, so that an already-started
3721 * distribute_cfs_runtime will not see us
3723 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3726 * If we're the first throttled task, make sure the bandwidth
3730 start_cfs_bandwidth(cfs_b);
3732 raw_spin_unlock(&cfs_b->lock);
3735 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3737 struct rq *rq = rq_of(cfs_rq);
3738 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3739 struct sched_entity *se;
3743 se = cfs_rq->tg->se[cpu_of(rq)];
3745 cfs_rq->throttled = 0;
3747 update_rq_clock(rq);
3749 raw_spin_lock(&cfs_b->lock);
3750 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3751 list_del_rcu(&cfs_rq->throttled_list);
3752 raw_spin_unlock(&cfs_b->lock);
3754 /* update hierarchical throttle state */
3755 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3757 if (!cfs_rq->load.weight)
3760 task_delta = cfs_rq->h_nr_running;
3761 for_each_sched_entity(se) {
3765 cfs_rq = cfs_rq_of(se);
3767 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3768 cfs_rq->h_nr_running += task_delta;
3770 if (cfs_rq_throttled(cfs_rq))
3775 add_nr_running(rq, task_delta);
3777 /* determine whether we need to wake up potentially idle cpu */
3778 if (rq->curr == rq->idle && rq->cfs.nr_running)
3782 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3783 u64 remaining, u64 expires)
3785 struct cfs_rq *cfs_rq;
3787 u64 starting_runtime = remaining;
3790 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3792 struct rq *rq = rq_of(cfs_rq);
3794 raw_spin_lock(&rq->lock);
3795 if (!cfs_rq_throttled(cfs_rq))
3798 runtime = -cfs_rq->runtime_remaining + 1;
3799 if (runtime > remaining)
3800 runtime = remaining;
3801 remaining -= runtime;
3803 cfs_rq->runtime_remaining += runtime;
3804 cfs_rq->runtime_expires = expires;
3806 /* we check whether we're throttled above */
3807 if (cfs_rq->runtime_remaining > 0)
3808 unthrottle_cfs_rq(cfs_rq);
3811 raw_spin_unlock(&rq->lock);
3818 return starting_runtime - remaining;
3822 * Responsible for refilling a task_group's bandwidth and unthrottling its
3823 * cfs_rqs as appropriate. If there has been no activity within the last
3824 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3825 * used to track this state.
3827 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3829 u64 runtime, runtime_expires;
3832 /* no need to continue the timer with no bandwidth constraint */
3833 if (cfs_b->quota == RUNTIME_INF)
3834 goto out_deactivate;
3836 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3837 cfs_b->nr_periods += overrun;
3840 * idle depends on !throttled (for the case of a large deficit), and if
3841 * we're going inactive then everything else can be deferred
3843 if (cfs_b->idle && !throttled)
3844 goto out_deactivate;
3846 __refill_cfs_bandwidth_runtime(cfs_b);
3849 /* mark as potentially idle for the upcoming period */
3854 /* account preceding periods in which throttling occurred */
3855 cfs_b->nr_throttled += overrun;
3857 runtime_expires = cfs_b->runtime_expires;
3860 * This check is repeated as we are holding onto the new bandwidth while
3861 * we unthrottle. This can potentially race with an unthrottled group
3862 * trying to acquire new bandwidth from the global pool. This can result
3863 * in us over-using our runtime if it is all used during this loop, but
3864 * only by limited amounts in that extreme case.
3866 while (throttled && cfs_b->runtime > 0) {
3867 runtime = cfs_b->runtime;
3868 raw_spin_unlock(&cfs_b->lock);
3869 /* we can't nest cfs_b->lock while distributing bandwidth */
3870 runtime = distribute_cfs_runtime(cfs_b, runtime,
3872 raw_spin_lock(&cfs_b->lock);
3874 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3876 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3880 * While we are ensured activity in the period following an
3881 * unthrottle, this also covers the case in which the new bandwidth is
3882 * insufficient to cover the existing bandwidth deficit. (Forcing the
3883 * timer to remain active while there are any throttled entities.)
3893 /* a cfs_rq won't donate quota below this amount */
3894 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3895 /* minimum remaining period time to redistribute slack quota */
3896 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3897 /* how long we wait to gather additional slack before distributing */
3898 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3901 * Are we near the end of the current quota period?
3903 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3904 * hrtimer base being cleared by hrtimer_start. In the case of
3905 * migrate_hrtimers, base is never cleared, so we are fine.
3907 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3909 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3912 /* if the call-back is running a quota refresh is already occurring */
3913 if (hrtimer_callback_running(refresh_timer))
3916 /* is a quota refresh about to occur? */
3917 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3918 if (remaining < min_expire)
3924 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3926 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3928 /* if there's a quota refresh soon don't bother with slack */
3929 if (runtime_refresh_within(cfs_b, min_left))
3932 hrtimer_start(&cfs_b->slack_timer,
3933 ns_to_ktime(cfs_bandwidth_slack_period),
3937 /* we know any runtime found here is valid as update_curr() precedes return */
3938 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3940 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3941 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3943 if (slack_runtime <= 0)
3946 raw_spin_lock(&cfs_b->lock);
3947 if (cfs_b->quota != RUNTIME_INF &&
3948 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3949 cfs_b->runtime += slack_runtime;
3951 /* we are under rq->lock, defer unthrottling using a timer */
3952 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3953 !list_empty(&cfs_b->throttled_cfs_rq))
3954 start_cfs_slack_bandwidth(cfs_b);
3956 raw_spin_unlock(&cfs_b->lock);
3958 /* even if it's not valid for return we don't want to try again */
3959 cfs_rq->runtime_remaining -= slack_runtime;
3962 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3964 if (!cfs_bandwidth_used())
3967 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3970 __return_cfs_rq_runtime(cfs_rq);
3974 * This is done with a timer (instead of inline with bandwidth return) since
3975 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3977 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3979 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3982 /* confirm we're still not at a refresh boundary */
3983 raw_spin_lock(&cfs_b->lock);
3984 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3985 raw_spin_unlock(&cfs_b->lock);
3989 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3990 runtime = cfs_b->runtime;
3992 expires = cfs_b->runtime_expires;
3993 raw_spin_unlock(&cfs_b->lock);
3998 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4000 raw_spin_lock(&cfs_b->lock);
4001 if (expires == cfs_b->runtime_expires)
4002 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4003 raw_spin_unlock(&cfs_b->lock);
4007 * When a group wakes up we want to make sure that its quota is not already
4008 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4009 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4011 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4013 if (!cfs_bandwidth_used())
4016 /* Synchronize hierarchical throttle counter: */
4017 if (unlikely(!cfs_rq->throttle_uptodate)) {
4018 struct rq *rq = rq_of(cfs_rq);
4019 struct cfs_rq *pcfs_rq;
4020 struct task_group *tg;
4022 cfs_rq->throttle_uptodate = 1;
4024 /* Get closest up-to-date node, because leaves go first: */
4025 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4026 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4027 if (pcfs_rq->throttle_uptodate)
4031 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4032 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4036 /* an active group must be handled by the update_curr()->put() path */
4037 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4040 /* ensure the group is not already throttled */
4041 if (cfs_rq_throttled(cfs_rq))
4044 /* update runtime allocation */
4045 account_cfs_rq_runtime(cfs_rq, 0);
4046 if (cfs_rq->runtime_remaining <= 0)
4047 throttle_cfs_rq(cfs_rq);
4050 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4051 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4053 if (!cfs_bandwidth_used())
4056 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4060 * it's possible for a throttled entity to be forced into a running
4061 * state (e.g. set_curr_task), in this case we're finished.
4063 if (cfs_rq_throttled(cfs_rq))
4066 throttle_cfs_rq(cfs_rq);
4070 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4072 struct cfs_bandwidth *cfs_b =
4073 container_of(timer, struct cfs_bandwidth, slack_timer);
4075 do_sched_cfs_slack_timer(cfs_b);
4077 return HRTIMER_NORESTART;
4080 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4082 struct cfs_bandwidth *cfs_b =
4083 container_of(timer, struct cfs_bandwidth, period_timer);
4087 raw_spin_lock(&cfs_b->lock);
4089 overrun = hrtimer_forward_now(timer, cfs_b->period);
4093 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4096 cfs_b->period_active = 0;
4097 raw_spin_unlock(&cfs_b->lock);
4099 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4102 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4104 raw_spin_lock_init(&cfs_b->lock);
4106 cfs_b->quota = RUNTIME_INF;
4107 cfs_b->period = ns_to_ktime(default_cfs_period());
4109 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4110 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4111 cfs_b->period_timer.function = sched_cfs_period_timer;
4112 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4113 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4116 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4118 cfs_rq->runtime_enabled = 0;
4119 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4122 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4124 lockdep_assert_held(&cfs_b->lock);
4126 if (!cfs_b->period_active) {
4127 cfs_b->period_active = 1;
4128 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4129 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4133 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4135 /* init_cfs_bandwidth() was not called */
4136 if (!cfs_b->throttled_cfs_rq.next)
4139 hrtimer_cancel(&cfs_b->period_timer);
4140 hrtimer_cancel(&cfs_b->slack_timer);
4143 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4145 struct cfs_rq *cfs_rq;
4147 for_each_leaf_cfs_rq(rq, cfs_rq) {
4148 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4150 raw_spin_lock(&cfs_b->lock);
4151 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4152 raw_spin_unlock(&cfs_b->lock);
4156 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4158 struct cfs_rq *cfs_rq;
4160 for_each_leaf_cfs_rq(rq, cfs_rq) {
4161 if (!cfs_rq->runtime_enabled)
4165 * clock_task is not advancing so we just need to make sure
4166 * there's some valid quota amount
4168 cfs_rq->runtime_remaining = 1;
4170 * Offline rq is schedulable till cpu is completely disabled
4171 * in take_cpu_down(), so we prevent new cfs throttling here.
4173 cfs_rq->runtime_enabled = 0;
4175 if (cfs_rq_throttled(cfs_rq))
4176 unthrottle_cfs_rq(cfs_rq);
4180 #else /* CONFIG_CFS_BANDWIDTH */
4181 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4183 return rq_clock_task(rq_of(cfs_rq));
4186 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4187 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4188 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4189 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4191 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4196 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4201 static inline int throttled_lb_pair(struct task_group *tg,
4202 int src_cpu, int dest_cpu)
4207 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4209 #ifdef CONFIG_FAIR_GROUP_SCHED
4210 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4213 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4217 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4218 static inline void update_runtime_enabled(struct rq *rq) {}
4219 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4221 #endif /* CONFIG_CFS_BANDWIDTH */
4223 /**************************************************
4224 * CFS operations on tasks:
4227 #ifdef CONFIG_SCHED_HRTICK
4228 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4230 struct sched_entity *se = &p->se;
4231 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4233 WARN_ON(task_rq(p) != rq);
4235 if (cfs_rq->nr_running > 1) {
4236 u64 slice = sched_slice(cfs_rq, se);
4237 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4238 s64 delta = slice - ran;
4245 hrtick_start(rq, delta);
4250 * called from enqueue/dequeue and updates the hrtick when the
4251 * current task is from our class and nr_running is low enough
4254 static void hrtick_update(struct rq *rq)
4256 struct task_struct *curr = rq->curr;
4258 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4261 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4262 hrtick_start_fair(rq, curr);
4264 #else /* !CONFIG_SCHED_HRTICK */
4266 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4270 static inline void hrtick_update(struct rq *rq)
4276 static bool cpu_overutilized(int cpu);
4277 unsigned long boosted_cpu_util(int cpu);
4279 #define boosted_cpu_util(cpu) cpu_util(cpu)
4283 static void update_capacity_of(int cpu)
4285 unsigned long req_cap;
4290 /* Convert scale-invariant capacity to cpu. */
4291 req_cap = boosted_cpu_util(cpu);
4292 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4293 set_cfs_cpu_capacity(cpu, true, req_cap);
4298 * The enqueue_task method is called before nr_running is
4299 * increased. Here we update the fair scheduling stats and
4300 * then put the task into the rbtree:
4303 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4305 struct cfs_rq *cfs_rq;
4306 struct sched_entity *se = &p->se;
4308 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4309 int task_wakeup = flags & ENQUEUE_WAKEUP;
4313 * If in_iowait is set, the code below may not trigger any cpufreq
4314 * utilization updates, so do it here explicitly with the IOWAIT flag
4318 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4320 for_each_sched_entity(se) {
4323 cfs_rq = cfs_rq_of(se);
4324 enqueue_entity(cfs_rq, se, flags);
4327 * end evaluation on encountering a throttled cfs_rq
4329 * note: in the case of encountering a throttled cfs_rq we will
4330 * post the final h_nr_running increment below.
4332 if (cfs_rq_throttled(cfs_rq))
4334 cfs_rq->h_nr_running++;
4335 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4337 flags = ENQUEUE_WAKEUP;
4340 for_each_sched_entity(se) {
4341 cfs_rq = cfs_rq_of(se);
4342 cfs_rq->h_nr_running++;
4343 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4345 if (cfs_rq_throttled(cfs_rq))
4348 update_load_avg(se, 1);
4349 update_cfs_shares(cfs_rq);
4353 add_nr_running(rq, 1);
4358 * Update SchedTune accounting.
4360 * We do it before updating the CPU capacity to ensure the
4361 * boost value of the current task is accounted for in the
4362 * selection of the OPP.
4364 * We do it also in the case where we enqueue a throttled task;
4365 * we could argue that a throttled task should not boost a CPU,
4367 * a) properly implementing CPU boosting considering throttled
4368 * tasks will increase a lot the complexity of the solution
4369 * b) it's not easy to quantify the benefits introduced by
4370 * such a more complex solution.
4371 * Thus, for the time being we go for the simple solution and boost
4372 * also for throttled RQs.
4374 schedtune_enqueue_task(p, cpu_of(rq));
4377 walt_inc_cumulative_runnable_avg(rq, p);
4378 if (!task_new && !rq->rd->overutilized &&
4379 cpu_overutilized(rq->cpu)) {
4380 rq->rd->overutilized = true;
4381 trace_sched_overutilized(true);
4385 * We want to potentially trigger a freq switch
4386 * request only for tasks that are waking up; this is
4387 * because we get here also during load balancing, but
4388 * in these cases it seems wise to trigger as single
4389 * request after load balancing is done.
4391 if (task_new || task_wakeup)
4392 update_capacity_of(cpu_of(rq));
4395 #endif /* CONFIG_SMP */
4399 static void set_next_buddy(struct sched_entity *se);
4402 * The dequeue_task method is called before nr_running is
4403 * decreased. We remove the task from the rbtree and
4404 * update the fair scheduling stats:
4406 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4408 struct cfs_rq *cfs_rq;
4409 struct sched_entity *se = &p->se;
4410 int task_sleep = flags & DEQUEUE_SLEEP;
4412 for_each_sched_entity(se) {
4413 cfs_rq = cfs_rq_of(se);
4414 dequeue_entity(cfs_rq, se, flags);
4417 * end evaluation on encountering a throttled cfs_rq
4419 * note: in the case of encountering a throttled cfs_rq we will
4420 * post the final h_nr_running decrement below.
4422 if (cfs_rq_throttled(cfs_rq))
4424 cfs_rq->h_nr_running--;
4425 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4427 /* Don't dequeue parent if it has other entities besides us */
4428 if (cfs_rq->load.weight) {
4429 /* Avoid re-evaluating load for this entity: */
4430 se = parent_entity(se);
4432 * Bias pick_next to pick a task from this cfs_rq, as
4433 * p is sleeping when it is within its sched_slice.
4435 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4439 flags |= DEQUEUE_SLEEP;
4442 for_each_sched_entity(se) {
4443 cfs_rq = cfs_rq_of(se);
4444 cfs_rq->h_nr_running--;
4445 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4447 if (cfs_rq_throttled(cfs_rq))
4450 update_load_avg(se, 1);
4451 update_cfs_shares(cfs_rq);
4455 sub_nr_running(rq, 1);
4460 * Update SchedTune accounting
4462 * We do it before updating the CPU capacity to ensure the
4463 * boost value of the current task is accounted for in the
4464 * selection of the OPP.
4466 schedtune_dequeue_task(p, cpu_of(rq));
4469 walt_dec_cumulative_runnable_avg(rq, p);
4472 * We want to potentially trigger a freq switch
4473 * request only for tasks that are going to sleep;
4474 * this is because we get here also during load
4475 * balancing, but in these cases it seems wise to
4476 * trigger as single request after load balancing is
4480 if (rq->cfs.nr_running)
4481 update_capacity_of(cpu_of(rq));
4482 else if (sched_freq())
4483 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4487 #endif /* CONFIG_SMP */
4495 * per rq 'load' arrray crap; XXX kill this.
4499 * The exact cpuload at various idx values, calculated at every tick would be
4500 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4502 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4503 * on nth tick when cpu may be busy, then we have:
4504 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4505 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4507 * decay_load_missed() below does efficient calculation of
4508 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4509 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4511 * The calculation is approximated on a 128 point scale.
4512 * degrade_zero_ticks is the number of ticks after which load at any
4513 * particular idx is approximated to be zero.
4514 * degrade_factor is a precomputed table, a row for each load idx.
4515 * Each column corresponds to degradation factor for a power of two ticks,
4516 * based on 128 point scale.
4518 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4519 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4521 * With this power of 2 load factors, we can degrade the load n times
4522 * by looking at 1 bits in n and doing as many mult/shift instead of
4523 * n mult/shifts needed by the exact degradation.
4525 #define DEGRADE_SHIFT 7
4526 static const unsigned char
4527 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4528 static const unsigned char
4529 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4530 {0, 0, 0, 0, 0, 0, 0, 0},
4531 {64, 32, 8, 0, 0, 0, 0, 0},
4532 {96, 72, 40, 12, 1, 0, 0},
4533 {112, 98, 75, 43, 15, 1, 0},
4534 {120, 112, 98, 76, 45, 16, 2} };
4537 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4538 * would be when CPU is idle and so we just decay the old load without
4539 * adding any new load.
4541 static unsigned long
4542 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4546 if (!missed_updates)
4549 if (missed_updates >= degrade_zero_ticks[idx])
4553 return load >> missed_updates;
4555 while (missed_updates) {
4556 if (missed_updates % 2)
4557 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4559 missed_updates >>= 1;
4566 * Update rq->cpu_load[] statistics. This function is usually called every
4567 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4568 * every tick. We fix it up based on jiffies.
4570 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4571 unsigned long pending_updates)
4575 this_rq->nr_load_updates++;
4577 /* Update our load: */
4578 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4579 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4580 unsigned long old_load, new_load;
4582 /* scale is effectively 1 << i now, and >> i divides by scale */
4584 old_load = this_rq->cpu_load[i];
4585 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4586 new_load = this_load;
4588 * Round up the averaging division if load is increasing. This
4589 * prevents us from getting stuck on 9 if the load is 10, for
4592 if (new_load > old_load)
4593 new_load += scale - 1;
4595 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4598 sched_avg_update(this_rq);
4601 /* Used instead of source_load when we know the type == 0 */
4602 static unsigned long weighted_cpuload(const int cpu)
4604 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4607 #ifdef CONFIG_NO_HZ_COMMON
4609 * There is no sane way to deal with nohz on smp when using jiffies because the
4610 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4611 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4613 * Therefore we cannot use the delta approach from the regular tick since that
4614 * would seriously skew the load calculation. However we'll make do for those
4615 * updates happening while idle (nohz_idle_balance) or coming out of idle
4616 * (tick_nohz_idle_exit).
4618 * This means we might still be one tick off for nohz periods.
4622 * Called from nohz_idle_balance() to update the load ratings before doing the
4625 static void update_idle_cpu_load(struct rq *this_rq)
4627 unsigned long curr_jiffies = READ_ONCE(jiffies);
4628 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4629 unsigned long pending_updates;
4632 * bail if there's load or we're actually up-to-date.
4634 if (load || curr_jiffies == this_rq->last_load_update_tick)
4637 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4638 this_rq->last_load_update_tick = curr_jiffies;
4640 __update_cpu_load(this_rq, load, pending_updates);
4644 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4646 void update_cpu_load_nohz(void)
4648 struct rq *this_rq = this_rq();
4649 unsigned long curr_jiffies = READ_ONCE(jiffies);
4650 unsigned long pending_updates;
4652 if (curr_jiffies == this_rq->last_load_update_tick)
4655 raw_spin_lock(&this_rq->lock);
4656 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4657 if (pending_updates) {
4658 this_rq->last_load_update_tick = curr_jiffies;
4660 * We were idle, this means load 0, the current load might be
4661 * !0 due to remote wakeups and the sort.
4663 __update_cpu_load(this_rq, 0, pending_updates);
4665 raw_spin_unlock(&this_rq->lock);
4667 #endif /* CONFIG_NO_HZ */
4670 * Called from scheduler_tick()
4672 void update_cpu_load_active(struct rq *this_rq)
4674 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4676 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4678 this_rq->last_load_update_tick = jiffies;
4679 __update_cpu_load(this_rq, load, 1);
4683 * Return a low guess at the load of a migration-source cpu weighted
4684 * according to the scheduling class and "nice" value.
4686 * We want to under-estimate the load of migration sources, to
4687 * balance conservatively.
4689 static unsigned long source_load(int cpu, int type)
4691 struct rq *rq = cpu_rq(cpu);
4692 unsigned long total = weighted_cpuload(cpu);
4694 if (type == 0 || !sched_feat(LB_BIAS))
4697 return min(rq->cpu_load[type-1], total);
4701 * Return a high guess at the load of a migration-target cpu weighted
4702 * according to the scheduling class and "nice" value.
4704 static unsigned long target_load(int cpu, int type)
4706 struct rq *rq = cpu_rq(cpu);
4707 unsigned long total = weighted_cpuload(cpu);
4709 if (type == 0 || !sched_feat(LB_BIAS))
4712 return max(rq->cpu_load[type-1], total);
4716 static unsigned long cpu_avg_load_per_task(int cpu)
4718 struct rq *rq = cpu_rq(cpu);
4719 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4720 unsigned long load_avg = weighted_cpuload(cpu);
4723 return load_avg / nr_running;
4728 static void record_wakee(struct task_struct *p)
4731 * Rough decay (wiping) for cost saving, don't worry
4732 * about the boundary, really active task won't care
4735 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4736 current->wakee_flips >>= 1;
4737 current->wakee_flip_decay_ts = jiffies;
4740 if (current->last_wakee != p) {
4741 current->last_wakee = p;
4742 current->wakee_flips++;
4746 static void task_waking_fair(struct task_struct *p)
4748 struct sched_entity *se = &p->se;
4749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4752 #ifndef CONFIG_64BIT
4753 u64 min_vruntime_copy;
4756 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4758 min_vruntime = cfs_rq->min_vruntime;
4759 } while (min_vruntime != min_vruntime_copy);
4761 min_vruntime = cfs_rq->min_vruntime;
4764 se->vruntime -= min_vruntime;
4768 #ifdef CONFIG_FAIR_GROUP_SCHED
4770 * effective_load() calculates the load change as seen from the root_task_group
4772 * Adding load to a group doesn't make a group heavier, but can cause movement
4773 * of group shares between cpus. Assuming the shares were perfectly aligned one
4774 * can calculate the shift in shares.
4776 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4777 * on this @cpu and results in a total addition (subtraction) of @wg to the
4778 * total group weight.
4780 * Given a runqueue weight distribution (rw_i) we can compute a shares
4781 * distribution (s_i) using:
4783 * s_i = rw_i / \Sum rw_j (1)
4785 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4786 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4787 * shares distribution (s_i):
4789 * rw_i = { 2, 4, 1, 0 }
4790 * s_i = { 2/7, 4/7, 1/7, 0 }
4792 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4793 * task used to run on and the CPU the waker is running on), we need to
4794 * compute the effect of waking a task on either CPU and, in case of a sync
4795 * wakeup, compute the effect of the current task going to sleep.
4797 * So for a change of @wl to the local @cpu with an overall group weight change
4798 * of @wl we can compute the new shares distribution (s'_i) using:
4800 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4802 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4803 * differences in waking a task to CPU 0. The additional task changes the
4804 * weight and shares distributions like:
4806 * rw'_i = { 3, 4, 1, 0 }
4807 * s'_i = { 3/8, 4/8, 1/8, 0 }
4809 * We can then compute the difference in effective weight by using:
4811 * dw_i = S * (s'_i - s_i) (3)
4813 * Where 'S' is the group weight as seen by its parent.
4815 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4816 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4817 * 4/7) times the weight of the group.
4819 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4821 struct sched_entity *se = tg->se[cpu];
4823 if (!tg->parent) /* the trivial, non-cgroup case */
4826 for_each_sched_entity(se) {
4827 struct cfs_rq *cfs_rq = se->my_q;
4828 long W, w = cfs_rq_load_avg(cfs_rq);
4833 * W = @wg + \Sum rw_j
4835 W = wg + atomic_long_read(&tg->load_avg);
4837 /* Ensure \Sum rw_j >= rw_i */
4838 W -= cfs_rq->tg_load_avg_contrib;
4847 * wl = S * s'_i; see (2)
4850 wl = (w * (long)tg->shares) / W;
4855 * Per the above, wl is the new se->load.weight value; since
4856 * those are clipped to [MIN_SHARES, ...) do so now. See
4857 * calc_cfs_shares().
4859 if (wl < MIN_SHARES)
4863 * wl = dw_i = S * (s'_i - s_i); see (3)
4865 wl -= se->avg.load_avg;
4868 * Recursively apply this logic to all parent groups to compute
4869 * the final effective load change on the root group. Since
4870 * only the @tg group gets extra weight, all parent groups can
4871 * only redistribute existing shares. @wl is the shift in shares
4872 * resulting from this level per the above.
4881 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4889 * Returns the current capacity of cpu after applying both
4890 * cpu and freq scaling.
4892 unsigned long capacity_curr_of(int cpu)
4894 return cpu_rq(cpu)->cpu_capacity_orig *
4895 arch_scale_freq_capacity(NULL, cpu)
4896 >> SCHED_CAPACITY_SHIFT;
4899 static inline bool energy_aware(void)
4901 return sched_feat(ENERGY_AWARE);
4905 struct sched_group *sg_top;
4906 struct sched_group *sg_cap;
4913 struct task_struct *task;
4928 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4929 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4930 * energy calculations. Using the scale-invariant util returned by
4931 * cpu_util() and approximating scale-invariant util by:
4933 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4935 * the normalized util can be found using the specific capacity.
4937 * capacity = capacity_orig * curr_freq/max_freq
4939 * norm_util = running_time/time ~ util/capacity
4941 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4943 int util = __cpu_util(cpu, delta);
4945 if (util >= capacity)
4946 return SCHED_CAPACITY_SCALE;
4948 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4951 static int calc_util_delta(struct energy_env *eenv, int cpu)
4953 if (cpu == eenv->src_cpu)
4954 return -eenv->util_delta;
4955 if (cpu == eenv->dst_cpu)
4956 return eenv->util_delta;
4961 unsigned long group_max_util(struct energy_env *eenv)
4964 unsigned long max_util = 0;
4966 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4967 delta = calc_util_delta(eenv, i);
4968 max_util = max(max_util, __cpu_util(i, delta));
4975 * group_norm_util() returns the approximated group util relative to it's
4976 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4977 * energy calculations. Since task executions may or may not overlap in time in
4978 * the group the true normalized util is between max(cpu_norm_util(i)) and
4979 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4980 * latter is used as the estimate as it leads to a more pessimistic energy
4981 * estimate (more busy).
4984 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4987 unsigned long util_sum = 0;
4988 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4990 for_each_cpu(i, sched_group_cpus(sg)) {
4991 delta = calc_util_delta(eenv, i);
4992 util_sum += __cpu_norm_util(i, capacity, delta);
4995 if (util_sum > SCHED_CAPACITY_SCALE)
4996 return SCHED_CAPACITY_SCALE;
5000 static int find_new_capacity(struct energy_env *eenv,
5001 const struct sched_group_energy * const sge)
5004 unsigned long util = group_max_util(eenv);
5006 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5007 if (sge->cap_states[idx].cap >= util)
5011 eenv->cap_idx = idx;
5016 static int group_idle_state(struct sched_group *sg)
5018 int i, state = INT_MAX;
5020 /* Find the shallowest idle state in the sched group. */
5021 for_each_cpu(i, sched_group_cpus(sg))
5022 state = min(state, idle_get_state_idx(cpu_rq(i)));
5024 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5031 * sched_group_energy(): Computes the absolute energy consumption of cpus
5032 * belonging to the sched_group including shared resources shared only by
5033 * members of the group. Iterates over all cpus in the hierarchy below the
5034 * sched_group starting from the bottom working it's way up before going to
5035 * the next cpu until all cpus are covered at all levels. The current
5036 * implementation is likely to gather the same util statistics multiple times.
5037 * This can probably be done in a faster but more complex way.
5038 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5040 static int sched_group_energy(struct energy_env *eenv)
5042 struct sched_domain *sd;
5043 int cpu, total_energy = 0;
5044 struct cpumask visit_cpus;
5045 struct sched_group *sg;
5047 WARN_ON(!eenv->sg_top->sge);
5049 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5051 while (!cpumask_empty(&visit_cpus)) {
5052 struct sched_group *sg_shared_cap = NULL;
5054 cpu = cpumask_first(&visit_cpus);
5057 * Is the group utilization affected by cpus outside this
5060 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5064 * We most probably raced with hotplug; returning a
5065 * wrong energy estimation is better than entering an
5071 sg_shared_cap = sd->parent->groups;
5073 for_each_domain(cpu, sd) {
5076 /* Has this sched_domain already been visited? */
5077 if (sd->child && group_first_cpu(sg) != cpu)
5081 unsigned long group_util;
5082 int sg_busy_energy, sg_idle_energy;
5083 int cap_idx, idle_idx;
5085 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5086 eenv->sg_cap = sg_shared_cap;
5090 cap_idx = find_new_capacity(eenv, sg->sge);
5092 if (sg->group_weight == 1) {
5093 /* Remove capacity of src CPU (before task move) */
5094 if (eenv->util_delta == 0 &&
5095 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5096 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5097 eenv->cap.delta -= eenv->cap.before;
5099 /* Add capacity of dst CPU (after task move) */
5100 if (eenv->util_delta != 0 &&
5101 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5102 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5103 eenv->cap.delta += eenv->cap.after;
5107 idle_idx = group_idle_state(sg);
5108 group_util = group_norm_util(eenv, sg);
5109 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5110 >> SCHED_CAPACITY_SHIFT;
5111 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5112 * sg->sge->idle_states[idle_idx].power)
5113 >> SCHED_CAPACITY_SHIFT;
5115 total_energy += sg_busy_energy + sg_idle_energy;
5118 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5120 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5123 } while (sg = sg->next, sg != sd->groups);
5126 cpumask_clear_cpu(cpu, &visit_cpus);
5130 eenv->energy = total_energy;
5134 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5136 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5140 * energy_diff(): Estimate the energy impact of changing the utilization
5141 * distribution. eenv specifies the change: utilisation amount, source, and
5142 * destination cpu. Source or destination cpu may be -1 in which case the
5143 * utilization is removed from or added to the system (e.g. task wake-up). If
5144 * both are specified, the utilization is migrated.
5146 static inline int __energy_diff(struct energy_env *eenv)
5148 struct sched_domain *sd;
5149 struct sched_group *sg;
5150 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5153 struct energy_env eenv_before = {
5155 .src_cpu = eenv->src_cpu,
5156 .dst_cpu = eenv->dst_cpu,
5157 .nrg = { 0, 0, 0, 0},
5161 if (eenv->src_cpu == eenv->dst_cpu)
5164 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5165 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5168 return 0; /* Error */
5173 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5174 eenv_before.sg_top = eenv->sg_top = sg;
5176 if (sched_group_energy(&eenv_before))
5177 return 0; /* Invalid result abort */
5178 energy_before += eenv_before.energy;
5180 /* Keep track of SRC cpu (before) capacity */
5181 eenv->cap.before = eenv_before.cap.before;
5182 eenv->cap.delta = eenv_before.cap.delta;
5184 if (sched_group_energy(eenv))
5185 return 0; /* Invalid result abort */
5186 energy_after += eenv->energy;
5188 } while (sg = sg->next, sg != sd->groups);
5190 eenv->nrg.before = energy_before;
5191 eenv->nrg.after = energy_after;
5192 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5195 trace_sched_energy_diff(eenv->task,
5196 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5197 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5198 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5199 eenv->nrg.delta, eenv->payoff);
5202 * Dead-zone margin preventing too many migrations.
5205 margin = eenv->nrg.before >> 6; /* ~1.56% */
5207 diff = eenv->nrg.after - eenv->nrg.before;
5209 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5211 return eenv->nrg.diff;
5214 #ifdef CONFIG_SCHED_TUNE
5216 struct target_nrg schedtune_target_nrg;
5219 * System energy normalization
5220 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5221 * corresponding to the specified energy variation.
5224 normalize_energy(int energy_diff)
5227 #ifdef CONFIG_SCHED_DEBUG
5230 /* Check for boundaries */
5231 max_delta = schedtune_target_nrg.max_power;
5232 max_delta -= schedtune_target_nrg.min_power;
5233 WARN_ON(abs(energy_diff) >= max_delta);
5236 /* Do scaling using positive numbers to increase the range */
5237 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5239 /* Scale by energy magnitude */
5240 normalized_nrg <<= SCHED_LOAD_SHIFT;
5242 /* Normalize on max energy for target platform */
5243 normalized_nrg = reciprocal_divide(
5244 normalized_nrg, schedtune_target_nrg.rdiv);
5246 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5250 energy_diff(struct energy_env *eenv)
5252 int boost = schedtune_task_boost(eenv->task);
5255 /* Conpute "absolute" energy diff */
5256 __energy_diff(eenv);
5258 /* Return energy diff when boost margin is 0 */
5260 return eenv->nrg.diff;
5262 /* Compute normalized energy diff */
5263 nrg_delta = normalize_energy(eenv->nrg.diff);
5264 eenv->nrg.delta = nrg_delta;
5266 eenv->payoff = schedtune_accept_deltas(
5272 * When SchedTune is enabled, the energy_diff() function will return
5273 * the computed energy payoff value. Since the energy_diff() return
5274 * value is expected to be negative by its callers, this evaluation
5275 * function return a negative value each time the evaluation return a
5276 * positive payoff, which is the condition for the acceptance of
5277 * a scheduling decision
5279 return -eenv->payoff;
5281 #else /* CONFIG_SCHED_TUNE */
5282 #define energy_diff(eenv) __energy_diff(eenv)
5286 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5287 * A waker of many should wake a different task than the one last awakened
5288 * at a frequency roughly N times higher than one of its wakees. In order
5289 * to determine whether we should let the load spread vs consolodating to
5290 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5291 * partner, and a factor of lls_size higher frequency in the other. With
5292 * both conditions met, we can be relatively sure that the relationship is
5293 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5294 * being client/server, worker/dispatcher, interrupt source or whatever is
5295 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5297 static int wake_wide(struct task_struct *p)
5299 unsigned int master = current->wakee_flips;
5300 unsigned int slave = p->wakee_flips;
5301 int factor = this_cpu_read(sd_llc_size);
5304 swap(master, slave);
5305 if (slave < factor || master < slave * factor)
5310 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5311 int prev_cpu, int sync)
5313 s64 this_load, load;
5314 s64 this_eff_load, prev_eff_load;
5316 struct task_group *tg;
5317 unsigned long weight;
5321 this_cpu = smp_processor_id();
5322 load = source_load(prev_cpu, idx);
5323 this_load = target_load(this_cpu, idx);
5326 * If sync wakeup then subtract the (maximum possible)
5327 * effect of the currently running task from the load
5328 * of the current CPU:
5331 tg = task_group(current);
5332 weight = current->se.avg.load_avg;
5334 this_load += effective_load(tg, this_cpu, -weight, -weight);
5335 load += effective_load(tg, prev_cpu, 0, -weight);
5339 weight = p->se.avg.load_avg;
5342 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5343 * due to the sync cause above having dropped this_load to 0, we'll
5344 * always have an imbalance, but there's really nothing you can do
5345 * about that, so that's good too.
5347 * Otherwise check if either cpus are near enough in load to allow this
5348 * task to be woken on this_cpu.
5350 this_eff_load = 100;
5351 this_eff_load *= capacity_of(prev_cpu);
5353 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5354 prev_eff_load *= capacity_of(this_cpu);
5356 if (this_load > 0) {
5357 this_eff_load *= this_load +
5358 effective_load(tg, this_cpu, weight, weight);
5360 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5363 balanced = this_eff_load <= prev_eff_load;
5365 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5370 schedstat_inc(sd, ttwu_move_affine);
5371 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5376 static inline unsigned long task_util(struct task_struct *p)
5378 #ifdef CONFIG_SCHED_WALT
5379 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5380 unsigned long demand = p->ravg.demand;
5381 return (demand << 10) / walt_ravg_window;
5384 return p->se.avg.util_avg;
5387 static inline unsigned long boosted_task_util(struct task_struct *task);
5389 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5391 unsigned long capacity = capacity_of(cpu);
5393 util += boosted_task_util(p);
5395 return (capacity * 1024) > (util * capacity_margin);
5398 static inline bool task_fits_max(struct task_struct *p, int cpu)
5400 unsigned long capacity = capacity_of(cpu);
5401 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5403 if (capacity == max_capacity)
5406 if (capacity * capacity_margin > max_capacity * 1024)
5409 return __task_fits(p, cpu, 0);
5412 static bool cpu_overutilized(int cpu)
5414 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5417 #ifdef CONFIG_SCHED_TUNE
5420 schedtune_margin(unsigned long signal, long boost)
5422 long long margin = 0;
5425 * Signal proportional compensation (SPC)
5427 * The Boost (B) value is used to compute a Margin (M) which is
5428 * proportional to the complement of the original Signal (S):
5429 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5430 * M = B * S, if B is negative
5431 * The obtained M could be used by the caller to "boost" S.
5434 margin = SCHED_LOAD_SCALE - signal;
5437 margin = -signal * boost;
5439 * Fast integer division by constant:
5440 * Constant : (C) = 100
5441 * Precision : 0.1% (P) = 0.1
5442 * Reference : C * 100 / P (R) = 100000
5445 * Shift bits : ceil(log(R,2)) (S) = 17
5446 * Mult const : round(2^S/C) (M) = 1311
5459 schedtune_cpu_margin(unsigned long util, int cpu)
5461 int boost = schedtune_cpu_boost(cpu);
5466 return schedtune_margin(util, boost);
5470 schedtune_task_margin(struct task_struct *task)
5472 int boost = schedtune_task_boost(task);
5479 util = task_util(task);
5480 margin = schedtune_margin(util, boost);
5485 #else /* CONFIG_SCHED_TUNE */
5488 schedtune_cpu_margin(unsigned long util, int cpu)
5494 schedtune_task_margin(struct task_struct *task)
5499 #endif /* CONFIG_SCHED_TUNE */
5502 boosted_cpu_util(int cpu)
5504 unsigned long util = cpu_util(cpu);
5505 long margin = schedtune_cpu_margin(util, cpu);
5507 trace_sched_boost_cpu(cpu, util, margin);
5509 return util + margin;
5512 static inline unsigned long
5513 boosted_task_util(struct task_struct *task)
5515 unsigned long util = task_util(task);
5516 long margin = schedtune_task_margin(task);
5518 trace_sched_boost_task(task, util, margin);
5520 return util + margin;
5523 static int cpu_util_wake(int cpu, struct task_struct *p);
5525 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5527 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5531 * find_idlest_group finds and returns the least busy CPU group within the
5534 static struct sched_group *
5535 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5536 int this_cpu, int sd_flag)
5538 struct sched_group *idlest = NULL, *group = sd->groups;
5539 struct sched_group *most_spare_sg = NULL;
5540 unsigned long min_load = ULONG_MAX, this_load = 0;
5541 unsigned long most_spare = 0, this_spare = 0;
5542 int load_idx = sd->forkexec_idx;
5543 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5545 if (sd_flag & SD_BALANCE_WAKE)
5546 load_idx = sd->wake_idx;
5549 unsigned long load, avg_load, spare_cap, max_spare_cap;
5553 /* Skip over this group if it has no CPUs allowed */
5554 if (!cpumask_intersects(sched_group_cpus(group),
5555 tsk_cpus_allowed(p)))
5558 local_group = cpumask_test_cpu(this_cpu,
5559 sched_group_cpus(group));
5562 * Tally up the load of all CPUs in the group and find
5563 * the group containing the CPU with most spare capacity.
5568 for_each_cpu(i, sched_group_cpus(group)) {
5569 /* Bias balancing toward cpus of our domain */
5571 load = source_load(i, load_idx);
5573 load = target_load(i, load_idx);
5577 spare_cap = capacity_spare_wake(i, p);
5579 if (spare_cap > max_spare_cap)
5580 max_spare_cap = spare_cap;
5583 /* Adjust by relative CPU capacity of the group */
5584 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5587 this_load = avg_load;
5588 this_spare = max_spare_cap;
5590 if (avg_load < min_load) {
5591 min_load = avg_load;
5595 if (most_spare < max_spare_cap) {
5596 most_spare = max_spare_cap;
5597 most_spare_sg = group;
5600 } while (group = group->next, group != sd->groups);
5603 * The cross-over point between using spare capacity or least load
5604 * is too conservative for high utilization tasks on partially
5605 * utilized systems if we require spare_capacity > task_util(p),
5606 * so we allow for some task stuffing by using
5607 * spare_capacity > task_util(p)/2.
5609 if (this_spare > task_util(p) / 2 &&
5610 imbalance*this_spare > 100*most_spare)
5612 else if (most_spare > task_util(p) / 2)
5613 return most_spare_sg;
5615 if (!idlest || 100*this_load < imbalance*min_load)
5621 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5624 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5626 unsigned long load, min_load = ULONG_MAX;
5627 unsigned int min_exit_latency = UINT_MAX;
5628 u64 latest_idle_timestamp = 0;
5629 int least_loaded_cpu = this_cpu;
5630 int shallowest_idle_cpu = -1;
5633 /* Check if we have any choice: */
5634 if (group->group_weight == 1)
5635 return cpumask_first(sched_group_cpus(group));
5637 /* Traverse only the allowed CPUs */
5638 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5640 struct rq *rq = cpu_rq(i);
5641 struct cpuidle_state *idle = idle_get_state(rq);
5642 if (idle && idle->exit_latency < min_exit_latency) {
5644 * We give priority to a CPU whose idle state
5645 * has the smallest exit latency irrespective
5646 * of any idle timestamp.
5648 min_exit_latency = idle->exit_latency;
5649 latest_idle_timestamp = rq->idle_stamp;
5650 shallowest_idle_cpu = i;
5651 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5652 rq->idle_stamp > latest_idle_timestamp) {
5654 * If equal or no active idle state, then
5655 * the most recently idled CPU might have
5658 latest_idle_timestamp = rq->idle_stamp;
5659 shallowest_idle_cpu = i;
5661 } else if (shallowest_idle_cpu == -1) {
5662 load = weighted_cpuload(i);
5663 if (load < min_load || (load == min_load && i == this_cpu)) {
5665 least_loaded_cpu = i;
5670 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5674 * Try and locate an idle CPU in the sched_domain.
5676 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5678 struct sched_domain *sd;
5679 struct sched_group *sg;
5680 int best_idle_cpu = -1;
5681 int best_idle_cstate = INT_MAX;
5682 unsigned long best_idle_capacity = ULONG_MAX;
5684 if (!sysctl_sched_cstate_aware) {
5685 if (idle_cpu(target))
5689 * If the prevous cpu is cache affine and idle, don't be stupid.
5691 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5696 * Otherwise, iterate the domains and find an elegible idle cpu.
5698 sd = rcu_dereference(per_cpu(sd_llc, target));
5699 for_each_lower_domain(sd) {
5703 if (!cpumask_intersects(sched_group_cpus(sg),
5704 tsk_cpus_allowed(p)))
5707 if (sysctl_sched_cstate_aware) {
5708 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5709 int idle_idx = idle_get_state_idx(cpu_rq(i));
5710 unsigned long new_usage = boosted_task_util(p);
5711 unsigned long capacity_orig = capacity_orig_of(i);
5713 if (new_usage > capacity_orig || !idle_cpu(i))
5716 if (i == target && new_usage <= capacity_curr_of(target))
5719 if (idle_idx < best_idle_cstate &&
5720 capacity_orig <= best_idle_capacity) {
5722 best_idle_cstate = idle_idx;
5723 best_idle_capacity = capacity_orig;
5727 for_each_cpu(i, sched_group_cpus(sg)) {
5728 if (i == target || !idle_cpu(i))
5732 target = cpumask_first_and(sched_group_cpus(sg),
5733 tsk_cpus_allowed(p));
5738 } while (sg != sd->groups);
5741 if (best_idle_cpu >= 0)
5742 target = best_idle_cpu;
5748 static int start_cpu(bool boosted)
5750 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
5752 RCU_LOCKDEP_WARN(rcu_read_lock_sched_held(),
5753 "sched RCU must be held");
5755 return boosted ? rd->max_cap_orig_cpu : rd->min_cap_orig_cpu;
5758 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5760 int target_cpu = -1;
5761 unsigned long target_util = prefer_idle ? ULONG_MAX : 0;
5762 unsigned long backup_capacity = ULONG_MAX;
5763 int best_idle_cpu = -1;
5764 int best_idle_cstate = INT_MAX;
5765 int backup_cpu = -1;
5766 unsigned long min_util = boosted_task_util(p);
5767 struct sched_domain *sd;
5768 struct sched_group *sg;
5769 int cpu = start_cpu(boosted);
5774 sd = rcu_dereference(per_cpu(sd_ea, cpu));
5784 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5785 unsigned long cur_capacity, new_util;
5791 * p's blocked utilization is still accounted for on prev_cpu
5792 * so prev_cpu will receive a negative bias due to the double
5793 * accounting. However, the blocked utilization may be zero.
5795 new_util = cpu_util(i) + task_util(p);
5798 * Ensure minimum capacity to grant the required boost.
5799 * The target CPU can be already at a capacity level higher
5800 * than the one required to boost the task.
5802 new_util = max(min_util, new_util);
5804 if (new_util > capacity_orig_of(i))
5807 #ifdef CONFIG_SCHED_WALT
5808 if (walt_cpu_high_irqload(i))
5813 * Unconditionally favoring tasks that prefer idle cpus to
5816 if (idle_cpu(i) && prefer_idle)
5819 cur_capacity = capacity_curr_of(i);
5821 if (new_util < cur_capacity) {
5822 if (cpu_rq(i)->nr_running) {
5824 * Find a target cpu with the lowest/highest
5825 * utilization if prefer_idle/!prefer_idle.
5827 if ((prefer_idle && target_util > new_util) ||
5828 (!prefer_idle && target_util < new_util)) {
5829 target_util = new_util;
5832 } else if (!prefer_idle) {
5833 int idle_idx = idle_get_state_idx(cpu_rq(i));
5835 if (best_idle_cpu < 0 ||
5836 (sysctl_sched_cstate_aware &&
5837 best_idle_cstate > idle_idx)) {
5838 best_idle_cstate = idle_idx;
5842 } else if (backup_capacity > cur_capacity) {
5843 /* Find a backup cpu with least capacity. */
5844 backup_capacity = cur_capacity;
5848 } while (sg = sg->next, sg != sd->groups);
5851 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5857 * cpu_util_wake: Compute cpu utilization with any contributions from
5858 * the waking task p removed.
5860 static int cpu_util_wake(int cpu, struct task_struct *p)
5862 unsigned long util, capacity;
5864 /* Task has no contribution or is new */
5865 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5866 return cpu_util(cpu);
5868 capacity = capacity_orig_of(cpu);
5869 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5871 return (util >= capacity) ? capacity : util;
5875 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5876 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5878 * In that case WAKE_AFFINE doesn't make sense and we'll let
5879 * BALANCE_WAKE sort things out.
5881 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5883 long min_cap, max_cap;
5885 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5886 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5888 /* Minimum capacity is close to max, no need to abort wake_affine */
5889 if (max_cap - min_cap < max_cap >> 3)
5892 /* Bring task utilization in sync with prev_cpu */
5893 sync_entity_load_avg(&p->se);
5895 return min_cap * 1024 < task_util(p) * capacity_margin;
5898 static int select_energy_cpu_brute(struct task_struct *p, int prev_cpu, int sync)
5900 struct sched_domain *sd;
5901 int target_cpu = prev_cpu, tmp_target;
5902 bool boosted, prefer_idle;
5904 if (sysctl_sched_sync_hint_enable && sync) {
5905 int cpu = smp_processor_id();
5907 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5912 #ifdef CONFIG_CGROUP_SCHEDTUNE
5913 boosted = schedtune_task_boost(p) > 0;
5914 prefer_idle = schedtune_prefer_idle(p) > 0;
5916 boosted = get_sysctl_sched_cfs_boost() > 0;
5920 sd = rcu_dereference(per_cpu(sd_ea, prev_cpu));
5921 /* Find a cpu with sufficient capacity */
5922 tmp_target = find_best_target(p, boosted, prefer_idle);
5926 if (tmp_target >= 0) {
5927 target_cpu = tmp_target;
5928 if ((boosted || prefer_idle) && idle_cpu(target_cpu))
5932 if (target_cpu != prev_cpu) {
5933 struct energy_env eenv = {
5934 .util_delta = task_util(p),
5935 .src_cpu = prev_cpu,
5936 .dst_cpu = target_cpu,
5940 /* Not enough spare capacity on previous cpu */
5941 if (cpu_overutilized(prev_cpu))
5944 if (energy_diff(&eenv) >= 0)
5945 target_cpu = prev_cpu;
5954 * select_task_rq_fair: Select target runqueue for the waking task in domains
5955 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5956 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5958 * Balances load by selecting the idlest cpu in the idlest group, or under
5959 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5961 * Returns the target cpu number.
5963 * preempt must be disabled.
5966 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5968 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5969 int cpu = smp_processor_id();
5970 int new_cpu = prev_cpu;
5971 int want_affine = 0;
5972 int sync = wake_flags & WF_SYNC;
5974 if (sd_flag & SD_BALANCE_WAKE)
5975 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5976 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5978 if (energy_aware() && !(cpu_rq(prev_cpu)->rd->overutilized))
5979 return select_energy_cpu_brute(p, prev_cpu, sync);
5982 for_each_domain(cpu, tmp) {
5983 if (!(tmp->flags & SD_LOAD_BALANCE))
5987 * If both cpu and prev_cpu are part of this domain,
5988 * cpu is a valid SD_WAKE_AFFINE target.
5990 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5991 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5996 if (tmp->flags & sd_flag)
5998 else if (!want_affine)
6003 sd = NULL; /* Prefer wake_affine over balance flags */
6004 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
6009 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
6010 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6013 struct sched_group *group;
6016 if (!(sd->flags & sd_flag)) {
6021 group = find_idlest_group(sd, p, cpu, sd_flag);
6027 new_cpu = find_idlest_cpu(group, p, cpu);
6028 if (new_cpu == -1 || new_cpu == cpu) {
6029 /* Now try balancing at a lower domain level of cpu */
6034 /* Now try balancing at a lower domain level of new_cpu */
6036 weight = sd->span_weight;
6038 for_each_domain(cpu, tmp) {
6039 if (weight <= tmp->span_weight)
6041 if (tmp->flags & sd_flag)
6044 /* while loop will break here if sd == NULL */
6052 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6053 * cfs_rq_of(p) references at time of call are still valid and identify the
6054 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
6055 * other assumptions, including the state of rq->lock, should be made.
6057 static void migrate_task_rq_fair(struct task_struct *p)
6060 * We are supposed to update the task to "current" time, then its up to date
6061 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6062 * what current time is, so simply throw away the out-of-date time. This
6063 * will result in the wakee task is less decayed, but giving the wakee more
6064 * load sounds not bad.
6066 remove_entity_load_avg(&p->se);
6068 /* Tell new CPU we are migrated */
6069 p->se.avg.last_update_time = 0;
6071 /* We have migrated, no longer consider this task hot */
6072 p->se.exec_start = 0;
6075 static void task_dead_fair(struct task_struct *p)
6077 remove_entity_load_avg(&p->se);
6080 #define task_fits_max(p, cpu) true
6081 #endif /* CONFIG_SMP */
6083 static unsigned long
6084 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6086 unsigned long gran = sysctl_sched_wakeup_granularity;
6089 * Since its curr running now, convert the gran from real-time
6090 * to virtual-time in his units.
6092 * By using 'se' instead of 'curr' we penalize light tasks, so
6093 * they get preempted easier. That is, if 'se' < 'curr' then
6094 * the resulting gran will be larger, therefore penalizing the
6095 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6096 * be smaller, again penalizing the lighter task.
6098 * This is especially important for buddies when the leftmost
6099 * task is higher priority than the buddy.
6101 return calc_delta_fair(gran, se);
6105 * Should 'se' preempt 'curr'.
6119 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6121 s64 gran, vdiff = curr->vruntime - se->vruntime;
6126 gran = wakeup_gran(curr, se);
6133 static void set_last_buddy(struct sched_entity *se)
6135 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6138 for_each_sched_entity(se)
6139 cfs_rq_of(se)->last = se;
6142 static void set_next_buddy(struct sched_entity *se)
6144 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6147 for_each_sched_entity(se)
6148 cfs_rq_of(se)->next = se;
6151 static void set_skip_buddy(struct sched_entity *se)
6153 for_each_sched_entity(se)
6154 cfs_rq_of(se)->skip = se;
6158 * Preempt the current task with a newly woken task if needed:
6160 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6162 struct task_struct *curr = rq->curr;
6163 struct sched_entity *se = &curr->se, *pse = &p->se;
6164 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6165 int scale = cfs_rq->nr_running >= sched_nr_latency;
6166 int next_buddy_marked = 0;
6168 if (unlikely(se == pse))
6172 * This is possible from callers such as attach_tasks(), in which we
6173 * unconditionally check_prempt_curr() after an enqueue (which may have
6174 * lead to a throttle). This both saves work and prevents false
6175 * next-buddy nomination below.
6177 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6180 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6181 set_next_buddy(pse);
6182 next_buddy_marked = 1;
6186 * We can come here with TIF_NEED_RESCHED already set from new task
6189 * Note: this also catches the edge-case of curr being in a throttled
6190 * group (e.g. via set_curr_task), since update_curr() (in the
6191 * enqueue of curr) will have resulted in resched being set. This
6192 * prevents us from potentially nominating it as a false LAST_BUDDY
6195 if (test_tsk_need_resched(curr))
6198 /* Idle tasks are by definition preempted by non-idle tasks. */
6199 if (unlikely(curr->policy == SCHED_IDLE) &&
6200 likely(p->policy != SCHED_IDLE))
6204 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6205 * is driven by the tick):
6207 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6210 find_matching_se(&se, &pse);
6211 update_curr(cfs_rq_of(se));
6213 if (wakeup_preempt_entity(se, pse) == 1) {
6215 * Bias pick_next to pick the sched entity that is
6216 * triggering this preemption.
6218 if (!next_buddy_marked)
6219 set_next_buddy(pse);
6228 * Only set the backward buddy when the current task is still
6229 * on the rq. This can happen when a wakeup gets interleaved
6230 * with schedule on the ->pre_schedule() or idle_balance()
6231 * point, either of which can * drop the rq lock.
6233 * Also, during early boot the idle thread is in the fair class,
6234 * for obvious reasons its a bad idea to schedule back to it.
6236 if (unlikely(!se->on_rq || curr == rq->idle))
6239 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6243 static struct task_struct *
6244 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6246 struct cfs_rq *cfs_rq = &rq->cfs;
6247 struct sched_entity *se;
6248 struct task_struct *p;
6252 #ifdef CONFIG_FAIR_GROUP_SCHED
6253 if (!cfs_rq->nr_running)
6256 if (prev->sched_class != &fair_sched_class)
6260 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6261 * likely that a next task is from the same cgroup as the current.
6263 * Therefore attempt to avoid putting and setting the entire cgroup
6264 * hierarchy, only change the part that actually changes.
6268 struct sched_entity *curr = cfs_rq->curr;
6271 * Since we got here without doing put_prev_entity() we also
6272 * have to consider cfs_rq->curr. If it is still a runnable
6273 * entity, update_curr() will update its vruntime, otherwise
6274 * forget we've ever seen it.
6278 update_curr(cfs_rq);
6283 * This call to check_cfs_rq_runtime() will do the
6284 * throttle and dequeue its entity in the parent(s).
6285 * Therefore the 'simple' nr_running test will indeed
6288 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6292 se = pick_next_entity(cfs_rq, curr);
6293 cfs_rq = group_cfs_rq(se);
6299 * Since we haven't yet done put_prev_entity and if the selected task
6300 * is a different task than we started out with, try and touch the
6301 * least amount of cfs_rqs.
6304 struct sched_entity *pse = &prev->se;
6306 while (!(cfs_rq = is_same_group(se, pse))) {
6307 int se_depth = se->depth;
6308 int pse_depth = pse->depth;
6310 if (se_depth <= pse_depth) {
6311 put_prev_entity(cfs_rq_of(pse), pse);
6312 pse = parent_entity(pse);
6314 if (se_depth >= pse_depth) {
6315 set_next_entity(cfs_rq_of(se), se);
6316 se = parent_entity(se);
6320 put_prev_entity(cfs_rq, pse);
6321 set_next_entity(cfs_rq, se);
6324 if (hrtick_enabled(rq))
6325 hrtick_start_fair(rq, p);
6327 rq->misfit_task = !task_fits_max(p, rq->cpu);
6334 if (!cfs_rq->nr_running)
6337 put_prev_task(rq, prev);
6340 se = pick_next_entity(cfs_rq, NULL);
6341 set_next_entity(cfs_rq, se);
6342 cfs_rq = group_cfs_rq(se);
6347 if (hrtick_enabled(rq))
6348 hrtick_start_fair(rq, p);
6350 rq->misfit_task = !task_fits_max(p, rq->cpu);
6355 rq->misfit_task = 0;
6357 * This is OK, because current is on_cpu, which avoids it being picked
6358 * for load-balance and preemption/IRQs are still disabled avoiding
6359 * further scheduler activity on it and we're being very careful to
6360 * re-start the picking loop.
6362 lockdep_unpin_lock(&rq->lock);
6363 new_tasks = idle_balance(rq);
6364 lockdep_pin_lock(&rq->lock);
6366 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6367 * possible for any higher priority task to appear. In that case we
6368 * must re-start the pick_next_entity() loop.
6380 * Account for a descheduled task:
6382 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6384 struct sched_entity *se = &prev->se;
6385 struct cfs_rq *cfs_rq;
6387 for_each_sched_entity(se) {
6388 cfs_rq = cfs_rq_of(se);
6389 put_prev_entity(cfs_rq, se);
6394 * sched_yield() is very simple
6396 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6398 static void yield_task_fair(struct rq *rq)
6400 struct task_struct *curr = rq->curr;
6401 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6402 struct sched_entity *se = &curr->se;
6405 * Are we the only task in the tree?
6407 if (unlikely(rq->nr_running == 1))
6410 clear_buddies(cfs_rq, se);
6412 if (curr->policy != SCHED_BATCH) {
6413 update_rq_clock(rq);
6415 * Update run-time statistics of the 'current'.
6417 update_curr(cfs_rq);
6419 * Tell update_rq_clock() that we've just updated,
6420 * so we don't do microscopic update in schedule()
6421 * and double the fastpath cost.
6423 rq_clock_skip_update(rq, true);
6429 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6431 struct sched_entity *se = &p->se;
6433 /* throttled hierarchies are not runnable */
6434 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6437 /* Tell the scheduler that we'd really like pse to run next. */
6440 yield_task_fair(rq);
6446 /**************************************************
6447 * Fair scheduling class load-balancing methods.
6451 * The purpose of load-balancing is to achieve the same basic fairness the
6452 * per-cpu scheduler provides, namely provide a proportional amount of compute
6453 * time to each task. This is expressed in the following equation:
6455 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6457 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6458 * W_i,0 is defined as:
6460 * W_i,0 = \Sum_j w_i,j (2)
6462 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6463 * is derived from the nice value as per prio_to_weight[].
6465 * The weight average is an exponential decay average of the instantaneous
6468 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6470 * C_i is the compute capacity of cpu i, typically it is the
6471 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6472 * can also include other factors [XXX].
6474 * To achieve this balance we define a measure of imbalance which follows
6475 * directly from (1):
6477 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6479 * We them move tasks around to minimize the imbalance. In the continuous
6480 * function space it is obvious this converges, in the discrete case we get
6481 * a few fun cases generally called infeasible weight scenarios.
6484 * - infeasible weights;
6485 * - local vs global optima in the discrete case. ]
6490 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6491 * for all i,j solution, we create a tree of cpus that follows the hardware
6492 * topology where each level pairs two lower groups (or better). This results
6493 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6494 * tree to only the first of the previous level and we decrease the frequency
6495 * of load-balance at each level inv. proportional to the number of cpus in
6501 * \Sum { --- * --- * 2^i } = O(n) (5)
6503 * `- size of each group
6504 * | | `- number of cpus doing load-balance
6506 * `- sum over all levels
6508 * Coupled with a limit on how many tasks we can migrate every balance pass,
6509 * this makes (5) the runtime complexity of the balancer.
6511 * An important property here is that each CPU is still (indirectly) connected
6512 * to every other cpu in at most O(log n) steps:
6514 * The adjacency matrix of the resulting graph is given by:
6517 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6520 * And you'll find that:
6522 * A^(log_2 n)_i,j != 0 for all i,j (7)
6524 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6525 * The task movement gives a factor of O(m), giving a convergence complexity
6528 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6533 * In order to avoid CPUs going idle while there's still work to do, new idle
6534 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6535 * tree itself instead of relying on other CPUs to bring it work.
6537 * This adds some complexity to both (5) and (8) but it reduces the total idle
6545 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6548 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6553 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6555 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6557 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6560 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6561 * rewrite all of this once again.]
6564 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6566 enum fbq_type { regular, remote, all };
6575 #define LBF_ALL_PINNED 0x01
6576 #define LBF_NEED_BREAK 0x02
6577 #define LBF_DST_PINNED 0x04
6578 #define LBF_SOME_PINNED 0x08
6581 struct sched_domain *sd;
6589 struct cpumask *dst_grpmask;
6591 enum cpu_idle_type idle;
6593 unsigned int src_grp_nr_running;
6594 /* The set of CPUs under consideration for load-balancing */
6595 struct cpumask *cpus;
6600 unsigned int loop_break;
6601 unsigned int loop_max;
6603 enum fbq_type fbq_type;
6604 enum group_type busiest_group_type;
6605 struct list_head tasks;
6609 * Is this task likely cache-hot:
6611 static int task_hot(struct task_struct *p, struct lb_env *env)
6615 lockdep_assert_held(&env->src_rq->lock);
6617 if (p->sched_class != &fair_sched_class)
6620 if (unlikely(p->policy == SCHED_IDLE))
6624 * Buddy candidates are cache hot:
6626 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6627 (&p->se == cfs_rq_of(&p->se)->next ||
6628 &p->se == cfs_rq_of(&p->se)->last))
6631 if (sysctl_sched_migration_cost == -1)
6633 if (sysctl_sched_migration_cost == 0)
6636 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6638 return delta < (s64)sysctl_sched_migration_cost;
6641 #ifdef CONFIG_NUMA_BALANCING
6643 * Returns 1, if task migration degrades locality
6644 * Returns 0, if task migration improves locality i.e migration preferred.
6645 * Returns -1, if task migration is not affected by locality.
6647 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6649 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6650 unsigned long src_faults, dst_faults;
6651 int src_nid, dst_nid;
6653 if (!static_branch_likely(&sched_numa_balancing))
6656 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6659 src_nid = cpu_to_node(env->src_cpu);
6660 dst_nid = cpu_to_node(env->dst_cpu);
6662 if (src_nid == dst_nid)
6665 /* Migrating away from the preferred node is always bad. */
6666 if (src_nid == p->numa_preferred_nid) {
6667 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6673 /* Encourage migration to the preferred node. */
6674 if (dst_nid == p->numa_preferred_nid)
6678 src_faults = group_faults(p, src_nid);
6679 dst_faults = group_faults(p, dst_nid);
6681 src_faults = task_faults(p, src_nid);
6682 dst_faults = task_faults(p, dst_nid);
6685 return dst_faults < src_faults;
6689 static inline int migrate_degrades_locality(struct task_struct *p,
6697 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6700 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6704 lockdep_assert_held(&env->src_rq->lock);
6707 * We do not migrate tasks that are:
6708 * 1) throttled_lb_pair, or
6709 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6710 * 3) running (obviously), or
6711 * 4) are cache-hot on their current CPU.
6713 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6716 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6719 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6721 env->flags |= LBF_SOME_PINNED;
6724 * Remember if this task can be migrated to any other cpu in
6725 * our sched_group. We may want to revisit it if we couldn't
6726 * meet load balance goals by pulling other tasks on src_cpu.
6728 * Also avoid computing new_dst_cpu if we have already computed
6729 * one in current iteration.
6731 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6734 /* Prevent to re-select dst_cpu via env's cpus */
6735 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6736 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6737 env->flags |= LBF_DST_PINNED;
6738 env->new_dst_cpu = cpu;
6746 /* Record that we found atleast one task that could run on dst_cpu */
6747 env->flags &= ~LBF_ALL_PINNED;
6749 if (task_running(env->src_rq, p)) {
6750 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6755 * Aggressive migration if:
6756 * 1) destination numa is preferred
6757 * 2) task is cache cold, or
6758 * 3) too many balance attempts have failed.
6760 tsk_cache_hot = migrate_degrades_locality(p, env);
6761 if (tsk_cache_hot == -1)
6762 tsk_cache_hot = task_hot(p, env);
6764 if (tsk_cache_hot <= 0 ||
6765 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6766 if (tsk_cache_hot == 1) {
6767 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6768 schedstat_inc(p, se.statistics.nr_forced_migrations);
6773 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6778 * detach_task() -- detach the task for the migration specified in env
6780 static void detach_task(struct task_struct *p, struct lb_env *env)
6782 lockdep_assert_held(&env->src_rq->lock);
6784 deactivate_task(env->src_rq, p, 0);
6785 p->on_rq = TASK_ON_RQ_MIGRATING;
6786 double_lock_balance(env->src_rq, env->dst_rq);
6787 set_task_cpu(p, env->dst_cpu);
6788 double_unlock_balance(env->src_rq, env->dst_rq);
6792 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6793 * part of active balancing operations within "domain".
6795 * Returns a task if successful and NULL otherwise.
6797 static struct task_struct *detach_one_task(struct lb_env *env)
6799 struct task_struct *p, *n;
6801 lockdep_assert_held(&env->src_rq->lock);
6803 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6804 if (!can_migrate_task(p, env))
6807 detach_task(p, env);
6810 * Right now, this is only the second place where
6811 * lb_gained[env->idle] is updated (other is detach_tasks)
6812 * so we can safely collect stats here rather than
6813 * inside detach_tasks().
6815 schedstat_inc(env->sd, lb_gained[env->idle]);
6821 static const unsigned int sched_nr_migrate_break = 32;
6824 * detach_tasks() -- tries to detach up to imbalance weighted load from
6825 * busiest_rq, as part of a balancing operation within domain "sd".
6827 * Returns number of detached tasks if successful and 0 otherwise.
6829 static int detach_tasks(struct lb_env *env)
6831 struct list_head *tasks = &env->src_rq->cfs_tasks;
6832 struct task_struct *p;
6836 lockdep_assert_held(&env->src_rq->lock);
6838 if (env->imbalance <= 0)
6841 while (!list_empty(tasks)) {
6843 * We don't want to steal all, otherwise we may be treated likewise,
6844 * which could at worst lead to a livelock crash.
6846 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6849 p = list_first_entry(tasks, struct task_struct, se.group_node);
6852 /* We've more or less seen every task there is, call it quits */
6853 if (env->loop > env->loop_max)
6856 /* take a breather every nr_migrate tasks */
6857 if (env->loop > env->loop_break) {
6858 env->loop_break += sched_nr_migrate_break;
6859 env->flags |= LBF_NEED_BREAK;
6863 if (!can_migrate_task(p, env))
6866 load = task_h_load(p);
6868 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6871 if ((load / 2) > env->imbalance)
6874 detach_task(p, env);
6875 list_add(&p->se.group_node, &env->tasks);
6878 env->imbalance -= load;
6880 #ifdef CONFIG_PREEMPT
6882 * NEWIDLE balancing is a source of latency, so preemptible
6883 * kernels will stop after the first task is detached to minimize
6884 * the critical section.
6886 if (env->idle == CPU_NEWLY_IDLE)
6891 * We only want to steal up to the prescribed amount of
6894 if (env->imbalance <= 0)
6899 list_move_tail(&p->se.group_node, tasks);
6903 * Right now, this is one of only two places we collect this stat
6904 * so we can safely collect detach_one_task() stats here rather
6905 * than inside detach_one_task().
6907 schedstat_add(env->sd, lb_gained[env->idle], detached);
6913 * attach_task() -- attach the task detached by detach_task() to its new rq.
6915 static void attach_task(struct rq *rq, struct task_struct *p)
6917 lockdep_assert_held(&rq->lock);
6919 BUG_ON(task_rq(p) != rq);
6920 p->on_rq = TASK_ON_RQ_QUEUED;
6921 activate_task(rq, p, 0);
6922 check_preempt_curr(rq, p, 0);
6926 * attach_one_task() -- attaches the task returned from detach_one_task() to
6929 static void attach_one_task(struct rq *rq, struct task_struct *p)
6931 raw_spin_lock(&rq->lock);
6934 * We want to potentially raise target_cpu's OPP.
6936 update_capacity_of(cpu_of(rq));
6937 raw_spin_unlock(&rq->lock);
6941 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6944 static void attach_tasks(struct lb_env *env)
6946 struct list_head *tasks = &env->tasks;
6947 struct task_struct *p;
6949 raw_spin_lock(&env->dst_rq->lock);
6951 while (!list_empty(tasks)) {
6952 p = list_first_entry(tasks, struct task_struct, se.group_node);
6953 list_del_init(&p->se.group_node);
6955 attach_task(env->dst_rq, p);
6959 * We want to potentially raise env.dst_cpu's OPP.
6961 update_capacity_of(env->dst_cpu);
6963 raw_spin_unlock(&env->dst_rq->lock);
6966 #ifdef CONFIG_FAIR_GROUP_SCHED
6967 static void update_blocked_averages(int cpu)
6969 struct rq *rq = cpu_rq(cpu);
6970 struct cfs_rq *cfs_rq;
6971 unsigned long flags;
6973 raw_spin_lock_irqsave(&rq->lock, flags);
6974 update_rq_clock(rq);
6977 * Iterates the task_group tree in a bottom up fashion, see
6978 * list_add_leaf_cfs_rq() for details.
6980 for_each_leaf_cfs_rq(rq, cfs_rq) {
6981 /* throttled entities do not contribute to load */
6982 if (throttled_hierarchy(cfs_rq))
6985 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
6987 update_tg_load_avg(cfs_rq, 0);
6989 raw_spin_unlock_irqrestore(&rq->lock, flags);
6993 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6994 * This needs to be done in a top-down fashion because the load of a child
6995 * group is a fraction of its parents load.
6997 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6999 struct rq *rq = rq_of(cfs_rq);
7000 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7001 unsigned long now = jiffies;
7004 if (cfs_rq->last_h_load_update == now)
7007 cfs_rq->h_load_next = NULL;
7008 for_each_sched_entity(se) {
7009 cfs_rq = cfs_rq_of(se);
7010 cfs_rq->h_load_next = se;
7011 if (cfs_rq->last_h_load_update == now)
7016 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7017 cfs_rq->last_h_load_update = now;
7020 while ((se = cfs_rq->h_load_next) != NULL) {
7021 load = cfs_rq->h_load;
7022 load = div64_ul(load * se->avg.load_avg,
7023 cfs_rq_load_avg(cfs_rq) + 1);
7024 cfs_rq = group_cfs_rq(se);
7025 cfs_rq->h_load = load;
7026 cfs_rq->last_h_load_update = now;
7030 static unsigned long task_h_load(struct task_struct *p)
7032 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7034 update_cfs_rq_h_load(cfs_rq);
7035 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7036 cfs_rq_load_avg(cfs_rq) + 1);
7039 static inline void update_blocked_averages(int cpu)
7041 struct rq *rq = cpu_rq(cpu);
7042 struct cfs_rq *cfs_rq = &rq->cfs;
7043 unsigned long flags;
7045 raw_spin_lock_irqsave(&rq->lock, flags);
7046 update_rq_clock(rq);
7047 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
7048 raw_spin_unlock_irqrestore(&rq->lock, flags);
7051 static unsigned long task_h_load(struct task_struct *p)
7053 return p->se.avg.load_avg;
7057 /********** Helpers for find_busiest_group ************************/
7060 * sg_lb_stats - stats of a sched_group required for load_balancing
7062 struct sg_lb_stats {
7063 unsigned long avg_load; /*Avg load across the CPUs of the group */
7064 unsigned long group_load; /* Total load over the CPUs of the group */
7065 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7066 unsigned long load_per_task;
7067 unsigned long group_capacity;
7068 unsigned long group_util; /* Total utilization of the group */
7069 unsigned int sum_nr_running; /* Nr tasks running in the group */
7070 unsigned int idle_cpus;
7071 unsigned int group_weight;
7072 enum group_type group_type;
7073 int group_no_capacity;
7074 int group_misfit_task; /* A cpu has a task too big for its capacity */
7075 #ifdef CONFIG_NUMA_BALANCING
7076 unsigned int nr_numa_running;
7077 unsigned int nr_preferred_running;
7082 * sd_lb_stats - Structure to store the statistics of a sched_domain
7083 * during load balancing.
7085 struct sd_lb_stats {
7086 struct sched_group *busiest; /* Busiest group in this sd */
7087 struct sched_group *local; /* Local group in this sd */
7088 unsigned long total_load; /* Total load of all groups in sd */
7089 unsigned long total_capacity; /* Total capacity of all groups in sd */
7090 unsigned long avg_load; /* Average load across all groups in sd */
7092 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7093 struct sg_lb_stats local_stat; /* Statistics of the local group */
7096 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7099 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7100 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7101 * We must however clear busiest_stat::avg_load because
7102 * update_sd_pick_busiest() reads this before assignment.
7104 *sds = (struct sd_lb_stats){
7108 .total_capacity = 0UL,
7111 .sum_nr_running = 0,
7112 .group_type = group_other,
7118 * get_sd_load_idx - Obtain the load index for a given sched domain.
7119 * @sd: The sched_domain whose load_idx is to be obtained.
7120 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7122 * Return: The load index.
7124 static inline int get_sd_load_idx(struct sched_domain *sd,
7125 enum cpu_idle_type idle)
7131 load_idx = sd->busy_idx;
7134 case CPU_NEWLY_IDLE:
7135 load_idx = sd->newidle_idx;
7138 load_idx = sd->idle_idx;
7145 static unsigned long scale_rt_capacity(int cpu)
7147 struct rq *rq = cpu_rq(cpu);
7148 u64 total, used, age_stamp, avg;
7152 * Since we're reading these variables without serialization make sure
7153 * we read them once before doing sanity checks on them.
7155 age_stamp = READ_ONCE(rq->age_stamp);
7156 avg = READ_ONCE(rq->rt_avg);
7157 delta = __rq_clock_broken(rq) - age_stamp;
7159 if (unlikely(delta < 0))
7162 total = sched_avg_period() + delta;
7164 used = div_u64(avg, total);
7167 * deadline bandwidth is defined at system level so we must
7168 * weight this bandwidth with the max capacity of the system.
7169 * As a reminder, avg_bw is 20bits width and
7170 * scale_cpu_capacity is 10 bits width
7172 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7174 if (likely(used < SCHED_CAPACITY_SCALE))
7175 return SCHED_CAPACITY_SCALE - used;
7180 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7182 raw_spin_lock_init(&mcc->lock);
7187 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7189 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7190 struct sched_group *sdg = sd->groups;
7191 struct max_cpu_capacity *mcc;
7192 unsigned long max_capacity;
7194 unsigned long flags;
7196 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7198 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7200 raw_spin_lock_irqsave(&mcc->lock, flags);
7201 max_capacity = mcc->val;
7202 max_cap_cpu = mcc->cpu;
7204 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7205 (max_capacity < capacity)) {
7206 mcc->val = capacity;
7208 #ifdef CONFIG_SCHED_DEBUG
7209 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7210 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7215 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7217 skip_unlock: __attribute__ ((unused));
7218 capacity *= scale_rt_capacity(cpu);
7219 capacity >>= SCHED_CAPACITY_SHIFT;
7224 cpu_rq(cpu)->cpu_capacity = capacity;
7225 sdg->sgc->capacity = capacity;
7226 sdg->sgc->max_capacity = capacity;
7227 sdg->sgc->min_capacity = capacity;
7230 void update_group_capacity(struct sched_domain *sd, int cpu)
7232 struct sched_domain *child = sd->child;
7233 struct sched_group *group, *sdg = sd->groups;
7234 unsigned long capacity, max_capacity, min_capacity;
7235 unsigned long interval;
7237 interval = msecs_to_jiffies(sd->balance_interval);
7238 interval = clamp(interval, 1UL, max_load_balance_interval);
7239 sdg->sgc->next_update = jiffies + interval;
7242 update_cpu_capacity(sd, cpu);
7248 min_capacity = ULONG_MAX;
7250 if (child->flags & SD_OVERLAP) {
7252 * SD_OVERLAP domains cannot assume that child groups
7253 * span the current group.
7256 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7257 struct sched_group_capacity *sgc;
7258 struct rq *rq = cpu_rq(cpu);
7261 * build_sched_domains() -> init_sched_groups_capacity()
7262 * gets here before we've attached the domains to the
7265 * Use capacity_of(), which is set irrespective of domains
7266 * in update_cpu_capacity().
7268 * This avoids capacity from being 0 and
7269 * causing divide-by-zero issues on boot.
7271 if (unlikely(!rq->sd)) {
7272 capacity += capacity_of(cpu);
7274 sgc = rq->sd->groups->sgc;
7275 capacity += sgc->capacity;
7278 max_capacity = max(capacity, max_capacity);
7279 min_capacity = min(capacity, min_capacity);
7283 * !SD_OVERLAP domains can assume that child groups
7284 * span the current group.
7287 group = child->groups;
7289 struct sched_group_capacity *sgc = group->sgc;
7291 capacity += sgc->capacity;
7292 max_capacity = max(sgc->max_capacity, max_capacity);
7293 min_capacity = min(sgc->min_capacity, min_capacity);
7294 group = group->next;
7295 } while (group != child->groups);
7298 sdg->sgc->capacity = capacity;
7299 sdg->sgc->max_capacity = max_capacity;
7300 sdg->sgc->min_capacity = min_capacity;
7304 * Check whether the capacity of the rq has been noticeably reduced by side
7305 * activity. The imbalance_pct is used for the threshold.
7306 * Return true is the capacity is reduced
7309 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7311 return ((rq->cpu_capacity * sd->imbalance_pct) <
7312 (rq->cpu_capacity_orig * 100));
7316 * Group imbalance indicates (and tries to solve) the problem where balancing
7317 * groups is inadequate due to tsk_cpus_allowed() constraints.
7319 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7320 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7323 * { 0 1 2 3 } { 4 5 6 7 }
7326 * If we were to balance group-wise we'd place two tasks in the first group and
7327 * two tasks in the second group. Clearly this is undesired as it will overload
7328 * cpu 3 and leave one of the cpus in the second group unused.
7330 * The current solution to this issue is detecting the skew in the first group
7331 * by noticing the lower domain failed to reach balance and had difficulty
7332 * moving tasks due to affinity constraints.
7334 * When this is so detected; this group becomes a candidate for busiest; see
7335 * update_sd_pick_busiest(). And calculate_imbalance() and
7336 * find_busiest_group() avoid some of the usual balance conditions to allow it
7337 * to create an effective group imbalance.
7339 * This is a somewhat tricky proposition since the next run might not find the
7340 * group imbalance and decide the groups need to be balanced again. A most
7341 * subtle and fragile situation.
7344 static inline int sg_imbalanced(struct sched_group *group)
7346 return group->sgc->imbalance;
7350 * group_has_capacity returns true if the group has spare capacity that could
7351 * be used by some tasks.
7352 * We consider that a group has spare capacity if the * number of task is
7353 * smaller than the number of CPUs or if the utilization is lower than the
7354 * available capacity for CFS tasks.
7355 * For the latter, we use a threshold to stabilize the state, to take into
7356 * account the variance of the tasks' load and to return true if the available
7357 * capacity in meaningful for the load balancer.
7358 * As an example, an available capacity of 1% can appear but it doesn't make
7359 * any benefit for the load balance.
7362 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7364 if (sgs->sum_nr_running < sgs->group_weight)
7367 if ((sgs->group_capacity * 100) >
7368 (sgs->group_util * env->sd->imbalance_pct))
7375 * group_is_overloaded returns true if the group has more tasks than it can
7377 * group_is_overloaded is not equals to !group_has_capacity because a group
7378 * with the exact right number of tasks, has no more spare capacity but is not
7379 * overloaded so both group_has_capacity and group_is_overloaded return
7383 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7385 if (sgs->sum_nr_running <= sgs->group_weight)
7388 if ((sgs->group_capacity * 100) <
7389 (sgs->group_util * env->sd->imbalance_pct))
7397 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7398 * per-cpu capacity than sched_group ref.
7401 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7403 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7404 ref->sgc->max_capacity;
7408 group_type group_classify(struct sched_group *group,
7409 struct sg_lb_stats *sgs)
7411 if (sgs->group_no_capacity)
7412 return group_overloaded;
7414 if (sg_imbalanced(group))
7415 return group_imbalanced;
7417 if (sgs->group_misfit_task)
7418 return group_misfit_task;
7424 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7425 * @env: The load balancing environment.
7426 * @group: sched_group whose statistics are to be updated.
7427 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7428 * @local_group: Does group contain this_cpu.
7429 * @sgs: variable to hold the statistics for this group.
7430 * @overload: Indicate more than one runnable task for any CPU.
7431 * @overutilized: Indicate overutilization for any CPU.
7433 static inline void update_sg_lb_stats(struct lb_env *env,
7434 struct sched_group *group, int load_idx,
7435 int local_group, struct sg_lb_stats *sgs,
7436 bool *overload, bool *overutilized)
7441 memset(sgs, 0, sizeof(*sgs));
7443 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7444 struct rq *rq = cpu_rq(i);
7446 /* Bias balancing toward cpus of our domain */
7448 load = target_load(i, load_idx);
7450 load = source_load(i, load_idx);
7452 sgs->group_load += load;
7453 sgs->group_util += cpu_util(i);
7454 sgs->sum_nr_running += rq->cfs.h_nr_running;
7456 nr_running = rq->nr_running;
7460 #ifdef CONFIG_NUMA_BALANCING
7461 sgs->nr_numa_running += rq->nr_numa_running;
7462 sgs->nr_preferred_running += rq->nr_preferred_running;
7464 sgs->sum_weighted_load += weighted_cpuload(i);
7466 * No need to call idle_cpu() if nr_running is not 0
7468 if (!nr_running && idle_cpu(i))
7471 if (cpu_overutilized(i)) {
7472 *overutilized = true;
7473 if (!sgs->group_misfit_task && rq->misfit_task)
7474 sgs->group_misfit_task = capacity_of(i);
7478 /* Adjust by relative CPU capacity of the group */
7479 sgs->group_capacity = group->sgc->capacity;
7480 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7482 if (sgs->sum_nr_running)
7483 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7485 sgs->group_weight = group->group_weight;
7487 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7488 sgs->group_type = group_classify(group, sgs);
7492 * update_sd_pick_busiest - return 1 on busiest group
7493 * @env: The load balancing environment.
7494 * @sds: sched_domain statistics
7495 * @sg: sched_group candidate to be checked for being the busiest
7496 * @sgs: sched_group statistics
7498 * Determine if @sg is a busier group than the previously selected
7501 * Return: %true if @sg is a busier group than the previously selected
7502 * busiest group. %false otherwise.
7504 static bool update_sd_pick_busiest(struct lb_env *env,
7505 struct sd_lb_stats *sds,
7506 struct sched_group *sg,
7507 struct sg_lb_stats *sgs)
7509 struct sg_lb_stats *busiest = &sds->busiest_stat;
7511 if (sgs->group_type > busiest->group_type)
7514 if (sgs->group_type < busiest->group_type)
7518 * Candidate sg doesn't face any serious load-balance problems
7519 * so don't pick it if the local sg is already filled up.
7521 if (sgs->group_type == group_other &&
7522 !group_has_capacity(env, &sds->local_stat))
7525 if (sgs->avg_load <= busiest->avg_load)
7528 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7532 * Candidate sg has no more than one task per CPU and
7533 * has higher per-CPU capacity. Migrating tasks to less
7534 * capable CPUs may harm throughput. Maximize throughput,
7535 * power/energy consequences are not considered.
7537 if (sgs->sum_nr_running <= sgs->group_weight &&
7538 group_smaller_cpu_capacity(sds->local, sg))
7542 /* This is the busiest node in its class. */
7543 if (!(env->sd->flags & SD_ASYM_PACKING))
7547 * ASYM_PACKING needs to move all the work to the lowest
7548 * numbered CPUs in the group, therefore mark all groups
7549 * higher than ourself as busy.
7551 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7555 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7562 #ifdef CONFIG_NUMA_BALANCING
7563 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7565 if (sgs->sum_nr_running > sgs->nr_numa_running)
7567 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7572 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7574 if (rq->nr_running > rq->nr_numa_running)
7576 if (rq->nr_running > rq->nr_preferred_running)
7581 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7586 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7590 #endif /* CONFIG_NUMA_BALANCING */
7593 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7594 * @env: The load balancing environment.
7595 * @sds: variable to hold the statistics for this sched_domain.
7597 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7599 struct sched_domain *child = env->sd->child;
7600 struct sched_group *sg = env->sd->groups;
7601 struct sg_lb_stats tmp_sgs;
7602 int load_idx, prefer_sibling = 0;
7603 bool overload = false, overutilized = false;
7605 if (child && child->flags & SD_PREFER_SIBLING)
7608 load_idx = get_sd_load_idx(env->sd, env->idle);
7611 struct sg_lb_stats *sgs = &tmp_sgs;
7614 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7617 sgs = &sds->local_stat;
7619 if (env->idle != CPU_NEWLY_IDLE ||
7620 time_after_eq(jiffies, sg->sgc->next_update))
7621 update_group_capacity(env->sd, env->dst_cpu);
7624 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7625 &overload, &overutilized);
7631 * In case the child domain prefers tasks go to siblings
7632 * first, lower the sg capacity so that we'll try
7633 * and move all the excess tasks away. We lower the capacity
7634 * of a group only if the local group has the capacity to fit
7635 * these excess tasks. The extra check prevents the case where
7636 * you always pull from the heaviest group when it is already
7637 * under-utilized (possible with a large weight task outweighs
7638 * the tasks on the system).
7640 if (prefer_sibling && sds->local &&
7641 group_has_capacity(env, &sds->local_stat) &&
7642 (sgs->sum_nr_running > 1)) {
7643 sgs->group_no_capacity = 1;
7644 sgs->group_type = group_classify(sg, sgs);
7648 * Ignore task groups with misfit tasks if local group has no
7649 * capacity or if per-cpu capacity isn't higher.
7651 if (sgs->group_type == group_misfit_task &&
7652 (!group_has_capacity(env, &sds->local_stat) ||
7653 !group_smaller_cpu_capacity(sg, sds->local)))
7654 sgs->group_type = group_other;
7656 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7658 sds->busiest_stat = *sgs;
7662 /* Now, start updating sd_lb_stats */
7663 sds->total_load += sgs->group_load;
7664 sds->total_capacity += sgs->group_capacity;
7667 } while (sg != env->sd->groups);
7669 if (env->sd->flags & SD_NUMA)
7670 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7672 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7674 if (!env->sd->parent) {
7675 /* update overload indicator if we are at root domain */
7676 if (env->dst_rq->rd->overload != overload)
7677 env->dst_rq->rd->overload = overload;
7679 /* Update over-utilization (tipping point, U >= 0) indicator */
7680 if (env->dst_rq->rd->overutilized != overutilized) {
7681 env->dst_rq->rd->overutilized = overutilized;
7682 trace_sched_overutilized(overutilized);
7685 if (!env->dst_rq->rd->overutilized && overutilized) {
7686 env->dst_rq->rd->overutilized = true;
7687 trace_sched_overutilized(true);
7694 * check_asym_packing - Check to see if the group is packed into the
7697 * This is primarily intended to used at the sibling level. Some
7698 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7699 * case of POWER7, it can move to lower SMT modes only when higher
7700 * threads are idle. When in lower SMT modes, the threads will
7701 * perform better since they share less core resources. Hence when we
7702 * have idle threads, we want them to be the higher ones.
7704 * This packing function is run on idle threads. It checks to see if
7705 * the busiest CPU in this domain (core in the P7 case) has a higher
7706 * CPU number than the packing function is being run on. Here we are
7707 * assuming lower CPU number will be equivalent to lower a SMT thread
7710 * Return: 1 when packing is required and a task should be moved to
7711 * this CPU. The amount of the imbalance is returned in *imbalance.
7713 * @env: The load balancing environment.
7714 * @sds: Statistics of the sched_domain which is to be packed
7716 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7720 if (!(env->sd->flags & SD_ASYM_PACKING))
7726 busiest_cpu = group_first_cpu(sds->busiest);
7727 if (env->dst_cpu > busiest_cpu)
7730 env->imbalance = DIV_ROUND_CLOSEST(
7731 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7732 SCHED_CAPACITY_SCALE);
7738 * fix_small_imbalance - Calculate the minor imbalance that exists
7739 * amongst the groups of a sched_domain, during
7741 * @env: The load balancing environment.
7742 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7745 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7747 unsigned long tmp, capa_now = 0, capa_move = 0;
7748 unsigned int imbn = 2;
7749 unsigned long scaled_busy_load_per_task;
7750 struct sg_lb_stats *local, *busiest;
7752 local = &sds->local_stat;
7753 busiest = &sds->busiest_stat;
7755 if (!local->sum_nr_running)
7756 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7757 else if (busiest->load_per_task > local->load_per_task)
7760 scaled_busy_load_per_task =
7761 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7762 busiest->group_capacity;
7764 if (busiest->avg_load + scaled_busy_load_per_task >=
7765 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7766 env->imbalance = busiest->load_per_task;
7771 * OK, we don't have enough imbalance to justify moving tasks,
7772 * however we may be able to increase total CPU capacity used by
7776 capa_now += busiest->group_capacity *
7777 min(busiest->load_per_task, busiest->avg_load);
7778 capa_now += local->group_capacity *
7779 min(local->load_per_task, local->avg_load);
7780 capa_now /= SCHED_CAPACITY_SCALE;
7782 /* Amount of load we'd subtract */
7783 if (busiest->avg_load > scaled_busy_load_per_task) {
7784 capa_move += busiest->group_capacity *
7785 min(busiest->load_per_task,
7786 busiest->avg_load - scaled_busy_load_per_task);
7789 /* Amount of load we'd add */
7790 if (busiest->avg_load * busiest->group_capacity <
7791 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7792 tmp = (busiest->avg_load * busiest->group_capacity) /
7793 local->group_capacity;
7795 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7796 local->group_capacity;
7798 capa_move += local->group_capacity *
7799 min(local->load_per_task, local->avg_load + tmp);
7800 capa_move /= SCHED_CAPACITY_SCALE;
7802 /* Move if we gain throughput */
7803 if (capa_move > capa_now)
7804 env->imbalance = busiest->load_per_task;
7808 * calculate_imbalance - Calculate the amount of imbalance present within the
7809 * groups of a given sched_domain during load balance.
7810 * @env: load balance environment
7811 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7813 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7815 unsigned long max_pull, load_above_capacity = ~0UL;
7816 struct sg_lb_stats *local, *busiest;
7818 local = &sds->local_stat;
7819 busiest = &sds->busiest_stat;
7821 if (busiest->group_type == group_imbalanced) {
7823 * In the group_imb case we cannot rely on group-wide averages
7824 * to ensure cpu-load equilibrium, look at wider averages. XXX
7826 busiest->load_per_task =
7827 min(busiest->load_per_task, sds->avg_load);
7831 * In the presence of smp nice balancing, certain scenarios can have
7832 * max load less than avg load(as we skip the groups at or below
7833 * its cpu_capacity, while calculating max_load..)
7835 if (busiest->avg_load <= sds->avg_load ||
7836 local->avg_load >= sds->avg_load) {
7837 /* Misfitting tasks should be migrated in any case */
7838 if (busiest->group_type == group_misfit_task) {
7839 env->imbalance = busiest->group_misfit_task;
7844 * Busiest group is overloaded, local is not, use the spare
7845 * cycles to maximize throughput
7847 if (busiest->group_type == group_overloaded &&
7848 local->group_type <= group_misfit_task) {
7849 env->imbalance = busiest->load_per_task;
7854 return fix_small_imbalance(env, sds);
7858 * If there aren't any idle cpus, avoid creating some.
7860 if (busiest->group_type == group_overloaded &&
7861 local->group_type == group_overloaded) {
7862 load_above_capacity = busiest->sum_nr_running *
7864 if (load_above_capacity > busiest->group_capacity)
7865 load_above_capacity -= busiest->group_capacity;
7867 load_above_capacity = ~0UL;
7871 * We're trying to get all the cpus to the average_load, so we don't
7872 * want to push ourselves above the average load, nor do we wish to
7873 * reduce the max loaded cpu below the average load. At the same time,
7874 * we also don't want to reduce the group load below the group capacity
7875 * (so that we can implement power-savings policies etc). Thus we look
7876 * for the minimum possible imbalance.
7878 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7880 /* How much load to actually move to equalise the imbalance */
7881 env->imbalance = min(
7882 max_pull * busiest->group_capacity,
7883 (sds->avg_load - local->avg_load) * local->group_capacity
7884 ) / SCHED_CAPACITY_SCALE;
7886 /* Boost imbalance to allow misfit task to be balanced. */
7887 if (busiest->group_type == group_misfit_task)
7888 env->imbalance = max_t(long, env->imbalance,
7889 busiest->group_misfit_task);
7892 * if *imbalance is less than the average load per runnable task
7893 * there is no guarantee that any tasks will be moved so we'll have
7894 * a think about bumping its value to force at least one task to be
7897 if (env->imbalance < busiest->load_per_task)
7898 return fix_small_imbalance(env, sds);
7901 /******* find_busiest_group() helpers end here *********************/
7904 * find_busiest_group - Returns the busiest group within the sched_domain
7905 * if there is an imbalance. If there isn't an imbalance, and
7906 * the user has opted for power-savings, it returns a group whose
7907 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7908 * such a group exists.
7910 * Also calculates the amount of weighted load which should be moved
7911 * to restore balance.
7913 * @env: The load balancing environment.
7915 * Return: - The busiest group if imbalance exists.
7916 * - If no imbalance and user has opted for power-savings balance,
7917 * return the least loaded group whose CPUs can be
7918 * put to idle by rebalancing its tasks onto our group.
7920 static struct sched_group *find_busiest_group(struct lb_env *env)
7922 struct sg_lb_stats *local, *busiest;
7923 struct sd_lb_stats sds;
7925 init_sd_lb_stats(&sds);
7928 * Compute the various statistics relavent for load balancing at
7931 update_sd_lb_stats(env, &sds);
7933 if (energy_aware() && !env->dst_rq->rd->overutilized)
7936 local = &sds.local_stat;
7937 busiest = &sds.busiest_stat;
7939 /* ASYM feature bypasses nice load balance check */
7940 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7941 check_asym_packing(env, &sds))
7944 /* There is no busy sibling group to pull tasks from */
7945 if (!sds.busiest || busiest->sum_nr_running == 0)
7948 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7949 / sds.total_capacity;
7952 * If the busiest group is imbalanced the below checks don't
7953 * work because they assume all things are equal, which typically
7954 * isn't true due to cpus_allowed constraints and the like.
7956 if (busiest->group_type == group_imbalanced)
7959 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7960 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7961 busiest->group_no_capacity)
7964 /* Misfitting tasks should be dealt with regardless of the avg load */
7965 if (busiest->group_type == group_misfit_task) {
7970 * If the local group is busier than the selected busiest group
7971 * don't try and pull any tasks.
7973 if (local->avg_load >= busiest->avg_load)
7977 * Don't pull any tasks if this group is already above the domain
7980 if (local->avg_load >= sds.avg_load)
7983 if (env->idle == CPU_IDLE) {
7985 * This cpu is idle. If the busiest group is not overloaded
7986 * and there is no imbalance between this and busiest group
7987 * wrt idle cpus, it is balanced. The imbalance becomes
7988 * significant if the diff is greater than 1 otherwise we
7989 * might end up to just move the imbalance on another group
7991 if ((busiest->group_type != group_overloaded) &&
7992 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7993 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7997 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7998 * imbalance_pct to be conservative.
8000 if (100 * busiest->avg_load <=
8001 env->sd->imbalance_pct * local->avg_load)
8006 env->busiest_group_type = busiest->group_type;
8007 /* Looks like there is an imbalance. Compute it */
8008 calculate_imbalance(env, &sds);
8017 * find_busiest_queue - find the busiest runqueue among the cpus in group.
8019 static struct rq *find_busiest_queue(struct lb_env *env,
8020 struct sched_group *group)
8022 struct rq *busiest = NULL, *rq;
8023 unsigned long busiest_load = 0, busiest_capacity = 1;
8026 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
8027 unsigned long capacity, wl;
8031 rt = fbq_classify_rq(rq);
8034 * We classify groups/runqueues into three groups:
8035 * - regular: there are !numa tasks
8036 * - remote: there are numa tasks that run on the 'wrong' node
8037 * - all: there is no distinction
8039 * In order to avoid migrating ideally placed numa tasks,
8040 * ignore those when there's better options.
8042 * If we ignore the actual busiest queue to migrate another
8043 * task, the next balance pass can still reduce the busiest
8044 * queue by moving tasks around inside the node.
8046 * If we cannot move enough load due to this classification
8047 * the next pass will adjust the group classification and
8048 * allow migration of more tasks.
8050 * Both cases only affect the total convergence complexity.
8052 if (rt > env->fbq_type)
8055 capacity = capacity_of(i);
8057 wl = weighted_cpuload(i);
8060 * When comparing with imbalance, use weighted_cpuload()
8061 * which is not scaled with the cpu capacity.
8064 if (rq->nr_running == 1 && wl > env->imbalance &&
8065 !check_cpu_capacity(rq, env->sd) &&
8066 env->busiest_group_type != group_misfit_task)
8070 * For the load comparisons with the other cpu's, consider
8071 * the weighted_cpuload() scaled with the cpu capacity, so
8072 * that the load can be moved away from the cpu that is
8073 * potentially running at a lower capacity.
8075 * Thus we're looking for max(wl_i / capacity_i), crosswise
8076 * multiplication to rid ourselves of the division works out
8077 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8078 * our previous maximum.
8080 if (wl * busiest_capacity > busiest_load * capacity) {
8082 busiest_capacity = capacity;
8091 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8092 * so long as it is large enough.
8094 #define MAX_PINNED_INTERVAL 512
8096 /* Working cpumask for load_balance and load_balance_newidle. */
8097 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8099 static int need_active_balance(struct lb_env *env)
8101 struct sched_domain *sd = env->sd;
8103 if (env->idle == CPU_NEWLY_IDLE) {
8106 * ASYM_PACKING needs to force migrate tasks from busy but
8107 * higher numbered CPUs in order to pack all tasks in the
8108 * lowest numbered CPUs.
8110 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8115 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8116 * It's worth migrating the task if the src_cpu's capacity is reduced
8117 * because of other sched_class or IRQs if more capacity stays
8118 * available on dst_cpu.
8120 if ((env->idle != CPU_NOT_IDLE) &&
8121 (env->src_rq->cfs.h_nr_running == 1)) {
8122 if ((check_cpu_capacity(env->src_rq, sd)) &&
8123 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8127 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8128 env->src_rq->cfs.h_nr_running == 1 &&
8129 cpu_overutilized(env->src_cpu) &&
8130 !cpu_overutilized(env->dst_cpu)) {
8134 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8137 static int active_load_balance_cpu_stop(void *data);
8139 static int should_we_balance(struct lb_env *env)
8141 struct sched_group *sg = env->sd->groups;
8142 struct cpumask *sg_cpus, *sg_mask;
8143 int cpu, balance_cpu = -1;
8146 * In the newly idle case, we will allow all the cpu's
8147 * to do the newly idle load balance.
8149 if (env->idle == CPU_NEWLY_IDLE)
8152 sg_cpus = sched_group_cpus(sg);
8153 sg_mask = sched_group_mask(sg);
8154 /* Try to find first idle cpu */
8155 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8156 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8163 if (balance_cpu == -1)
8164 balance_cpu = group_balance_cpu(sg);
8167 * First idle cpu or the first cpu(busiest) in this sched group
8168 * is eligible for doing load balancing at this and above domains.
8170 return balance_cpu == env->dst_cpu;
8174 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8175 * tasks if there is an imbalance.
8177 static int load_balance(int this_cpu, struct rq *this_rq,
8178 struct sched_domain *sd, enum cpu_idle_type idle,
8179 int *continue_balancing)
8181 int ld_moved, cur_ld_moved, active_balance = 0;
8182 struct sched_domain *sd_parent = sd->parent;
8183 struct sched_group *group;
8185 unsigned long flags;
8186 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8188 struct lb_env env = {
8190 .dst_cpu = this_cpu,
8192 .dst_grpmask = sched_group_cpus(sd->groups),
8194 .loop_break = sched_nr_migrate_break,
8197 .tasks = LIST_HEAD_INIT(env.tasks),
8201 * For NEWLY_IDLE load_balancing, we don't need to consider
8202 * other cpus in our group
8204 if (idle == CPU_NEWLY_IDLE)
8205 env.dst_grpmask = NULL;
8207 cpumask_copy(cpus, cpu_active_mask);
8209 schedstat_inc(sd, lb_count[idle]);
8212 if (!should_we_balance(&env)) {
8213 *continue_balancing = 0;
8217 group = find_busiest_group(&env);
8219 schedstat_inc(sd, lb_nobusyg[idle]);
8223 busiest = find_busiest_queue(&env, group);
8225 schedstat_inc(sd, lb_nobusyq[idle]);
8229 BUG_ON(busiest == env.dst_rq);
8231 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8233 env.src_cpu = busiest->cpu;
8234 env.src_rq = busiest;
8237 if (busiest->nr_running > 1) {
8239 * Attempt to move tasks. If find_busiest_group has found
8240 * an imbalance but busiest->nr_running <= 1, the group is
8241 * still unbalanced. ld_moved simply stays zero, so it is
8242 * correctly treated as an imbalance.
8244 env.flags |= LBF_ALL_PINNED;
8245 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8248 raw_spin_lock_irqsave(&busiest->lock, flags);
8251 * cur_ld_moved - load moved in current iteration
8252 * ld_moved - cumulative load moved across iterations
8254 cur_ld_moved = detach_tasks(&env);
8256 * We want to potentially lower env.src_cpu's OPP.
8259 update_capacity_of(env.src_cpu);
8262 * We've detached some tasks from busiest_rq. Every
8263 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8264 * unlock busiest->lock, and we are able to be sure
8265 * that nobody can manipulate the tasks in parallel.
8266 * See task_rq_lock() family for the details.
8269 raw_spin_unlock(&busiest->lock);
8273 ld_moved += cur_ld_moved;
8276 local_irq_restore(flags);
8278 if (env.flags & LBF_NEED_BREAK) {
8279 env.flags &= ~LBF_NEED_BREAK;
8284 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8285 * us and move them to an alternate dst_cpu in our sched_group
8286 * where they can run. The upper limit on how many times we
8287 * iterate on same src_cpu is dependent on number of cpus in our
8290 * This changes load balance semantics a bit on who can move
8291 * load to a given_cpu. In addition to the given_cpu itself
8292 * (or a ilb_cpu acting on its behalf where given_cpu is
8293 * nohz-idle), we now have balance_cpu in a position to move
8294 * load to given_cpu. In rare situations, this may cause
8295 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8296 * _independently_ and at _same_ time to move some load to
8297 * given_cpu) causing exceess load to be moved to given_cpu.
8298 * This however should not happen so much in practice and
8299 * moreover subsequent load balance cycles should correct the
8300 * excess load moved.
8302 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8304 /* Prevent to re-select dst_cpu via env's cpus */
8305 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8307 env.dst_rq = cpu_rq(env.new_dst_cpu);
8308 env.dst_cpu = env.new_dst_cpu;
8309 env.flags &= ~LBF_DST_PINNED;
8311 env.loop_break = sched_nr_migrate_break;
8314 * Go back to "more_balance" rather than "redo" since we
8315 * need to continue with same src_cpu.
8321 * We failed to reach balance because of affinity.
8324 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8326 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8327 *group_imbalance = 1;
8330 /* All tasks on this runqueue were pinned by CPU affinity */
8331 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8332 cpumask_clear_cpu(cpu_of(busiest), cpus);
8333 if (!cpumask_empty(cpus)) {
8335 env.loop_break = sched_nr_migrate_break;
8338 goto out_all_pinned;
8343 schedstat_inc(sd, lb_failed[idle]);
8345 * Increment the failure counter only on periodic balance.
8346 * We do not want newidle balance, which can be very
8347 * frequent, pollute the failure counter causing
8348 * excessive cache_hot migrations and active balances.
8350 if (idle != CPU_NEWLY_IDLE)
8351 if (env.src_grp_nr_running > 1)
8352 sd->nr_balance_failed++;
8354 if (need_active_balance(&env)) {
8355 raw_spin_lock_irqsave(&busiest->lock, flags);
8357 /* don't kick the active_load_balance_cpu_stop,
8358 * if the curr task on busiest cpu can't be
8361 if (!cpumask_test_cpu(this_cpu,
8362 tsk_cpus_allowed(busiest->curr))) {
8363 raw_spin_unlock_irqrestore(&busiest->lock,
8365 env.flags |= LBF_ALL_PINNED;
8366 goto out_one_pinned;
8370 * ->active_balance synchronizes accesses to
8371 * ->active_balance_work. Once set, it's cleared
8372 * only after active load balance is finished.
8374 if (!busiest->active_balance) {
8375 busiest->active_balance = 1;
8376 busiest->push_cpu = this_cpu;
8379 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8381 if (active_balance) {
8382 stop_one_cpu_nowait(cpu_of(busiest),
8383 active_load_balance_cpu_stop, busiest,
8384 &busiest->active_balance_work);
8388 * We've kicked active balancing, reset the failure
8391 sd->nr_balance_failed = sd->cache_nice_tries+1;
8394 sd->nr_balance_failed = 0;
8396 if (likely(!active_balance)) {
8397 /* We were unbalanced, so reset the balancing interval */
8398 sd->balance_interval = sd->min_interval;
8401 * If we've begun active balancing, start to back off. This
8402 * case may not be covered by the all_pinned logic if there
8403 * is only 1 task on the busy runqueue (because we don't call
8406 if (sd->balance_interval < sd->max_interval)
8407 sd->balance_interval *= 2;
8414 * We reach balance although we may have faced some affinity
8415 * constraints. Clear the imbalance flag if it was set.
8418 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8420 if (*group_imbalance)
8421 *group_imbalance = 0;
8426 * We reach balance because all tasks are pinned at this level so
8427 * we can't migrate them. Let the imbalance flag set so parent level
8428 * can try to migrate them.
8430 schedstat_inc(sd, lb_balanced[idle]);
8432 sd->nr_balance_failed = 0;
8435 /* tune up the balancing interval */
8436 if (((env.flags & LBF_ALL_PINNED) &&
8437 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8438 (sd->balance_interval < sd->max_interval))
8439 sd->balance_interval *= 2;
8446 static inline unsigned long
8447 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8449 unsigned long interval = sd->balance_interval;
8452 interval *= sd->busy_factor;
8454 /* scale ms to jiffies */
8455 interval = msecs_to_jiffies(interval);
8456 interval = clamp(interval, 1UL, max_load_balance_interval);
8462 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8464 unsigned long interval, next;
8466 interval = get_sd_balance_interval(sd, cpu_busy);
8467 next = sd->last_balance + interval;
8469 if (time_after(*next_balance, next))
8470 *next_balance = next;
8474 * idle_balance is called by schedule() if this_cpu is about to become
8475 * idle. Attempts to pull tasks from other CPUs.
8477 static int idle_balance(struct rq *this_rq)
8479 unsigned long next_balance = jiffies + HZ;
8480 int this_cpu = this_rq->cpu;
8481 struct sched_domain *sd;
8482 int pulled_task = 0;
8484 long removed_util=0;
8486 idle_enter_fair(this_rq);
8489 * We must set idle_stamp _before_ calling idle_balance(), such that we
8490 * measure the duration of idle_balance() as idle time.
8492 this_rq->idle_stamp = rq_clock(this_rq);
8494 if (!energy_aware() &&
8495 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8496 !this_rq->rd->overload)) {
8498 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8500 update_next_balance(sd, 0, &next_balance);
8506 raw_spin_unlock(&this_rq->lock);
8509 * If removed_util_avg is !0 we most probably migrated some task away
8510 * from this_cpu. In this case we might be willing to trigger an OPP
8511 * update, but we want to do so if we don't find anybody else to pull
8512 * here (we will trigger an OPP update with the pulled task's enqueue
8515 * Record removed_util before calling update_blocked_averages, and use
8516 * it below (before returning) to see if an OPP update is required.
8518 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8519 update_blocked_averages(this_cpu);
8521 for_each_domain(this_cpu, sd) {
8522 int continue_balancing = 1;
8523 u64 t0, domain_cost;
8525 if (!(sd->flags & SD_LOAD_BALANCE))
8528 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8529 update_next_balance(sd, 0, &next_balance);
8533 if (sd->flags & SD_BALANCE_NEWIDLE) {
8534 t0 = sched_clock_cpu(this_cpu);
8536 pulled_task = load_balance(this_cpu, this_rq,
8538 &continue_balancing);
8540 domain_cost = sched_clock_cpu(this_cpu) - t0;
8541 if (domain_cost > sd->max_newidle_lb_cost)
8542 sd->max_newidle_lb_cost = domain_cost;
8544 curr_cost += domain_cost;
8547 update_next_balance(sd, 0, &next_balance);
8550 * Stop searching for tasks to pull if there are
8551 * now runnable tasks on this rq.
8553 if (pulled_task || this_rq->nr_running > 0)
8558 raw_spin_lock(&this_rq->lock);
8560 if (curr_cost > this_rq->max_idle_balance_cost)
8561 this_rq->max_idle_balance_cost = curr_cost;
8564 * While browsing the domains, we released the rq lock, a task could
8565 * have been enqueued in the meantime. Since we're not going idle,
8566 * pretend we pulled a task.
8568 if (this_rq->cfs.h_nr_running && !pulled_task)
8572 /* Move the next balance forward */
8573 if (time_after(this_rq->next_balance, next_balance))
8574 this_rq->next_balance = next_balance;
8576 /* Is there a task of a high priority class? */
8577 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8581 idle_exit_fair(this_rq);
8582 this_rq->idle_stamp = 0;
8583 } else if (removed_util) {
8585 * No task pulled and someone has been migrated away.
8586 * Good case to trigger an OPP update.
8588 update_capacity_of(this_cpu);
8595 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8596 * running tasks off the busiest CPU onto idle CPUs. It requires at
8597 * least 1 task to be running on each physical CPU where possible, and
8598 * avoids physical / logical imbalances.
8600 static int active_load_balance_cpu_stop(void *data)
8602 struct rq *busiest_rq = data;
8603 int busiest_cpu = cpu_of(busiest_rq);
8604 int target_cpu = busiest_rq->push_cpu;
8605 struct rq *target_rq = cpu_rq(target_cpu);
8606 struct sched_domain *sd;
8607 struct task_struct *p = NULL;
8609 raw_spin_lock_irq(&busiest_rq->lock);
8611 /* make sure the requested cpu hasn't gone down in the meantime */
8612 if (unlikely(busiest_cpu != smp_processor_id() ||
8613 !busiest_rq->active_balance))
8616 /* Is there any task to move? */
8617 if (busiest_rq->nr_running <= 1)
8621 * This condition is "impossible", if it occurs
8622 * we need to fix it. Originally reported by
8623 * Bjorn Helgaas on a 128-cpu setup.
8625 BUG_ON(busiest_rq == target_rq);
8627 /* Search for an sd spanning us and the target CPU. */
8629 for_each_domain(target_cpu, sd) {
8630 if ((sd->flags & SD_LOAD_BALANCE) &&
8631 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8636 struct lb_env env = {
8638 .dst_cpu = target_cpu,
8639 .dst_rq = target_rq,
8640 .src_cpu = busiest_rq->cpu,
8641 .src_rq = busiest_rq,
8645 schedstat_inc(sd, alb_count);
8647 p = detach_one_task(&env);
8649 schedstat_inc(sd, alb_pushed);
8651 * We want to potentially lower env.src_cpu's OPP.
8653 update_capacity_of(env.src_cpu);
8656 schedstat_inc(sd, alb_failed);
8660 busiest_rq->active_balance = 0;
8661 raw_spin_unlock(&busiest_rq->lock);
8664 attach_one_task(target_rq, p);
8671 static inline int on_null_domain(struct rq *rq)
8673 return unlikely(!rcu_dereference_sched(rq->sd));
8676 #ifdef CONFIG_NO_HZ_COMMON
8678 * idle load balancing details
8679 * - When one of the busy CPUs notice that there may be an idle rebalancing
8680 * needed, they will kick the idle load balancer, which then does idle
8681 * load balancing for all the idle CPUs.
8684 cpumask_var_t idle_cpus_mask;
8686 unsigned long next_balance; /* in jiffy units */
8687 } nohz ____cacheline_aligned;
8689 static inline int find_new_ilb(void)
8691 int ilb = cpumask_first(nohz.idle_cpus_mask);
8693 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8700 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8701 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8702 * CPU (if there is one).
8704 static void nohz_balancer_kick(void)
8708 nohz.next_balance++;
8710 ilb_cpu = find_new_ilb();
8712 if (ilb_cpu >= nr_cpu_ids)
8715 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8718 * Use smp_send_reschedule() instead of resched_cpu().
8719 * This way we generate a sched IPI on the target cpu which
8720 * is idle. And the softirq performing nohz idle load balance
8721 * will be run before returning from the IPI.
8723 smp_send_reschedule(ilb_cpu);
8727 static inline void nohz_balance_exit_idle(int cpu)
8729 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8731 * Completely isolated CPUs don't ever set, so we must test.
8733 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8734 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8735 atomic_dec(&nohz.nr_cpus);
8737 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8741 static inline void set_cpu_sd_state_busy(void)
8743 struct sched_domain *sd;
8744 int cpu = smp_processor_id();
8747 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8749 if (!sd || !sd->nohz_idle)
8753 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8758 void set_cpu_sd_state_idle(void)
8760 struct sched_domain *sd;
8761 int cpu = smp_processor_id();
8764 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8766 if (!sd || sd->nohz_idle)
8770 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8776 * This routine will record that the cpu is going idle with tick stopped.
8777 * This info will be used in performing idle load balancing in the future.
8779 void nohz_balance_enter_idle(int cpu)
8782 * If this cpu is going down, then nothing needs to be done.
8784 if (!cpu_active(cpu))
8787 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8791 * If we're a completely isolated CPU, we don't play.
8793 if (on_null_domain(cpu_rq(cpu)))
8796 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8797 atomic_inc(&nohz.nr_cpus);
8798 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8801 static int sched_ilb_notifier(struct notifier_block *nfb,
8802 unsigned long action, void *hcpu)
8804 switch (action & ~CPU_TASKS_FROZEN) {
8806 nohz_balance_exit_idle(smp_processor_id());
8814 static DEFINE_SPINLOCK(balancing);
8817 * Scale the max load_balance interval with the number of CPUs in the system.
8818 * This trades load-balance latency on larger machines for less cross talk.
8820 void update_max_interval(void)
8822 max_load_balance_interval = HZ*num_online_cpus()/10;
8826 * It checks each scheduling domain to see if it is due to be balanced,
8827 * and initiates a balancing operation if so.
8829 * Balancing parameters are set up in init_sched_domains.
8831 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8833 int continue_balancing = 1;
8835 unsigned long interval;
8836 struct sched_domain *sd;
8837 /* Earliest time when we have to do rebalance again */
8838 unsigned long next_balance = jiffies + 60*HZ;
8839 int update_next_balance = 0;
8840 int need_serialize, need_decay = 0;
8843 update_blocked_averages(cpu);
8846 for_each_domain(cpu, sd) {
8848 * Decay the newidle max times here because this is a regular
8849 * visit to all the domains. Decay ~1% per second.
8851 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8852 sd->max_newidle_lb_cost =
8853 (sd->max_newidle_lb_cost * 253) / 256;
8854 sd->next_decay_max_lb_cost = jiffies + HZ;
8857 max_cost += sd->max_newidle_lb_cost;
8859 if (!(sd->flags & SD_LOAD_BALANCE))
8863 * Stop the load balance at this level. There is another
8864 * CPU in our sched group which is doing load balancing more
8867 if (!continue_balancing) {
8873 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8875 need_serialize = sd->flags & SD_SERIALIZE;
8876 if (need_serialize) {
8877 if (!spin_trylock(&balancing))
8881 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8882 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8884 * The LBF_DST_PINNED logic could have changed
8885 * env->dst_cpu, so we can't know our idle
8886 * state even if we migrated tasks. Update it.
8888 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8890 sd->last_balance = jiffies;
8891 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8894 spin_unlock(&balancing);
8896 if (time_after(next_balance, sd->last_balance + interval)) {
8897 next_balance = sd->last_balance + interval;
8898 update_next_balance = 1;
8903 * Ensure the rq-wide value also decays but keep it at a
8904 * reasonable floor to avoid funnies with rq->avg_idle.
8906 rq->max_idle_balance_cost =
8907 max((u64)sysctl_sched_migration_cost, max_cost);
8912 * next_balance will be updated only when there is a need.
8913 * When the cpu is attached to null domain for ex, it will not be
8916 if (likely(update_next_balance)) {
8917 rq->next_balance = next_balance;
8919 #ifdef CONFIG_NO_HZ_COMMON
8921 * If this CPU has been elected to perform the nohz idle
8922 * balance. Other idle CPUs have already rebalanced with
8923 * nohz_idle_balance() and nohz.next_balance has been
8924 * updated accordingly. This CPU is now running the idle load
8925 * balance for itself and we need to update the
8926 * nohz.next_balance accordingly.
8928 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8929 nohz.next_balance = rq->next_balance;
8934 #ifdef CONFIG_NO_HZ_COMMON
8936 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8937 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8939 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8941 int this_cpu = this_rq->cpu;
8944 /* Earliest time when we have to do rebalance again */
8945 unsigned long next_balance = jiffies + 60*HZ;
8946 int update_next_balance = 0;
8948 if (idle != CPU_IDLE ||
8949 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8952 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8953 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8957 * If this cpu gets work to do, stop the load balancing
8958 * work being done for other cpus. Next load
8959 * balancing owner will pick it up.
8964 rq = cpu_rq(balance_cpu);
8967 * If time for next balance is due,
8970 if (time_after_eq(jiffies, rq->next_balance)) {
8971 raw_spin_lock_irq(&rq->lock);
8972 update_rq_clock(rq);
8973 update_idle_cpu_load(rq);
8974 raw_spin_unlock_irq(&rq->lock);
8975 rebalance_domains(rq, CPU_IDLE);
8978 if (time_after(next_balance, rq->next_balance)) {
8979 next_balance = rq->next_balance;
8980 update_next_balance = 1;
8985 * next_balance will be updated only when there is a need.
8986 * When the CPU is attached to null domain for ex, it will not be
8989 if (likely(update_next_balance))
8990 nohz.next_balance = next_balance;
8992 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8996 * Current heuristic for kicking the idle load balancer in the presence
8997 * of an idle cpu in the system.
8998 * - This rq has more than one task.
8999 * - This rq has at least one CFS task and the capacity of the CPU is
9000 * significantly reduced because of RT tasks or IRQs.
9001 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9002 * multiple busy cpu.
9003 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9004 * domain span are idle.
9006 static inline bool nohz_kick_needed(struct rq *rq)
9008 unsigned long now = jiffies;
9009 struct sched_domain *sd;
9010 struct sched_group_capacity *sgc;
9011 int nr_busy, cpu = rq->cpu;
9014 if (unlikely(rq->idle_balance))
9018 * We may be recently in ticked or tickless idle mode. At the first
9019 * busy tick after returning from idle, we will update the busy stats.
9021 set_cpu_sd_state_busy();
9022 nohz_balance_exit_idle(cpu);
9025 * None are in tickless mode and hence no need for NOHZ idle load
9028 if (likely(!atomic_read(&nohz.nr_cpus)))
9031 if (time_before(now, nohz.next_balance))
9034 if (rq->nr_running >= 2 &&
9035 (!energy_aware() || cpu_overutilized(cpu)))
9039 sd = rcu_dereference(per_cpu(sd_busy, cpu));
9040 if (sd && !energy_aware()) {
9041 sgc = sd->groups->sgc;
9042 nr_busy = atomic_read(&sgc->nr_busy_cpus);
9051 sd = rcu_dereference(rq->sd);
9053 if ((rq->cfs.h_nr_running >= 1) &&
9054 check_cpu_capacity(rq, sd)) {
9060 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9061 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
9062 sched_domain_span(sd)) < cpu)) {
9072 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9076 * run_rebalance_domains is triggered when needed from the scheduler tick.
9077 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9079 static void run_rebalance_domains(struct softirq_action *h)
9081 struct rq *this_rq = this_rq();
9082 enum cpu_idle_type idle = this_rq->idle_balance ?
9083 CPU_IDLE : CPU_NOT_IDLE;
9086 * If this cpu has a pending nohz_balance_kick, then do the
9087 * balancing on behalf of the other idle cpus whose ticks are
9088 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9089 * give the idle cpus a chance to load balance. Else we may
9090 * load balance only within the local sched_domain hierarchy
9091 * and abort nohz_idle_balance altogether if we pull some load.
9093 nohz_idle_balance(this_rq, idle);
9094 rebalance_domains(this_rq, idle);
9098 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9100 void trigger_load_balance(struct rq *rq)
9102 /* Don't need to rebalance while attached to NULL domain */
9103 if (unlikely(on_null_domain(rq)))
9106 if (time_after_eq(jiffies, rq->next_balance))
9107 raise_softirq(SCHED_SOFTIRQ);
9108 #ifdef CONFIG_NO_HZ_COMMON
9109 if (nohz_kick_needed(rq))
9110 nohz_balancer_kick();
9114 static void rq_online_fair(struct rq *rq)
9118 update_runtime_enabled(rq);
9121 static void rq_offline_fair(struct rq *rq)
9125 /* Ensure any throttled groups are reachable by pick_next_task */
9126 unthrottle_offline_cfs_rqs(rq);
9129 #endif /* CONFIG_SMP */
9132 * scheduler tick hitting a task of our scheduling class:
9134 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9136 struct cfs_rq *cfs_rq;
9137 struct sched_entity *se = &curr->se;
9139 for_each_sched_entity(se) {
9140 cfs_rq = cfs_rq_of(se);
9141 entity_tick(cfs_rq, se, queued);
9144 if (static_branch_unlikely(&sched_numa_balancing))
9145 task_tick_numa(rq, curr);
9148 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9149 rq->rd->overutilized = true;
9150 trace_sched_overutilized(true);
9153 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9159 * called on fork with the child task as argument from the parent's context
9160 * - child not yet on the tasklist
9161 * - preemption disabled
9163 static void task_fork_fair(struct task_struct *p)
9165 struct cfs_rq *cfs_rq;
9166 struct sched_entity *se = &p->se, *curr;
9167 int this_cpu = smp_processor_id();
9168 struct rq *rq = this_rq();
9169 unsigned long flags;
9171 raw_spin_lock_irqsave(&rq->lock, flags);
9173 update_rq_clock(rq);
9175 cfs_rq = task_cfs_rq(current);
9176 curr = cfs_rq->curr;
9179 * Not only the cpu but also the task_group of the parent might have
9180 * been changed after parent->se.parent,cfs_rq were copied to
9181 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9182 * of child point to valid ones.
9185 __set_task_cpu(p, this_cpu);
9188 update_curr(cfs_rq);
9191 se->vruntime = curr->vruntime;
9192 place_entity(cfs_rq, se, 1);
9194 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9196 * Upon rescheduling, sched_class::put_prev_task() will place
9197 * 'current' within the tree based on its new key value.
9199 swap(curr->vruntime, se->vruntime);
9203 se->vruntime -= cfs_rq->min_vruntime;
9205 raw_spin_unlock_irqrestore(&rq->lock, flags);
9209 * Priority of the task has changed. Check to see if we preempt
9213 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9215 if (!task_on_rq_queued(p))
9219 * Reschedule if we are currently running on this runqueue and
9220 * our priority decreased, or if we are not currently running on
9221 * this runqueue and our priority is higher than the current's
9223 if (rq->curr == p) {
9224 if (p->prio > oldprio)
9227 check_preempt_curr(rq, p, 0);
9230 static inline bool vruntime_normalized(struct task_struct *p)
9232 struct sched_entity *se = &p->se;
9235 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9236 * the dequeue_entity(.flags=0) will already have normalized the
9243 * When !on_rq, vruntime of the task has usually NOT been normalized.
9244 * But there are some cases where it has already been normalized:
9246 * - A forked child which is waiting for being woken up by
9247 * wake_up_new_task().
9248 * - A task which has been woken up by try_to_wake_up() and
9249 * waiting for actually being woken up by sched_ttwu_pending().
9251 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9257 static void detach_task_cfs_rq(struct task_struct *p)
9259 struct sched_entity *se = &p->se;
9260 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9262 if (!vruntime_normalized(p)) {
9264 * Fix up our vruntime so that the current sleep doesn't
9265 * cause 'unlimited' sleep bonus.
9267 place_entity(cfs_rq, se, 0);
9268 se->vruntime -= cfs_rq->min_vruntime;
9271 /* Catch up with the cfs_rq and remove our load when we leave */
9272 detach_entity_load_avg(cfs_rq, se);
9275 static void attach_task_cfs_rq(struct task_struct *p)
9277 struct sched_entity *se = &p->se;
9278 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9280 #ifdef CONFIG_FAIR_GROUP_SCHED
9282 * Since the real-depth could have been changed (only FAIR
9283 * class maintain depth value), reset depth properly.
9285 se->depth = se->parent ? se->parent->depth + 1 : 0;
9288 /* Synchronize task with its cfs_rq */
9289 attach_entity_load_avg(cfs_rq, se);
9291 if (!vruntime_normalized(p))
9292 se->vruntime += cfs_rq->min_vruntime;
9295 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9297 detach_task_cfs_rq(p);
9300 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9302 attach_task_cfs_rq(p);
9304 if (task_on_rq_queued(p)) {
9306 * We were most likely switched from sched_rt, so
9307 * kick off the schedule if running, otherwise just see
9308 * if we can still preempt the current task.
9313 check_preempt_curr(rq, p, 0);
9317 /* Account for a task changing its policy or group.
9319 * This routine is mostly called to set cfs_rq->curr field when a task
9320 * migrates between groups/classes.
9322 static void set_curr_task_fair(struct rq *rq)
9324 struct sched_entity *se = &rq->curr->se;
9326 for_each_sched_entity(se) {
9327 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9329 set_next_entity(cfs_rq, se);
9330 /* ensure bandwidth has been allocated on our new cfs_rq */
9331 account_cfs_rq_runtime(cfs_rq, 0);
9335 void init_cfs_rq(struct cfs_rq *cfs_rq)
9337 cfs_rq->tasks_timeline = RB_ROOT;
9338 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9339 #ifndef CONFIG_64BIT
9340 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9343 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9344 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9348 #ifdef CONFIG_FAIR_GROUP_SCHED
9349 static void task_move_group_fair(struct task_struct *p)
9351 detach_task_cfs_rq(p);
9352 set_task_rq(p, task_cpu(p));
9355 /* Tell se's cfs_rq has been changed -- migrated */
9356 p->se.avg.last_update_time = 0;
9358 attach_task_cfs_rq(p);
9361 void free_fair_sched_group(struct task_group *tg)
9365 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9367 for_each_possible_cpu(i) {
9369 kfree(tg->cfs_rq[i]);
9372 remove_entity_load_avg(tg->se[i]);
9381 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9383 struct cfs_rq *cfs_rq;
9384 struct sched_entity *se;
9387 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9390 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9394 tg->shares = NICE_0_LOAD;
9396 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9398 for_each_possible_cpu(i) {
9399 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9400 GFP_KERNEL, cpu_to_node(i));
9404 se = kzalloc_node(sizeof(struct sched_entity),
9405 GFP_KERNEL, cpu_to_node(i));
9409 init_cfs_rq(cfs_rq);
9410 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9411 init_entity_runnable_average(se);
9422 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9424 struct rq *rq = cpu_rq(cpu);
9425 unsigned long flags;
9428 * Only empty task groups can be destroyed; so we can speculatively
9429 * check on_list without danger of it being re-added.
9431 if (!tg->cfs_rq[cpu]->on_list)
9434 raw_spin_lock_irqsave(&rq->lock, flags);
9435 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9436 raw_spin_unlock_irqrestore(&rq->lock, flags);
9439 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9440 struct sched_entity *se, int cpu,
9441 struct sched_entity *parent)
9443 struct rq *rq = cpu_rq(cpu);
9447 init_cfs_rq_runtime(cfs_rq);
9449 tg->cfs_rq[cpu] = cfs_rq;
9452 /* se could be NULL for root_task_group */
9457 se->cfs_rq = &rq->cfs;
9460 se->cfs_rq = parent->my_q;
9461 se->depth = parent->depth + 1;
9465 /* guarantee group entities always have weight */
9466 update_load_set(&se->load, NICE_0_LOAD);
9467 se->parent = parent;
9470 static DEFINE_MUTEX(shares_mutex);
9472 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9475 unsigned long flags;
9478 * We can't change the weight of the root cgroup.
9483 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9485 mutex_lock(&shares_mutex);
9486 if (tg->shares == shares)
9489 tg->shares = shares;
9490 for_each_possible_cpu(i) {
9491 struct rq *rq = cpu_rq(i);
9492 struct sched_entity *se;
9495 /* Propagate contribution to hierarchy */
9496 raw_spin_lock_irqsave(&rq->lock, flags);
9498 /* Possible calls to update_curr() need rq clock */
9499 update_rq_clock(rq);
9500 for_each_sched_entity(se)
9501 update_cfs_shares(group_cfs_rq(se));
9502 raw_spin_unlock_irqrestore(&rq->lock, flags);
9506 mutex_unlock(&shares_mutex);
9509 #else /* CONFIG_FAIR_GROUP_SCHED */
9511 void free_fair_sched_group(struct task_group *tg) { }
9513 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9518 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9520 #endif /* CONFIG_FAIR_GROUP_SCHED */
9523 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9525 struct sched_entity *se = &task->se;
9526 unsigned int rr_interval = 0;
9529 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9532 if (rq->cfs.load.weight)
9533 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9539 * All the scheduling class methods:
9541 const struct sched_class fair_sched_class = {
9542 .next = &idle_sched_class,
9543 .enqueue_task = enqueue_task_fair,
9544 .dequeue_task = dequeue_task_fair,
9545 .yield_task = yield_task_fair,
9546 .yield_to_task = yield_to_task_fair,
9548 .check_preempt_curr = check_preempt_wakeup,
9550 .pick_next_task = pick_next_task_fair,
9551 .put_prev_task = put_prev_task_fair,
9554 .select_task_rq = select_task_rq_fair,
9555 .migrate_task_rq = migrate_task_rq_fair,
9557 .rq_online = rq_online_fair,
9558 .rq_offline = rq_offline_fair,
9560 .task_waking = task_waking_fair,
9561 .task_dead = task_dead_fair,
9562 .set_cpus_allowed = set_cpus_allowed_common,
9565 .set_curr_task = set_curr_task_fair,
9566 .task_tick = task_tick_fair,
9567 .task_fork = task_fork_fair,
9569 .prio_changed = prio_changed_fair,
9570 .switched_from = switched_from_fair,
9571 .switched_to = switched_to_fair,
9573 .get_rr_interval = get_rr_interval_fair,
9575 .update_curr = update_curr_fair,
9577 #ifdef CONFIG_FAIR_GROUP_SCHED
9578 .task_move_group = task_move_group_fair,
9582 #ifdef CONFIG_SCHED_DEBUG
9583 void print_cfs_stats(struct seq_file *m, int cpu)
9585 struct cfs_rq *cfs_rq;
9588 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9589 print_cfs_rq(m, cpu, cfs_rq);
9593 #ifdef CONFIG_NUMA_BALANCING
9594 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9597 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9599 for_each_online_node(node) {
9600 if (p->numa_faults) {
9601 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9602 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9604 if (p->numa_group) {
9605 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9606 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9608 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9611 #endif /* CONFIG_NUMA_BALANCING */
9612 #endif /* CONFIG_SCHED_DEBUG */
9614 __init void init_sched_fair_class(void)
9617 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9619 #ifdef CONFIG_NO_HZ_COMMON
9620 nohz.next_balance = jiffies;
9621 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9622 cpu_notifier(sched_ilb_notifier, 0);